{"file_name": "audio/val_000000.wav", "transcription": "[0.0][S08] What can you see right now, Gemini?[2.33] [2.33][S03] OK.[3.12] [3.12][S03]I see a woman with blonde hair sitting[4.85] [4.85][S03]at a desk with a microphone in front of her.[7.11] [7.11][S03]She has her arms crossed and is smiling.[9.45] [9.45][S03]There's a computer keyboard and mouse on the desk[11.78] [11.78][S03]and a lamp in the background.[13.457] [13.457][S03]What else can you tell me?[14.54] [14.54][S08] That has been so useful in diagnosing[17.57] [17.57][S08]my sick houseplants, I can't tell you.[19.47] [19.47][S09] And so if I describe[22.82] [22.82][S09]that I want to hear a song that reminds me[26.3] [26.3][S09]of being in a rainy cafe in Japan at nighttime,[29.97] [29.97][S09]that's something that you would really[32.6] [32.6][S09]want to have all of these different modalities coming[35.69] [35.69][S09]together to understand, as opposed to just guessing[41.63] [41.63][S09]at what a rainy cafe might sound like.[43.512] [43.512][S08] It's like a two-dimensional version[45.47] [45.47][S08]of a fully immersive experience.[46.893] [46.893][S09] Exactly.[47.81] [47.81][S08] And this is the notoriously difficult challenge[50.18] [50.18][S08]that people have used to see how good--[51.78] [51.78][S09] So instead of having one person eating[53.6] [53.6][S09]spaghetti, we just had 10 people eating spaghetti[55.85] [55.85][S09]all at the same table.[57.32] [57.32][S08] I mean, there was a lot of Will Smith eating[58.77] [58.77][S08]spaghetti for a number of years, I seem to remember.[61.06] [61.06][S09] Yes.[61.81] [61.81][S09]And so Will Smith eating spaghetti, that is so last year.[67.57] [67.57][S09]Let's try multiple people eating spaghetti.[69.88] [69.88][MUSIC PLAYING][73.044] [74.4][S08] Welcome to \"Google DeepMind, The Podcast.\"[76.65] [76.65][S08]I'm Professor Hannah Fry.[77.833] [77.833][S08]Now, one of the things that we've always[79.5] [79.5][S08]done with this podcast is to bring you[81.93] [81.93][S08]access to the people who are working on some of the biggest[85.05] [85.05][S08]breakthroughs in AI.[86.44] [86.44][S08]And a lot of the time, the researchers,[88.45] [88.45][S08]they're talking about techniques and technology[90.81] [90.81][S08]that underpins big ideas.[93.64] [93.64][S08]But now, we are at a stage where more and more[96.81] [96.81][S08]of the tools that we have seen the early iterations of here[100.65] [100.65][S08]are now live.[102.1] [102.1][S08]They are out there in the world for you to interact with.[105.13] [105.13][S08]So what we wanted to do in this episode[106.98] [106.98][S08]is just to pause, to look at the array of tools[110.07] [110.07][S08]that have been released, talk about how they have changed[113.1] [113.1][S08]since we first encountered them, and to explore the myriad ways[116.7] [116.7][S08]that they can be used.[118.23] [118.23][S08]And if that is our objective for today,[120.22] [120.22][S08]well, there is no one better to show[121.98] [121.98][S08]us this progress than Paige Bailey, AI Developer Relations[126.06] [126.06][S08]Engineering Lead at Google DeepMind.[128.56] [128.56][S08]Paige, welcome to the podcast.[130.03] [130.03][S09] Thank you so much for having me.[131.97] [131.97][S08] The thing is is that we[133.23] [133.23][S08]get to see a lot of the early iterations of this stuff.[135.57] [135.57][S09] Yes.[136.32] [136.32][S08] And last year, we had Doug Eck on the show.[139.42] [139.42][S08]And he was showing us, I think, the very first iteration of Veo,[144.12] [144.12][S08]which now, with the launch of Veo 3, I mean,[147.03] [147.03][S08]it's quite a different beast.[148.843] [148.843][S09] Exactly.[149.76] [149.76][S09]The first implementation of the Veo model[152.16] [152.16][S09]was still just visual only, not including[156.48] [156.48][S09]all of the really enriching sound qualities that we[159.21] [159.21][S09]see from the Veo 3 model.[160.98] [160.98][S09]And you also had to give it pretty significant guidance[163.333] [163.333][S09]in order to get the model to produce something that looked[165.75] [165.75][S09]photorealistic, or even something that you[167.64] [167.64][S09]might see in a cinematic film.[169.41] [169.41][S09]But we've come a long way.[171.018] [171.018][S09]I would be really curious to see what[172.56] [172.56][S09]that first video looked like.[173.74] [173.74][S08] Yeah, so I think we have it, actually.[175.29] [175.29][S09] Yeah, let's do it.[176.69] [176.69][S08] We have this car racing through the streets.[179.88] [179.88][S08]You can see the neon lights reflected in the wet pavement[183.8] [183.8][S08]below.[184.56] [184.56][S08]There's other cars jostling for position around,[187.38] [187.38][S08]and it's almost like everything is blurred, because you're just[190.49] [190.49][S08]going so fast.[192.27] [192.27][S08]But it's really consistent.[193.53] [193.53][S08]Now, it's gone through a tunnel.[195.03] [195.03][S08]There are these big lights overhead,[197.97] [197.97][S08]and it's come out of the tunnel into an extremely realistic[203.06] [203.06][S08]modern scene.[203.608] [203.608][S02] It's a wow moment.[204.9] [204.9][S08] That's incredible.[205.77] [205.77][S02] That's a wow moment.[206.61] [206.61][S09] That is incredible.[208.02] [208.02][S08] I mean, that is really good, isn't it?[209.64] [209.64][S09] It is so good.[210.6] [210.6][S08] So some things that I[211.975] [211.975][S08]notice now, it is quite blurry.[214.86] [214.86][S08]And I guess that's part of the vibe that it's going for.[218.28] [218.28][S08]But you're not seeing this pristine detail on the car.[220.748] [220.748][S09] Not at all.[221.79] [221.79][S09]I think if we also looked really closely,[223.62] [223.62][S09]we might also see that some of the physics that's expressing[226.76] [226.76][S09]the shots is not quite right.[229.14] [229.14][S09]And the way that the light gets reflected on things[231.83] [231.83][S09]is also not necessarily quite consistent.[234.92] [234.92][S09]So let's see how well the new Veo 3 model does for this.[239.43] [239.43][S09]Yeah.[239.99] [239.99][S08] I mean, Doug was demonstrating the real limits[243.29] [243.29][S08]of what Veo was able to do.[245.245] [245.245][S08]And I suppose it's no coincidence[246.62] [246.62][S08]that it was a car that was racing through a scene.[249.66] [249.66][S08]If you tried to test Veo on, I don't know,[252.62] [252.62][S08]like a human face that was moving,[255.5] [255.5][S08]conveying a particular emotion, it might struggle a bit more.[258.12] [258.12][S09] Definitely.[259.279] [259.279][S09]It's very famous, the videos of people eating spaghetti,[263.75] [263.75][S09]or some of these things that are very,[266.52] [266.52][S09]very clearly not human characters[269.33] [269.33][S09]or not real life characters.[271.67] [271.67][S09]So it might be interesting also to just try with another prompt,[274.91] [274.91][S09]too, that is a little bit more human-centric,[277.82] [277.82][S09]but let's try with this.[278.82] [278.82][S08] Let's try this one.[279.62] [279.62][S08]So you're using exactly the same prompt here.[281.3] [281.3][S09] Exactly the same prompt.[282.883] [282.883][S09]And here we're in the Gemini app.[284.88] [284.88][S09]So you can see the exact same prompt generating this video.[288.36] [288.36][S09]It can take up to two to three minutes[290.39] [290.39][S09]in order to get it to have the right outputs.[293.29] [293.29][S08] One thing that was really noticeable[295.29] [295.29][S08]was the way that Doug's prompt was very poetic.[297.913] [297.913][S09] It was lovely.[299.08] [299.08][S08] It was gorgeous.[300.33] [300.33][S08]It was a little mini movie in its own right.[302.775] [302.775][S09] Yes.[303.525] [303.525][S08] What are your tips for generating these prompts?[305.71] [305.71][S09] So interestingly, we[307.23] [307.23][S09]have a new feature with the Veo 3 model called prompt rewriting,[311.62] [311.62][S09]which gives you the ability to give your input[314.49] [314.49][S09]sentence to the API, the way that we interact with the model,[318.87] [318.87][S09]and to get a response back that makes the prompt a lot more[322.23] [322.23][S09]detailed, a lot more aligned with perhaps[325.2] [325.2][S09]what you're imagining, so you don't necessarily[327.66] [327.66][S09]have to go through all of the mental work[330.9] [330.9][S09]and all of the terminology that would[332.97] [332.97][S09]be appropriate to describe this thing that you just imagined.[337.8] [337.8][S09]And I often use Gemini for something quite similar[341.28] [341.28][S09]is you can give Gemini the prompt that you're thinking[344.52] [344.52][S09]in your objective, and then ask it[346.59] [346.59][S09]to craft a prompt for a large language[348.9] [348.9][S09]model, or a video generation model,[351.4] [351.4][S09]in a way that would make the prompt much more[354.64] [354.64][S09]likely to produce the optimal output that you were expecting.[357.715] [357.715][S08] That's quite a good tip then, actually.[359.84] [359.84][S08]If you are not that good at writing prompts,[362.39] [362.39][S08]get Gemini to write the prompt better.[364.52] [364.52][S09] Yeah.[365.472] [365.472][S09]And so you can do that, kind of just[368.56] [368.56][S09]invoke it naturally through the Gemini app, which is probably[371.86] [371.86][S09]easiest for people to try.[373.6] [373.6][S09]But we also have the prompt rewriter hyper[376.84] [376.84][S09]specified for Veo 3 available through the API,[380.1] [380.1][S09]if people are a little bit more comfortable with programming[382.6] [382.6][S09]languages.[383.54] [383.54][S08] So how long are the Veo clips?[385.67] [385.67][S09] The Veo clips for the ones[387.4] [387.4][S09]that we released publicly, so the ones that people can try,[390.98] [390.98][S09]they're all around eight seconds in size, so very short.[394.01] [394.01][S08] Why does it need to be eight seconds long?[396.26] [396.26][S09] So it's eight seconds made available publicly.[399.73] [399.73][S09]Internally, we have models that are[401.44] [401.44][S09]capable of producing much more long-form content,[404.68] [404.68][S09]but we find that eight seconds is really good[407.23] [407.23][S09]to give you full creative control over that first clip.[411.52] [411.52][S09]And for the eight seconds, it's also useful in the sense[416.14] [416.14][S09]that you can get an idea of the style,[418.37] [418.37][S09]you can start experimenting with language.[420.44] [420.44][S09]And it's also wonderful in the sense[422.38] [422.38][S09]that you can start putting to life things[425.77] [425.77][S09]that you might have been imagining before.[428.14] [428.14][S09]I know that all of the internet is very enchanted with memes,[433.34] [433.34][S09]and now you can have memes that are much more long form,[436.49] [436.49][S09]that are actual videos, as opposed[438.76] [438.76][S09]to just the single snapshots.[441.522] [441.522][S08] Absolutely.[442.48] [442.48][S08]OK, it's ready.[443.105] [443.105][S09] Amazing.[444.022] [444.022][S09]Let's do it.[444.71] [444.71][S09]Oh, my gosh.[445.52] [445.52][S09]This is already--[446.47] [446.47][S08] All of a sudden.[447.32] [447.32][S09] This is so cool.[448.64] [448.64][S08] OK, now-- oh, wow.[450.05] [450.05][S08]Goodness me.[452.5] [452.5][S08]OK, it's like suddenly someone has switched on HD.[456.58] [456.58][S09] It absolutely is.[457.88] [457.88][S08] Especially that first shot.[460.34] [460.34][S08]So what we're looking at here, let me watch it again.[465.47] [465.47][S08]It looks like you're in \"Blade Runner.\"[467.395] [467.395][S09] It does look like--[468.77] [468.77][S08] It's really evocative of that.[470.52] [470.52][S08]So the neon lights have changed from these sort of pink,[473.58] [473.58][S08]sort of horrifying neons to full on billboards, which you can't[477.62] [477.62][S08]totally see the detail of as you go past them,[479.97] [479.97][S08]but there's structure to them that it's included.[484.94] [484.94][S08]The car is now sort of painted with light, almost,[492.05] [492.05][S08]running through the scene.[494.1] [494.1][S08]But also look at the lighting on that bonnet.[496.14] [496.14][S08]Oh, my gosh, that's extraordinary.[498.142] [498.142][S09] It is gorgeous.[499.35] [499.35][S09]And the level of detail in all of the buildings.[502.55] [502.55][S09]Actually, this really does look like a city in a futuristic Hong[506.99] [506.99][S09]Kong.[507.63] [507.63][S08] What I'm noticing here is there's a spotlight.[510.09] [510.09][S08]A car comes out of the tunnel, there's a lamppost above it,[514.35] [514.35][S08]and the spotlight perfectly tracks[516.35] [516.35][S08]where you would expect it to be along the bonnet of the car.[518.85] [518.85][S09] That's amazing.[520.058] [520.058][S09]Do you hear the sound?[521.184] [521.184][S08] The screeching tires.[522.559] [522.559][S09] Yes.[524.268] [524.268][S08] Oh, sirens in the background.[525.976] [525.976][GASPS][526.85] [526.85][S09] This is so cool.[528.633] [528.633][S08] Oh, it's very evocative, isn't it?[530.55] [530.55][S09] It is.[532.91] [532.91][S08] Oh, yeah.[533.94] [533.94][S09] Yeah.[534.18] [534.18][S08] But the sound perfectly[535.31] [535.31][S08]matches the frames, then.[536.52] [536.52][S09] It definitely does.[537.895] [537.895][S09]And the background tracks the sounds[539.81] [539.81][S09]for these videos that get generated with Veo 3.[542.695] [545.85][S09]They're actually able to match not just things[549.68] [549.68][S09]like the cars, as they're kind of coming into the scene,[555.17] [555.17][S09]but also capable of giving you audio.[558.03] [558.03][S09]So audio is outputs, background music,[561.12] [561.12][S09]if you wanted to have cinematic music coupled with the audio[565.07] [565.07][S09]or coupled with the background noises, all[567.65] [567.65][S09]of this kind of stitched together into a single video.[570.39] [570.39][S09]It's pretty magical.[571.563] [571.563][S08] Let's try one of the hard ones, the spaghetti thing.[574.23] [574.23][S09] OK, so let's try a new video output.[578.19] [578.19][S08] Yeah.[578.9] [578.9][S09] What kind of spaghetti eating[581.243] [581.243][S09]would you like to see?[582.16] [582.16][S08] I would like to see somebody at a spaghetti eating[584.4] [584.4][S08]contest.[585.22] [585.22][S08]I would like their face to be dirty with sauce.[587.51] [587.51][S09] Excellent.[588.51] [588.51][S08] And I'd a whole hunk of spaghetti to be essentially[592.23] [592.23][S08]hanging from their mouth.[593.47] [593.47][S08]Sauce everywhere.[595.29] [595.29][S09] Sauce everywhere.[598.02] [598.02][S09]And camera pans in--[605.49] [605.49][S09]or camera pans out to see all of the contestants eating spaghetti[613.26] [613.26][S09]simultaneously.[614.292] [614.292][S08] Nice.[615.0] [615.0][S09] Yeah.[616.44] [616.44][S09]And we'll see how well the model can do at this one.[620.37] [620.37][S09]You can also use--[621.88] [621.88][S09]as I'm sure you can see here, you[623.61] [623.61][S09]can ground the initial model frame with an image.[628.78] [628.78][S09]So if you wanted to use something like Imagen 4[631.2] [631.2][S09]to generate an image as the seed frame for the video itself,[635.92] [635.92][S09]you would be able to do that as well.[638.73] [638.73][S08] So if you had, for example,[640.46] [640.46][S08]a picture of a spaghetti eating contest[642.61] [642.61][S08]and you wanted it to replicate where[644.23] [644.23][S08]the tables were, where the scenery was,[646.09] [646.09][S08]you could do that with this?[647.257] [647.257][S09] Absolutely.[648.298] [648.298][S09]And then there are some other capabilities and characteristics[650.95] [650.95][S09]of our latest Gemini models that allow[652.81] [652.81][S09]you to do really nice editing of images as well.[657.1] [657.1][S09]So perhaps you have a spaghetti eating contest,[659.87] [659.87][S09]but you only want to have a certain number of people[662.56] [662.56][S09]in the frame or you want to have a specific style of bench[667.06] [667.06][S09]or a specific kind of lighting.[668.66] [668.66][S09]You would also be able to give that image input to Gemini,[672.5] [672.5][S09]ask for it to transform it, and then use that as the seed[675.01] [675.01][S09]frame for the Veo 3.[676.12] [676.12][S08] So let me understand then.[677.703] [677.703][S08]What is new in Veo 3 that we didn't have before?[680.57] [680.57][S08]What makes it better?[681.92] [681.92][S09] Yeah.[682.712] [682.712][S09]So Veo 3 is better for a few reasons.[684.65] [684.65][S09]One is that it has the ability to produce[687.88] [687.88][S09]sounds, which has been really enchanting a lot of people,[690.68] [690.68][S09]I think.[691.51] [691.51][S09]Not just background noises, but also realistic music, dialogue,[697.75] [697.75][S09]all sorts of things that can really[699.49] [699.49][S09]construct videos that feel like they might[701.68] [701.68][S09]have been taken in real life.[704.56] [704.56][S09]Another way that we've improved Veo 3[707.14] [707.14][S09]is that the video outputs are grounded[709.45] [709.45][S09]in more physics understanding.[711.08] [711.08][S09]So as we look at videos, we can spot[714.22] [714.22][S09]ways in which the light or gravity really[717.19] [717.19][S09]does seem to align with the physical world.[719.83] [719.83][S09]And then there's also been a lot of improvements[722.32] [722.32][S09]around character consistency.[724.34] [724.34][S08] How are these additional features possible?[728.18] [728.18][S09] Yeah, I think DeepMind is really,[730.4] [730.4][S09]really paying attention to how the data is curated[733.69] [733.69][S09]to use to train the models.[735.16] [735.16][S09]When you think about it, you can see the word tree.[739.6] [739.6][S09]You can see a picture of a tree.[741.95] [741.95][S09]You could have the 3D representation of trees or any[749.41] [749.41][S09]of the other things that you would expect.[751.19] [751.19][S08] A sound of wind blowing through leaves.[753.893] [753.893][S09] Exactly.[754.81] [754.81][S09]And it's like a video of somebody panning around.[758.07] [758.07][S09]All of these things are still associated with that one entity,[762.78] [762.78][S09]but it's all very different modalities[765.05] [765.05][S09]describing the same thing.[766.73] [766.73][S09]And so I think historically folks have been concentrating[772.16] [772.16][S09]on just one modality-- so text, or code, or something similar--[776.03] [776.03][S09]when as humans, we experience the entire world[779.21] [779.21][S09]in very different ways.[781.14] [781.14][S09]Everything from seeing to hearing to touching,[784.14] [784.14][S09]all of these things.[785.52] [785.52][S09]And so I think that the team has put a lot of time,[789.11] [789.11][S09]and energy, and effort into being able to couple[791.63] [791.63][S09]together, not just the video footage, but also the sound that[796.22] [796.22][S09]composes the video footage, detailed descriptions, even[799.19] [799.19][S09]at the frame by frame level, and then also stitching all of that[804.02] [804.02][S09]together into a full representation of the training.[809.377] [809.377][S08] So whereas a language-only model might[811.46] [811.46][S08]have the word tree and its closely[813.84] [813.84][S08]associated with the word branch or twig.[816.13] [816.13][S08]The multimodal version not only has[818.58] [818.58][S08]all of those embedded within it, but additionally has[821.7] [821.7][S08]audio, images, video, all of these different layers.[825.648] [825.648][S09] Absolutely.[826.69] [826.69][S09]I think this is one of the things that[828.72] [828.72][S09]makes me most excited about the Gemini models is that we're[831.54] [831.54][S09]really the only model family that also allows you to output[835.41] [835.41][S09]text and code, but also images, to edit images,[838.65] [838.65][S09]to have output audio, which we'll see in a second,[841.83] [841.83][S09]as well as steerable audio.[843.67] [843.67][S09]So being able to say, speak softer, or speak more loud,[847.44] [847.44][S09]or speak in a different language.[849.81] [849.81][S09]All of these other model families[851.58] [851.58][S09]kind of relied on stitching together different trained[855.84] [855.84][S09]experiences, as opposed to baking it all into one[858.66] [858.66][S09]innate model.[860.31] [860.31][S09]And I think that's really powerful about Gemini.[864.868] [864.868][S08] Oh, OK.[865.66] [865.66][S08]It's ready.[866.28] [866.28][S09] Oh, my gosh.[866.98] [866.98][S09]Let's see it.[867.64] [867.64][S08] Go for it.[868.916] [868.916][CHATTER][870.66] [870.66][S08]I regret.[871.63] [871.63][S08]I regret.[872.37] [872.37][SLURPING IN VIDEO][873.25] [873.25][S09] Oh, my gosh.[876.364] [876.364][S09]So these do look actual people, though.[879.53] [879.53][S08] They do look actual people.[881.98] [881.98][S08]The spaghetti is-- I mean, I would say--[884.69] [884.69][SCREAMS IN VIDEO][885.97] [885.97][LAUGHTER][887.39] [887.623][S09] OK, we're going to pause for a second,[889.79] [889.79][S09]but look at the faces.[892.51] [892.51][S09]This does look a little bit more photorealistic than some[898.09] [898.09][S09]of the experiences that we've seen previously.[900.23] [900.23][S08] The hair especially.[901.563] [901.563][S08]If you just go to about there, two seconds in,[903.64] [903.64][S08]maybe, as it zooms out, this girl's hair.[906.98] [906.98][S08]I mean, look, she's even got a little bit of a--[909.46] [909.46][S08]you see the little parting bit, and tufts of baby hair.[912.478] [912.478][S09] Yeah.[913.27] [913.27][S08] But yeah, as she moves, that is really realistic.[916.13] [916.13][S09] This is so cool.[917.93] [917.93][S09]And then also the noises are something that I would imagine.[921.73] [921.73][S09]Yes, obviously disgusting, but things[923.59] [923.59][S09]that I would expect to see at a spaghetti eating contest.[926.77] [926.77][S09]And also how the sauce looks, how the noodles look,[933.75] [933.75][S09]that as you're gulping, some of the noodles[936.67] [936.67][S09]are falling back to the plate.[939.115] [939.115][S09]This is incredible.[940.752] [940.752][S08] And this is the notoriously difficult challenge[943.21] [943.21][S08]that people have used to see how good Veo is.[944.81] [944.81][S09] So instead of having one person eating[946.63] [946.63][S09]spaghetti, we just had 10 people eating spaghetti[948.88] [948.88][S09]all at the same table.[950.32] [950.32][S08] I mean, there was a lot of Will Smith eating[951.76] [951.76][S08]spaghetti for a number of years, I seem to remember.[954.08] [954.08][S09] Yes.[954.83] [954.83][S09]And so Will Smith eating spaghetti, that is so last year.[960.59] [960.59][S09]Let's try multiple people eating spaghetti.[962.51] [962.51][S08] The thing that I really[963.968] [963.968][S08]find impressive is how an individual object tracks[967.81] [967.81][S08]through the frames, and keeps its consistency.[971.41] [971.41][S08]Is that something that you have to impart afterwards?[974.6] [974.6][S09] I do think that the way that the model is[977.29] [977.29][S09]constructed, which is a diffusion model,[980.17] [980.17][S09]is attempting to hold that consistency throughout frames.[984.71] [984.71][S09]But a lot of this logic is baked into the model itself.[987.59] [987.59][S09]So in the before times, you would have to explicitly program[991.85] [991.85][S09]each one of the behaviors.[994.01] [994.01][S09]For now, I think that a lot of the magic behind the scenes[997.55] [997.55][S09]is just Veo and Gemini kind of figuring out[1000.94] [1000.94][S09]what they need to do in order to make the experiences[1004.15] [1004.15][S09]and make the outputs look photorealistic.[1006.07] [1006.07][S08] Because it has read the internet,[1008.24] [1008.24][S08]but it's also watched all of YouTube.[1009.86] [1009.86][S09] Well, not all of YouTube, and in addition[1012.22] [1012.22][S09]to a lot of synthetic data.[1013.55] [1013.55][S09]So as an example, we're incredibly[1018.01] [1018.01][S09]benefited by having so many game designers[1020.5] [1020.5][S09]and people from the game industry working at DeepMind.[1023.93] [1023.93][S09]And Demis, of course, has a game background.[1028.22] [1028.22][S09]But these environments are really[1030.97] [1030.97][S09]great synthetic data generators for all sorts of things.[1034.7] [1034.7][S09]You can have figures running through games,[1038.54] [1038.54][S09]accomplishing tasks, like, doing all sorts of behaviors[1042.412] [1042.412][S09]and monitoring them along the way, which[1044.079] [1044.079][S09]can be used as an excellent source of training data.[1046.589] [1046.589][S09]It's also nice in the sense that as you do gameplay,[1049.02] [1049.02][S09]you can have the camera orient itself in various ways[1052.46] [1052.46][S09]around to the characters.[1053.88] [1053.88][S08] So, OK, I know a lot of the buzz around this,[1056.255] [1056.255][S08]a lot of the new thing about Veo 3 is the audio.[1058.88] [1058.88][S08]Just, I mean, how is it being generated in order[1062.03] [1062.03][S08]to correspond?[1063.06] [1063.06][S08]Is it that it's generating something[1064.67] [1064.67][S08]that matches the visuals, or is it[1067.16] [1067.16][S08]like, there's a context which produces[1070.07] [1070.07][S08]both the visuals and the audio?[1071.667] [1071.667][S09] I think it's because the training data has[1074.0] [1074.0][S09]all of the different modalities associated with the thing[1078.47] [1078.47][S09]that it sees.[1079.14] [1079.14][S09]So it's not just seeing a video.[1082.07] [1082.07][S09]It also has the transcript.[1083.55] [1083.55][S09]It also has the frame by frame level description of the video[1086.24] [1086.24][S09]and what's happening.[1087.18] [1087.18][S09]It's also has the description of the audio,[1089.6] [1089.6][S09]if there's any background tracks.[1091.29] [1091.29][S09]And so all of that kind of brought together simultaneously[1095.63] [1095.63][S09]is capable of generating these much more immersive and natural[1099.74] [1099.74][S09]sounds and natural responses.[1102.77] [1102.77][S09]Because there are certainly instances[1104.87] [1104.87][S09]where if you listen to a song, you could read the sheet music[1112.79] [1112.79][S09]and you could hear the different tones[1116.12] [1116.12][S09]displayed, but it could also make you feel a certain way.[1119.28] [1119.28][S09]And so if I describe that I want to hear a song that reminds me[1126.29] [1126.29][S09]of being in a rainy café in Japan at nighttime,[1130.396] [1130.396][S09]that's something that you would really[1132.56] [1132.56][S09]want to have all of these different modalities coming[1135.65] [1135.65][S09]together to understand, as opposed to just guessing[1141.62] [1141.62][S09]at what a rainy café might sound like.[1143.502] [1143.502][S08] It's like a two-dimensional version[1145.46] [1145.46][S08]of a fully immersive experience.[1147.27] [1147.27][S09] Exactly.[1148.43] [1148.43][S08] That's nice.[1149.94] [1149.94][S09] I love that description of it as well.[1152.435] [1152.435][S09]We're getting closer and closer towards something that[1159.41] [1159.41][S09]feels very close to reality.[1161.158] [1161.158][S08] Like a simulated reality.[1162.7] [1162.7][S09] Yes.[1163.504] [1163.504][S09]And I don't think that was possible previously.[1166.93] [1166.93][S08] So thus far, then, is the Gemini app[1170.22] [1170.22][S08]that we know and love.[1171.16] [1171.16][S09] Yes.[1171.45] [1171.45][S08] But if you are a professional filmmaker,[1173.893] [1173.893][S08]or you want to take this a bit more seriously,[1175.81] [1175.81][S08]there is another place you can go to, correct?[1178.468] [1178.468][S09] Yes, it is called Flow.[1180.01] [1180.01][S09]It's built by our colleagues over in the Google Labs team,[1182.61] [1182.61][S09]and they've been partnering directly[1184.11] [1184.11][S09]with filmmakers to really build an experience that[1187.41] [1187.41][S09]aligns with their expectations.[1188.902] [1188.902][S08] So go on, show me Flow.[1190.36] [1190.36][S09] Yeah.[1192.39] [1192.39][S09]And so this is Flow, which is the first place that you would[1197.49] [1197.49][S09]go if you were a filmmaker.[1199.06] [1199.06][S09]You can also see some of the projects[1201.45] [1201.45][S09]that have been happening recently,[1204.48] [1204.48][S09]as well as this thing called Flow TV, which[1207.72] [1207.72][S09]is a really, really cool way to experience[1211.2] [1211.2][S09]some of the videos that have been generated by Flow recently.[1216.168] [1216.168][S08] So is the idea here then,[1217.71] [1217.71][S08]you still have the eight second videos,[1219.34] [1219.34][S08]but you can stitch them together?[1221.308] [1221.308][S09] You can stitch them together,[1223.1] [1223.1][S09]you can style them.[1224.51] [1224.51][S09]There are even camera controls associated.[1226.607] [1226.607][S08] Oh, wow.[1227.44] [1227.44][S09] So it really does give you a lot more creative[1230.23] [1230.23][S09]control as a filmmaker without necessarily having to--[1237.21] [1240.31][S09]oh, that's so cool, too.[1241.88] [1241.88][S08] These are--[1243.108] [1243.108][S09] But look at that.[1244.4] [1244.4][S09]The physics understanding of that video.[1246.7] [1246.7][S08] The fluid dynamics, right?[1249.04] [1249.04][S08]I mean, wow, extraordinary.[1252.608] [1252.608][S08]Look at that.[1253.15] [1253.15][S08]Oh, lord.[1253.65] [1253.65][S09] Yeah.[1254.442] [1254.442][S09]And then if you take a look at the prompt,[1256.22] [1256.22][S09]you can also see some of the sentences associated.[1260.12] [1260.12][S09]So \"clean metal plastic injection opens releasing[1262.96] [1262.96][S09]a pinkish yellow jellyfish.\"[1266.26] [1266.26][S09]And a lot of the others along the way.[1269.11] [1269.11][S09]Yeah.[1269.71] [1269.71][S09]So detailed.[1270.89] [1270.89][S08] So detailed.[1272.12] [1272.12][S09] Yeah.[1272.29] [1272.29][S08] Show me the ways in which this differs[1274.373] [1274.373][S08]from the Gemini app, then.[1275.84] [1275.84][S09] Absolutely.[1276.882] [1276.882][S09]So this is a hyper specialized experience.[1280.27] [1280.27][S09]So if you click Create with Flow,[1282.24] [1282.24][S09]you're put into this development environment that you can use.[1286.48] [1286.48][S09]You can start a new project, where you can use Text to Video,[1291.3] [1291.3][S09]as well as Frames to Video, or Ingredients to Video.[1293.897] [1293.897][S08] So hold on, Text to Video,[1295.48] [1295.48][S08]that's the thing that we were already doing?[1296.65] [1296.65][S09] That's the thing that we were just doing.[1298.942] [1298.942][S09]There's also the frames to video.[1301.62] [1301.62][S09]So let's go ahead and do frames to video.[1305.94] [1305.94][S09]And you can stitch together multiple frames[1308.34] [1308.34][S09]to create a final scene.[1310.0] [1310.0][S09]You can also click this Camera Control button[1312.9] [1312.9][S09]and see these different examples to dollying in, dollying out,[1318.63] [1318.63][S09]static, or tilt down.[1320.97] [1320.97][S09]And all of those are things that can[1324.42] [1324.42][S09]build the takeaways that you would like[1327.81] [1327.81][S09]to see in the video footage.[1328.977] [1328.977][S08] So essentially this is just--[1330.685] [1330.685][S08]I mean, it's kind of the same thing, but just a more[1332.88] [1332.88][S08]specialized environment?[1334.0] [1334.0][S09] Definitely.[1335.042] [1335.042][S09]And we find that, just as you would[1336.68] [1336.68][S09]have specialized environments for musicians to create[1340.73] [1340.73][S09]their electronic tracks or CAD designers,[1344.72] [1344.72][S09]you would probably want a really,[1346.47] [1346.47][S09]really dedicated and focused UI for each one of these use cases[1350.0] [1350.0][S09]that can really hyper optimize for the things[1351.92] [1351.92][S09]that you would care about as a filmmaker.[1353.7] [1353.7][S08] I think the point about this[1355.367] [1355.367][S08]is that you then have this absolute open door[1357.86] [1357.86][S08]for creativity.[1358.83] [1358.83][S09] Yes.[1359.0] [1359.0][S08] So I've seen people take videos[1360.86] [1360.86][S08]as though the Spartans were Instagram influencers.[1363.98] [1363.98][S09] Yes.[1364.73] [1364.73][S09]Absolutely.[1365.23] [1365.23][S08] Like, reporting on their siege.[1368.758] [1368.758][S09] Absolutely.[1369.8] [1369.8][S09]And also character consistency across different experiences,[1374.91] [1374.91][S09]no matter what the lighting might be.[1376.58] [1376.58][S09]Like you might have a little monster character[1379.58] [1379.58][S09]that you want to have swimming through the ocean.[1381.918] [1381.918][S09]And then you also want to have him climbing a mountain,[1384.21] [1384.21][S09]and you want to have him singing on a stage.[1386.34] [1386.34][S09]And it's able to keep that same character consistent,[1389.04] [1389.04][S09]but to change all of the dynamics[1390.65] [1390.65][S09]around it, which is pretty magical.[1393.39] [1393.39][S08] I do wonder, though, about putting these tools[1396.03] [1396.03][S08]in the hands of people.[1397.26] [1397.26][S08]There are also concerns about it, too, right?[1399.73] [1399.73][S08]I mean, deepfakes, but also scams,[1404.0] [1404.0][S08]tricking people into thinking that news events are happening[1406.5] [1406.5][S08]that perhaps aren't.[1408.12] [1408.12][S08]Where do you stand on that?[1409.32] [1409.32][S09] Yes.[1410.07] [1410.07][S09]So we do have safety filters introduced within the Veo[1413.73] [1413.73][S09]models themselves.[1415.03] [1415.03][S09]And relatedly, for all of the Veo models[1417.66] [1417.66][S09]that are generated through the Gemini app,[1419.41] [1419.41][S09]there's a specialized watermark that gives you the ability[1423.0] [1423.0][S09]to know that this was AI authored, as opposed[1426.81] [1426.81][S09]to being something that was just shot via raw footage[1430.35] [1430.35][S09]out in the world.[1431.91] [1431.91][S09]But we also have special constraints in place[1434.82] [1434.82][S09]around not being able to generate images of things[1440.25] [1440.25][S09]like children or special entities.[1443.92] [1443.92][S09]There's also a constraint in place[1446.1] [1446.1][S09]such that government officials, or people[1448.77] [1448.77][S09]who are significantly present in the public sphere for policy,[1453.22] [1453.22][S09]or for science, or for any of the notable figures[1457.51] [1457.51][S09]in the world, you can't generate video content about them.[1462.4] [1462.4][S09]And even the kind of models that we experiment with internally,[1470.92] [1470.92][S09]they still have these constraints.[1472.61] [1472.61][S08] Well, if one of the key things about Veo 3[1474.95] [1474.95][S08]is the audio, can we go into the audio a bit more?[1477.19] [1477.19][S09] Yes.[1477.94] [1477.94][S09]We even just released a Gemini Text[1480.13] [1480.13][S09]to Speech API that allows you to generate audio,[1483.14] [1483.14][S09]including steerable audio, in multiple languages.[1485.205] [1485.205][S08] So without the images[1486.58] [1486.58][S08]to go with it, just audio only?[1488.27] [1488.27][S09] Just audio only, but really,[1490.1] [1490.1][S09]really expressive audio.[1491.78] [1491.78][S09]And you can also have multiple speakers in different languages.[1496.51] [1496.51][S09]So I believe you might have seen the NotebookLM[1500.05] [1500.05][S09]before with the podcasts that were generated.[1503.6] [1503.6][S09]This allows you to create customizable and similar[1508.24] [1508.24][S09]experiences using multiple speakers or a single speaker.[1511.173] [1511.173][S08] Well, let me just bring you back for a second,[1513.59] [1513.59][S08]because we did actually get to talk to the researchers who[1517.03] [1517.03][S08]were working on WaveNet.[1518.21] [1518.21][S09] Oh, yes.[1519.127] [1519.127][S08] This is like, now, I mean, only four years ago.[1522.12] [1522.12][S08]And this is where they were at that point, because what they[1524.62] [1524.62][S08]did is they trained a model, this was neural networks, right,[1528.85] [1528.85][S08]using my voice.[1530.19] [1530.19][S08]And this is where they got to.[1531.44] [1531.44][S01] Hi, there.[1532.01] [1532.01][S01]I'm a mathematician, author, and podcaster[1534.7] [1534.7][S01]who's fascinated by artificial intelligence.[1537.003] [1537.003][S08] It's very breathy, isn't it?[1538.67] [1538.67][S09] It is very breathy.[1539.72] [1539.72][S08] How have things changed since then?[1542.0] [1542.0][S09] Yes.[1542.75] [1542.75][S09]So things have changed significantly.[1544.54] [1544.54][S09]When WaveNet, which was pioneering at the time,[1546.86] [1546.86][S09]was first created, you would need[1549.49] [1549.49][S09]dedicated single-task models for each one of the things[1552.67] [1552.67][S09]that you were trying to do.[1554.0] [1554.0][S09]And so I started doing machine learning,[1556.4] [1556.4][S09]I think, around 2009, 2010, and it was extraordinarily painful[1560.71] [1560.71][S09]because you had to acquire all of these special purpose data[1564.13] [1564.13][S09]sets.[1564.8] [1564.8][S09]You had to get them cleaned up.[1566.72] [1566.72][S09]You would have to monitor for things like data drift.[1569.31] [1569.31][S09]If anything changed over time, you[1570.79] [1570.79][S09]would have to retrain the model from scratch.[1572.67] [1572.67][S09]And so WaveNet was a dedicated single-task model[1576.65] [1576.65][S09]for generating these really realistic, at this time,[1580.58] [1580.58][S09]sounding voices.[1582.11] [1582.11][S09]But it couldn't do other things.[1584.04] [1584.04][S09]So it couldn't have steerable audio.[1588.03] [1588.03][S09]Like, you couldn't say, \"please, give me this audio clip in this[1595.04] [1595.04][S09]kind of style, and do it in German.\"[1598.11] [1598.11][S09]Whereas with our recent models, they[1600.89] [1600.89][S09]are a lot more steerable by design.[1603.96] [1603.96][S09]So you can give instructions about style,[1606.36] [1606.36][S09]about the language that you're speaking,[1610.22] [1610.22][S09]about pause instructions, or speak quickly, speak slower,[1615.98] [1615.98][S09]all sorts of things.[1617.25] [1617.25][S08] So how much of the WaveNet, I mean, code, even,[1620.72] [1620.72][S08]has actually ended up feeding into this model?[1623.49] [1623.49][S08]Or is it you started again once large language models[1627.202] [1627.202][S08]and transformers came on the scene?[1628.66] [1628.66][S09] Yeah.[1629.452] [1629.452][S09]So a lot of the code that was used to create the WaveNet[1632.52] [1632.52][S09]model, the architectures are a little bit[1636.81] [1636.81][S09]different for our Gemini family, but the data that was used[1641.34] [1641.34][S09]is definitely repurposed.[1642.94] [1642.94][S09]And so all of the examples created[1645.03] [1645.03][S09]for this is the text input, this is the audio output.[1648.39] [1648.39][S09]You can also enrich those kinds of data sets[1650.64] [1650.64][S09]with the descriptions of the style of the audio or the tone[1655.8] [1655.8][S09]or the temper of the voice.[1657.66] [1657.66][S09]All of that is incredibly useful for Gemini.[1660.22] [1660.22][S09]And that's been one of the really interesting things[1662.49] [1662.49][S09]to see over the years as well is that previously,[1666.31] [1666.31][S09]like with the first generation of Gemini models,[1668.8] [1668.8][S09]we had to do significant work to fine tune them.[1672.25] [1672.25][S09]So you had things like MetaLimb.[1675.1] [1675.1][S09]You had things like special purpose models for generating[1679.29] [1679.29][S09]code that were kind of variations of Gemini's.[1683.64] [1683.64][S09]And all of those training data have been now incorporated[1687.15] [1687.15][S09]into the base model itself.[1689.41] [1689.41][S09]So Gemini has just gotten better and better natively[1691.782] [1691.782][S09]at all of the tasks that people had previously needed[1693.99] [1693.99][S09]to fine tune to do.[1695.425] [1695.425][S08] So and then show me how it works.[1697.3] [1697.3][S08]Show me, give me some examples.[1698.71] [1698.71][S09] Exactly.[1699.627] [1699.627][S09]So if we go over to AI Studio, I'll open up a new tab[1703.59] [1703.59][S09]and pull us into the UI.[1705.64] [1705.64][S09]You can see here this great playground[1708.33] [1708.33][S09]for experimenting with and trying out the latest Gemini[1710.67] [1710.67][S09]models as soon as they're released.[1712.3] [1712.3][S08] And so to create audio,[1714.028] [1714.028][S08]it's the one that looks like an audio wave on the side.[1716.32] [1716.32][S09] So interestingly, for the text-to-speech model,[1719.32] [1719.32][S09]you would go to generate media, and then you[1721.5] [1721.5][S09]would go to Gemini speech generation,[1723.79] [1723.79][S09]and you're launched into this text-to-speech UI where[1726.57] [1726.57][S09]you can specify the different speakers, the different voices,[1730.12] [1730.12][S09]and then also the style instructions[1732.18] [1732.18][S09]for each one of the speakers.[1733.6] [1733.6][S08] Let's think of a prompt[1734.31] [1734.31][S08]then, because I'm particularly interested[1736.018] [1736.018][S08]in this different emotions.[1737.62] [1737.62][S08]So what about if you get it to say something like,[1739.95] [1739.95][S08]\"I was waiting for you,\" and then we try different emotion.[1742.845] [1742.845][S09] All right, so we're[1744.22] [1744.22][S09]going to specify in the system instructions,[1747.04] [1747.04][S09]speak in a friendly tone like you're greeting[1754.27] [1754.27][S09]a relative who just came home.[1757.672] [1757.672][S08] Nice.[1758.38] [1758.38][S09] Yeah, and then start typing a prompt.[1760.87] [1760.87][S09]\"I was waiting for you.\"[1763.09] [1763.09][S09]And what we're going to do is speak in a friendly tone,[1768.59] [1768.59][S09]only say what the person prompts you to say.[1778.03] [1778.03][S09]Yeah, and then hit Run.[1782.26] [1782.26][S05] I was waiting for you.[1784.48] [1784.48][S09] Amazing.[1784.84] [1784.84][S08] Friendly.[1785.36] [1785.36][S09] Yeah.[1786.01] [1786.01][S08] Can we try something different, though?[1787.18] [1787.18][S08]Can you make it more romantic?[1789.29] [1789.29][S09] Yes.[1790.21] [1790.21][S09]So speak in a romantic, hushed tone, very breathy.[1796.9] [1796.9][S09]And only say the words that the person puts into the prompt.[1805.94] [1805.94][S09]And so we'll see how this goes.[1808.5] [1808.5][S09]And it was, \"I was waiting for you.\"[1810.66] [1810.66][S08] Yeah.[1811.52] [1811.52][S09] And hit Run.[1812.93] [1812.93][S09]And OK, good.[1815.13] [1815.13][S09]Zephyr is selected.[1817.94] [1817.94][S06] I was waiting for you.[1819.95] [1819.95][S09] Ooh.[1820.7] [1820.7][S08] Saucy.[1821.7] [1821.7][S09] Very saucy.[1822.8] [1822.8][S09]I also love that you can see in this UI the thoughts associated.[1827.39] [1827.39][S09]So Gemini's thinking process as it's[1830.48] [1830.48][S09]kind of going through the path of creating this audio response.[1833.46] [1833.46][S08] What has it said?[1834.27] [1834.27][S09] It says, \"I've processed the input.[1836.312] [1836.312][S09]I've pinpointed the exact phrase I need to use.[1838.29] [1838.29][S09]The task core now centers on delivering the specific phrase.[1841.47] [1841.47][S09]I've meticulously identified the romantic, hushed,[1843.72] [1843.72][S09]and very breathy tone required, and the next step is generating[1846.56] [1846.56][S09]the audio.\"[1847.17] [1847.17][S09]So it really gives you step by step instructions[1850.82] [1850.82][S09]of how to incorporate all of these responses.[1852.695] [1852.695][S08] Can I do some more?[1853.987] [1853.987][S08]Can we do angry?[1854.66] [1854.66][S09] Yes.[1856.11] [1856.11][S09]You can definitely do angry.[1857.4] [1857.4][S09]So let's change the system instructions again.[1860.58] [1860.58][S09]And so let's say--[1862.823] [1862.823][S08] Someone's late for the date.[1864.49] [1864.49][S09] Yes.[1865.24] [1865.24][S09]So someone is late for the date.[1867.09] [1867.09][S09]I'm going to copy the system instructions again, turn off[1870.27] [1870.27][S09]the stream and start a new one.[1871.81] [1871.81][S09]And we will do this.[1874.96] [1874.96][S09]So speak in an annoyed, angry tone, like a person has just[1885.54] [1885.54][S09]arrived late for a date, and only say the words[1892.92] [1892.92][S09]that you see here.[1894.94] [1894.94][S09]And so let's try.[1895.648] [1895.648][S08] This is so much fun.[1896.982] [1896.982][S09] Yes, it is.[1898.08] [1898.08][S04] I was waiting for you.[1899.86] [1899.86][S09] Oh, that--[1901.0] [1901.0][S08] She's angry.[1901.78] [1901.78][S09] She is very angry.[1903.113] [1906.03][S08] Can you do grieving?[1907.632] [1907.632][S09] Grieving.[1908.59] [1908.59][S09]Yes.[1908.67] [1908.67][S08] Like a lost loved one?[1909.82] [1909.82][S09] Yes.[1910.57] [1910.57][S09]So let's try another stream and speak in a grieving tone.[1919.11] [1919.11][S09]Grieving tone, like you have just lost a loved relative.[1927.84] [1927.84][S09]Only say the words that the person puts into the prompt.[1930.67] [1930.67][S09]And \"I was waiting for you.\"[1934.5] [1934.5][S09]And hit Run.[1935.09] [1938.97][S07] I was waiting for you.[1940.992] [1940.992][S09] Aw.[1941.7] [1941.7][S09]And so I feel like that was a little bit less grieving.[1946.57] [1946.57][S09]So you could probably experiment a little with the prompt.[1951.54] [1951.54][S08] Now, can you make them French?[1953.61] [1953.61][S09] So we can modify the system instructions.[1956.26] [1956.26][S09]So within this streaming real-time feature,[1958.98] [1958.98][S09]we can say something to the effect of only respond.[1962.46] [1962.46][S09]So only respond to the user in French.[1968.55] [1968.55][S09]Make sure to be bubbly and excited.[1972.672] [1972.672][S08] Nice.[1973.38] [1973.38][S09] Bubbly and excited.[1975.79] [1975.79][S09]And only say the text that the user adds in the prompt.[1982.72] [1982.72][S09]And so we can see what the output might be.[1987.375] [1987.375][S09]And we're still going to go with \"I was waiting for you?\"[1989.75] [1989.75][S08] Absolutely.[1990.31] [1990.31][S09] All right.[1991.31] [1991.31][S09]\"I was waiting for you.\"[1993.85] [1993.85][S09]And we've got Zephyr selected.[1997.63] [1997.63][S09]We've got the Run button here, and let's try.[2001.52] [2004.36][S03] [SPEAKING FRENCH][2006.947] [2006.947][S09] I don't speak French.[2008.405] [2008.405][S08] [SPEAKING FRENCH][2010.29] [2010.29][S08]Yeah, I mean, sure.[2011.6] [2011.6][S09] Yeah, excellent.[2012.85] [2012.85][S08] So this is all in AI Studio?[2014.735] [2014.735][S09] This is all in AI Studio.[2016.36] [2016.36][S08] Which is available for me to play with?[2017.98] [2017.98][S09] Yes, it is available today for free[2020.022] [2020.022][S09]for people to experiment, to try things out.[2022.6] [2022.6][S09]And even cooler, if you're a developer,[2025.11] [2025.11][S09]there's a little button here called[2026.67] [2026.67][S09]Get SDK Code that whenever you click it,[2029.14] [2029.14][S09]it gives you everything that you just did in the UI,[2032.29] [2032.29][S09]but in code form.[2033.83] [2033.83][S09]So if you wanted it in Python, in TypeScript,[2037.36] [2037.36][S09]in any of the languages that you're[2039.04] [2039.04][S09]using in your personal project, all you[2040.96] [2040.96][S09]have to do to replicate what you just[2042.52] [2042.52][S09]did in AI Studio is click Get SDK Code,[2045.5] [2045.5][S09]and you're kind of off to the races.[2047.12] [2047.12][S08] Absolutely extraordinary.[2048.34] [2048.34][S08]Are there other things that you can do in AI Studio[2050.465] [2050.465][S08]that you can't do in the Gemini app?[2053.3] [2053.3][S09] So Gemini Live is kind[2055.69] [2055.69][S09]of Project Astra baked directly within the UI of AI Studio.[2059.527] [2059.527][S08] And Project Astra, of course, we[2061.36] [2061.36][S08]did a whole episode on this with Greg Wayne.[2063.949] [2063.949][S08]This is the real-time visual understanding, the first attempt[2068.8] [2068.8][S08]at a universal AI assistant.[2070.608] [2070.608][S09] Absolutely.[2071.65] [2071.65][S09]It can see anything that you see.[2073.297] [2073.297][S09]It can talk to you in real time.[2074.63] [2074.63][S09]It can talk to you in multiple languages.[2076.989] [2076.989][S09]And you can also add additional tool calls to this process.[2082.73] [2082.73][S09]So if you wanted Gemini to be able to pull out and extract[2085.63] [2085.63][S09]up-to-date information, if you wanted[2087.699] [2087.699][S09]it to be able to interact with any of the apps[2090.13] [2090.13][S09]or any of the kind of products that you[2092.739] [2092.739][S09]use every day, something like Calendar, or Docs, or Sheets,[2095.679] [2095.679][S09]or Gmail, you can give it access to be[2098.29] [2098.29][S09]able to use all of those features[2102.07] [2102.07][S09]within the context of the Gemini Live API.[2105.26] [2105.26][S08] OK, show me then, give me an example[2107.26] [2107.26][S08]then of this working.[2108.135] [2108.135][S09] Yeah.[2108.927] [2108.927][S09]So let's turn on Grounding with Google Search,[2110.95] [2110.95][S09]and then also Share Screen with Gemini.[2114.32] [2114.32][S09]So you can share your screen, you can share your Webcam.[2117.4] [2117.4][S09]And of course, you can Talk with the model.[2119.38] [2119.38][S09]But let's go ahead and share our screen.[2121.88] [2121.88][S09]And I'm going to select one of these Chrome tabs[2124.57] [2124.57][S09]that we currently have open.[2127.03] [2127.03][S09]And let's take a look at that video that we saw with Doug Eck[2130.84] [2130.84][S09]and see what it has to say.[2134.11] [2134.11][S09]Hey there, Gemini.[2135.14] [2135.14][S09]Could you tell me what you see on the screen?[2137.355] [2141.37][S03] I see a YouTube video playing.[2143.59] [2143.59][S03]A woman with red hair is speaking and subtitles are[2146.38] [2146.38][S03]displayed on the screen saying, \"Hannah Fry,[2149.12] [2149.12][S03]welcome to 'Google DeepMind, The Podcast,' with me.\"[2151.98] [2151.98][S03]The video appears to be about AI and creativity.[2155.78] [2155.78][S09] Amazing.[2156.71] [2156.71][S09]So tell me a little bit about what you[2159.53] [2159.53][S09]think the video might cover.[2161.25] [2161.25][S09]And also what is Google DeepMind's latest[2164.9] [2164.9][S09]models for creative teams?[2166.48] [2170.9][S09]And so you can see here it's expanding out its thoughts,[2174.47] [2174.47][S09]and it's realizing that it needs to make a search.[2176.72] [2176.72][S03] --AI, your new creative muse,[2178.34] [2178.34][S03]and the speakers introduction of the \"Google DeepMind,[2180.59] [2180.59][S03]The Podcast,\" it's likely the video explores the potential[2183.86] [2183.86][S03]of AI in creative fields.[2185.58] [2185.58][S03]It may feature discussions on how AI can inspire or assist[2188.78] [2188.78][S03]creative professionals.[2190.23] [2190.23][S03]Regarding Google DeepMind's latest models[2192.17] [2192.17][S03]for creative teams, they've introduced various tools,[2195.33] [2195.33][S03]including Veo 3 for video generation,[2197.72] [2197.72][S03]Imagen 4 for image creation, and Lyria 2 for music composition.[2201.96] [2201.96][S03]They've also developed Flow, an AI filmmaking tool,[2205.29] [2205.29][S03]and are working on world modeling AI models[2207.33] [2207.33][S03]with applications in gaming and entertainment.[2209.8] [2209.8][S03]These advancements aim to empower artists and creators.[2212.963] [2212.963][S09] Amazing.[2213.88] [2213.88][S08] Amazing.[2214.3] [2214.3][S09] Yeah.[2215.092] [2215.092][S09]And so one of the things that you can see here[2217.2] [2217.2][S09]is the thinking trace has been exposed.[2220.87] [2220.87][S09]So it's kind of walking through all of Gemini's thought process[2224.73] [2224.73][S09]on how to answer the question.[2226.39] [2226.39][S09]It realizes that it needs to invoke one of the tools[2229.23] [2229.23][S09]that it has access to, which is Google Search.[2231.73] [2231.73][S09]So it needs to look up some information.[2233.8] [2233.8][S09]Once it gets the responses back from its search,[2236.98] [2236.98][S09]it incorporates that back into its summary,[2239.14] [2239.14][S09]and then it generates this wonderful audio clip[2242.16] [2242.16][S09]with insights that it's captured around all of the products[2246.42] [2246.42][S09]that Google has recently released.[2248.11] [2248.11][S08] So this stuff, this Gemini Live stuff,[2249.79] [2249.79][S08]you also have it on Android phones, right?[2251.55] [2251.55][S08]So actually, I spotted this on my phone a little while ago[2254.85] [2254.85][S08]and have been using it extensively.[2258.15] [2258.15][S08]Let's give it a go.[2259.3] [2259.3][S08]Hang on.[2260.82] [2260.82][S08]What can you see right now, Gemini?[2265.12] [2265.12][S03] OK.[2265.82] [2265.82][S03]I see a woman with blonde hair sitting[2267.55] [2267.55][S03]at a desk with a microphone in front of her.[2269.81] [2269.81][S03]She has her arms crossed and is smiling.[2272.18] [2272.18][S03]There's a computer keyboard and mouse on the desk,[2274.48] [2274.48][S03]and a lamp in the background.[2275.96] [2275.96][S09] It's amazing.[2276.32] [2276.32][S03] What else can you tell me about--[2278.028] [2278.028][S08] That has been so useful in diagnosing[2280.27] [2280.27][S08]my sick houseplants, I can't tell you.[2281.993] [2281.993][S09] Oh, wow.[2282.91] [2282.91][S09]And it's also just been so cool to have something[2286.06] [2286.06][S09]that feels like a real life commentator[2290.74] [2290.74][S09]on as you go about your daily life.[2293.57] [2293.57][S09]There's been something really, really[2297.13] [2297.13][S09]fascinating too about using it as a pair programmer.[2300.82] [2300.82][S09]One thing that I commonly use this[2302.86] [2302.86][S09]for is we have a tool at Google called Google Colab,[2307.43] [2307.43][S09]which is a notebook-spaced environment with Python code.[2310.72] [2310.72][S09]You can run code, you can execute it.[2314.86] [2314.86][S09]But Gemini also understands Colab features.[2319.97] [2319.97][S09]So if we want to share this tab instead, you could even say,[2326.44] [2326.44][S09]\"hey, Gemini, what do you see on the screen now?[2329.3] [2329.3][S09]And if you see any code, please explain what it does.\"[2332.51] [2332.51][S08] Oh, nice.[2333.5] [2333.5][S09] Yeah.[2334.292] [2337.48][S03] I see a Google Colab notebook on the screen titled,[2340.84] [2340.84][S03]\"Making the Most of Your Colab Subscription.\"[2343.31] [2343.31][S03]It seems to provide tips for optimizing Colab usage,[2346.52] [2346.52][S03]focusing on faster GPUs and more memory.[2349.37] [2349.37][S03]There are a couple of code snippets visible.[2351.62] [2351.62][S03]The first one appears to check if a GPU is connected[2354.49] [2354.49][S03]and display GPU information.[2356.54] [2356.54][S03]The second code uses the psutil library[2359.5] [2359.5][S03]to determine the total amount of RAM available in the runtime,[2362.48] [2362.48][S03]and presents that in gigabytes.[2363.795] [2363.795][S09] So it's like having[2365.17] [2365.17][S09]a helpful assistant that somehow understands every single thing--[2368.317] [2368.317][S08] That you could possibly be looking at.[2370.4] [2370.4][S09] Absolutely.[2371.442] [2371.442][S09]And that also is capable of answering,[2375.85] [2375.85][S09]in an empathetic way, all of the questions that you might have[2380.038] [2380.038][S09]or that a curious human might have about the things that they[2382.58] [2382.58][S09]see every day.[2383.4] [2383.4][S08] So this is good.[2384.06] [2384.06][S08]I mean, I already use Google Colab with my students.[2386.19] [2386.19][S09] Oh, amazing.[2387.273] [2387.273][S08] So thank you for that.[2388.7] [2388.7][S08]That's going to be very helpful.[2390.033] [2390.033][S08]But for people who can't necessarily code themselves,[2394.22] [2394.22][S08]I mean, there's another offering within AI Studio.[2396.428] [2396.428][S09] Absolutely.[2397.47] [2397.47][S09]So in addition to being able to take the Gemini APIs[2401.48] [2401.48][S09]and embed them within tools like Cursor, or Windsurf, or Copilot,[2406.64] [2406.64][S09]or any of the other many coding IDEs[2409.432] [2409.432][S09]that you might have access to if you're attempting to write code,[2412.14] [2412.14][S09]we also have something called the new Build[2414.32] [2414.32][S09]feature that allows you to build apps directly with Gemini.[2418.19] [2418.19][S09]So the Build Apps with Gemini section[2421.7] [2421.7][S09]is unique in that all of the code that[2424.37] [2424.37][S09]gets generated within this UI is hyper[2426.56] [2426.56][S09]optimized for the latest SDKs, the latest code[2429.83] [2429.83][S09]that we have for interfacing with the Gemini models.[2432.677] [2432.677][S08] So just to put that another way,[2434.51] [2434.51][S08]even if you have never written a line of code in your life,[2438.33] [2438.33][S08]you are going to have slick code if you just prompt Gemini here[2441.827] [2441.827][S08]to build you an app.[2442.66] [2442.66][S09] Absolutely.[2443.13] [2443.13][S09]And it will be using the latest models, the latest[2445.38] [2445.38][S09]features from the models.[2447.12] [2447.12][S09]It will be generating really, really robust TypeScript code,[2451.33] [2451.33][S09]and it will also be resolving any errors along the way.[2454.99] [2454.99][S09]So if the model hits any problems, it's able to cycle[2457.89] [2457.89][S09]back through, fix the error to get the models[2460.62] [2460.62][S09]kind of implementation of an app and a clear and coherent state.[2463.94] [2463.94][S08] Self-healing code.[2465.19] [2465.19][S09] Self-healing code, and all with you just[2468.12] [2468.12][S09]having to describe what you would like to see.[2470.61] [2470.61][S09]This is interesting too in the sense[2472.5] [2472.5][S09]that a lot of the vibe coding demos[2475.59] [2475.59][S09]that you might have seen externally[2477.06] [2477.06][S09]are really, really focused on just creating a simple app.[2480.07] [2480.07][S09]So something that's not AI-enriched.[2482.31] [2482.31][S09]This is hyper optimized for creating[2484.65] [2484.65][S09]apps that are also using the Gemini[2486.42] [2486.42][S09]APIs, or Imagen, or the like.[2488.82] [2488.82][S09]So it's really, really cool to see[2490.77] [2490.77][S09]what you can cook up using the Build Apps with Gemini.[2494.29] [2494.29][S09]If we have an idea for an app around creating a--[2498.742] [2498.742][S08] Spaghetti competitions.[2500.2] [2500.2][S09] Spaghetti competitions.[2501.742] [2501.742][S09]This is also correct.[2502.65] [2502.65][S09]So maybe we say, like, generate an application that[2509.61] [2509.61][S09]showcases a spaghetti competition or spaghetti eating[2514.95] [2514.95][S09]contest.[2516.42] [2516.42][S09]Make sure to use Imagen to generate[2523.08] [2523.08][S09]example images for the website landing photo,[2528.97] [2528.97][S09]as well as other photos about the competition.[2534.6] [2534.6][S09]Also, make sure to include a sign-up form for people[2541.65] [2541.65][S09]to indicate interest.[2543.263] [2543.263][S08] So this is essentially[2544.68] [2544.68][S08]going to make a website where all of the images[2547.44] [2547.44][S08]are AI-generated.[2549.91] [2549.91][S09] All of the images are AI-generated.[2552.39] [2552.39][S09]It should have a form where people[2554.1] [2554.1][S09]can submit their information, if they would like to see it.[2557.59] [2557.59][S09]And it also has a kind of immediately launched us[2561.64] [2561.64][S09]into this new IDE with a code assistant, kind of a directory[2565.78] [2565.78][S09]structure here in the middle, and then[2567.64] [2567.64][S09]also a place for you to visualize the app as it gets[2571.12] [2571.12][S09]created here off to the right.[2572.63] [2572.63][S09]And so this thinking trace box, if we expand it out,[2575.72] [2575.72][S09]you can see that Gemini is doing a lot of work behind the scenes[2579.25] [2579.25][S09]to build the kind of core features of the app,[2583.03] [2583.03][S09]to map out the structure, to build the components.[2587.78] [2587.78][S09]And it's writing every single one of these code files[2591.37] [2591.37][S09]into being as we sit here watching, including everything[2596.92] [2596.92][S09]from design patterns, to things like the latest React libraries,[2604.09] [2604.09][S09]to even doing error handling for you,[2607.03] [2607.03][S09]so you don't have to think through all of the things[2609.28] [2609.28][S09]that you would have to consider if building out[2612.13] [2612.13][S09]a native application.[2613.295] [2613.295][S08] OK, so I am someone who has built[2615.17] [2615.17][S08]websites in the past.[2616.52] [2616.52][S08]And honestly, even with the tools[2619.478] [2619.478][S08]that are supposed to make it quicker,[2621.02] [2621.02][S08]weeks and weeks, weeks and weeks and weeks this would take.[2624.14] [2624.14][S08]And my websites were rubbish.[2625.495] [2625.495][LAUGHTER][2626.96] [2626.96][S08]Here we go.[2627.777] [2627.777][S09] Oh, my gosh.[2628.86] [2628.86][S08] It's chosen Comic Sans.[2630.42] [2630.42][S08]I love it.[2631.02] [2631.02][S09] It absolutely has chosen Comic Sans.[2633.103] [2633.103][S09]And so it has the contest.[2635.93] [2635.93][S09]It's got some images that it looks like are[2639.35] [2639.35][S09]in the process of loading.[2640.89] [2640.89][S08] Oh, look at this.[2642.18] [2642.18][S09] Oh, my gosh.[2643.41] [2643.41][S09]And then if I give it a little bit more room,[2645.853] [2645.853][S09]so let's hide this code assistant.[2647.27] [2647.27][S08] A fake trophy.[2647.94] [2647.94][S09] Yeah, it's got a fake trophy.[2649.98] [2649.98][S09]It's got the contest, including a background image[2652.43] [2652.43][S09]that it generated of the contest,[2653.9] [2653.9][S09]and then also changed the transparency of the image[2658.58] [2658.58][S09]so that it matched this very interesting stylistic choice[2662.3] [2662.3][S09]that it's created.[2664.35] [2664.35][S09]And it also has some contest highlights[2666.8] [2666.8][S09]for people who it looks have previously[2671.42] [2671.42][S09]won the contest before.[2673.14] [2673.14][S09]And all of these images, by the way, they're using Imagen 3.[2676.31] [2676.31][S09]So not Imagen 4.[2678.215] [2678.215][S08] But you could also, at this point[2680.09] [2680.09][S08]then, let's say that there was, I don't know,[2681.965] [2681.965][S08]some video of a previous spaghetti eating competition,[2685.01] [2685.01][S08]and you wanted to have a page, which[2686.84] [2686.84][S08]was a story about that moment.[2688.91] [2688.91][S08]You could upload a YouTube video,[2691.16] [2691.16][S08]have Gemini Live essentially watch it for you,[2695.03] [2695.03][S08]summarize it, and then embed it into the website?[2697.443] [2697.443][S09] Absolutely.[2698.485] [2698.485][S09]You could do all of that work.[2700.44] [2700.44][S09]You could even use the Gemini search capabilities.[2703.53] [2703.53][S09]So finding examples of related information[2707.93] [2707.93][S09]to spaghetti eating competitions and citing[2710.12] [2710.12][S09]all of those sources for the video itself.[2711.88] [2711.88][S08] And if you wanted to turn this[2713.63] [2713.63][S08]into an actual website, what do you do then?[2717.24] [2717.24][S09] So to turn it into an actual website, something[2719.93] [2719.93][S09]that you could share with your friends[2721.513] [2721.513][S09]and not just share the file itself, what you can do[2724.85] [2724.85][S09]is there's this little rocket here off[2726.83] [2726.83][S09]to the side that says Deploy to Cloud Run.[2730.53] [2730.53][S09]Since this has been integrated so closely with the Google Cloud[2734.64] [2734.64][S09]kind of development environment, what you can do[2737.97] [2737.97][S09]is say Deploy to Cloud Run.[2740.46] [2740.46][S09]And so after the app gets created,[2743.89] [2743.89][S09]you can select a project.[2745.66] [2745.66][S09]So just one of the projects that you have access[2748.26] [2748.26][S09]to via Cloud Run.[2752.1] [2752.1][S09]It verifies the project that you could actually use, and then[2755.7] [2755.7][S09]says that you can deploy it, and gives you[2758.04] [2758.04][S09]a unique URL that you can share with friends that's scalable.[2762.13] [2762.13][S09]And that, even more importantly, is hiding all of those APIs keys[2765.45] [2765.45][S09]so that you can use them and your app can use them,[2767.65] [2767.65][S09]but nobody else would be able to capture them.[2770.188] [2770.188][S08] What does this mean for developers, though?[2772.48] [2772.48][S08]I mean, you've just done some really extraordinary things[2775.29] [2775.29][S08]very, very quickly.[2776.49] [2776.49][S09] Yes.[2777.24] [2777.24][S09]Like, I do feel like for developers, this[2780.27] [2780.27][S09]means that they can focus more on building,[2782.22] [2782.22][S09]and ideation, and the product experience, as opposed to--[2786.33] [2786.33][S09]the daily life of a developer right[2789.28] [2789.28][S09]now is a lot of things that aren't necessarily[2792.82] [2792.82][S09]the most exciting in the world.[2794.69] [2794.69][S09]So you might be upgrading a code base[2796.48] [2796.48][S09]from one version to another, or you[2799.3] [2799.3][S09]might be adding typing to a repository,[2802.52] [2802.52][S09]or you might need to periodically check[2805.063] [2805.063][S09]for security vulnerabilities and make sure your code base aren't[2807.73] [2807.73][S09]susceptible to them.[2809.53] [2809.53][S09]And all of these things, they aren't joyful.[2812.93] [2812.93][S09]It's kind of tidying up an apartment[2815.17] [2815.17][S09]to make sure that your app is kind of sustainable[2818.53] [2818.53][S09]and maintained.[2820.0] [2820.0][S09]And there's that saying that everybody wants to build[2822.7] [2822.7][S09]and nobody wants to do maintenance.[2824.27] [2824.27][S09]I think one of the promises of these models[2826.87] [2826.87][S09]and these kinds of capabilities is[2828.97] [2828.97][S09]that there are so many more opportunities for software[2831.91] [2831.91][S09]developers to build and to really[2835.09] [2835.09][S09]build more ambitious systems.[2837.52] [2837.52][S09]And even more importantly, it opens up[2839.8] [2839.8][S09]the door for more people to learn and get[2842.65] [2842.65][S09]inspired by this whole process of creating[2845.9] [2845.9][S09]and getting something out into the world.[2847.88] [2847.88][S08] Do you think all of these tools[2849.02] [2849.02][S08]together then, do you think that they will fundamentally[2851.6] [2851.6][S08]change the way we think about human creativity?[2853.978] [2853.978][S09] Absolutely.[2855.02] [2855.02][S09]Like, human creativity is about to have[2858.14] [2858.14][S09]this explosion of progress.[2861.45] [2861.45][S09]And there's this promise of everyone[2863.36] [2863.36][S09]being able to become a creator, and not just a creator in one[2867.8] [2867.8][S09]specific discipline, but to also be[2869.57] [2869.57][S09]able to expand to that out into many other disciplines.[2872.82] [2872.82][S09]So I think that there's real magic in having people[2876.8] [2876.8][S09]who are scientists maybe in the physical sciences,[2879.56] [2879.56][S09]or chemistry or biology, or people who are historians[2883.67] [2883.67][S09]or musicians, being able to suddenly get[2887.03] [2887.03][S09]all of their ideas or their projects[2890.0] [2890.0][S09]into a digital form that could be shared with others.[2892.85] [2892.85][S08] Paige, absolutely fascinating.[2894.6] [2894.6][S08]Thank you so much for joining me.[2895.95] [2895.95][S09] It was awesome to be here.[2897.617] [2897.617][S09]Thank you so much.[2898.44] [2898.44][S08] With these new tools,[2900.06] [2900.06][S08]it feels like suddenly everything has clicked together.[2902.96] [2902.96][S08]Until now, we've spoken to researchers[2904.64] [2904.64][S08]about every individual element of these new releases, audio,[2908.82] [2908.82][S08]video, language.[2910.8] [2910.8][S08]But there is something so different about having[2913.52] [2913.52][S08]all of those elements integrated and working seamlessly together.[2917.19] [2917.19][S08]And I have been coming here for years.[2919.67] [2919.67][S08]I have spoken to the people all along the way.[2922.43] [2922.43][S08]But even still, seeing these tools come to life[2927.11] [2927.11][S08]and imagining the possibilities, I still feel completely wowed.[2930.69] [2930.69][S08]And to be honest with you, I'm now[2932.78] [2932.78][S08]itching to get on the train home and just[2934.488] [2934.488][S08]try out all of the things that I've always wanted to build,[2936.947] [2936.947][S08]but never had time to.[2937.95] [2937.95][S08]You have been listening to \"Google DeepMind, The Podcast,\"[2940.55] [2940.55][S08]with me, Professor Hannah Fry.[2942.2] [2942.2][S08]If you enjoyed this episode, then[2943.94] [2943.94][S08]do subscribe to our YouTube channel[2945.8] [2945.8][S08]or leave a review on your favorite podcast platform.[2948.45] [2948.45][S08]And of course, we have plenty more episodes[2951.32] [2951.32][S08]on a whole range of topics to come, so do--[2954.29] [2954.29][MUSIC PLAYING][2957.94]"}
{"file_name": "audio/val_000001.wav", "transcription": "[5.439][S01] Artificial intelligence is slowly appearing[8.208] [8.208][S01]in every aspect of our modern lives.[10.878] [10.878][S01]It’s in our smart phones, our central heating,[13.847] [13.847][S01]on our sideboards and in our cars.[17.084] [17.084][S01]But what about artificial general intelligence?[20.354] [20.354][S01]That is the real quest.[22.656] [22.656][S01]The aim to build an agent - an algorithm -[25.759] [25.759][S01]that can learn to solve any problem from scratch without being taught how.[31.365] [34.968][S01]Welcome to Deep Mind - the podcast. I’m Hannah Fry,[39.706] [39.706][S01]I’m a mathematician who has worked with algorithms for almost a decade.[44.077] [45.012][S01]In this series of podcasts,[46.98] [46.98][S01]we’re following the fast moving story of artificial intelligence.[51.852] [51.852][S01]For the past 12 months we’ve been tracking the latest work[55.088] [55.088][S01]of scientists, researchers and engineers at DeepMind in London.[60.194] [60.194][S01]We’re looking at how they’re approaching the science of AI[63.864] [63.864][S01]and some of the tricky decisions[65.566] [65.566][S01]the whole field is wrestling with at the moment.[68.836] [68.836][S01]So whether you want to know more about where technology is headed,[73.106] [73.106][S01]or want to be inspired on your own AI journey,[76.243] [76.243][S01]then you’ve come to the right place.[78.345] [79.479][S01]Now in the last episode we were talking about[82.082] [82.082][S01]how pitting artificial intelligence[84.384] [84.384][S01]against world class players in the game of Chess,[88.055] [88.055][S01]and the game of Go,[89.59] [89.59][S01]is about much more than just showing off what a computer can do.[93.66] [93.66][S01]Human players can learn from how the AI plays[97.097] [97.097][S01]and improve their own play as a result. And there’s also a bigger picture -[102.336] [102.336][S01]the world of games provides the perfect mini universe[105.772] [105.772][S01]to try out everything we want our artificial intelligence to do.[109.81] [110.577][S01]But intelligence is much more than just championing raw logic,[114.548] [114.548][S01]intelligence requires other skills like the ability to collaborate.[120.053] [121.221][S01]I want to introduce research scientist Max Jaderburg.[124.691] [124.691][S01]Max and his colleagues are trying to work out how to train agents[128.629] [128.629][S01]to work together as a team.[130.531] [130.531][S01]02:11[131.732] [131.732][S03] So imagine it’s a few decades in the future -[133.967] [133.967][S03]we have all these AI systems out in the world doing different things[138.071] [138.071][S03]but they maybe have never seen each other before.[139.84] [139.84][S03]There’s thousands of these things, hundreds of thousands,[142.242] [142.242][S03]each have their own objectives[143.41] [143.41][S03]but somehow they have to cooperate and compete in a sensible way[147.214] [147.214][S03]and in a very ad hoc way, in a way that they’ve never seen each other before.[150.517] [150.517][S01] Humans are really good at this when we want to be, anyway.[154.655] [154.655][S01]Even when we haven’t encountered another person before,[157.09] [157.09][S01]we still know how to understand their intentions[160.327] [160.327][S01]and how to interact with them.[162.529] [162.529][S01]Our agents of the future need to be able to do the same thing with each other.[167.034] [167.034][S03] We already have things like Google Home[169.469] [169.469][S03]and these sort of smart devices out there.[171.305] [171.305][S03]We probably have more and more of those and you can imagine them[174.208] [174.208][S03]having to interact and work with each other[177.644] [177.644][S03]and one device may not have ever seen another device before,[180.781] [180.781][S03]but they still somehow have to interact and get things done for you.[183.951] [183.951][S01] Are we talking about like your Google Home[187.287] [187.287][S01]and your dishwasher here - this kind of stuff or. -[189.556] [189.556][S03] Yeah, potentially. You know your dishwasher[191.291] [191.291][S03]might want to actually go on its cleaning cycle[193.627] [193.627][S03]but Google Home wants it to you know clean all the dishes and so [laughs][197.631] [197.631][S03]what’s best for you as a person - I don’t know?[199.566] [199.566][S01] And who gets to decide? [S03] Who gets to decide?[201.502] [201.502][S01] Who rules supreme - your dishwasher or your Google assistant?[204.805] [204.805][S03] yeah. I don’t know.[206.106] [206.106][S01] There’s an important distinction here.[208.141] [208.141][S01]If you’ve got a smart light bulb that you can program to come on[211.278] [211.278][S01]at six o’clock in the evening, that is an algorithm.[215.382] [215.382][S01]If you’ve got one that can learn your preferences,[217.584] [217.584][S01]that can understand when you like the lights to be dimmed,[221.121] [221.121][S01]what kind of mood lighting you like when you’re reading, that is AI.[226.527] [226.527][S01]But as we switch away from building things that do rigid, pre-decided tasks,[231.665] [231.665][S01]we are asking our technology to read the situation and react[235.435] [235.435][S01]to what’s going on around it.[237.604] [237.604][S01]And in the long-term, that’s going to require collaboration.[242.276] [242.276][S01]So in the spirit of trying things out in a toy universe,[246.079] [246.079][S01]the team at DeepMind have been trying to find inspiration[248.749] [248.749][S01]in another kind of game.[250.784] [250.784][S01]One taken straight from the school playground.[253.921] [256.49][S01] this is Capture The Flag - you know the deal -[259.993] [259.993][S01]the first team to steal the Flag of their opponent[262.396] [262.396][S01]and bring it back to their own base woods.[264.998] [264.998][S01]You get tagged by the opposition, then you’re out of the game.[269.069] [269.069][S01]Oh, come on! Don’t cry.[271.405] [271.405][S01]Max dropped whole populations of AI agents[274.708] [274.708][S01]into a digital version of the game.[277.144] [277.144][S03] This is an onscreen version,[278.345] [278.345][S03]you sort of you just see your first person point of view,[281.281] [281.281][S03]so you have to sort of look around and move through this 3D world[286.353] [286.353][S03]from your own first person perspective, but interact with these other things[289.156] [289.156][S03]which suit their own first person perspective.[291.525] [291.525][S03]So here there’s no centralized entity or being that can see everything.[295.462] [295.462][S01] No army commander. [S03] No army commander.[296.997] [296.997][S03]Every player acts independently.[299.166] [299.166][S03]They only see their own observation and the way that we train[302.569] [302.569][S03]these things we actually train whole populations of teammates -[305.706] [305.706][S03]let’s say 30 agents in parallel[308.075] [308.075][S03]and they are all playing with and against each other.[311.478] [313.146][S01] Rather than just creating a single agent,[315.916] [315.916][S01]Max and his team build an entire classroom of them - 30 in total.[321.255] [321.255][S01]And for each round of the game, he randomly selects[324.057] [324.057][S01]a few of the agents from the class to play together on the team.[328.295] [328.295][S01]By doing this thousands and thousands of times,[331.298] [331.298][S01]each agent will learn from their own experience,[333.834] [333.834][S01]but because they are playing with each other too -[336.203] [336.203][S01]with their classmates as it were - they have to learn to interact[339.806] [339.806][S01]with someone who’s different from themselves.[342.075] [342.075][S03] The problem is when we start it’s actually just very random.[344.578] [344.578][S01] yeah.[344.878] [344.878][S03] They’re just bouncing about the place.[347.047] [347.047][S03]Without a clue [Hannah laughs][348.315] [348.315][S03]and then one of them will discover something[350.951] [350.951][S03]and will start actually let’s say taking control of the flag[355.189] [355.189][S03]and actually scoring points,[356.924] [356.924][S03]and at that point there’s evolutionary pressure on this population.[360.494] [364.865][S01] And here’s the clever bit - Max and his team[367.434] [367.434][S01]aren’t just letting the agents in the classroom play on and on forever,[371.371] [371.371][S01]they’re also using something called a genetic algorithm.[375.242] [375.242][S01]A way to make sure the whole culture of the population of agents evolves.[379.98] [379.98][S03] So actually some of the weaker ones[382.015] [382.015][S03]will be removed from this population.[383.417] [383.417][S01] So it’s almost like you’re making that population of 30 have children.[387.955] [387.955][S03] Yeah, absolutely.[388.488] [388.488][S01] You’re sort of breeding them together.[389.656] [389.656][S03] yeah.[391.358] [393.46][S01] The original classroom of agents breeds together[396.663] [396.663][S01]and have kids of their own.[398.799] [398.799][S01]And as you go down the generations, the strongest traits survive.[403.67] [403.67][S03] But unlike human children, when an agent has children[406.974] [406.974][S03]in this set-up, they inherit everything,[408.942] [408.942][S03]they inherit the knowledge that’s been gained from their parent.[412.045] [412.045][S01] But you’re mixing up the characteristics[414.114] [414.114][S01]as you go from one generation to the next.[415.649] [415.649][S03] Yeah, so this agent has to learn to play[417.551] [417.551][S03]a 5 minute game of Capture The Flag - which is really -[420.587] [420.587][S03]you play 5 minutes, you do thousands of actions,[423.056] [423.056][S03]and you just get a win or a loss whether you have won or lost the game.[426.56] [426.56][S03]Somehow we have to learn what to do with that,[429.296] [429.296][S03]and so to help bridge that problem, we have this idea of internal rewards[433.133] [433.133][S03]where the events in the game such as picking up a flag,[437.738] [437.738][S03]or dropping a flag, or your teammate tagging an opponent[441.742] [441.742][S03]or an opponent tagging you, all these sort of things,[445.612] [445.612][S03]and we allow the agents to individually evolve their own internal rewards,[450.851] [450.851][S03]which is the reward they have assigned to each one of these events.[453.453] [453.453][S01] So some agents are going to care a great deal[455.756] [455.756]about grabbing hold of the flag - [S03] Yeah, exactly -[457.791] [457.791][S01] and other agents are going to care a lot about teammate tagging someone.[461.328] [461.328][S03] Yeah.[463.03] [463.53][S01] This kind of evolutionary group training means[466.4] [466.4][S01]that they can assume different roles.[468.969] [468.969][S01]Producing better results for stealing a flag.[472.306] [472.306][S01]And with a bit of practice, after a few thousand rounds say,[475.776] [475.776][S01]teams of agents become really rather good at this game.[480.48] [481.048][S03] They absolutely smashed it.[482.049] [482.049][S03]And the great thing about training an agent[484.418] [484.418][S03]in this manner is that the robust - yes they can play themselves[487.521] [487.521][S03]but they can play other agents that have been trained[490.023] [490.023][S03]in completely different regimes, they can also play these in-game bots[494.394] [494.394][S03]which are sort of these hard coded bots that shift with the game,[497.598] [497.598][S03]but most interestingly, they can also play with people -[500.167] [500.167][S03]so you can drop people into these games[502.402] [502.402][S03]and have you know an AI teammate or AI opponents.[505.873] [505.873][S01] What was it actually like to play with an agent then - do they -[509.71] [509.71][S01]does it feel like they are guessing what you are going to do[513.447] [513.447][S01]as well as doing their own thing?[515.415] [515.415][S03] it feels less like they are guessing what you are doing[517.417] [517.417][S03]and more like they completely ignore you and they are very ruthless.[521.021] [521.021][S03]Humans pay a lot of attention to other humans -[523.59] [523.59][S03]even in game scenarios like humans will fixate[525.559] [525.559][S03]on the other players of the game,[526.894] [526.894][S03]but these agents have been trained completely unbiased[529.763] [529.763][S03]without these sort of human biases -[532.165] [532.165][S03]your opponent will run right past you and not even try and tag you[535.402] [535.402][S03]because they’re so fixated on actually getting the flag as quickly as possible[538.605] [538.605][S03]because that’s what’s going to maximize their number of flag captures[541.475] [541.475][S03]and win them the game.[542.709] [542.709][S03]Things that really annoying human players would do.[545.279] [546.713][S01] There’s a kind of magic going on here.[549.383] [549.383][S01]Initially researchers are working on these agents[551.718] [551.718][S01]trying to see a way through the muddle.[554.087] [554.087][S01]Then there is that breakthrough moment when the agent gets it.[558.025] [558.025][S01]When they start to behave like you think they should.[561.228] [562.329][S01] Let me tease you with Koray Kavukcuoglu,[565.032] [565.032][S01]director of research at DeepMind.[567.301] [567.301][S04] I remember training agents in the early days -[571.104] [571.104][S04]the first time actually those agents started behaving like okay,[576.777] [576.777][S04]it’s an environment, it’s trying to navigate,[578.946] [578.946][S04]it’s trying to avoid certain obstacles and what not -[581.715] [581.715][S04]the first time it starts doing that it’s actually-[584.852] [584.852][S04]it is nice, it’s like it’s quite fun to see that[587.821] [587.821][S04]because you know that it makes a decision for itself.[591.925] [591.925][S04]It think knowing that you have created an algorithm that can take decisions[597.464] [597.464][S04]I think that aspect is quite enjoyable.[599.132] [599.132][S04]That is very satisfactory. 10:00[600.534] [600.534][S01] It’s worth remembering that these games aren’t just a trivial pursuit for[604.304] [604.304][S01]DeepMind, they’ve invested in this rigorous training for a reason.[608.842] [608.842][S01]They want to see how an AI develops these kinds of skills for itself.[614.648] [614.648][S03] We spent a lot of time in this[616.383] [616.383][S03]Capture The Flag work looking into the the neural networks of these agents[620.521] [620.521][S03]to try and understand what they care about[621.989] [621.989][S03]and how they represent the game world.[624.157] [624.157][S03]And what was really cool is that we found that the agents[626.026] [626.026][S03]actually have a really, really rich representation of this game world[629.296] [629.296][S03]without being told anything about the game world itself.[631.231] [631.231][S03]You know it’s just look at the pixels of the screen[634.168] [634.168][S03]yet somehow they have clustered[635.569] [635.569][S03]the you know internal activations into things like oh[639.94] [639.94][S03]I’m in my home base, I’m in my opponent base,[642.943] [642.943][S03]I’ve got the flag and I can see my teammates ahead of me.[646.58] [646.58][S03]I’m looking at the opponent flag carrier while my teammates are holding my flag.[650.584] [650.584][S03]And you can even find individual neurons[652.319] [652.319][S03]which just activate if for example your teammates are holding the flag.[655.956] [655.956][S01] You can totally understand[657.824] [657.824][S01]how the agent is seeing the game as you go through.[661.328] [661.328][S03] I’m not sure about totally understand[663.063] [663.063][S03]but we’re really getting an idea of what is being represented strongly[667.1] [667.1][S03]and what isn’t being represented strongly.[669.203] [672.539][S01] Max’s agents are using something called neural networks.[676.009] [676.009][S01]It’s a type of machine learning algorithm that is loosely based[680.08] [680.08][S01]on a simplified version of the human brain.[683.383] [683.383][S01]Layers on layers of artificial neurons are connected together in a vast network[689.523] [689.523][S01]and fire information between themselves.[692.593] [692.593][S01]By looking inside an agent’s electronic brain,[696.763] [696.763][S01]Max can work out which micro level connections are responsible[701.068] [701.068][S01]for what macro level behavior. And this can be hugely beneficial[706.306] [706.306][S01]as AI becomes more integrated in our everyday lives.[710.878] [710.878][S03] The hope is that well into the future[713.78] [713.78][S03]we can start actually having agents which can go out into the real world[716.65] [716.65][S03]that can interact with humans, with other agents -[720.921] [720.921][S01] Without fighting -[722.022] [722.022][S03] Without fighting - being sensible, yeah.[724.992] [724.992][S03][Hannah chuckles] Not squabbling too much.[726.493] [726.493][S01] Unlike humans. [S03] yeah, exactly.[729.53] [731.899][S01] Games without frontiers team work without tears.[736.336] [736.336][S01]But there is a big leap between board games[738.438] [738.438][S01]or simple games like Capture The Flag[741.108] [741.108][S01]and the big bad world with all of its complexity and messiness.[746.446] [747.781][S01]You’ll remember David Silver - the man who brought us AlphaGo -[751.618] [751.618][S01]the agent that defeated the world champion[754.288] [754.288][S01]at the ancient board game of Go.[756.857] [756.857][S01]While he’s also involved in pushing DeepMind’s AI[760.727] [760.727][S01]into ever-more perplexing environments.[763.363] [763.363][S06] In the context of games, I think there is a further challenge[766.366] [766.366][S06]which is many people in the community are moving towards,[769.67] [769.67][S06]which is to take the most challenging computer game -[773.44] [773.44][S06]in this case it’s the game of StarCraft - and many people in the AI community[777.678] [777.678][S06]are viewing this as the next Grand Challenge -[780.013] [780.013][S06]now how can we actually devise agents[782.349] [782.349][S06]which can play in this very rich environment[785.485] [785.485][S06]which has challenges which are not only different[788.722] [788.722][S06]but many times vaster than Go in other ways.[791.992] [793.66][S01] This is DeepMind the podcast - an introduction to AI -[797.931] [797.931][S01]one of the most fascinating fields in science today.[801.134] [803.837][S01]Have you ever seen footage of those vast e-tournaments[807.174] [807.174][S01]where an entire arena of dedicated fans[809.843] [809.843][S01]excitedly watches on in support of highly skilled players,[814.615] [814.615][S01]sat on stage in their gaming chairs.[817.317] [817.317][S01]Armed only with a keyboard, a mouse and a computer screen.[821.355] [821.355][S01]Well chances are, they are playing something like[824.224] [824.224][S01]StarCraft II, created by the American video game[827.728] [827.728][S01]developer Blizzard entertainment.[829.496] [829.496][S01]It is a monumentally tricky tactical game[837.204] [837.204][S01]where you play as one of three races -[840.174] [840.174][S01]the enigmatically named Zerg, Protoss or Terrins.[845.212] [846.346][S01]Each player has to mine resources, build an economy[849.616] [849.616][S01]and acquire increasingly sophisticated technology,[852.92] [852.92][S01]all the time trying to defeat your alien opponents in a futuristic[858.125] [858.125][S01]rather bleak looking landscape.[860.294] [862.663][S01]Your field of view of the simulated game is limited by a moving camera[866.8] [866.8][S01]that you have to operate[868.302] [868.302][S01]and so there’s no way to see everything at once,[871.138] [871.138][S01]often you can’t see your opponent at all.[873.874] [873.874][S01]And it is played by 10s of thousands of people -[877.511] [877.511][S01]sometimes for hefty cash prizes.[880.247] [881.415][S01]And the human players are staggeringly fast.[884.751] [884.751][S01]The best in the world can manage up to 800 clicks in a minute.[890.09] [890.09][S01]Feeling inadequate?[891.692] [891.692][S10] Definitely super cool that I can work on one thing[895.362] [895.362][S10]that has been certainly a passion of mine in my teenage days.[900.0] [900.0][S01] Meet Oriol Vinyals, a research scientist at DeepMind.[903.904] [903.904][S01]He is an ex-pro StarCraft player and co-leads[907.241] [907.241][S01]the StarCraft effort at DeepMind.[909.61] [909.61][S10] as you develop a new algorithm, or a new idea, when you test it,[913.247] [913.247][S10]you actually see it play better the game you like,[916.216] [916.216][S10]so that’s very rewarding and very visual right,[918.519] [918.519][S10]that you try something new and you really see[921.221] [921.221][S10]oh my god, it’s really understands how this unit works.[925.626] [925.626][S01] StarCraft is a serious business - so serious in fact[928.829] [928.829][S01]that it has now been professionalized,[931.532] [931.532][S01]and for Oriol that proves that is a game that pushes human intelligence.[936.003] [936.003][S10] Humans found it interesting, so that means it’s an interesting game[940.24] [940.24][S10]that challenges intelligence and creativity[942.476] [942.476][S10]in ways that we like that we spend many hours playing.[945.579] [945.579][S01] So how good is the AI at the moment then?[948.148] [948.148][S01]How well can it play StarCraft?[950.017] [950.017][S10] It’s better than any AI anyone has ever built[953.487] [953.487][S10]and it obviously has learned from experience[956.39] [956.39][S10]not from someone knowing the game and encoding some set of rules.[960.494] [960.494][S10]This is I mean one of the most complicated games we’ve ever tackled -[964.498] [964.498][S10]it’s challenging kind of our understanding[967.367] [967.367][S10]and our algorithms quite a bit.[969.636] [969.636][S01] The DeepMind team started to see how good their work in progress really[974.141] [974.141][S01]was by inviting two of the world’s best[977.044] [977.044][S01]StarCraft II players to take on their own algorithm.[980.948] [981.982][S01]So let me introduce DeepMind’s AlphaStar -[984.885] [984.885][S01]the first artificial intelligence to ever take on top professional players.[990.524] [990.524][S01]It plays the full game of StarCraft II[993.126] [993.126][S01]by using a deep neural network trained directly from raw game data,[998.665] [998.665][S01]by supervised learning and reinforcement learning.[1001.568] [1002.336][S01]Your commentators are Dan Stemkoski - aka Artosis,[1007.407] [1007.407][S01]and Kevin van der Kooi - aka Rotterdam.[1011.645] [1011.645][S09] well first of all it’s really awesome to be here[1013.413] [1013.413][S09]together with you Dan, we’re both I think incredibly excited[1016.416] [1016.416][S09]to see how this evening unfolds.[1018.352] [1018.352][S07] I mean this is just so exciting that[1020.888] [1020.888][S07]DeepMind is doing all this -[1022.489] [1022.489][S01] Taking on AlphaStar in this benchmarking match is German champion[1026.66] [1026.66][S01]Dario Wünsch, better known as TLO. He’s normally a Zerg player,[1032.266] [1032.266][S01]but he’s playing as Protoss for this match.[1034.601] [1034.601][S01]Kevin and Dan are excited. Maybe even a tad over excited.[1039.706] [1039.706][S09] I’m so incredibly excited.[1041.308] [1041.308][S07] Oh my god - this is like the most exciting[1044.111] [1044.111][S07]I have personally ever been for an event.[1045.679] [1045.679][S09] I can’t wait to break down some -[1047.915] [1048.849][S07] So this is AlphaStar -[1050.784] [1050.784][S07]this is an AI that we don’t know how good it is yet,[1054.655] [1054.655][S07]but already we have some interesting things happening.[1058.225] [1058.225][S01] Now I’m not entirely conversant with the StarCraft playbook lingo here,[1063.43] [1063.43][S01]so I’ll just say that AlphaStar’s stalkers[1066.834] [1066.834][S01]are laying down some sharp moves.[1068.735] [1068.735][S07] it feels to me like so far these attacks have been very well[1072.406] [1072.406][S07]planned by AlphaStar.[1074.374] [1074.374][S09] and ah relentless - the last two attacks -[1076.677] [1076.677][S01] And in a matter of minutes, it’s all over.[1079.279] [1079.279][S07] Well that is it! [laughs][1081.615] [1081.615][S07]the GG is called the Good Game here from TLO[1085.519] [1085.519][S07]and the first game from AlphaStar against a pro gamer goes to AlphaStar.[1090.924] [1090.924][S01] David Silver was there, ringside.[1093.327] [1093.327][S06] We have a team that has been working on this[1094.962] [1094.962][S06]and ramping up our development over the last few months and this represents[1097.831] [1097.831][S06]ah you know a milestone where we actually for the first time[1100.701] [1100.701][S06]we saw an AI that was actually able to defeat a professional player.[1104.705] [1104.705][S01] Shall we have a quick word with our defeated challenger - TLO?[1107.908] [1107.908][S08] When I was practicing most of the humans[1109.71] [1109.71][S08]I played against played very standard StarCraft.[1112.246] [1112.246][S08]I once again I assumed after the first match[1115.182] [1115.182][S08]I probably have a good idea how to play against this agent, I did not.[1119.486] [1121.555][S01] Next up, the main event. AlphaStar versus Poland’s finest-[1126.46] [1126.46][S01]Grzegorz Komincz - better known as MaNa -[1129.63] [1129.63][S01]one of the world’s strongest professional[1131.832] [1131.832][S01]StarCraft players.[1133.934] [1133.934][S07] MaNa - I need to hear what you’re thinking here[1136.303] [1136.303][S07]- cause that looks scary.[1138.639] [1138.639][S05] Yeah, AlphaStar like he’s not scared about[1142.543] [1142.543][S05]ah the ramp, so if I would be playing against a human player[1146.58] [1146.58][S05]right there, nobody’s going up that ramp.[1149.316] [1149.316][S01] I should point for those of you who play[1151.051] [1151.051][S01]StarCraft that these matches are taking place under professional[1154.321] [1154.321][S01]match conditions on a competitive ladder map[1157.024] [1157.024][S01]and without any game restrictions. This version of AlphaStar[1160.494] [1160.494][S01]could see the whole of the game map at any one time,[1163.83] [1163.83][S01]but otherwise played in a comparable way to humans.[1167.267] [1167.267][S06] Our goal is not just to defeat these players.[1171.071] [1171.071]Our goal is to do it in the right way. [S10] two seconds guys![1175.142] [1175.142][S01] And the result - AlphaStar: 5; [S05] nil.[1179.947] [1182.649][S01] I should tell you that MaNa played a later version of the algorithm[1185.085] [1185.085][S01]in the end and won, so all in all, 5 - 1.[1189.523] [1191.592][S01]Now to understand how an AI could learn to play[1194.328] [1194.328][S01]StarCraft, Oriol Vinyals put me to the test.[1198.432] [1198.432][S01]A match to the end - mathematician versus machine.[1202.369] [1202.369][S01]I’ve got a quite a funky looking mouse in front of me and a normal keyboard,[1208.075] [1208.075][S01]and on the screen there is a very mean looking alien…[1212.479] [1212.479][S10] Yeah, Protoss.[1213.58] [1213.58][S01] I mean sort of sort of like an elephant meets -[1216.35] [1216.35][S01]um well he’s got fists.[1217.918] [1217.918][S01]I wouldn’t want to meet him on a dark night.[1219.186] [1219.186][S10] No. [S01] Is he my friend or not?[1221.154] [1221.154][S10] He is you. You’re going to be the commander of this particular race -[1225.859] [1225.859][S01] I quickly found out there is a lot to take in.[1229.73] [1229.73][S01]StarCraft is perhaps not for beginners.[1233.1] [1233.1][S01]You have your worker bees collecting resources for you.[1236.003] [1236.003][S01]Are these - are these all -[1237.271] [1237.271][S01]they’re almost like ant creatures running out and and grabbing crystals.[1241.475] [1241.475][S10] Right, exactly. [S01] And you need to try and work out[1243.21] [1243.21][S01]how your actions will affect your game in the future.[1245.779] [1245.779][S01]This is not easy for humans to learn let alone agents[1248.682] [1248.682][S01]who have absolutely no context, no object recognition,[1252.686] [1252.686][S01]and definitely no former StarCraft champion to hold their hand.[1257.224] [1257.224][S10] Hah! Look this is the enemy. [S01] oh no![1259.026] [1259.026][S10] Um it’s going to be pleasant - it just came to kind of find you[1262.863] [1262.863][S10]and now see what you’re doing which is absolutely nothing so far[1266.733] [1266.733][S10]so far we have done nothing at all -[1268.335] [1268.335][S01] Part of the challenge of StarCraft[1269.736] [1269.736][S01]is that there isn’t an idea strategy that wins every time.[1272.94] [1272.94][S01]it’s a bit like rock, paper, scissors in that way.[1275.209] [1275.209][S01]The winning tactic will depend on how your opponent plays.[1279.246] [1279.246][S01]But remember you only have a very narrow field of vision[1282.549] [1282.549][S01]- outside of where your camera is pointing,[1284.985] [1284.985][S01]your opponent could be up to anything.[1287.054] [1287.054][S10] Because you don’t see the other player, you must decide[1290.691] [1290.691][S10]when am I going to see it - do I already know what’s going on[1293.994] [1293.994][S10]and should I not go and scout what it’s doing[1296.997] [1296.997][S10]but maybe if I do that,[1298.398] [1298.398][S10]he knows that I know and so on and so forth,[1300.334] [1300.334][S10]so these kind of imperfect information aspect of StarCraft[1304.238] [1304.238][S10]is extremely interesting as a player and it’s going to be testing our agents[1309.243] [1309.243][S10]to levels that we haven’t seen in any other game.[1311.812] [1311.812][S10]And then of course there’s sort of details that have happening in the game[1315.215] [1315.215][S10]that you must remember for a long time.[1317.384] [1317.384][S01] Advice I should have listened to more carefully perhaps -[1320.654] [1320.654][S10] Ah - we’re being attacked and we’re probably going to die.[1322.456] [1322.456][S01] Oh no! Oh no! [laughs][1323.357] [1323.357][S10] Um, so that’s okay.[1324.892] [1324.892][S01] I didn’t last very long.[1326.126] [1326.126][S10] So but I think that this, this discovery phase[1328.095] [1328.095][S10]right where you would now basically you would lose,[1330.764] [1330.764][S10]you get the reward of minus one, and you start again.[1334.334] [1334.334][S01] If I was an algorithm, I wouldn’t be upset by losing,[1337.504] [1337.504][S01]I would just reset and go again.[1339.473] [1339.473][S01]Each time armed with a little more knowledge.[1342.709] [1342.709][S01]But to even be able to play[1344.178] [1344.178][S01]StarCraft in the first place, to even be able to operate the controls,[1348.615] [1348.615][S01]the AI had to master quite a few transferrable skills.[1352.586] [1352.586][S10] You’ve noticed when you were playing[1354.388] [1354.388][S10]that there were some movements that were resembling what it was like[1358.392] [1358.392][S10]to maybe navigate the web or like operate[1361.628] [1361.628][S10]um your laptop, namely click, drag and click,[1365.933] [1365.933][S10]drag and drop, like select rectangles and moving the mouse,[1370.404] [1370.404][S10]and maybe combining mouse with keyboard and so on,[1374.174] [1374.174][S10]and we tried exactly the same agent,[1376.343] [1376.343][S10]the same architecture absolutely everything the same the way,[1379.479] [1379.479][S10]the same code almost, and we changed, we changed the environment[1382.916] [1382.916][S10]instead of saying now here is the[1384.785] [1384.785][S10]StarCraft, please play to win, we said here is paint -[1390.157] [1390.157][S10]Microsoft Paint as an environment - interact with it[1393.293] [1393.293][S10]and I’ll reward you if what you paint looks like a face,[1396.73] [1396.73][S10]and it actually worked[1398.265] [1398.265][S10]so I think that’s just learning this basic skills of point[1402.436] [1402.436][S10]and click interfaces that applies so in so many places.[1405.739] [1405.739][S01] the same agent that plays[1407.674] [1407.674][S01]StarCraft can draw real faces in Microsoft Paint.[1411.812] [1411.812][S10] Right and here that the point to be clarified[1415.315] [1415.315][S10]is not the same agent that was trained to play[1418.185] [1418.185][S10]StarCraft, it’s the same algorithm that can train to play[1420.787] [1420.787][S10]StarCraft, can also train to do Paint.[1424.057] [1425.526][S01] Put that same algorithm to work drawing celebrities in Paint,[1429.796] [1429.796][S01]and it can capture all the main traits of the face.[1432.499] [1432.499][S01]Clicking and dragging the mouse to recreate shape[1435.569] [1435.569][S01]and tone and hairstyle, much like a street artist would.[1439.173] [1439.173][S10] It’s the same technique, but if you will,[1442.075] [1442.075][S10]it’s kind of a brain that is blank[1444.011] [1444.011][S10]and this brain can learn to do this and that and that[1446.313] [1446.313][S10]and then we kind of by acting in the environment repeatedly[1450.717] [1450.717][S10]and getting reward, the brain waits or gets shaped to do this task[1455.522] [1455.522][S10]or that task or that task -[1457.191] [1457.191][S10]we are not yet at the point where the same brain does both like we do[1461.695] [1461.695][S10]but obviously that’s one of the things[1463.363] [1463.363][S10]that we would be very interested in tackling next as well.[1466.767] [1466.767][S01] Because that’s stepping towards artificial general intelligence I guess.[1469.937] [1469.937][S10] Exactly. And that’s what we do every day.[1472.606] [1476.343][S01] That is the ultimate goal and it’s a topic of conversation[1479.713] [1479.713][S01]that’s never far away whoever in this building[1482.416] [1482.416][S01]you find yourself talking to[1484.384] [1484.384][S01]because the point of getting AI to play games like StarCraft[1487.521] [1487.521][S01]or Go is to enhance our understanding of what intelligence actually is.[1493.46] [1493.46][S01]Here’s Raia Hadsell from the Deep Learning team[1496.163] [1496.163][S02] we write programs, we run those programs,[1498.532] [1498.532][S02]those experiments where we might train an agent to play a game for instance[1503.504] [1503.504][S02]or to solve a puzzle in a simulated world[1506.907] [1506.907][S02]and then we look at the results of that.[1508.742] [1508.742][S02]It really is trying to understand this puzzle of learning and representation,[1513.847] [1513.847][S02]memory, control in terms of actions that a robot would take.[1517.851] [1517.851][S02]There’s so many complex parts to this big puzzle[1522.122] [1522.122][S02]of what is an intelligent being, what is an intelligent agent.[1526.193] [1527.294][S01] But if you ask people what they think the future of AI looks like,[1531.698] [1531.698][S01]it tends to be wrapped up in something a bit more physical,[1535.435] [1535.435][S01]something that comes complete with moving arms and everything.[1539.573] [1539.573][S06] I think one natural challenge for AI[1544.044] [1544.044][S06]which many people are centering upon[1547.014] [1547.014][S06]would be to actually have an impact on the real world[1551.251] [1551.251][S06]in the guise of robotics to actually see a robot[1554.354] [1554.354][S06]which is able to move, to grip, to manipulate,[1557.524] [1557.524][S06]to even have locomotion in anything approaching[1560.894] [1560.894][S06]not even what a human does - maybe even an animal.[1563.23] [1563.23][S06]I think this this would represent a major stride forwards.[1566.667] [1566.667][S01] More on that next time.[1568.635] [1570.504][S01]If you would like to find out more about the themes in this episode,[1574.408] [1574.408][S01]or explore the world of AI research beyond[1577.077] [1577.077][S01]DeepMind, you’ll find plenty of useful links[1579.646] [1579.646][S01]in the show notes for each episode - and if there are stories or resources[1583.784] [1583.784][S01]that you think other listeners would find helpful, then let us know.[1587.054] [1587.054][S01]You can message us on Twitter or email the team at podcast@deepmind.com.[1592.593] [1592.593][S01]You can also use the address to send us your questions[1595.429] [1595.429][S01]or feedback on the series.[1597.464] [1598.198][S01]But for now, let’s nip out for a bit of air.[1600.601]"}
{"file_name": "audio/val_000002.wav", "transcription": "[0.0][S02] Welcome back to \"Google DeepMind--[2.13] [2.13][S02]The Podcast,\" with me, your host, Professor Hannah Fry.[5.745] [5.745][MUSIC PLAYING][9.705] [13.18][S02]I think it's fair to say that generative AI, as we have it[16.059] [16.059][S02]now, is a bit of a mismatch for our science fiction fantasies[20.5] [20.5][S02]about what technology might become.[23.18] [23.18][S02]But it is worth remembering that this isn't the only possibility.[26.33] [26.33][S02]There are other forms of AI forms,[28.61] [28.61][S02]forms which are by their very design, autonomous entities.[32.509] [32.509][S02]These AI are firmly in the realm of robot butlers[35.92] [35.92][S02]or a virtual concierge that anticipates your every whim[39.43] [39.43][S02]before you even have it.[41.51] [41.51][S02]And OK, we might have to wait a little while for those two[44.53] [44.53][S02]particular examples.[46.01] [46.01][S02]But I'm talking here about artificial intelligence[49.24] [49.24][S02]with agency, AI that has built-in wants and objectives[54.28] [54.28][S02]and the ability to independently make decision after decision[58.54] [58.54][S02]in pursuit of its own goals.[61.25] [61.25][S02]This kind of AI are known as agents,[64.42] [64.42][S02]and while agents haven't yet exploded[66.46] [66.46][S02]into the public consciousness quite as much[68.84] [68.84][S02]as generative AI has, trust us, they[71.84] [71.84][S02]are going to be a very big deal.[74.64] [74.64][S02]And in the meantime, a very good training ground for agents[77.72] [77.72][S02]is the world of video games.[79.74] [79.74][S02]Think about it-- perfectly packaged,[82.08] [82.08][S02]neatly constrained environments where the agents can run wild,[85.84] [85.84][S02]work out the rules for themselves,[87.44] [87.44][S02]and learn how to handle autonomy.[90.19] [90.19][S02]If you've been following us since the beginning,[92.19] [92.19][S02]then you will know that AI in games[94.1] [94.1][S02]is something that DeepMind has a proud heritage in.[97.44] [97.44][S02]And so to explain today, we are joined[100.25] [100.25][S02]by Frederic Besse, who is a Senior Staff Research[103.01] [103.01][S02]Engineer at Google DeepMind.[104.76] [104.76][S02]Frederic has a PhD in Computer Vision[106.94] [106.94][S02]and was one of the earliest people[108.59] [108.59][S02]to join the DeepMind team back in 2015,[111.18] [111.18][S02]when there were only about 100 researchers here.[114.08] [114.08][S02]And before the decade he spent at the cutting-edge[116.33] [116.33][S02]of artificial intelligence, he also[118.31] [118.31][S02]spent some time working in the visual effects[120.38] [120.38][S02]industry on how to integrate CGI into real footage.[124.71] [124.71][S02]Frederic, thank you so much for joining us.[126.95] [126.95][S02]With a CV like that, I can see why they wanted you.[129.83] [129.83][S01] Well, thanks.[131.038] [131.038][S02] OK, well, so I think it's probably[132.955] [132.955][S02]best if we start off with some definitions.[134.86] [134.86][S01] Sure.[135.21] [135.21][S02] So, I mean, I've talked briefly about[137.54] [137.54][S02]what agents are there, but how do you define them?[140.16] [140.16][S01] So, in my mind, an agent[142.4] [142.4][S01]is a very general concept.[144.33] [144.33][S01]And I define it as an entity that can act in an environment.[149.13] [149.13][S01]I think action is what defines an agent.[151.73] [151.73][S01]And an environment, on the other side,[153.41] [153.41][S01]is a place which can provide observations for the agent[159.05] [159.05][S01]and which the agent can interact with.[161.16] [161.16][S01]And so what's important to understand[165.59] [165.59][S01]is that the agent, by acting into the environment,[167.85] [167.85][S01]will change the environment, which[169.94] [169.94][S01]is crucial for developing AI systems that[173.33] [173.33][S01]can do useful things and understand[175.4] [175.4][S01]the consequences of their actions[177.32] [177.32][S01]in environments, in games, in the world,[180.65] [180.65][S01]on the internet, et cetera.[181.858] [181.858][S02] This sounds like something that[183.65] [183.65][S02]exists beyond just AI, though.[185.19] [185.19][S02]I mean, autopilot on planes, would that count as an agent?[187.757] [187.757][S01] Yes, absolutely.[189.09] [189.09][S01]So maybe we can give a few examples of agents[191.19] [191.19][S01]because it's a very general concept.[192.69] [192.69][S01]So humans are agents because we can act.[195.56] [195.56][S01]We observe the world with our eyes and our ears.[198.18] [198.18][S01]We have proprioception.[199.38] [199.38][S01]We know where our limbs are.[200.61] [200.61][S01]So we are agents.[201.39] [201.39][S01]We can affect the world.[202.83] [202.83][S01]Similarly, robots as well.[204.36] [204.36][S01]But if you go back to maybe decades ago[208.1] [208.1][S01]where we had agents such as Autopilot, which we handcrafted[212.78] [212.78][S01]the logic of those agents, we were[215.51] [215.51][S01]programming them to do what we wanted them to do.[218.46] [218.46][S01]So they were different from the modern agents[221.06] [221.06][S01]that we can talk about later on, but these were indeed[224.57] [224.57][S01]examples of agents.[225.567] [225.567][S02] So, all right.[226.65] [226.65][S02]If that's an agent then, is that the same thing[229.46] [229.46][S02]as autonomy or having agency?[232.04] [232.04][S02]Are these all slightly, subtly different?[233.94] [233.94][S01] So I think they are subtly different.[236.148] [236.148][S01]I think an agent can act and do things[239.51] [239.51][S01]that it has been programmed to do,[244.07] [244.07][S01]and that's agency, being able to take actions.[246.9] [246.9][S01]Now autonomy is more the skill or the property[251.71] [251.71][S01]of acting by itself to accomplish a certain task.[257.06] [257.06][S01]And there are various levels of autonomy.[259.089] [259.089][S01]So if we have a self-driving car, which will take decision[264.04] [264.04][S01]to best navigate the roads, you could[266.98] [266.98][S01]say that there is a certain level of autonomy.[269.8] [269.8][S01]Some agents are much less autonomous[272.41] [272.41][S01]and will require a lot more of your input.[274.94] [274.94][S01]And then on the other end of the spectrum,[276.703] [276.703][S01]you could have an agent that goes in the world[278.62] [278.62][S01]and learn new things, explore, which we don't really[283.63] [283.63][S01]have to instruct much.[286.19] [286.19][S01]So there are a spectrum in the space of autonomy.[289.053] [289.053][S02] Is part of that almost[290.47] [290.47][S02]a hierarchy of the type of objectives that you're setting?[293.7] [293.7][S02]Like, I don't know, the autonomous action[295.595] [295.595][S02]of picking up a cup and putting it down[297.22] [297.22][S02]is slightly different to deciding what's for lunch.[299.8] [299.8][S01] Yes, because we are training agents[302.68] [302.68][S01]to do simple things because this is a starting point.[306.72] [306.72][S01]But having an agent that can cook for you your whole lunch--[309.462] [309.462][S02] And decide what you want as well would be nice.[311.92] [311.92][S01] Yes, as well.[313.128] [313.128][S01]It would be a recommender system will require a lot of sub steps[315.9] [315.9][S01]and sub tasks too, and solve those tasks properly[318.66] [318.66][S01]as well first.[320.01] [320.01][S01]So we call those high-level tasks[323.19] [323.19][S01]compared to short-horizon tasks, which will be very basic,[327.33] [327.33][S01]almost mechanical actions, like grab an object,[330.85] [330.85][S01]put the objects on the table, compared to cook lunch.[335.105] [335.105][S02] Yeah, like the robots[336.48] [336.48][S02]you see in factories, for example, versus the type of ones[338.94] [338.94][S02]that we have in science fiction.[340.42] [340.42][S02]But is that the aim?[342.04] [342.04][S02]Are we talking about robots eventually[344.52] [344.52][S02]that will be in our homes?[345.805] [345.805][S01] I think so.[346.93] [346.93][S01]So what we are trying to do is to develop general agents.[350.22] [350.22][S01]And we want to reach a point, which[353.19] [353.19][S01]we call AGI, Artificial General Intelligence, which[356.34] [356.34][S01]is a point where we will have agents and AI systems that will[361.17] [361.17][S01]be truly general, as general as humans,[364.03] [364.03][S01]will be able to adapt to new situations, reason like we do,[367.57] [367.57][S01]and even surpass our capabilities in doing tasks.[372.17] [372.17][S01]Once we reach this level of competency in AI systems,[379.02] [379.02][S01]this will unlock a lot of potential[381.13] [381.13][S01]for what we can do as humans.[383.71] [383.71][S01]Those AI systems will be able to accelerate[387.22] [387.22][S01]our scientific progress, for example,[389.29] [389.29][S01]or do a lot of tasks which maybe we are not very good for.[393.49] [393.49][S01]For example, maybe driving, or help you in your home,[398.83] [398.83][S01]help you do shopping online.[400.63] [400.63][S01]I think there are some applications which we can't even[403.18] [403.18][S01]begin to conceive because this is such[405.52] [405.52][S01]a technological leap that I think we will just discover it[410.02] [410.02][S01]as we go.[410.9] [410.9][S01]But we do believe that building such general agents, which[414.85] [414.85][S01]can act into an environment, is key to achieve AGI.[418.662] [418.662][S02] That's interesting in the way[420.37] [420.37][S02]that you're describing it, that some of these agents[423.04] [423.04][S02]are going to be virtual, essentially,[424.778] [424.778][S02]and then some of them will be actually physically[426.82] [426.82][S02]embedded in an environment.[428.3] [428.3][S01] Yes.[429.133] [429.133][S01]So I think it all goes back to what's your observation space[434.26] [434.26][S01]and what's your action space.[435.94] [435.94][S01]So we can have physical agents, embodied physical agents,[440.08] [440.08][S01]like robots, which will be able to do things in the real world,[443.15] [443.15][S01]which will be very useful.[444.5] [444.5][S01]And we already, as you--[445.595] [445.595][S02] My robot butler, thank you.[447.22] [447.22][S01] Exactly.[448.9] [448.9][S01]But also virtual agents, which work not in the real world,[453.2] [453.2][S01]but in an abstract virtual space.[456.7] [456.7][S01]This could be still a physical space, for example, games,[459.77] [459.77][S01]which we use in our research to also demonstrate and understand[464.71] [464.71][S01]what those agents can do.[466.66] [466.66][S01]But also, you could have an agent that[468.52] [468.52][S01]doesn't have any physical, even virtual action space,[473.18] [473.18][S01]but instead it can make programs.[476.15] [476.15][S01]For example, write Python code, build software for you,[479.89] [479.89][S01]which is not a physical space, but it's still an agent[482.41] [482.41][S01]because it can take actions, it can compile the function,[485.93] [485.93][S01]it can look on the internet.[488.26] [488.26][S01]And so, this is still very valuable.[490.978] [490.978][S02] How is this different to the chatbots[493.02] [493.02][S02]that we've seen from large language models?[494.65] [494.65][S02]Because they can do a lot of that stuff too, right?[496.69] [496.69][S02]They can search the internet, they[498.107] [498.107][S02]can understand your instructions and make decisions, sort of?[501.647] [501.647][S01] Yes.[502.48] [502.48][S01]So chatbots, what you get out of chatbots generally[506.73] [506.73][S01]is a chat output.[509.11] [509.11][S01]So they produce natural language.[511.29] [511.29][S01]Now, there are some capabilities in some chatbots[515.37] [515.37][S01]which could go and search the internet for you.[519.0] [519.0][S01]But this space is still fairly constrained[523.559] [523.559][S01]to just being either natural language or specific functions[530.34] [530.34][S01]that the programmer adds to the chatbot to do things.[534.79] [534.79][S01]Now those chatbots, when they speak with you,[537.82] [537.82][S01]I mean, in a way, you are the environment of the chatbot,[540.25] [540.25][S01]right?[540.75] [540.75][S01]What the chatbot says will affect you[542.49] [542.49][S01]and will affect what you ask the chatbot next.[545.74] [545.74][S01]But maybe there is not as much consequence[550.06] [550.06][S01]as if the chatbot or the agent was[551.86] [551.86][S01]acting in an environment which drastically changes when[556.36] [556.36][S01]the agent takes some action.[557.93] [557.93][S01]So it's a bit more of a passive type of agent.[561.11] [561.11][S01]We could call it an agent, but it's[562.75] [562.75][S01]very different from an agent that acts in the real world.[566.96] [566.96][S01]And the way also those models are trained,[570.28] [570.28][S01]they are trained with a lot of data that[573.19] [573.19][S01]are scraped from the internet.[575.89] [575.89][S01]But they are not trained in an environment[578.14] [578.14][S01]to try to do things, fail, understand why they failed,[581.65] [581.65][S01]and try again.[582.41] [582.41][S01]So this is a difference between some[585.64] [585.64][S01]of the way those models are trained with human data.[588.77] [588.77][S02] So then, should we be thinking of these as separate[592.095] [592.095][S02]or can you combine the two?[593.22] [593.22][S02]Can you combine large language models and an agent[595.4] [595.4][S02]that has an objective?[596.4] [596.4][S01] So you can actually,[598.28] [598.28][S01]there are some robotics paper from Google DeepMind[601.73] [601.73][S01]that is called RT-2 and RT-X, which are doing that.[605.185] [605.185][S01]So they are training.[606.06] [606.06][S01]So a language model has a lot of knowledge[609.14] [609.14][S01]from the internet scale data that it has been trained on.[612.2] [612.2][S01]So if you want, it has a lot of common sense[614.81] [614.81][S01]and it understands a bunch of concepts,[616.47] [616.47][S01]but it doesn't have the ability to actuate joints of a robot.[620.13] [620.13][S01]But you can text this model and train[622.76] [622.76][S01]it to produce joint activation for your robotic arm,[625.65] [625.65][S01]for example.[626.15] [626.15][S01]That's what they've done in that paper.[628.64] [628.64][S02] I guess almost the language model gives[631.52] [631.52][S02]the algorithm some sort of inverted commas,[634.83] [634.83][S02]conceptual understanding of things.[636.63] [636.63][S01] Yes, I think that's what is great[639.17] [639.17][S01]about those big models is that they have such knowledge[642.72] [642.72][S01]and understand concepts and words[646.16] [646.16][S01]and what you chat to them about that they can make plans.[650.74] [650.74][S01]But all of this is in language space.[652.35] [652.35][S01]And we need to bridge the gap towards acting in the real world[658.95] [658.95][S01]or in the virtual world.[660.027] [660.027][S02] To go off and actually enact it?[661.86] [661.86][S01] Yes.[662.958] [662.958][S02] What is it about games, though?[664.75] [664.75][S02]Why is it useful to start with games?[666.97] [666.97][S01] So I think games offer a very rich experience[672.51] [672.51][S01]and games are very varied.[674.05] [674.05][S01]So if you look at how many games there[675.72] [675.72][S01]are on computers, for example, you have tens of thousands[679.74] [679.74][S01]of games.[680.41] [680.41][S01]And each game offers you different scenarios,[683.203] [683.203][S01]different environments, different things[684.87] [684.87][S01]you could be doing in the game.[686.185] [686.185][S01]So we think it's a great proving ground[687.81] [687.81][S01]to train AI systems to match human performance.[692.61] [692.61][S01]There are other advantages in games.[694.87] [694.87][S01]So they are virtual, so they are easy to scale.[697.21] [697.21][S01]You could instantiate a lot of different games in parallel[701.13] [701.13][S01]compared to physical systems, which are[703.5] [703.5][S01]constraints to the real world.[705.9] [705.9][S01]They are also safe.[707.55] [707.55][S01]So if things go wrong in games, it doesn't really matter.[710.53] [710.53][S01]It's just a game.[712.5] [712.5][S01]Games are a place where humans can[714.72] [714.72][S01]interact with agents, for example, in multiplayer games.[719.31] [719.31][S01]And games are fun as well for humans to play.[723.76] [723.76][S01]And I think it's a great source of experience data[726.6] [726.6][S01]that we can collect.[727.602] [727.602][S02] Because you've already got agents already[729.81] [729.81][S02]in non-player characters, NPCs floating around the place.[732.738] [732.738][S01] Yes, yes.[733.78] [733.78][S01]So I think that's a very interesting point.[739.03] [739.03][S01]Historically, and actually it's still the case,[741.75] [741.75][S01]when game developers create agents inside of games,[745.18] [745.18][S01]it's a very different beast to the type of agents that we make.[749.14] [749.14][S01]Those agents have access to the internal code of the game.[752.62] [752.62][S01]And the game developers just write the logic, which[756.6] [756.6][S01]works great, and they are fast and they can[758.85] [758.85][S01]do pretty impressive things.[761.218] [761.218][S02] Things like stand around,[762.76] [762.76][S02]and then if someone approaches you, attack them.[764.79] [764.79][S01] Yeah, there is some pretty complicated logic[768.42] [768.42][S01]you could put in the game, in the agents in the game,[770.77] [770.77][S01]in the NPCs.[771.55] [771.55][S01]But there is a limit in how complicated you can describe[775.68] [775.68][S01]a behavior in terms of code.[777.52] [777.52][S01]And that's where deep learning and modern agents come in,[781.21] [781.21][S01]because those are not-- the logic inside those agents[784.41] [784.41][S01]is not handcrafted by humans.[786.46] [786.46][S01]Instead, we subject the agent to a lot of experience data,[790.29] [790.29][S01]and the agent will learn the behavior for itself[793.98] [793.98][S01]in a softer way with more nuances.[797.41] [797.41][S01]And it can pick up on very complex behavior, which[800.64] [800.64][S01]maybe wouldn't be feasible to write down in code,[804.71] [804.71][S01]in programming language.[805.71] [805.71][S02] Rather than just following rules,[806.95] [806.95][S02]it's sort of working out how to achieve its objective itself?[809.492] [809.492][S01] Yeah.[810.45] [810.45][S02] DeepMind's got quite a long history[812.73] [812.73][S02]of playing around with games.[815.62] [815.62][S02]Maybe that a bit trivializes it slightly[817.47] [817.47][S02]considering the things that they've achieved by doing so.[819.9] [819.9][S01] Yes.[820.733] [820.733][S01]So back in 2015, there was the first breakthrough[826.05] [826.05][S01]in neural network agents, called DQN, for Deep Q Network,[831.15] [831.15][S01]in which the agent was trained to play Atari.[835.05] [835.05][S01]And the breakthrough here is that the agent was only[838.5] [838.5][S01]using pixels as the observation space.[841.6] [841.6][S02] So Atari is like \"Space Invaders,\"[844.17] [844.17][S02]break out the one with the bricks where you've got a paddle[847.245] [847.245][S02]and a ball bouncing around.[848.37] [848.37][S02]Like, the games you had in the 1980s.[850.78] [850.78][S01] Yeah, so the arcade games, exactly.[853.987] [853.987][S02] And the only thing it would take as an input[856.32] [856.32][S02]were the pixels on the screen?[857.98] [857.98][S01] So the pixels of the screen.[859.06] [859.06][S01]And there was also the concept of reward, which is[861.48] [861.48][S01]like how many points you get.[862.87] [862.87][S01]And that is very important for those agents to learn.[865.5] [865.5][S01]Because what those agents are doing and how they are trained,[868.93] [868.93][S01]they are trained using a method called reinforcement learning,[871.9] [871.9][S01]which aims at maximizing this score, maximizing the reward.[876.85] [876.85][S01]And so combining deep learning with deep neural network[880.14] [880.14][S01]and reinforcement learning yielded the DQN agent,[883.51] [883.51][S01]which was first of its kind.[885.49] [885.49][S01]And that was a very big milestone for DeepMind.[890.27] [890.27][S01]Then another important agent that was built[895.34] [895.34][S01]was AlphaGo, which was an agent that[898.79] [898.79][S01]tries to learn to play the game of Go, which is a board game.[903.89] [903.89][S02] Notoriously difficult.[905.39] [905.39][S01] Notoriously difficult.[906.973] [906.973][S01]I think it was uncertain whether AI could ever[910.85] [910.85][S01]beat the game of Go.[912.5] [912.5][S01]But then that was proven to be wrong.[916.58] [916.58][S02] When AlphaGo beat Lee Sedol.[918.36] [918.36][S01] Yes, exactly.[919.568] [919.568][S01]So that was a huge result. Then if you continue[922.645] [922.645][S01]along the history, so this is a non-exhaustive list[924.77] [924.77][S01]because there is a lot of research that[925.91] [925.91][S01]has been happening also in between all[927.95] [927.95][S01]these important milestones.[930.32] [930.32][S01]There was AlphaZero, which is taking the agents that play Go,[935.9] [935.9][S01]and also it could play chess and Shogi, to the next level.[939.39] [939.39][S01]And this agent was trained without any human data.[942.12] [942.12][S01]So while AlphaGo was trained using[943.7] [943.7][S01]a mix of human data and self-play and reinforcement[946.13] [946.13][S01]learning, AlphaZero was trained without any human data.[948.98] [948.98][S01]And just starting from scratch with self-play,[951.31] [951.31][S01]it discovered the best chess moves,[954.36] [954.36][S01]and that was very interesting to see.[957.523] [957.523][S02] Extraordinary, that taking the human input out[959.94] [959.94][S02]made it better.[961.44] [961.44][S01] Well, yes.[963.87] [963.87][S01]So it's interesting because you can actually[966.15] [966.15][S01]look at what are the most optimal moves for chess,[969.81] [969.81][S01]which is a fascinating find.[972.898] [972.898][S02] But then it moved into computer games[974.94] [974.94][S02]again, didn't it?[976.027] [976.027][S01] Yes.[976.86] [976.86][S01]So I think one next very important paper was[981.09] [981.09][S01]the AlphaStar paper, which is an agent that is playing \"StarCraft[986.91] [986.91][S01]II,\" which is a game, competitive game.[989.91] [989.91][S01]That's very different environment[991.44] [991.44][S01]from, say, Go or Chess.[993.49] [993.49][S01]So one of the challenge here is that in \"StarCraft II,\"[997.33] [997.33][S01]you only see partial information from your environment.[1000.215] [1000.215][S01]There is this thing called fog of war where you cannot see what[1002.84] [1002.84][S01]your enemy is doing.[1004.23] [1004.23][S01]And so you really have to take this into account[1007.01] [1007.01][S01]during your strategy, while in Go, in chess, you[1012.38] [1012.38][S01]see the whole board in front of you.[1013.88] [1013.88][S01]So you have access to the whole state of the game.[1016.07] [1016.07][S01]Also, the agent has to play in real time because \"StarCraft[1019.454] [1019.454][S01]II,\" once you play against a human at test time,[1021.98] [1021.98][S01]you really have to play as fast as you can,[1024.19] [1024.19][S01]so that you cannot take minutes to think.[1027.13] [1027.13][S01]And the \"StarCraft II\" agent was able to beat very,[1030.079] [1030.079][S01]very good humans at the game.[1032.14] [1032.14][S02] So now this is one step on even further.[1034.46] [1034.46][S02]So it's like at every step, you remove[1037.06] [1037.06][S02]something that made it easy.[1038.42] [1038.42][S02]So Atari, you get points for every action.[1040.579] [1040.579][S02]chess, you don't.[1041.66] [1041.66][S02]In chess, you get to see the entire board.[1045.17] [1045.17][S02]In \"StarCraft,\" you don't.[1046.54] [1046.54][S02]In \"StarCraft,\" it's competitive,[1048.46] [1048.46][S02]in what you're doing now, it isn't.[1050.537] [1050.537][S01] Yes.[1051.37] [1051.37][S01]So those agents were trying to maximize the score[1054.97] [1054.97][S01]and to win the game, while what we do in SIMA is very different.[1058.88] [1058.88][S01]We train agents to follow instructions.[1061.16] [1061.16][S02] What's this project called?[1062.83] [1062.83][S01] Our project is called SIMA.[1064.622] [1064.622][S01]That stands for Scalable Instructable Multi-world Agent.[1068.83] [1068.83][S01]And we used various different video games and also research[1075.31] [1075.31][S01]environments to train an agent to follow instructions[1079.69] [1079.69][S01]that we ask it to do.[1081.26] [1081.26][S01]Those instructions are fairly simple that you can typically[1084.34] [1084.34][S01]carry out in 5 to 10 seconds, such as pick up the apple,[1087.97] [1087.97][S01]or turn around, climb up the ladder, et cetera.[1091.46] [1091.46][S01]And so there is no score.[1093.19] [1093.19][S01]And we use the games really as containers,[1095.21] [1095.21][S01]as walls in which we can carry out[1097.42] [1097.42][S01]a whole bunch of instructions.[1098.885] [1098.885][S01]And the amount of instruction, the number of instructions[1101.26] [1101.26][S01]that you can come up with is very large due to the fact[1105.52] [1105.52][S01]that we use natural language to formulate those instructions.[1109.84] [1109.84][S01]So this poses other challenges.[1113.89] [1113.89][S01]We really want to make our agents to be general.[1116.54] [1116.54][S01]So while a Go agent only learns to beat the game of Go,[1121.48] [1121.48][S01]we are interested in making an agent that can really[1124.63] [1124.63][S01]understand what you're saying, understand the context,[1127.28] [1127.28][S01]understand the environment, and then carry out this instruction.[1130.06] [1130.06][S01]And so we are aiming for generality.[1132.0] [1132.0][S01]And I think generality is a key property that we want[1135.98] [1135.98][S01]in our agents to pursue AGI.[1138.693] [1138.693][S02] What do you mean by sandbox games?[1140.61] [1140.61][S01] So a sandbox game[1141.985] [1141.985][S01]is a game where the player can basically[1144.38] [1144.38][S01]do whatever they want within the space of what the game offers.[1148.53] [1148.53][S01]So it's a game which has natural landscape.[1151.62] [1151.62][S01]You can craft, you can build, you can cook, you can explore.[1155.45] [1155.45][S02] Like \"Minecraft.\"[1156.685] [1156.685][S01] Like \"Minecraft.\"[1158.06] [1158.06][S01]That's what we call a sandbox game because there is not really[1160.7] [1160.7][S01]a goal to the game.[1162.81] [1162.81][S01]Some games are sandbox with a goal, so it's a bit of a mix.[1166.82] [1166.82][S01]But yeah, you just have fun and do a bunch of things.[1170.077] [1170.077][S02] Like in an actual sandbox.[1171.66] [1171.66][S01] Exactly, yes.[1172.76] [1172.76][S02] Well, how about you?[1173.64] [1173.64][S02]Are you a gamer?[1174.6] [1174.6][S01] Yes, very much so.[1176.73] [1176.73][S01]I was given a Gameboy for my sixth birthday.[1178.805] [1178.805][S02] Sixth birthday?[1179.93] [1179.93][S01] Yes.[1180.92] [1180.92][S02] Lucky thing.[1181.92] [1181.92][LAUGHTER][1182.72] [1182.72][S01] Yeah, it was amazing.[1184.262] [1184.262][S01]I think my first game was \"Tetris.\"[1186.11] [1186.11][S01]And yeah, I was amazed by being able to play[1191.0] [1191.0][S01]those games in the car.[1192.65] [1192.65][S01]Car trips were never the same.[1194.74] [1194.74][S02] Yeah, that's true.[1195.99] [1195.99][S02]And as an adult, did you go beyond \"Tetris?\"[1197.908] [1197.908][S01] Yes, of course.[1199.2] [1199.2][S01]I remember when we at home got DSL,[1203.27] [1203.27][S01]which could give you basically constant internet access.[1206.463] [1206.463][S01]I was very happy because I could finally[1208.13] [1208.13][S01]play online games without being worried[1210.14] [1210.14][S01]about the counter, how much time do I spend on the internet.[1215.51] [1215.51][S01]So yeah, online games, competitive games,[1218.22] [1218.22][S01]sandbox games, and [INAUDIBLE] games, all of this I[1221.93] [1221.93][S01]do quite like.[1223.02] [1223.02][S02] I mean, so you've got[1224.395] [1224.395][S02]basically the perfect experience then to be in this space.[1226.98] [1226.98][S01] Looks like it.[1228.863] [1228.863][S02] The new projects that you're[1230.53] [1230.53][S02]working on now, these sandbox games,[1232.34] [1232.34][S02]I mean, they are quite a big departure.[1234.47] [1234.47][S02]How hard is it to get an agent to achieve things[1237.497] [1237.497][S02]in an environment where you don't have[1239.08] [1239.08][S02]that really clearly defined metric of success?[1241.14] [1241.14][S01] Yes, so it is a very different style of building[1245.71] [1245.71][S01]an agent training from before.[1247.39] [1247.39][S01]There is some prior work, which we were very inspired by,[1253.19] [1253.19][S01]which was the interactive agents project,[1255.01] [1255.01][S01]there is a blog post on the DeepMind blog post, where[1259.0] [1259.0][S01]they use the Playhouse environment[1261.43] [1261.43][S01]with a bunch of objects, and the goal[1263.29] [1263.29][S01]was to get the agent to carry out instructions in that house.[1267.173] [1267.173][S02] It's like a simulated environment?[1269.09] [1269.09][S01] It's a simulated environment, yes.[1270.25] [1270.25][S01]It's like a children's house with colorful objects.[1272.99] [1272.99][S01]And the goal was to take the teapots and put it on the beds,[1277.07] [1277.07][S01]and there was no reward as well.[1279.86] [1279.86][S01]So we wanted to take this to the next level.[1282.89] [1282.89][S01]And we also wanted to use games that are not games[1286.93] [1286.93][S01]that we make ourselves, that are available for everyone to play.[1290.235] [1290.235][S02] So basically it's the games[1291.86] [1291.86][S02]that you guys are playing?[1293.678] [1293.678][S02]Is that what it is?[1294.47] [1294.47][S01] Some of them are, indeed.[1296.178] [1296.178][LAUGHTER][1296.888] [1296.888][S02] I just want more time to spend on this game[1299.18] [1299.18][S02]that I've been doing, so I'm going[1300.597] [1300.597][S02]to build an agent to do it.[1301.795] [1301.795][S02]That's essentially what it sounds like.[1303.23] [1303.23][S01] Yeah, you could do it.[1304.813] [1304.813][S01]And so, yeah, we got into partnership[1309.5] [1309.5][S01]with game developers and game studios[1312.47] [1312.47][S01]to be able to use their games in our research.[1314.76] [1314.76][S01]And we picked games which are of a certain genre.[1319.52] [1319.52][S01]So we wanted 3D games because we were interested in these[1323.81] [1323.81][S01]embodied 3D setup.[1326.495] [1326.495][S01]We wanted to be able to define our own objective with language.[1329.12] [1329.12][S01]So if you go into \"Valheim,\" for example,[1331.71] [1331.71][S01]which is a Viking survival simulation,[1334.01] [1334.01][S01]you can make up a lot of tasks for the agents.[1336.03] [1336.03][S01]You can say, go get some mud, build a house,[1339.51] [1339.51][S01]craft jackets, et cetera.[1342.32] [1342.32][S02] Then how on Earth do you train an agent[1344.69] [1344.69][S02]when you don't have rewards?[1345.99] [1345.99][S01] So what we did is that we collected data[1348.35] [1348.35][S01]from humans playing the game.[1350.27] [1350.27][S01]And then we used a technique called imitation learning, which[1354.62] [1354.62][S01]teaches the agent to just mimic and predict[1357.56] [1357.56][S01]what a human would do next.[1359.16] [1359.16][S01]So given all of the observation, given what you've done before,[1362.25] [1362.25][S01]what would the human do?[1364.4] [1364.4][S01]And that's how we train the behavior of the agent.[1367.44] [1367.44][S01]So there is no reward.[1369.26] [1369.26][S01]We just imitate humans, which-- yeah.[1371.21] [1371.21][S02] Is it probabilistic, then?[1372.793] [1372.793][S02]Is it sort of like, in this sort of situation,[1375.05] [1375.05][S02]humans acted in this way, so we'll[1376.987] [1376.987][S02]act in that way with this probability?[1378.57] [1378.57][S01] Yes, exactly.[1379.11] [1379.11][S01]So all the actions of the agent are probabilistic.[1381.51] [1381.51][S01]So before the agent acts, there is a probability distribution,[1386.19] [1386.19][S01]which needs to sample and pick an action from.[1388.86] [1388.86][S01]Often it's the most likely action,[1390.387] [1390.387][S01]but sometimes you can take an unlikely action, which can[1392.72] [1392.72][S01]lead to unforeseen behavior.[1396.15] [1396.15][S01]Sometimes it's a good behavior.[1397.79] [1397.79][S01]Sometimes it's not.[1399.765] [1399.765][S02] Go on, give you an example.[1401.39] [1401.39][S01] Well, so for example,[1402.932] [1402.932][S01]we have a game such as \"Goat Simulator.\"[1405.9] [1405.9][S02] Tell us the objective of \"Goat Simulator.\"[1408.15] [1408.15][S01] So the objective of \"Goat Simulator\"[1410.317] [1410.317][S01]is to basically cause mayhem as a goat in the world.[1413.78] [1413.78][S01]And you can drive cars, you can jump on objects.[1416.373] [1416.373][S02] As a goat.[1417.29] [1417.29][S01] As a goat.[1417.89] [1417.89][S01]Yeah, you're a goat.[1418.97] [1418.97][S01]You can have wings and you can fly as well.[1421.57] [1421.57][S01]But it's a perfect sandbox game because you can[1424.03] [1424.03][S01]do so many different things.[1426.52] [1426.52][S01]And so in \"Goat Simulator,\" there is a key which causes[1430.06] [1430.06][S01]the goat to become all floppy and ragdoll on the floor.[1433.45] [1433.45][S01]I'm not too sure what's the--[1435.202] [1435.202][S01]when you would actually as a player[1436.66] [1436.66][S01]use this, maybe to pretend.[1439.403] [1439.403][S02] Play dead.[1440.32] [1440.32][S01] Exactly.[1441.32] [1441.32][S01]But sometimes when we play test the agents,[1443.98] [1443.98][S01]it ragdolls on the floor, which brings[1448.0] [1448.0][S01]some funny moments in the team when we do the playtest.[1450.408] [1450.408][S02] Just expecting it to barge into a crowd of people[1452.95] [1452.95][S02]and it just ragdolls.[1453.83] [1453.83][S01] Yes.[1454.12] [1454.12][S01]Yes.[1454.62] [1454.62][S01]That can happen.[1455.287] [1455.287][S02] Can I see what it looks like?[1456.995] [1456.995][S01] Yeah, I'll show you.[1458.72] [1458.72][S01]So here, let's start by \"Goat Simulator 3.\"[1464.217] [1464.217][S02] My fave.[1465.05] [1465.05][S01] So we have a video of the agent[1468.1] [1468.1][S01]here, which the instruction is to go to a green object.[1472.123] [1472.123][LAUGHTER][1473.62] [1474.97][S02] We've got a goat.[1476.44] [1476.44][S02]And we've got quite realistic environment.[1479.69] [1479.69][S02]And we've got the goat trotting forwards.[1481.977] [1481.977][S02]There's plenty of green, which the goat is walking past,[1484.31] [1484.31][S02]but it's doing a big jump into a pond.[1486.11] [1486.11][S02]And in the pond, there is an inflatable[1488.53] [1488.53][S02]in the shape of a snake, which the goat has successfully[1491.49] [1491.49][S02]reached.[1491.99] [1491.99][S01] It's green.[1493.115] [1493.115][S02] It's definitely green.[1494.66] [1494.66][S02]And it is an object.[1495.95] [1495.95][S01] It's an object, yes.[1496.55] [1496.55][S02] And it did a jump there to get there.[1498.65] [1498.65][S01] Yeah.[1499.525] [1499.525][S01]So sometimes the agent picks up on some human peculiarities.[1503.43] [1503.43][S01]For example, humans, when they play games,[1505.51] [1505.51][S01]are often bored to take the longest path.[1509.3] [1509.3][S01]For example, climbing up the stairs.[1510.8] [1510.8][S01]If you have a jetpack, you're just[1512.05] [1512.05][S01]going to jetpack over the stairs.[1513.5] [1513.5][S01]So one of the first agents where we asked,[1516.04] [1516.04][S01]get up to the platform, instead of taking[1518.26] [1518.26][S01]the stairs the long way, the agent[1519.97] [1519.97][S01]just jetpacked over the stairs.[1522.37] [1522.37][S01]So it does learn about human behavior.[1525.973] [1525.973][S02] The strange shortcuts.[1527.39] [1527.39][S02]I also noticed that rather than just walking into the pond,[1529.89] [1529.89][S02]it jumped, which is exactly what I would do as a human.[1532.68] [1532.68][S01] Yeah.[1533.93] [1533.93][S02] You want to make a splash.[1535.02] [1535.02][S01] Yes.[1535.4] [1535.4][S02] OK.[1535.89] [1535.89][S02]All right.[1536.13] [1536.13][S02]Well, I mean, it succeeded.[1537.255] [1537.255][S02]It succeeded.[1538.08] [1538.08][S01] So this is the game \"Valheim,\"[1541.04] [1541.04][S01]which is a Viking survival sandbox game.[1544.74] [1544.74][S01]And the task here that we asked the agent to carry out[1547.46] [1547.46][S01]is to pick up mushrooms.[1549.5] [1549.5][S02] OK, so again, another very beautifully crafted[1553.91] [1553.91][S02]environment.[1554.82] [1554.82][S02]The camera is moving around, which[1556.963] [1556.963][S02]I guess is the agent doing that, moving the camera around.[1559.38] [1559.38][S02]It has found quite quickly some mushrooms, shot them out[1562.745] [1562.745][S02]of the ground--[1563.37] [1563.37][S02]I don't know whether the Vikings had that particular ability--[1567.95] [1567.95][S02]and just carrying on wandering through the forest.[1570.63] [1570.63][S01] Yeah.[1570.86] [1570.86][S02] OK.[1571.485] [1571.485][S02]So definitely like camera control and movement in that 3D[1574.01] [1574.01][S02]environment, it's mastered.[1575.517] [1575.517][S01] Yes.[1576.35] [1576.35][S01]So I think because all the games that we use are 3D games,[1581.4] [1581.4][S01]the camera control is very similar from all those games.[1584.82] [1584.82][S01]So I there is a lot of experience[1586.378] [1586.378][S01]to learn from in terms of controlling the camera.[1588.42] [1588.42][S01]You can see that the movement of the camera is human-like.[1592.14] [1592.14][S01]So it hesitates.[1593.33] [1593.33][S01]It's not very robotic, just aim on the mushroom and go forward.[1596.08] [1596.08][S01]And that's because it's trained from human data.[1598.08] [1598.08][S01]If you train an agent with reinforcement[1600.41] [1600.41][S01]learning to do the same task, you could detect--[1602.99] [1602.99][S02] As opposed to imitation learning?[1604.22] [1604.22][S01] As opposed to imitation learning,[1606.262] [1606.262][S01]the agent would have very linear and robotic camera motion.[1610.83] [1610.83][S01]It would just go to the mushroom.[1612.47] [1612.47][S01]It needs to optimize its reward.[1614.43] [1614.43][S01]So it will not spend any time hesitating.[1617.43] [1617.43][S02] When you're instructing the agent,[1619.71] [1619.71][S02]you're not explaining what a mushroom is[1621.59] [1621.59][S02]or what a mushroom might look like in the game?[1623.79] [1623.79][S01] Yeah, so it gets that from the training data.[1626.58] [1626.58][S01]So in this training data, there is a bunch[1630.35] [1630.35][S01]of sequences of humans interacting with mushrooms,[1633.24] [1633.24][S01]picking up mushrooms.[1634.76] [1634.76][S01]And so it just learns from the data.[1637.02] [1637.02][S01]So an interesting task we could ask the agent here[1641.42] [1641.42][S01]is, if you hold out, if the agent has never[1643.85] [1643.85][S01]seen mushrooms or maybe never seen red mushrooms,[1646.65] [1646.65][S01]but maybe it has seen blue mushrooms,[1648.365] [1648.365][S01]and you ask the agent to pick up the red mushrooms,[1650.49] [1650.49][S01]is the agent able to generalize color and object type?[1655.547] [1655.547][S01]So those type of questions are the questions[1657.38] [1657.38][S01]we are interested in.[1658.58] [1658.58][S02] Which is definitely demonstrating[1660.47] [1660.47][S02]a sort of conceptual understanding of the environment[1662.33] [1662.33][S02]that it's in.[1662.97] [1662.97][S01] Yeah, it would.[1664.41] [1664.41][S01]Yes.[1665.39] [1665.39][S01]We're looking for this generality and transfer[1667.55] [1667.55][S01]between concepts and trying to put things together,[1670.65] [1670.65][S01]new concepts together to carry out a task.[1673.98] [1673.98][S01]And yeah, this will show understanding[1676.73] [1676.73][S01]of what a mushroom is and what the color red is.[1679.048] [1679.048][S02] In a way, the way that you describe the imitation[1681.59] [1681.59][S02]part of it, the imitation learning,[1683.42] [1683.42][S02]it sounds a little bit like you're[1685.49] [1685.49][S02]doing a sort of computer game version of predictive text,[1688.26] [1688.26][S02]almost.[1688.76] [1688.76][S02]Like, what has been done in this environment before,[1691.17] [1691.17][S02]how do you repeat it?[1692.682] [1692.682][S01] It is exactly that.[1694.14] [1694.14][S02] Is it?[1694.94] [1694.94][S01] Yes, so when you train language models[1697.52] [1697.52][S01]on language, the objective in which you train,[1700.8] [1700.8][S01]at least the initial model, is to predict[1703.85] [1703.85][S01]what's the next token.[1707.54] [1707.54][S01]I mean, if it's assuming that it's all human-generated text,[1711.44] [1711.44][S01]what would the human say next?[1713.16] [1713.16][S01]What word would the human say next?[1714.69] [1714.69][S01]Which is very similar to this, which is,[1716.55] [1716.55][S01]what would the human do next in terms of the keyboard and mouse[1720.17] [1720.17][S01]action?[1720.84] [1720.84][S01]So you're right, it is pretty much the same technique.[1724.245] [1724.245][S02] The same ideas.[1725.37] [1725.37][S02]But the key difference here is that you're instructing it.[1727.83] [1727.83][S02]You're saying, this is the objective for you[1730.55] [1730.55][S02]and go and do that in a way that a human would do?[1733.017] [1733.017][S01] Yes.[1733.85] [1733.85][S01]So it's given this instruction, what would the human do next?[1736.59] [1736.59][S01]But the language models are also,[1738.96] [1738.96][S01]in a way, prompted with some prior text.[1740.8] [1740.8][S01]So they would say, given that this text is[1742.55] [1742.55][S01]in your prompt, what is the next most probable word that you[1747.68] [1747.68][S01]should output?[1748.56] [1748.56][S01]So it's not very different.[1751.5] [1751.5][S01]The maybe infrastructure and architecture of the agent[1755.18] [1755.18][S01]itself is a bit different, because we here[1757.52] [1757.52][S01]need to produce a very different modality, which is action[1762.98] [1762.98][S01]in keyboard and mouse space.[1764.34] [1764.34][S02] It sounds like it's kind[1764.99] [1764.99][S02]of exploring the network of actions[1766.58] [1766.58][S02]that people have taken in the past.[1768.65] [1768.65][S02]Does it do original things, then?[1771.18] [1771.18][S01] So it will not do anything[1773.51] [1773.51][S01]that humans haven't done, per se,[1776.76] [1776.76][S01]but the agent is able to generalize[1778.76] [1778.76][S01]to unseen environments, so in a way that's kind of novel.[1782.18] [1782.18][S01]So it's like, you're a tennis player.[1785.12] [1785.12][S01]And suddenly, you're asked to do a badminton match,[1788.1] [1788.1][S01]right, as a human.[1788.93] [1788.93][S01]Because you've learned tennis, you[1790.61] [1790.61][S01]have some reflexes and mechanics that you know.[1794.17] [1794.17][S01]You know how to hold the racket, how to strike a ball.[1799.05] [1799.05][S01]And so you're going to do fairly well at badminton.[1802.02] [1802.02][S01]Not the best, but you're not going to be beginner level.[1805.08] [1805.08][S01]And in our agent, we see the same.[1807.24] [1807.24][S01]So what we've done is that we trained the agent[1809.99] [1809.99][S01]on all environments except one, and then we[1812.185] [1812.185][S01]test the agent on that one environment.[1813.81] [1813.81][S01]And we can see that the agent is able to do[1817.71] [1817.71][S01]a few things at a reasonable level in that environment.[1820.66] [1820.66][S01]So in a way, I don't know if you can say those are novel,[1824.62] [1824.62][S01]but it definitely shows some promise with respect[1827.28] [1827.28][S01]to generalization and generalizing to unseen[1829.56] [1829.56][S01]environments.[1830.52] [1830.52][S02] So, OK, tell me how well does it perform overall,[1833.19] [1833.19][S02]then?[1833.38] [1833.38][S01] Maybe I can tell you[1833.91] [1833.91][S01]about two interesting results.[1835.0] [1835.0][S02] Yes, please.[1836.0] [1836.0][S01] So first of all, we[1837.87] [1837.87][S01]observed that when we train an agent on all[1840.39] [1840.39][S01]of the environments, it performs better[1842.85] [1842.85][S01]than an agent that is only trained[1844.47] [1844.47][S01]on each singular individual environment.[1846.803] [1846.803][S02] So you train an agent on \"No Man's Sky,\"[1848.97] [1848.97][S02]\"Goat Simulator 3,\" \"Valheim,\" and then it does better[1853.74] [1853.74][S02]in, for example, \"Goat Simulator 3\"[1857.4] [1857.4][S02]than an agent that is trained only in \"Goat Simulator 3?\"[1860.452] [1860.452][S01] Exactly, yes.[1861.66] [1861.66][S01]So that was a very good result for us[1866.64] [1866.64][S01]because it demonstrates that the experience that you[1869.85] [1869.85][S01]get from training on variety of game[1872.13] [1872.13][S01]make your skill in each individual game stronger.[1875.97] [1875.97][S01]So what I like to think of is, let's say as a human,[1879.34] [1879.34][S01]maybe to go back to the tennis example, maybe I train on tennis[1882.15] [1882.15][S01]and I train on badminton.[1883.35] [1883.35][S01]And maybe I'll be better at both of those compared[1887.82] [1887.82][S01]to if I had only trained on, say, tennis[1890.73] [1890.73][S01]because maybe I gained some extra skills and subtlety[1893.85] [1893.85][S01]in the type of actions I need to do from being exposed to a wider[1898.56] [1898.56][S01]range of skills and experience.[1900.433] [1900.433][S02] Is it also partly they're just exposed to more[1902.85] [1902.85][S02]data, so more of how humans manipulate 3D environments,[1906.4] [1906.4][S02]more of different kinds of things that humans might do?[1909.07] [1909.07][S01] Yeah.[1909.945] [1909.945][S01]So games, different games offer different situations.[1912.55] [1912.55][S01]So the agent can extract knowledge from those situations[1916.11] [1916.11][S01]and maybe relate them to other situations from other games.[1919.54] [1919.54][S01]And so it just accumulates and consolidates the behavior[1922.86] [1922.86][S01]into something that is better than not having seen[1927.64] [1927.64][S01]this other experience, which in a way, our hypothesis when we[1931.41] [1931.41][S01]started the project was the same hypothesis as the language[1934.78] [1934.78][S01]model.[1935.28] [1935.28][S01]Training a lot of data will make some interesting properties[1940.05] [1940.05][S01]emerge, such as being able to generalize and transfer.[1943.72] [1943.72][S01]And so, yeah, that's what we also observed here.[1948.413] [1948.413][S02] But then I think also,[1949.83] [1949.83][S02]that result isn't immediately obvious because it could well[1952.707] [1952.707][S02]have been the case that actually all the things that it learns[1955.29] [1955.29][S02]from \"No Man's Sky,\" like, confuse it when it goes back[1958.29] [1958.29][S02]to \"Goat Simulator 3.\"[1959.29] [1959.29][S01] Yes.[1960.123] [1960.123][S01]So we were worried about that because we thought maybe[1962.49] [1962.49][S01]there will be a destructive interference between games.[1965.98] [1965.98][S01]That's true.[1967.23] [1967.23][S01]We picked those games to all be 3D because we thought there was[1972.33] [1972.33][S01]a chance that some of the experience would transfer[1975.297] [1975.297][S01]between games.[1975.88] [1975.88][S01]If we had trained on \"Tetris\" and \"No Man's Sky,\"[1978.04] [1978.04][S01]we wouldn't expect any positive interference here.[1983.362] [1983.362][S02] Slightly different environment, isn't it?[1985.57] [1985.57][S01] Yes.[1986.403] [1986.403][S01]But because all of those games have some similar concepts,[1990.33] [1990.33][S01]same way of navigating around, moving the camera,[1993.53] [1993.53][S01]we thought maybe there would be some interesting and positive[1999.16] [1999.16][S01]outcome to training on multiple games.[2002.917] [2002.917][S02] You told me that there were two,[2004.75] [2004.75][S02]there were two big results.[2005.77] [2005.77][S01] Yes.[2006.45] [2006.45][S01]So the other game, the other result,[2008.02] [2008.02][S01]sorry, which I mentioned before, was[2010.68] [2010.68][S01]that when you train an agent on all environments but one,[2015.66] [2015.66][S01]it performs roughly as well on this held-out environment[2020.73] [2020.73][S01]as the agent that has only been trained on that environment.[2023.445] [2023.445][S02] OK, wait, wait, wait.[2024.82] [2024.82][S02]So, right, you've got eight games.[2027.6] [2027.6][S02]You train an agent on them and leave off \"Goat Simulator 3.\"[2031.98] [2031.98][S02]It's never seen \"Goat Simulator 3\" before.[2034.86] [2034.86][S02]And then you compare it against an agent that's seen nothing[2039.39] [2039.39][S02]else but \"Goat Simulator 3.\"[2040.557] [2040.557][S01] Exactly.[2041.557] [2041.557][S02] And then you test them in \"Goat Simulator 3.\"[2044.17] [2044.17][S01] Yes, and the SIMA agent here[2046.89] [2046.89][S01]performs nearly as well as the environment.[2049.445] [2049.445][S01]We call it the environment expert[2050.82] [2050.82][S01]because it's an expert in goat expert.[2053.409] [2053.409][S02] Goal expert, sure.[2055.23] [2055.23][S01] So, yes, it doesn't[2056.94] [2056.94][S01]perform as well, on average, but it does perform reasonably well.[2062.83] [2062.83][S01]So it has learned to generalize to unseen.[2065.79] [2065.79][S01]All the other games have different avatars, for example.[2068.73] [2068.73][S01]The agent has not--[2069.69] [2069.69][S02] There's no goats anywhere else.[2071.482] [2071.482][S02]There's no goats in space.[2072.788] [2072.788][S01] But the agent is able to move the goat around[2075.33] [2075.33][S01]and perform a few tasks.[2076.989] [2076.989][S01]Of course, there are some actions in \"Goat Simulator\"[2080.489] [2080.489][S01]which are specific and unique to \"Goat Simulator,\"[2082.86] [2082.86][S01]and that we don't--[2083.739] [2083.739][S02] There's no ragdolling astronauts.[2085.0] [2085.0][S01] No, and we don't expect the agent[2087.042] [2087.042][S01]to be able to get that because it has never seen it,[2090.81] [2090.81][S01]or it has never seen experience from it.[2093.659] [2093.659][S01]But other things like navigation, maybe concepts[2097.44] [2097.44][S01]about objects and colors, that is something[2101.49] [2101.49][S01]that it can achieve.[2102.348] [2102.348][S02] So how do you expand that[2103.89] [2103.89][S02]beyond these sort of tasks that last[2105.96] [2105.96][S02]for 10 seconds, 15 seconds of jumping and picking up[2109.14] [2109.14][S02]a mushroom?[2109.74] [2109.74][S01] We are working currently[2111.407] [2111.407][S01]on trying to find ways of doing that.[2113.96] [2113.96][S01]First of all, there is a lot to improve because our agent is not[2118.87] [2118.87][S01]at human level, even for those short horizon tasks.[2122.53] [2122.53][S01]For longer horizon tasks, we are looking into methods which[2129.76] [2129.76][S01]would allow us to do that.[2131.59] [2131.59][S01]You mentioned large language models before,[2134.15] [2134.15][S01]which are able to maybe reason on a higher time horizon, longer[2138.94] [2138.94][S01]time horizon.[2139.88] [2139.88][S01]So maybe combining those models with our agent could be a way,[2143.44] [2143.44][S01]but that's future work.[2144.562] [2144.562][S02] Can I play alongside one of these things,[2146.77] [2146.77][S02]just to make my experience of the game better?[2148.91] [2148.91][S02]Can I cheat, basically?[2150.315] [2150.315][S01] So some agents do[2151.69] [2151.69][S01]find ways of exploiting the games to do things that were not[2158.02] [2158.02][S01]originally intended.[2159.17] [2159.17][S01]So in a way, you could call that cheating, I guess.[2162.82] [2162.82][S01]But playing, like, once we have very competent AI systems, which[2168.873] [2168.873][S01]could do all sorts of things, playing games with you[2171.04] [2171.04][S01]could also be one of the things you[2172.54] [2172.54][S01]want to do with those systems.[2175.12] [2175.12][S01]That would be quite fun, I think, especially[2177.45] [2177.45][S01]in these sandbox games.[2178.57] [2178.57][S01]If you had an agent, which you could say, OK,[2180.445] [2180.445][S01]now I want to play hide-and-seek in \"Goat Simulator,\"[2183.33] [2183.33][S01]and let's say it's multiplayer, you have two goats,[2185.91] [2185.91][S01]and you can play whatever you want.[2188.012] [2188.012][S01]I think it would be quite fun to interact[2189.72] [2189.72][S01]with an agent in exactly the way you want.[2192.55] [2192.55][S01]So the way we currently sometimes use large language[2196.11] [2196.11][S01]models is we prompt them to do all sorts of things,[2198.58] [2198.58][S01]write a poem, do some role play, write code.[2202.23] [2202.23][S01]If we had the same ability, but also[2203.88] [2203.88][S01]with an agent that has an avatar in game, which you could prompt,[2208.75] [2208.75][S01]I think that could be lots of fun for gamers.[2211.36] [2211.36][S02] Yeah, it's like, level, level,[2213.36] [2213.36][S02]level up on an NPC, isn't it?[2215.53] [2215.53][S01] Yes.[2216.43] [2216.43][S02] OK, so what is the big objective here, then?[2218.77] [2218.77][S01] So I think what we[2221.01] [2221.01][S01]are interested in in the SIMA team is[2225.03] [2225.03][S01]to build truly general agents.[2226.87] [2226.87][S01]So we use games as a platform to drive innovation[2231.93] [2231.93][S01]and breakthrough into this particular space[2234.49] [2234.49][S01]so that we can get a step closer to AGI[2236.96] [2236.96][S01]and understand how to build those truly general agents.[2239.99] [2239.99][S01]So we are really in for the research towards AGI.[2242.553] [2242.553][S02] What's the definition of AGI[2244.22] [2244.22][S02]that DeepMind work to?[2246.295] [2246.295][S01] I think the definition of AGI[2248.17] [2248.17][S01]is an agent that is as general and capable as a human.[2254.56] [2254.56][S01]And I think the general is key here.[2257.23] [2257.23][S01]It could be better than a human, but at least at human level.[2260.39] [2260.39][S02] So you can pick it up and drop it[2261.28] [2261.28][S02]into different environments and it[2262.697] [2262.697][S02]can do as well as a human can?[2264.38] [2264.38][S01] Yes.[2265.213] [2265.213][S01]So, I think if you think about gaming, as a human,[2268.868] [2268.868][S01]maybe there is a new game coming out.[2270.41] [2270.41][S01]I install it and very quickly I will be able to play this game.[2274.48] [2274.48][S01]Maybe I need to go through the tutorial[2276.13] [2276.13][S01]because there are some specificity to the game that I[2278.38] [2278.38][S01]need to learn, but in general, I think[2280.54] [2280.54][S01]we are very good at adapting to those new situations,[2283.13] [2283.13][S01]to those new environments, and be competent at it.[2285.77] [2285.77][S01]And we really want the same for our agents.[2287.953] [2287.953][S01]We want to drop the agent in a new environment[2289.87] [2289.87][S01]or show it a new game and observe that they can, indeed,[2294.31] [2294.31][S01]also adapt and become competent at those new situations.[2297.753] [2297.753][S02] Well, you've sort of done that already, haven't you?[2300.42] [2300.42][S01] There is still a long way to go.[2302.42] [2302.42][S01]So, our agents have been trained to only perform simple actions,[2307.89] [2307.89][S01]but really, there is so much more skills and scenarios[2315.5] [2315.5][S01]which we could use to train our agents.[2317.49] [2317.49][S01]How do we make an agent that could handle 1,000 different[2322.25] [2322.25][S01]games?[2323.15] [2323.15][S01]What does it take?[2324.8] [2324.8][S01]There are some breakthroughs that needs to happen.[2327.035] [2327.035][S02] So then let's say that you manage[2328.91] [2328.91][S02]that, which I'm sure you will.[2331.71] [2331.71][S02]What happens then?[2334.692] [2334.692][S02]Is the idea that you can take that learning and then apply[2338.48] [2338.48][S02]it to really generalist agents that exist in the environment[2342.44] [2342.44][S02]that we all live in?[2343.7] [2343.7][S01] Yes, there are methods.[2345.41] [2345.41][S01]How did we build the agent?[2346.83] [2346.83][S01]I think the understanding what method[2348.44] [2348.44][S01]we use to achieve this result is very important, that we can then[2351.32] [2351.32][S01]apply the same method to other domains, which[2354.37] [2354.37][S01]we might be interested in.[2355.61] [2355.61][S01]Understanding technical details of how to build those agents,[2361.05] [2361.05][S01]understanding how to collect the data, which data is useful,[2365.72] [2365.72][S01]algorithms, how do we train the agent,[2368.27] [2368.27][S01]I think that's very important as well.[2371.29] [2371.29][S01]So, yeah, any kind of innovation and progress[2373.87] [2373.87][S01]we can make towards understanding the[2375.73] [2375.73][S01]how to make progress in AGI is what we are after.[2380.505] [2380.505][S02] But these are lessons[2381.88] [2381.88][S02]that can apply elsewhere?[2383.467] [2383.467][S01] They could apply to other domains.[2385.55] [2385.55][S01]Yes.[2385.6] [2385.6][S02] OK.[2386.225] [2386.225][S02]So, give me an example, then.[2388.87] [2388.87][S02]What kind of agents should I be hoping for?[2392.8] [2392.8][S01] I mean, so are we talking about the far future?[2395.51] [2395.51][S02] Yeah, go on.[2396.51] [2396.51][S01] OK, so I think I could say an example of agents[2401.98] [2401.98][S01]would be self-driving cars that are safe and reliable.[2405.88] [2405.88][S01]I think those are agents which take[2408.64] [2408.64][S01]actions that have consequences.[2411.08] [2411.08][S01]And so having very robust and reliable self-driving car, which[2415.99] [2415.99][S01]can generalize to unseen circumstances,[2418.9] [2418.9][S01]would be very important.[2419.93] [2419.93][S01]And I think the generality here is very important.[2422.055] [2422.055][S01]Because when you drive, you never[2423.43] [2423.43][S01]know what's going to happen.[2424.597] [2424.597][S01]I mean, if we had agents in the real world,[2427.15] [2427.15][S01]one of the things we would like would be for agents to carry out[2431.47] [2431.47][S01]instructions for us.[2432.38] [2432.38][S01]For example, would be a robot in your house[2435.64] [2435.64][S01]where you say, maybe carry these heavy objects to the table.[2440.54] [2440.54][S01]That's a task which is defined in language,[2443.64] [2443.64][S01]which would be very useful.[2444.8] [2444.8][S01]Maybe virtual agents, which can code and write software[2450.32] [2450.32][S01]would be very useful for us.[2453.38] [2453.38][S01]Maybe agents that can help you in everyday tasks which you do[2458.66] [2458.66][S01]online, for example, shopping.[2461.27] [2461.27][S01]I mean, we spend a lot of time researching what to buy.[2465.3] [2465.3][S01]If I want a new pair of headphones,[2467.61] [2467.61][S01]I might spend hours to understand[2469.58] [2469.58][S01]which one is good for me.[2471.15] [2471.15][S01]But if I could have an agent that[2472.76] [2472.76][S01]can do this research for me, that would be very useful,[2476.76] [2476.76][S01]I think.[2477.26] [2477.26][S02] I'm just thinking about,[2478.19] [2478.19][S02]I'm just going back to your AGI stuff, actually, if I may,[2480.607] [2480.607][S02]just to finish on.[2481.92] [2481.92][S02]Do you think that the sparks of AGI will happen in your lab?[2486.76] [2486.76][S01] Well, it's very difficult to say,[2489.02] [2489.02][S01]but we do hope, yes, that the methods we develop[2493.45] [2493.45][S01]will help to make progress towards AGI.[2497.32] [2497.32][S01]I think if we had an agent that could play any games[2502.84] [2502.84][S01]and carry out any instruction in any game--[2505.66] [2505.66][S02] And perform a superhuman ability?[2507.64] [2507.64][S01] Yes, I think that would be a good milestone[2513.38] [2513.38][S01]to achieve because that would show human-level intelligence,[2516.99] [2516.99][S01]at least in games.[2518.51] [2518.51][S01]But games are very rich and sometimes are[2522.057] [2522.057][S01]a bit like the real world.[2523.14] [2523.14][S01]You need to understand your surroundings.[2525.058] [2525.058][S01]You need to understand objects.[2526.35] [2526.35][S01]You need to understand the consequence of your actions.[2529.28] [2529.28][S01]Especially if you add language into the mix[2531.24] [2531.24][S01]because if you add language into the mix for tasks, then really[2535.01] [2535.01][S01]you could ask the agent to do anything.[2536.67] [2536.67][S01]So if such agent is able to carry out[2538.28] [2538.28][S01]any instruction in any game, I think[2540.29] [2540.29][S01]it would be one step closer to AGI.[2543.262] [2543.262][S02] Fred, thank you very much for joining me.[2545.47] [2545.47][S01] Thank you so much for having me.[2547.61] [2547.61][S02] There have been a few moments now[2549.485] [2549.485][S02]while making this podcast where we have been here[2551.66] [2551.66][S02]at the early stages of a project, before the results were[2555.41] [2555.41][S02]grand or headline grabbing, when there was only the tiniest[2558.38] [2558.38][S02]little spark of something to see.[2560.36] [2560.36][S02]Like, remember when we first visited the robot lab[2563.09] [2563.09][S02]to watch them flailing around in circles, or the first time[2566.78] [2566.78][S02]that we saw AlphaFold when it was only just starting[2570.59] [2570.59][S02]to unravel the mysteries of proteins?[2573.4] [2573.4][S02]Well, I think it feels a lot like this is[2575.61] [2575.61][S02]the stage we're at with SIMA.[2577.78] [2577.78][S02]Because sure, picking up a mushroom[2579.9] [2579.9][S02]isn't quite the vision of the AI future[2582.09] [2582.09][S02]that we've been waiting for, but it's the direction[2585.15] [2585.15][S02]that those agents are moving in that[2588.12] [2588.12][S02]really holds the secret promise to what might be possible.[2591.1] [2591.1][S02]Because if we want AI that can make its own decisions,[2594.79] [2594.79][S02]achieve its own objectives, operate independently[2598.65] [2598.65][S02]of our instructions, then this is a vital step along the way.[2603.76] [2603.76][S02]So today, it might be a mushroom.[2606.24] [2606.24][S02]The question is, what will it be next time we come back?[2610.18] [2610.18][S02]You've been listening to \"Google DeepMind--[2612.07] [2612.07][S02]The Podcast,\" with me, Professor Hannah Fry.[2614.73] [2614.73][S02]If you like what you just heard, hit Subscribe on YouTube[2618.09] [2618.09][S02]or follow us on your favorite podcast platform.[2621.29] [2621.29][S02]We've got an exciting line up of conversations to come,[2623.65] [2623.65][S02]including an interview with Pushmeet Kohli, who[2627.0] [2627.0][S02]oversees Science Research at Google DeepMind.[2629.95] [2629.95][S02]If you have someone in mind that you'd[2631.89] [2631.89][S02]love to hear from on this podcast, do leave us a comment.[2635.37] [2635.37][S02]See you next time.[2636.5] [2636.5][MUSIC PLAYING][2640.75]"}
{"file_name": "audio/val_000003.wav", "transcription": "[0.0][MUSIC PLAYING][4.275] [6.65][S01] Welcome to \"Google DeepMind-- the Podcast.\"[9.15] [9.15][S01]I'm Professor Hannah Fry.[10.68] [10.68][S01]Right, agents, they are here, or almost here,[15.21] [15.21][S01]and they are probably all anyone is going[18.41] [18.41][S01]to be talking about in 2025.[20.97] [20.97][S01]But they are definitely not new.[24.42] [24.42][S01]My guest on today's episode is someone[26.84] [26.84][S01]who last came on the podcast in 2019[30.17] [30.17][S01]to talk to me about the multi-agent system he[33.08] [33.08][S01]was working on.[34.11] [34.11][S01]It could beat the professional StarCraft players[36.71] [36.71][S01]at their own game, and eventually went on[38.81] [38.81][S01]to achieve Grandmaster status.[40.98] [40.98][S01]But how have agents evolved since then?[44.64] [44.64][S01]What can they do now?[46.02] [46.02][S01]How have the advances in language models[48.35] [48.35][S01]and multimodal AI change things?[51.06] [51.06][S01]And how do you possibly go about building something[55.37] [55.37][S01]that can make autonomous decisions on behalf of its user?[59.73] [59.73][S01]Now, I should tell you, if you want a primer on agents,[62.84] [62.84][S01]you can watch our episode with Frederic Besse[64.989] [64.989][S01]that we recorded over the summer.[67.07] [67.07][S01]But for now, Oriol Vinyals is Vice President[70.39] [70.39][S01]of Drastic Research and co-tech-lead of Gemini.[73.91] [73.91][S01]And it's fair to say we've got quite a lot to catch up on.[76.88] [76.88][S01]Oriol, welcome back to the podcast.[78.56] [78.56][S02] Hi, thank you for having me.[79.88] [79.88][S01] What is Drastic Research?[81.56] [81.56][S02] Well, I keep telling[83.2] [83.2][S02]my team they have to think drastic,[85.03] [85.03][S02]meaning don't just do the incremental stuff[88.12] [88.12][S02]that everyone is thinking about.[89.75] [89.75][S02]Try to drastically think what will happen in a few years time,[94.19] [94.19][S02]and then try to backport those ideas[96.97] [96.97][S02]and then execute today with that mindset in mind.[99.95] [99.95][S02]So that's what drastic means.[101.205] [101.205][S02]But yeah, it's a word that I use a lot.[102.83] [102.83][S01] I think when I last got to see you,[104.788] [104.788][S01]you had been working on an agent that[107.14] [107.14][S01]could use a keyboard and a mouse to do things draw pictures[111.37] [111.37][S01]in Paint or play StarCraft.[114.71] [114.71][S01]And, well, things have moved on quite a bit since then.[117.85] [117.85][S02] So those agents, at the time,[120.19] [120.19][S02]you took sort of a very generic set of principles,[123.61] [123.61][S02]very simple principles in the field of machine learning,[127.75] [127.75][S02]and you would basically specialize a model on one task.[134.29] [134.29][S02]And what we were doing at the time[136.62] [136.62][S02]is have a curriculum of tasks that were more and more and more[140.46] [140.46][S02]difficult, right?[141.25] [141.25][S02]So when we last spoke, for instance, in video games,[144.58] [144.58][S02]we were looking at StarCraft, which[146.28] [146.28][S02]is one of the most complex modern strategy games out there.[150.4] [150.4][S02]And of course, DeepMind is notorious[151.98] [151.98][S02]for having started the trend with Atari, which[154.38] [154.38][S02]is a fairly simple game of left, right, hit the paddle[157.17] [157.17][S02]and hit the ball and off you go.[159.85] [159.85][S02]So that is what the algorithms themselves, we[164.94] [164.94][S02]tried to push for them to be very general so we can keep[168.42] [168.42][S02]climbing this ladder of difficulty, curriculum of games,[172.05] [172.05][S02]and doing more and more complex things.[174.91] [174.91][S02]And right now what has happened is even[178.62] [178.62][S02]the models we train are broadly applicable to many more things[182.76] [182.76][S02]than the models we developed back then were.[185.92] [185.92][S02]So think about the process of creating this digital brain[189.54] [189.54][S02]hasn't changed that much.[191.11] [191.11][S02]But what that brain was able to do[193.38] [193.38][S02]was reasonably narrow, although very complex,[196.18] [196.18][S02]like playing StarCraft or playing Go.[198.27] [198.27][S02]Right now, these models can do quite a lot more broad[203.64] [203.64][S02]applications and, of course, talking to us, chatbots, et[207.39] [207.39][S02]cetera, et cetera.[208.39] [208.39][S01] So back then, reinforcement learning[210.57] [210.57][S01]was your main lever, I guess.[213.04] [213.04][S01]How different are things now?[214.457] [214.457][S02] Yeah, so algorithmically, actually,[216.54] [216.54][S02]the process of AlphaGo and actually AlphaStar, those two[221.49] [221.49][S02]had the same set of sequence of algorithms applied[224.88] [224.88][S02]to creating this digital brain.[226.81] [226.81][S02]And it's not actually that different from how[229.8] [229.8][S02]current large language models or multimodal models[232.74] [232.74][S02]are created today.[233.74] [233.74][S02]There's two basic steps that have[235.47] [235.47][S02]been pretty constant throughout many years[238.58] [238.58][S02]in many of the projects that we've worked on,[241.11] [241.11][S02]which we can call the first one pretraining or imitation[244.7] [244.7][S02]learning.[245.52] [245.52][S02]That is, you start with random weights,[248.07] [248.07][S02]you have an algorithm that will try to imitate lots of data[252.98] [252.98][S02]that humans have created to either play a game[256.25] [256.25][S02]or, in this case, all of the internet, all of the knowledge[258.8] [258.8][S02]available to us.[260.22] [260.22][S02]And in that first stage, you just[263.06] [263.06][S02]adapt the weights to try to imitate that data as well[267.41] [267.41][S02]as possible.[268.23] [268.23][S01] And these weights are, essentially[270.68] [270.68][S01]inside each of the neurons is a series of numbers[275.15] [275.15][S01]that kind of describes how it's connected to everything else?[278.1] [278.1][S02] Yeah, so basically there[280.61] [280.61][S02]are units of computations that are neurons.[283.35] [283.35][S02]And the connections between neurons[286.16] [286.16][S02]are what actually you have as weights.[288.63] [288.63][S02]So you can imagine that there's a neuron,[291.9] [291.9][S02]there's a few neurons connected to it.[294.21] [294.21][S02]And you're basically adding all the activations[297.79] [297.79][S02]from the incoming neurons there multiplied by the weights.[302.06] [302.06][S02]And those weights are the only things that move.[304.81] [304.81][S02]And the inputs excite the neurons.[306.92] [306.92][S02]It's pretty much how a brain works with some freedom[311.14] [311.14][S02]of, yeah, creativity.[312.425] [312.425][S01] OK, if we were to do an analogy,[315.22] [315.22][S01]it's almost like you've got the neurons and you're like,[319.39] [319.39][S01]water is flowing through it and the weights[322.48] [322.48][S01]is like the width of the pipes between the neurons?[324.98] [324.98][S02] Yeah, that's right.[326.397] [326.397][S02]And then you can imagine having millions of neurons and billions[332.8] [332.8][S02]or even trillions of pipes.[334.55] [334.55][S02]And that is what we spend most to compute actually[339.79] [339.79][S02]training these models, especially language models,[342.11] [342.11][S02]is in this pretraining or imitating all the data[344.98] [344.98][S02]that we have available to us.[346.61] [346.61][S01] OK, so you've now got this gigantic network[350.02] [350.02][S01]with loads of pipes going between all the neurons.[352.3] [352.3][S01]And that's your imitation phase done.[355.02] [355.02][S01]Next bit, if you were doing, say, AlphaGo or AlphaZero,[360.88] [360.88][S01]you would then get it to play itself.[363.76] [363.76][S02] Yeah, so this model now[365.64] [365.64][S02]is reasonably good at playing moves that look human-like.[369.4] [369.4][S02]So that means, of course, the sentences[373.5] [373.5][S02]are very plausible sentences in English.[376.27] [376.27][S02]Or if it was playing a game, it would click things reasonably[379.89] [379.89][S02]to move pieces on the board and whatnot.[383.07] [383.07][S02]But what this model hasn't done is[384.99] [384.99][S02]learn that these actions yield reward.[389.62] [389.62][S02]That's the bit of reinforcement learning or post-training,[393.07] [393.07][S02]which is the second phase of training.[395.26] [395.26][S02]So you can write a poem by just, hey,[400.12] [400.12][S02]just how does a poem on the internet look like on average?[403.72] [403.72][S02]But then the question is, well, I want only the good poems.[407.46] [407.46][S02]So how can I further adjust these pipes based on a signal[414.0] [414.0][S02]that now, having written now a whole poem,[417.01] [417.01][S02]would give a score of 0 or 1, let's say.[420.85] [420.85][S02]And if it's a mediocre poem, you get a 0.[424.15] [424.15][S02]If it's a good poem, you get a 1.[425.8] [425.8][S02]Again, for a game analogy, which is[427.83] [427.83][S02]what we use reinforcement learning traditionally, if you[431.22] [431.22][S02]win at the game, you get a 1.[432.73] [432.73][S02]If you lose, you get a 0.[433.96] [433.96][S02]And then you further adjust the weights.[437.17] [437.17][S02]But now, instead of imitating humans,[439.39] [439.39][S02]you're just saying forget, I want[441.57] [441.57][S02]to go beyond what humans could do[443.46] [443.46][S02]and try to really get all my poems to be the perfect poem[447.43] [447.43][S02]or all my chess games to be the perfect game.[449.8] [449.8][S02]And in language models, this second phase,[453.7] [453.7][S02]which is reinforcement learning post-training,[455.85] [455.85][S02]tends to be fairly short-lived because we do not have access[458.76] [458.76][S02]to super clean reward as you've won the game[462.18] [462.18][S02]or you lost the game, when you do[464.31] [464.31][S02]self-play in traditional board games, for example.[467.5] [467.5][S01] So once that's done, that's[470.64] [470.64][S01]all the stuff that goes on behind the scenes.[472.53] [472.53][S01]And then you're like, hold it right there.[475.447] [475.447][S02] Yeah.[476.28] [476.28][S01] Stay exactly where you are, everybody.[478.5] [478.5][S01]We're going to take just basically a snapshot[481.16] [481.16][S01]of this entire network, and that is what you actually[485.54] [485.54][S01]get to access as a user.[488.37] [488.37][S02] Yeah, so now this amazing process finished.[492.4] [492.4][S02]These weights are super precious, right?[494.73] [494.73][S02]So this configuration you found, you've[497.39] [497.39][S02]really spent months to finesse it, to tweak everything.[501.03] [501.03][S02]And now you will never move it anymore.[505.04] [505.04][S02]So training is over.[506.37] [506.37][S02]You're not changing the configuration anymore.[508.92] [508.92][S02]You might want to make it super efficient.[510.9] [510.9][S02]So say you find that, oh, look, this neuron is not that useful.[515.565] [515.565][S02]It's not used for anything.[516.69] [516.69][S02]You remove it, so everything becomes faster and cheaper[520.01] [520.01][S02]to run it at scale.[521.46] [521.46][S02]And then as a user, you just get the same weights.[524.73] [524.73][S02]Everyone gets the same weights we've trained.[526.95] [526.95][S02]That's what we call Gemini 1.5 Flash.[530.12] [530.12][S02]That just means a set of weights that[532.27] [532.27][S02]are frozen, will not change, will not[534.73] [534.73][S02]further train or anything.[536.42] [536.42][S02]So those two steps actually pretty much[538.51] [538.51][S02]are identical from AlphaGo to AlphaStar[542.83] [542.83][S02]to current large language models.[545.208] [545.208][S02]And of course, there's details that matter.[547.0] [547.0][S02]And the field has evolved, certainly,[549.35] [549.35][S02]but the principle is pretty much unchanged, actually.[552.67] [552.67][S01] Because under the hood, as it were,[554.75] [554.75][S01]there are differences between--[556.33] [556.33][S01]I don't know, I'm thinking like DQN here,[558.29] [558.29][S01]which was the Atari example, or the types of algorithms[562.3] [562.3][S01]that are used in AlphaGo or then again, in the large language[565.66] [565.66][S01]models, the architecture is different, right?[567.85] [567.85][S02] Yeah, so there's a few components[569.85] [569.85][S02]that go into then what the digital brain is.[573.14] [573.14][S02]One is the architecture.[574.37] [574.37][S02]So there are these neural networks.[577.6] [577.6][S02]Now we have the transformers, which we certainly didn't[581.2] [581.2][S02]have back in the DQN days.[582.86] [582.86][S02]So there's always some sort of breakthroughs[585.37] [585.37][S02]in architectures that are better at learning from the data.[589.6] [589.6][S02]But then from transformers to today,[592.54] [592.54][S02]it's almost all about little tweaks.[595.55] [595.55][S02]Even if you look at AlphaFold, which[597.22] [597.22][S02]also is fueled by a transformer, what the teams do for years[602.14] [602.14][S02]sometimes is just to find little tweaks on,[604.97] [604.97][S02]hey, let's remove this set of neurons,[607.07] [607.07][S02]let's add another layer, let's make this a bit wider.[610.79] [610.79][S02]So the brain shape changes a little bit,[613.79] [613.79][S02]and that makes it or breaks it sometimes,[615.91] [615.91][S02]in terms of the performance achieved.[618.02] [618.02][S01] So if these are all the things that[619.978] [619.978][S01]have been achieved so far, the goal, as I understand it,[624.56] [624.56][S01]is to create more agentic behavior,[626.8] [626.8][S01]to get these things to make autonomous decisions.[630.83] [630.83][S01]How did these help to achieve that end?[634.053] [634.053][S02] Yeah, so let's zoom[635.47] [635.47][S02]in a little bit on the current trend.[637.79] [637.79][S02]We call it large language models, but they're multimodal.[640.49] [640.49][S02]I think we had an episode earlier covering heavily[643.15] [643.15][S02]the multimodality aspect, how good[644.86] [644.86][S02]it is to be able to add an image,[647.15] [647.15][S02]then ask something, a follow-up question, and so on.[649.74] [649.74][S02]So this score, we will still improve it,[653.12] [653.12][S02]this set of weights that do these amazing inferences[658.67] [658.67][S02]about the input.[659.89] [659.89][S02]What's this image about?[661.08] [661.08][S02]What's the user asking?[662.43] [662.43][S02]Can I write a better poem?[664.02] [664.02][S02]Can I just make it longer or whatever?[665.9] [665.9][S02]Like all these interactions we all get to play with these days.[670.55] [670.55][S02]But this is just a component.[673.25] [673.25][S02]And we can think, hey, this is now our CPU.[677.18] [677.18][S02]And we can add more to it, around it.[680.79] [680.79][S02]What if the model could go off and do research for you,[685.58] [685.58][S02]for example?[686.49] [686.49][S02]One example, we were already thinking[688.61] [688.61][S02]about that back in the day, I could ask a model, a language[691.85] [691.85][S02]model or a visual language model to learn[695.24] [695.24][S02]to play the game of StarCraft.[697.07] [697.07][S02]That's a very different approach to say, create one agent that[700.25] [700.25][S02]does play the game.[701.36] [701.36][S02]In this other example, it could go online, watch videos[706.07] [706.07][S02]about the game.[707.28] [707.28][S02]It could, of course, download the game[709.25] [709.25][S02]to start interacting with it to learn, oh, yeah, I know,[713.06] [713.06][S02]I get it.[714.5] [714.5][S02]Do research online, go to forums, read the forums.[717.92] [717.92][S02]Go play and figure out that it's weak at this thing[721.4] [721.4][S02]and improve and so on.[722.73] [722.73][S02]And after, literally, it could be weeks,[725.27] [725.27][S02]it sends you an email that says, I now know how to play the game.[728.94] [728.94][S02]Let's play.[730.14] [730.14][S02]That's not a reality that's that far away.[732.57] [732.57][S02]But these models all of a sudden actually do something,[737.07] [737.07][S02]take some actions, and learn anything new[739.76] [739.76][S02]that is available to them.[741.18] [741.18][S02]And that's pretty powerful to think about.[744.03] [744.03][S02]It's what pushes the generality the most.[746.55] [746.55][S02]And that's what makes the AGI, as many people call it,[751.76] [751.76][S02]feel closer.[752.748] [752.748][S01] So if I understand it correctly then,[754.79] [754.79][S01]it's almost like the stuff that we[756.44] [756.44][S01]have at the moment, the large language models,[757.98] [757.98][S01]the multimodal models, whatever you want to call them,[760.23] [760.23][S01]that's the central core.[762.18] [762.18][S01]But the next step is that you build stuff[765.13] [765.13][S01]on top of that central core, that it can go off[767.92] [767.92][S01]and take off the stabilizers and go off and do its own thing.[771.972] [771.972][S02] Yeah, exactly.[773.18] [773.18][S02]If it has access to all the knowledge[775.75] [775.75][S02]and it can use its time to do some proper research,[780.71] [780.71][S02]write hypotheses, write some code[782.32] [782.32][S02]and so on, and take its time to really answer very, very, very[785.92] [785.92][S02]complex questions, then, yeah, the possibilities[789.89] [789.89][S02]now have broadened quite drastically.[792.32] [792.32][S02]Although of course, we're not going[794.11] [794.11][S02]to need that for everything.[795.53] [795.53][S02]I mean, if we ask a question like, hey, I like rice.[800.24] [800.24][S02]What should I prepare tonight?[801.53] [801.53][S02]Probably no need to do a very deep dive into thinking or just[806.08] [806.08][S02]going off for three weeks.[807.74] [807.74][S02]Then you'll probably not be very happy about the waiting time,[810.595] [810.595][S02]right?[811.6] [811.6][S02]But I think to push the frontier,[815.27] [815.27][S02]you're giving a digital body to the computer so it can not[819.04] [819.04][S02]only just think and give an instruction or a word output,[823.52] [823.52][S02]but it can also go off and do things online[826.98] [826.98][S02]or on documents that you might upload or whatever.[830.5] [830.5][S02]And ask very, very complex questions and personalized[832.89] [832.89][S02]to you, et cetera, et cetera.[834.13] [834.13][S01] I like that idea, this central core and then[836.07] [836.07][S01]you're giving it a digital body.[837.4] [837.4][S01]You've got the electric brain, and now you're[839.275] [839.275][S01]giving it a digital body.[840.43] [840.43][S02] Exactly.[841.05] [841.05][S01] OK, so in terms of the electric brain[843.092] [843.092][S01]and in terms of this core, this processor,[844.858] [844.858][S01]let me just ask you a bit about that.[846.4] [846.4][S01]I guess we should be thinking about Gemini here, right?[849.013] [849.013][S02] Mm-hmm.[849.93] [849.93][S01] Which is essentially what we're talking about,[851.32] [851.32][S01]the multimodal model that you guys have.[853.78] [853.78][S01]I know that one of the big ideas for large models[856.95] [856.95][S01]was just to scale it up, to get them bigger and bigger[859.26] [859.26][S01]and bigger and bigger and bigger.[860.74] [860.74][S01]Do you think that the results that we've seen from scaling[864.33] [864.33][S01]have plateaued by now?[866.26] [866.26][S02] Yeah, it's a very important question.[868.88] [868.88][S02]We have studied how as you make the models larger,[872.83] [872.83][S02]that is how many neurons literally these models have,[877.33] [877.33][S02]how do they become better at certain tasks[881.16] [881.16][S02]that we have clear metrics from the whole machine[883.68] [883.68][S02]learning community?[884.53] [884.53][S02]For example, one that is very simple to understand[887.52] [887.52][S02]is machine translation.[888.9] [888.9][S02]So how good the models are at translating[892.17] [892.17][S02]between two languages as you scale,[894.64] [894.64][S02]as you go from millions to billions[897.03] [897.03][S02]to potentially trillions of neurons,[900.03] [900.03][S02]you can see the performance keep improving.[902.64] [902.64][S02]Now, even when you do those studies,[905.83] [905.83][S02]one trick is that it looks linear,[909.43] [909.43][S02]but you have to plot logarithmic axis.[913.26] [913.26][S02]What that means in lay terms is that, let's say[917.58] [917.58][S02]from the last three years, we had some improvement,[920.83] [920.83][S02]you shouldn't expect the same improvement[922.92] [922.92][S02]in the next three years.[924.15] [924.15][S02]It's actually exponentially hard to get there.[926.74] [926.74][S02]So that means the compute investment, which[930.0] [930.0][S02]of course, also advances at the super linear rate,[932.74] [932.74][S02]but perhaps not as good as these sort of trends suggest,[937.485] [937.485][S02]you would just see some diminishing returns.[940.05] [940.05][S02]Because simply scaling the x-axis,[943.08] [943.08][S02]the number of parameters, you need to go 10x to see the same[947.03] [947.03][S02]improvement.[948.03] [948.03][S02]And that just creates some pressure to,[951.71] [951.71][S02]hey, maybe we can't scale as much[954.78] [954.78][S02]and we need to think about other ways[956.39] [956.39][S02]to scale to make the models better.[957.965] [957.965][S01] The example I give to my students[959.84] [959.84][S01]is like, if you've got a room that's really messy,[962.55] [962.55][S01]the first 10 minutes that you spend tidying,[964.74] [964.74][S01]it's going to make a massive difference.[966.03] [966.03][S01]You pick up all the dirty plates,[967.05] [967.05][S01]put away all the dirty washing, fine.[968.91] [968.91][S01]But once you're like 7 hours in, that extra 10 minutes,[972.69] [972.69][S01]it's not going to make any difference at all.[974.698] [974.698][S01]And that's essentially where we are, right?[976.49] [976.49][S02] Yeah, that's exactly a very good analogy.[978.9] [978.9][S02]And in fact, that analogy can even[980.63] [980.63][S02]apply to then the performance of the models.[982.92] [982.92][S02]Even if you have extremely good performance,[985.47] [985.47][S02]if you want these models to be 100% factual,[988.962] [988.962][S02]they will never make something up--[990.42] [990.42][S02]we know that if you probe them, you can make them say some[993.17] [993.17][S02]things that are not real--[995.27] [995.27][S02]even that last mile also is super hard,[998.43] [998.43][S02]which creates some interesting challenges to deploy them[1001.45] [1001.45][S02]at scale.[1002.21] [1002.21][S01] So OK, I hear what you're[1003.16] [1003.16][S01]saying about how there's diminishing returns[1004.993] [1004.993][S01]in all of this.[1005.62] [1005.62][S01]But in terms of how you make these things better,[1008.9] [1008.9][S01]how you make these models better,[1010.37] [1010.37][S01]is it just data, computational power, and size?[1014.42] [1014.42][S01]Are those the only things, the levers that you have to pull?[1017.14] [1017.14][S02] Yeah, so certainly,[1019.54] [1019.54][S02]if you froze the architecture, let's say[1022.93] [1022.93][S02]for the next year, no innovation,[1025.19] [1025.19][S02]we just scale because there's better hardware coming out--[1027.607] [1027.607][S01] Just make it bigger, yeah.[1029.19] [1029.19][S02] --make it bigger, that certainly[1031.329] [1031.329][S02]would have a trend that would look OK.[1033.74] [1033.74][S02]But what's happened, and certainly in Gemini,[1036.88] [1036.88][S02]we have other innovations, other tricks, techniques,[1042.91] [1042.91][S02]details from how to order the data that you present[1048.099] [1048.099][S02]the model with, to the details of the architecture,[1051.62] [1051.62][S02]to how to run the training process, how long to run it for.[1056.14] [1056.14][S02]What kind of data do we actually present the model?[1059.34] [1059.34][S02]How do we filter?[1060.6] [1060.6][S02]Do we present more data that's high quality, less data that's[1063.29] [1063.29][S02]low quality?[1063.96] [1063.96][S02]All sorts of different what we call hyperparameters[1068.25] [1068.25][S02]and, of course other algorithmic advances,[1070.71] [1070.71][S02]we also investigate fairly carefully.[1073.55] [1073.55][S02]Because the process of training a model is expensive.[1076.65] [1076.65][S02]So we need to be extremely careful with piling up[1080.15] [1080.15][S02]innovation so that eventually when we are ready,[1083.22] [1083.22][S02]we have enough innovation.[1084.63] [1084.63][S02]And also probably we'll have a better scale[1087.59] [1087.59][S02]to run for the next iteration of models.[1090.12] [1090.12][S02]We run it, and then we get algorithmic, not[1092.87] [1092.87][S02]only breakthroughs through data and compute.[1095.56] [1095.56][S01] I guess the other thing about this scaling stuff[1098.06] [1098.06][S01]is that there's no limit, really, to the number of nodes[1101.63] [1101.63][S01]that you can put in.[1102.84] [1102.84][S01]Maybe there's no limit, in theory,[1105.308] [1105.308][S01]to the computational power that you put in.[1107.1] [1107.1][S01]But there is a limit to the data that you can put in.[1109.08] [1109.08][S01]There's a limit to the number of human words that are out there.[1111.75] [1111.75][S02] Good point.[1112.833] [1112.833][S02]So I think there is a limit on the nodes.[1115.91] [1115.91][S02]Because how you scale these models is,[1118.07] [1118.07][S02]well, they don't fit on one single chip, hardware chip.[1120.59] [1120.59][S02]So now you have a mesh of chips that are communicating.[1123.8] [1123.8][S02]There's certain limits like speed of light,[1125.775] [1125.775][S02]et cetera, et cetera.[1126.65] [1126.65][S01] [LAUGHS][1127.405] [1127.405][S02] So there starts to be[1128.905] [1128.905][S02]a time where the efficiency of training such a big model[1131.8] [1131.8][S02]is also just not worth it, even from a utilization[1135.01] [1135.01][S02]of the hardware at your disposal, but very good point.[1138.585] [1138.585][S02]The other bit that is critical on this pretraining, imitate[1142.57] [1142.57][S02]all the data, is that we do not have what[1145.06] [1145.06][S02]we call infinite data regime.[1147.67] [1147.67][S02]There's finite data.[1149.48] [1149.48][S02]And so as soon as the models need to--[1152.86] [1152.86][S02]you can think, well, let's train on all the data.[1155.183] [1155.183][S02]If you want to train--[1156.1] [1156.1][S01] Everything humans have ever read--[1157.21] [1157.21][S02] Everything.[1157.7] [1157.7][S01] --or written.[1158.14] [1158.14][S02] All of the internet.[1159.598] [1159.598][S02]So we're just starting to think, OK, we're running out of data.[1163.16] [1163.16][S02]There are techniques like synthetic data.[1165.2] [1165.2][S02]Can we write or rewrite existing data in many different ways?[1169.45] [1169.45][S02]I mean, languages would be obvious ways to think, hey,[1171.7] [1171.7][S02]you could write the internet.[1173.84] [1173.84][S02]It's mostly in English, I mean, 60%.[1175.75] [1175.75][S02]I don't know what's the exact percentage.[1177.81] [1177.81][S02]But there are ways to rewrite the same knowledge[1180.8] [1180.8][S02]in different ways.[1181.98] [1181.98][S02]We're exploring those.[1183.42] [1183.42][S02]That's kind of a research area that many people[1185.6] [1185.6][S02]are starting to invest.[1187.02] [1187.02][S02]Because if you run out of data, the scaling laws[1190.85] [1190.85][S02]punish you even more.[1192.5] [1192.5][S01] So for example, then, you[1194.81] [1194.81][S01]could get Gemini to write its own version of the internet,[1199.05] [1199.05][S01]and then use that to train a new version of Gemini?[1201.428] [1201.428][S02] Yes.[1202.22] [1202.22][S01] Is there a danger, though,[1204.09] [1204.09][S01]that if you start feeding in the output of the same model,[1208.89] [1208.89][S01]that you can end up creating these little, well, unhelpful[1212.45] [1212.45][S01]feedback loops?[1213.87] [1213.87][S02] They certainly can[1215.45] [1215.45][S02]do some interesting experiments to test ideas[1218.9] [1218.9][S02]like this one you just mentioned.[1220.74] [1220.74][S02]And indeed, that is, on the surface, not a good idea.[1225.74] [1225.74][S02]The model suffers if you just ask it to recreate[1229.55] [1229.55][S02]all of the internet.[1230.66] [1230.66][S02]And indeed, a priori from information[1233.82] [1233.82][S02]an content point of view, look, this data set[1236.62] [1236.62][S02]has the information that it has, how could you[1239.56] [1239.56][S02]create new information?[1241.05] [1241.05][S02]I don't know.[1241.93] [1241.93][S02]These ideas might help a little bit[1244.735] [1244.735][S02]because there's machine learning deficiencies[1246.61] [1246.61][S02]that we're not at that fundamental ability[1250.06] [1250.06][S02]to extract all the information truly from the internet.[1252.823] [1252.823][S02]We have good algorithms, but they're not perfect.[1254.865] [1254.865][S01] Well, yeah.[1255.67] [1255.67][S02] So we'll see, yeah.[1256.55] [1256.55][S01] I guess I just want to think about that a little bit[1257.89] [1257.89][S01]more because it's a really interesting idea.[1259.18] [1259.18][S01]Because of course, naively, if you did it without thinking,[1261.8] [1261.8][S01]then it's like, the new version would have the biases in it.[1267.06] [1267.06][S01]And then the new version on top of that would be more biased.[1270.005] [1270.005][S01]And you'd end up sort of spiraling away[1271.63] [1271.63][S01]from the original human one.[1273.44] [1273.44][S01]But then what you're saying is as[1274.99] [1274.99][S01]though in the original human internet[1277.57] [1277.57][S01]is sort of embedded these conceptual connections.[1281.42] [1281.42][S01]And if you can extract those--[1282.92] [1282.92][S01]I'm thinking almost like E equals mc squared.[1285.13] [1285.13][S01]If you can find the E equals mc squared for human concepts[1290.29] [1290.29][S01]and then just generate new data using that alone,[1294.45] [1294.45][S01]then that seems more realistic.[1295.95] [1295.95][S02] Yeah, exactly.[1297.158] [1297.158][S02]And I think that's where you start hitting--[1299.57] [1299.57][S02]I mean, are these language models[1301.76] [1301.76][S02]just repeating what's online and not[1304.28] [1304.28][S02]being able to create anything new?[1306.06] [1306.06][S02]Or are they learning a world model[1308.45] [1308.45][S02]truly that you can then from the principles it extracts, possibly[1313.55] [1313.55][S02]generalize beyond what the data has?[1315.69] [1315.69][S02]And under the more optimistic version,[1318.0] [1318.0][S02]which I tend to believe more, we can push the limits of data[1322.82] [1322.82][S02]little bit more than the current limits that we have.[1325.98] [1325.98][S02]That being said, there are some data sources[1329.51] [1329.51][S02]that we haven't quite seen a breakthrough, like video data.[1334.35] [1334.35][S02]There's a lot of it.[1335.43] [1335.43][S02]And we haven't quite seen a moment[1338.12] [1338.12][S02]of take all the video data where you probably[1341.54] [1341.54][S02]can derive a lot of knowledge, a lot of laws of physics,[1344.91] [1344.91][S02]a lot of how the world works, even if there are[1347.12] [1347.12][S02]no words associated with the videos necessarily,[1350.28] [1350.28][S02]and extract that knowledge.[1351.96] [1351.96][S02]Even that I don't think we tapped into that source.[1354.58] [1354.58][S01] And it doesn't work that way?[1356.44] [1356.44][S02] I mean--[1356.86] [1356.86][S01] Or you don't know?[1358.11] [1358.11][S02] Yeah, it feels like it should,[1360.59] [1360.59][S02]even how we learn.[1363.558] [1363.558][S02]There's some language learning in the early days,[1365.6] [1365.6][S02]but we learn by also observing three dimensions and so[1368.98] [1368.98][S02]on and so forth.[1369.65] [1369.65][S02]So there probably is more knowledge[1373.6] [1373.6][S02]that we haven't extracted.[1375.43] [1375.43][S02]What obviously we've gotten pretty well at,[1378.47] [1378.47][S02]and you can see by testing the models,[1380.12] [1380.12][S02]is connecting the concepts present in the video.[1384.08] [1384.08][S02]And then you can do amazing things like, hey,[1386.03] [1386.03][S02]take this full hour video and just extract me[1388.9] [1388.9][S02]three interesting moments.[1390.53] [1390.53][S02]But the model itself, has it actually used that--[1393.995] [1393.995][S01] Oh, yeah.[1394.87] [1394.87][S02] --information directly?[1397.16] [1397.16][S02]Probably not.[1397.76] [1397.76][S01] Oh, I like this so much.[1399.26] [1399.26][S01]We were talking to Jeff about with multimodal models,[1402.14] [1402.14][S01]if you get these models to just watch[1404.8] [1404.8][S01]all of the videos that have ever been created,[1407.03] [1407.03][S01]can it quite literally extract what gravity means as a concept?[1411.513] [1411.513][S01]But what you're describing here, if I understand it,[1413.68] [1413.68][S01]is that at the moment, it can tell you what's in the video[1416.79] [1416.79][S01]that it's seen, but it can't then say,[1419.16] [1419.16][S01]and E equals mc squared.[1421.81] [1421.81][S01]Or if you showed it pictures of the night sky,[1424.33] [1424.33][S01]it wouldn't suddenly be able to predict the planetary motion[1427.53] [1427.53][S01]in the same way that human astronomers did.[1429.692] [1429.692][S02] Yeah, exactly.[1430.9] [1430.9][S02]The shortcut we're taking here is[1432.275] [1432.275][S02]that the data we train when we train on images and/or videos,[1437.23] [1437.23][S02]we almost always have a text representation[1442.26] [1442.26][S02]associated with that modality.[1443.77] [1443.77][S02]So it could be a caption explaining[1446.04] [1446.04][S02]what this image has or this video has[1447.96] [1447.96][S02]and so on and so forth.[1449.35] [1449.35][S02]And, of course, then it's incredible, right?[1452.86] [1452.86][S02]You can put a picture of homework[1454.71] [1454.71][S02]and with a little drawing conceptual[1456.6] [1456.6][S02]and it will connect and do quite a lot of good logic just based[1460.56] [1460.56][S02]on that.[1461.32] [1461.32][S02]But what I'm saying here is, could I just[1464.4] [1464.4][S02]take videos, no language, and then train[1467.34] [1467.34][S02]a model to then understand what's happening?[1470.98] [1470.98][S02]Maybe even, in a way, derive a language--[1473.91] [1473.91][S02]it's obviously not going to be our language--[1476.88] [1476.88][S02]and extract those concepts.[1478.99] [1478.99][S02]And that has not happened.[1481.63] [1481.63][S02]And it probably will.[1486.18] [1486.18][S01] Just going back to what[1488.07] [1488.07][S01]you said at the beginning about those two phases[1490.29] [1490.29][S01]to basically all the models that the DeepMind have built.[1493.507] [1493.507][S02] Yeah.[1494.34] [1494.34][S01] The imitation phase, which[1496.05] [1496.05][S01]is what we've been talking about right[1497.633] [1497.633][S01]here, but then the reinforcement learning phase on top.[1501.1] [1501.1][S01]And I know that AlphaGo and AlphaZero and many more[1505.29] [1505.29][S01]got better by playing themselves.[1507.73] [1507.73][S01]Does that apply here as well?[1510.16] [1510.16][S02] Yeah, that's one of the main open challenges--[1514.11] [1514.11][S02]scaling, not only pretraining but postraining or reinforcement[1518.04] [1518.04][S02]learning.[1518.926] [1518.926][S02]So the beauty about reinforcement learning in games[1522.9] [1522.9][S02]is that there is a set of rules that are coded.[1526.96] [1526.96][S02]And if you've won, you know you've won.[1529.34] [1529.34][S02]There is kind of a program that if you play chess[1532.22] [1532.22][S02]and you've won, it will check everything.[1534.66] [1534.66][S02]OK, that's a check mate.[1535.998] [1535.998][S02]Congratulations, you've won the game.[1537.54] [1537.54][S01] A clear metric of success.[1538.68] [1538.68][S02] A clear metric.[1540.11] [1540.11][S02]Now, in language, much trickier.[1544.232] [1544.232][S02]Is this a better poem than this?[1546.21] [1546.21][S02]Good luck discussing this, even amongst us.[1548.28] [1548.28][S02]So the generality again makes computing exactness very hard.[1556.44] [1556.44][S02]Is this a better summary of the movie?[1558.39] [1558.39][S02]Is this the most interesting bit of this video?[1561.21] [1561.21][S02]It's very hard to quantify.[1562.92] [1562.92][S02]But we can try, and we do try.[1565.17] [1565.17][S02]You train a model, and based on some human preferences,[1569.27] [1569.27][S02]roughly you just say, OK, try to now generalize.[1573.06] [1573.06][S02]So if I ask a model to criticize its own output,[1575.91] [1575.91][S02]it's not going to do that bad.[1577.62] [1577.62][S02]It's going to be good maybe 80% of the time, which[1581.33] [1581.33][S02]is not terrible.[1582.77] [1582.77][S02]But it can give you some signal.[1584.64] [1584.64][S02]But at that point, you start saying, well,[1586.68] [1586.68][S02]now you climb this metric.[1588.62] [1588.62][S02]You have this imperfect way to assess performance.[1591.98] [1591.98][S02]But now we're going to start training[1595.63] [1595.63][S02]against this reward that is not perfect.[1599.11] [1599.11][S02]What the model is going to do is exploit[1601.99] [1601.99][S02]the weaknesses of the reward.[1605.2] [1605.2][S02]Maybe using the chess example, imagine that I had a bug,[1609.14] [1609.14][S02]and if a pawn is in a certain position, you always win.[1612.7] [1612.7][S02]And it's a position that no one would play ever.[1616.43] [1616.43][S02]So maybe no one knows this exists.[1618.26] [1618.26][S02]But now you ask an algorithm to please[1621.49] [1621.49][S02]explore everything and try to discover[1623.44] [1623.44][S02]how to win at this game.[1625.04] [1625.04][S02]All of a sudden, you're going to find, hey, oh,[1627.53] [1627.53][S02]if I move the first pawn to this position,[1629.9] [1629.9][S02]no one plays this opening, You've.[1631.78] [1631.78][S02]Won the game.[1632.69] [1632.69][S02]Certainly, you the algorithm has nailed the game.[1636.74] [1636.74][S02]And then a researcher goes and sees how you play chess,[1639.49] [1639.49][S02]and it's just terrible.[1640.76] [1640.76][S01] Naughty AI, basically[1642.19] [1642.19][S01]that's what we're talking about.[1642.98] [1642.98][S02] So that's the challenge.[1644.87] [1644.87][S02]Basically you're finding exploits[1646.44] [1646.44][S02]rather than really learning what a good poem means truthfully.[1651.465] [1651.465][S01] Can you not just add in another player?[1653.59] [1653.59][S01]So add in another model, which is the ultimate arbiter?[1660.36] [1660.36][S02] Good suggestion, but then the problem[1662.79] [1662.79][S02]is, how do you train that model?[1665.782] [1665.782][S02]We have only a finite notion of what's[1669.3] [1669.3][S02]a good poem by some experts that we might ask, hey, compare[1672.51] [1672.51][S02]these two poems and so on.[1673.78] [1673.78][S02]So there's just a limited amount of data we[1677.19] [1677.19][S02]have to train these arbiters.[1679.57] [1679.57][S02]So the ground truth might be to ask someone[1683.01] [1683.01][S02]that is the expert, of course.[1685.327] [1685.327][S02]And if we could, we would.[1686.41] [1686.41][S02]But that's not scalable.[1688.12] [1688.12][S02]Then imagine how slow it would be to say,[1691.06] [1691.06][S02]OK, I have a parameter update that in three seconds, I found.[1696.88] [1696.88][S02]Now please review these 10,000 things by an expert.[1700.08] [1700.08][S02]Because that's the source of truth.[1702.98] [1702.98][S02]And we don't have enough data to train a good enough reward[1705.44] [1705.44][S02]model.[1705.94] [1705.94][S02]So again, there are some ideas.[1708.03] [1708.03][S02]But intuitively, you clearly understand,[1711.82] [1711.82][S02]but the problem is we don't have access to the ground truth.[1715.103] [1715.103][S01] That's it.[1716.02] [1716.02][S01]It's like you're feeling around in the dark with oven gloves on.[1718.83] [1718.83][S02] Yeah.[1719.25] [1719.25][S01] [LAUGHS][1719.83] [1719.83][S02] Yeah.[1720.085] [1720.085][S01] [LAUGHS] You're not even completely sure that there[1722.85] [1722.85][S01]are solid objects--[1723.9] [1723.9][S02] Yeah.[1723.99] [1723.99][S01] --to grab onto.[1725.115] [1725.115][S01]OK, so if that's the core then, that's the electronic brain,[1729.96] [1729.96][S01]and now we're building the digital body, what kind[1734.34] [1734.34][S01]of capabilities do you want that digital body to have,[1736.8] [1736.8][S01]like reasoning, for example?[1738.193] [1738.193][S01]Because there's been quite a lot of work[1739.86] [1739.86][S01]on that too, hasn't there?[1741.22] [1741.22][S02] Yeah, so when you start thinking, well,[1743.76] [1743.76][S02]what are the main surfaces that we would[1748.71] [1748.71][S02]be able to give these models limited access to so they can[1752.88] [1752.88][S02]see beyond what's in their weights, which are frozen,[1756.42] [1756.42][S02]to be able to gather knowledge or do maybe something a bit more[1760.5] [1760.5][S02]complex than just predicting the next word from just what they[1764.57] [1764.57][S02]have in context, plus what they have in their weights?[1768.12] [1768.12][S02]And so obvious things that come to mind[1771.53] [1771.53][S02]is giving them access to a search engine.[1774.09] [1774.09][S02]That's what we do very well at Google.[1777.38] [1777.38][S02]Another one is to give them the ability[1780.02] [1780.02][S02]to run the code they write.[1781.885] [1781.885][S02]And then, of course, maybe even more broadly, that[1785.33] [1785.33][S02]could be more general is giving them the ability[1788.57] [1788.57][S02]to maybe interact with a browser that has access to the internet.[1792.77] [1792.77][S02]With all of these, you always have[1794.66] [1794.66][S02]to be careful to sand box, that just means protect[1798.26] [1798.26][S02]these environments so that the models, even if they're not[1800.78] [1800.78][S02]that advanced, wouldn't do something that is unintended.[1804.32] [1804.32][S02]So there's the whole safety aspect of this[1807.06] [1807.06][S02]that as you move beyond the model,[1809.22] [1809.22][S02]it starts to be quite interesting.[1811.17] [1811.17][S02]But if we're just dreaming what would be possible,[1815.1] [1815.1][S02]by having these tools available to the models, all of a sudden[1818.99] [1818.99][S02]they can start doing much more advanced things[1822.6] [1822.6][S02]beyond what was the training corpus that we used at the time.[1828.7] [1828.7][S02]They can rely on the latest news to explain us[1832.9] [1832.9][S02]or to summarize what was the main thing[1835.36] [1835.36][S02]yesterday that happened.[1836.9] [1836.9][S02]All these kind of things, you need to give them these tools.[1840.5] [1840.5][S01] OK, so how does reasoning fit into all of this[1843.01] [1843.01][S01]then?[1843.65] [1843.65][S02] Yeah, reasoning is interesting, right?[1845.858] [1845.858][S02]So what I described could be summarized as,[1848.86] [1848.86][S02]hey, I want to know what happened yesterday,[1852.46] [1852.46][S02]then I could just say, look, maybe personalize it[1856.158] [1856.158][S02]a little bit.[1856.7] [1856.7][S02]So I could describe in words, I could say, hey hey, model,[1860.32] [1860.32][S02]I'm Oriol, I'm interested in this and that.[1862.85] [1862.85][S02]My political views are this or that.[1864.82] [1864.82][S02]Give me a positive take on the news yesterday.[1867.53] [1867.53][S02]And the model would probably search, retrieve all the news,[1871.37] [1871.37][S02]and then, given what I asked it to do,[1873.5] [1873.5][S02]just do it in a way that I like and I find it enjoyable.[1876.8] [1876.8][S02]Maybe if I don't like it, I can even go[1878.47] [1878.47][S02]and then say, I didn't like this or this is not a good joke.[1881.62] [1881.62][S02]And then we could iterate a little bit in a conversation.[1884.62] [1884.62][S02]Now, reasoning is a bit of a different axis of scaling.[1890.95] [1890.95][S02]And so you could imagine the model[1894.99] [1894.99][S02]deciding what kind of intermediate steps[1897.51] [1897.51][S02]to do to give me a better answer.[1899.8] [1899.8][S02]So imagine there's 100 news outlets[1903.24] [1903.24][S02]that Google search retrieves.[1905.4] [1905.4][S02]Maybe the model decides, hey, I'm[1907.32] [1907.32][S02]just not going to read this and just[1909.18] [1909.18][S02]try to summarize it all at once, I'm[1911.22] [1911.22][S02]going to summarize each of the 100 articles first.[1914.71] [1914.71][S02]So that means the model decided I'm[1917.25] [1917.25][S02]going to write a summary for each of the 100 pages.[1920.86] [1920.86][S02]I'm going to write it not to the user, to myself.[1924.16] [1924.16][S02]And now it has hundreds summaries.[1926.773] [1926.773][S02]And maybe the next step it decides to do[1928.44] [1928.44][S02]is, I'm going to group these by topics.[1932.19] [1932.19][S02]Then it decides one of the articles looks suspicious,[1935.74] [1935.74][S02]so maybe it goes online and checks if in any forums[1939.06] [1939.06][S02]someone discusses like, oh, this might[1940.91] [1940.91][S02]be not truthful because of the author and so on and so forth.[1945.03] [1945.03][S02]So it can do a lot of steps to do research.[1948.48] [1948.48][S02]And it could do this for quite a while.[1951.06] [1951.06][S02]And only when the model says, well, I[1953.99] [1953.99][S02]think now I have a much better quality answer,[1956.87] [1956.87][S02]then it will give you the few-word summary.[1960.48] [1960.48][S02]But now it had all this time to do much more processing[1963.62] [1963.62][S02]on the information that was available to it.[1966.65] [1966.65][S02]And that inference time compute, we[1970.88] [1970.88][S02]hope that the more time we give to the model,[1973.98] [1973.98][S02]the better it's going to summarize[1975.43] [1975.43][S02]the news, the better it's going to write a poem,[1977.43] [1977.43][S02]the better it's going to, of course, do math.[1979.305] [1979.305][S02]But that's certainly another axis of scaling[1981.41] [1981.41][S02]which we're starting to quite unlock, we hope to unlock.[1985.23] [1985.23][S02]And again, we'll break a bit of the scaling laws and the limits[1988.85] [1988.85][S02]that we see in pure pretraining.[1991.44] [1991.44][S01] Does this also include planning?[1994.656] [1994.656][S01]Could it look at your calendar, work out when your payday was,[1999.81] [1999.81][S01]maybe know that the January sales are coming up soon,[2002.74] [2002.74][S01]and tell you to postpone booking your holiday for a few days?[2008.39] [2008.39][S02] That can get very complex,[2010.19] [2010.19][S02]but of course, when you factor in things like personalization[2014.23] [2014.23][S02]and when to do things because of all the other things[2017.83] [2017.83][S02]that are ongoing, you have more sources of information.[2021.89] [2021.89][S02]You need to collect them and then give the best answer.[2025.73] [2025.73][S02]And it stops being what color is the sky, which[2029.02] [2029.02][S02]is not that simple to answer.[2030.89] [2030.89][S02]I was thinking about that example.[2032.42] [2032.42][S02]We had a very early paper where we have that example,[2035.363] [2035.363][S02]as, oh, something that language models can do.[2037.28] [2037.28][S02]Amazing, right?[2037.94] [2037.94][S02]You don't program them to answer, but they answer.[2040.28] [2040.28][S02]But then, actually even the answer[2042.61] [2042.61][S02]is quite nuanced if you start thinking, oh, yeah,[2045.83] [2045.83][S02]planets, and what time of the day?[2048.38] [2048.38][S02]Is it cloudy or not?[2049.4] [2049.4][S02]So the thinking and the planning,[2052.01] [2052.01][S02]yeah, that's something that these models can do.[2059.472] [2059.472][S01] I'm reminded, I had a conversation with Dennis[2061.889] [2061.889][S01]probably back in 2019.[2063.489] [2063.489][S01]And he was talking about the Kahneman and Tversky[2067.32] [2067.32][S01]ideas of how the human brain has almost two systems of thinking,[2071.409] [2071.409][S01]the sort of quick, instinctive, intuition based,[2074.02] [2074.02][S01]and then the much slower, calculated, the way that you do[2077.88] [2077.88][S01]maths and chess.[2079.0] [2079.0][S01]And Demis was saying that that second one[2082.679] [2082.679][S01]has been traditionally easier for us to do with computers.[2087.06] [2087.06][S01]But now we're seeing the much quicker instinctive stuff.[2091.12] [2091.12][S01]But you're sort of talking about putting the two together, right?[2094.65] [2094.65][S02] Yeah, right.[2096.382] [2096.382][S02]Probably what Demis was talking about[2097.95] [2097.95][S02]is systems too, which is indeed one that you reflect a bit more.[2102.36] [2102.36][S02]And in games it's very clear.[2103.75] [2103.75][S02]You just could say, oh, this move feels right, you just move.[2106.99] [2106.99][S02]But if you think and ponder, you might get to a better move.[2111.96] [2111.96][S02]The challenge is that, now because we[2115.38] [2115.38][S02]are in such general direction-- these models can do[2118.73] [2118.73][S02]anything, anything, literally.[2120.12] [2120.12][S02]You just do whatever you want, upload[2122.15] [2122.15][S02]an image, talk about the news.[2124.14] [2124.14][S02]So what it means to have this deeper[2129.05] [2129.05][S02]thinking is so domain specific that,[2131.81] [2131.81][S02]how are you going to do that?[2133.11] [2133.11][S02]And there's a few answers.[2135.03] [2135.03][S02]But the one I like is like, well, these models[2137.81] [2137.81][S02]are very general.[2138.98] [2138.98][S02]To add the ability to think on top[2141.53] [2141.53][S02]of a very general set of capabilities,[2143.97] [2143.97][S02]you probably need a general way to think.[2146.13] [2146.13][S02]And so you use the model itself to generate how[2151.19] [2151.19][S02]it should think about anything.[2152.85] [2152.85][S02]And the model will come up with, oh, I'm[2155.42] [2155.42][S02]going to summarize each article.[2156.81] [2156.81][S02]I'm going to do this and that and that.[2158.435] [2158.435][S02]And it is not us programming it.[2160.08] [2160.08][S02]That's a very deep insight.[2161.9] [2161.9][S02]Now is it the only way to do it and is it the best way to do it?[2165.96] [2165.96][S02]Early days.[2166.592] [2166.592][S01] Yeah.[2167.3] [2167.3][S02] Five years.[2167.91] [2167.91][S02]We'll see.[2168.02] [2168.02][S01] [LAUGHS] Exactly, I'll talk to you in 2029.[2170.065] [2170.065][S02] Yeah.[2170.55] [2170.55][S01] OK, I'm thinking now though[2172.175] [2172.175][S01]also about lots of the things that felt very important back,[2176.98] [2176.98][S01]five years ago.[2178.45] [2178.45][S01]And a lot of it was about inspiration from neuroscience.[2181.73] [2181.73][S01]So I suppose in a way here, you're[2183.22] [2183.22][S01]talking about planning and reasoning.[2184.85] [2184.85][S01]But memory was the other really big one.[2187.34] [2187.34][S01]And has that come through?[2189.47] [2189.47][S01]People talk about long context and short context a lot.[2192.53] [2192.53][S01]I suppose that is working memory in a way, isn't it?[2195.53] [2195.53][S02] Yeah, there's techniques[2199.03] [2199.03][S02]that are out there that you can apply to a language model.[2202.67] [2202.67][S02]There are at the very least three.[2206.02] [2206.02][S02]And they're reasonably simple to explain.[2209.512] [2209.512][S02]The first way in which we have a system that[2212.71] [2212.71][S02]memorizes all of the internet is by literally doing[2216.19] [2216.19][S02]the pretraining step.[2217.51] [2217.51][S02]That's literally a memorization step[2220.3] [2220.3][S02]in a particular format, which is we have these weights,[2223.55] [2223.55][S02]they're random, and then we assemble them[2226.12] [2226.12][S02]in these amazing architectures.[2228.13] [2228.13][S02]Now, the second level is maybe I explained a little bit[2233.17] [2233.17][S02]how you would give the tool of a search engine such as Google[2236.56] [2236.56][S02]to the model.[2237.61] [2237.61][S02]That, you could claim, is what neuroscientists[2242.86] [2242.86][S02]would call episodic memory, which, as a human,[2246.16] [2246.16][S02]maybe it's like we have these memories from a long time ago.[2250.61] [2250.61][S02]They're not very precise.[2251.75] [2251.75][S02]So they tend to be a bit more fuzzy.[2253.34] [2253.34][S02]If I have to think, oh, what was my first day at Google?[2257.12] [2257.12][S02]I remember bits and pieces, being in a room or someone[2259.99] [2259.99][S02]I met or whatnot.[2260.955] [2260.955][S01] The gist.[2261.83] [2261.83][S02] Yeah, the gist, right?[2263.372] [2263.372][S02]Now, interestingly, these models maybe[2265.18] [2265.18][S02]don't have that limitation.[2266.53] [2266.53][S02]You can literally get an article written many years ago online,[2270.32] [2270.32][S02]and it's going to have all the images, everything[2272.53] [2272.53][S02]will be perfect, reconstructed perfectly.[2275.06] [2275.06][S02]So that second mode called episodic memory, clearly[2278.89] [2278.89][S02]we're seeing that when you integrate especially powerful[2281.56] [2281.56][S02]search engines into our models.[2284.12] [2284.12][S02]And then the third one is what you[2287.11] [2287.11][S02]could call working memory, which actually the whole of thinking[2290.95] [2290.95][S02]that I described is one of.[2292.7] [2292.7][S02]If we take every news article, but then we[2296.43] [2296.43][S02]want to create summaries, find how they relate to each other,[2299.17] [2299.17][S02]criticize some of them, this starts[2302.01] [2302.01][S02]to combine working memory, meaning[2304.68] [2304.68][S02]I'm going to have a scratchpad of the summaries, the issues[2308.67] [2308.67][S02]that I think I'm finding.[2310.11] [2310.11][S02]And that, when we call short or long context,[2312.96] [2312.96][S02]generally, we mean this last bit, the working memory,[2316.36] [2316.36][S02]whether you have a thousand tokens, which means[2319.29] [2319.29][S02]I couldn't possibly do much.[2321.04] [2321.04][S02]I can retrieve articles.[2322.78] [2322.78][S02]It's already over thousand words.[2324.37] [2324.37][S02]Not much I can do to summarize them.[2326.71] [2326.71][S02]Or it can be massive, in which case[2329.22] [2329.22][S02]you have many more possibilities to do reasoning on top of that,[2334.06] [2334.06][S02]and so on and so forth.[2335.02] [2335.02][S02]And so one of the breakthroughs of the year,[2338.47] [2338.47][S02]actually, we're still in 2024.[2341.38] [2341.38][S01] Just.[2342.25] [2342.25][S02] Yes, but it was just[2343.708] [2343.708][S02]to enable millions of tokens in context,[2345.85] [2345.85][S02]which enables many things.[2348.64] [2348.64][S02]You can retrieve something from the past,[2351.25] [2351.25][S02]but then bring it forward and then do[2352.88] [2352.88][S02]a very detailed analysis.[2354.57] [2354.57][S02]That's a bit of the examples of, we have a movie,[2357.56] [2357.56][S02]we can upload a movie or some very long video[2360.53] [2360.53][S02]and start doing summarization.[2362.37] [2362.37][S02]The fact we upload it is more episodic memory,[2365.4] [2365.4][S02]but then now we have it in memory.[2367.17] [2367.17][S02]It all fits in memory.[2368.22] [2368.22][S02]We can do quite a lot of associations within each frame,[2372.89] [2372.89][S02]each object in the movie, and so on and so forth.[2375.92] [2375.92][S01] Is a longer context window better always?[2379.17] [2379.17][S01]Because I'm just thinking about, I[2380.685] [2380.685][S01]don't know how much you guys are still using neuroscience[2383.06] [2383.06][S01]as an inspiration for what you're doing.[2384.727] [2384.727][S01]But the human memory, there's a limit to the working memory.[2388.582] [2388.582][S01]There's certainly some times where you're like,[2390.54] [2390.54][S01]my brain is full and I'm done.[2392.52] [2392.52][S02] Sometimes brain is an inspiration,[2395.22] [2395.22][S02]but computers certainly have advantages.[2397.26] [2397.26][S02]We should build on its strengths, right?[2399.36] [2399.36][S02]So perhaps the fact that they can have literally in memory[2403.944] [2403.944][S02]every Wikipedia article, whatever it is,[2407.88] [2407.88][S02]we can't, but if the model can, well, there you go.[2411.6] [2411.6][S02]You have new capabilities.[2412.8] [2412.8][S02]But also, it might be too confusing to have[2415.31] [2415.31][S02]too much information even for these neural networks.[2418.89] [2418.89][S02]So it might be a good idea to compress.[2421.44] [2421.44][S02]So that's where you probably want[2423.14] [2423.14][S02]to push for getting some inspiration for how we might[2426.86] [2426.86][S02]do what we do, which is quite amazing,[2428.45] [2428.45][S02]in terms of memory retrieval and so on.[2430.382] [2430.382][S01] Yeah.[2431.09] [2431.09][S01]This is why you're leading up Drastic Research.[2433.33] [2433.33][S02] Yes.[2434.88] [2434.88][S02]What we want to do with the models[2436.82] [2436.82][S02]should be definitely inspiring and forward looking.[2440.61] [2440.61][S02]And then what are the main limits of the technology?[2445.44] [2445.44][S02]And then try to, of course, place the bets[2447.59] [2447.59][S02]and inspire the teams to finding solutions[2449.84] [2449.84][S02]around the critical components.[2451.755] [2451.755][S01] But some of the bets that you've already made[2454.13] [2454.13][S01]have come off.[2454.92] [2454.92][S01]I know there's been a big announcement[2456.98] [2456.98][S01]of a dizzying number of new features[2461.36] [2461.36][S01]that have just come out.[2462.36] [2462.36][S01]Can we talk through some of them?[2463.97] [2463.97][S01]But then maybe also, talk to me about the different skills[2466.935] [2466.935][S01]that we've already spoken about and how[2468.56] [2468.56][S01]they appear in each of these.[2470.51] [2470.51][S02] Yeah, so we have quite a few systems[2476.77] [2476.77][S02]around our best Gemini models.[2479.6] [2479.6][S02]So one of the things that we've done is update to 2.0.[2483.71] [2483.71][S02]We're seeing a generational leap.[2485.42] [2485.42][S02]Even if you say, look, let's not scale anymore,[2487.73] [2487.73][S02]can we get better quality?[2488.96] [2488.96][S02]So we've done it sort of again.[2490.97] [2490.97][S02]These models are faster, they're cheaper,[2493.45] [2493.45][S02]and they're actually better.[2494.695] [2494.695][S01] Basically, Gemini has got better.[2496.57] [2496.57][S01]That's it.[2496.93] [2496.93][S02] Yeah, Gemini has got better, but not[2498.43] [2498.43][S02]only because we scaled.[2499.61] [2499.61][S01] Sure.[2500.318] [2500.318][S02] That's one of the main messages.[2502.58] [2502.58][S01] Tell me more about the agentic capabilities[2505.17] [2505.17][S01]that you've brought to Gemini.[2506.42] [2506.42][S02] Yeah, so we're releasing a companion in Chrome[2510.22] [2510.22][S02]where you can just type to do a task[2512.95] [2512.95][S02]that maybe some of these tasks are tricky.[2517.0] [2517.0][S02]Because I partly enjoy them, but also I partly don't like them.[2519.77] [2519.77][S02]So I'm thinking now very clearly about trips.[2522.136] [2522.136][S02]So you travel and you look for hotels or flights or whatnot,[2526.88] [2526.88][S02]and a lot of it feels like, oh, I wish this could be automated.[2531.13] [2531.13][S02]But at the same time, I wouldn't just[2533.34] [2533.34][S02]want to not be part of this journey.[2536.172] [2536.172][S02]So I guess the kind of thing we're releasing[2539.64] [2539.64][S02]is something that hopefully will automate parts of the more[2545.67] [2545.67][S02]trivial steps or repeated or things[2548.22] [2548.22][S02]that need automation because I can't be[2550.532] [2550.532][S02]bothered to click everything.[2551.74] [2551.74][S02]So we're adding an agent that you can ask it to do something[2557.64] [2557.64][S02]for you, and then it's going to, again, through thinking[2562.02] [2562.02][S02]and through acting on the basic clicking on links[2567.24] [2567.24][S02]and so on, try to solve the task for you.[2570.04] [2570.04][S02]And that's quite an exciting, both research challenge,[2574.47] [2574.47][S02]opportunity.[2575.1] [2575.1][S02]Because it's a very general environment for a very general[2578.01] [2578.01][S02]agent and model, ultimately.[2580.18] [2580.18][S02]And some examples that we've had again in the early prototypes,[2583.69] [2583.69][S02]we can ask to play a game, which of course, goes back[2586.38] [2586.38][S02]to the roots of DeepMind, on the browser.[2590.55] [2590.55][S02]And it did OK.[2592.14] [2592.14][S02]It finds a website, it starts playing the game.[2594.69] [2594.69][S02]It's kind of a cool connection to the more general you are,[2598.41] [2598.41][S02]then the more you can treat environments[2601.88] [2601.88][S02]where you had specialization in the past, as now,[2605.1] [2605.1][S02]oh, I just can type it and it just[2606.86] [2606.86][S02]goes and learn to play this game.[2608.79] [2608.79][S02]We're not quite there, but this is[2610.67] [2610.67][S02]a glimpse of maybe where we could go[2612.74] [2612.74][S02]with this kind of technology.[2615.0] [2615.0][S01] You're right that it does really[2616.88] [2616.88][S01]bring us back to that thing that you were doing[2619.52] [2619.52][S01]so many years ago, which was something that could[2621.62] [2621.62][S01]use a keyboard and a mouse.[2623.54] [2623.54][S01]It's a really similar thing.[2625.02] [2625.02][S02] Yeah, even the actions are very similar.[2627.312] [2627.312][S02]They understand the screen and, given[2629.6] [2629.6][S02]what you ask, where would you click and so on,[2632.13] [2632.13][S02]that's the same sort of actions even that very general games[2638.045] [2638.045][S02]you would have to interact with.[2639.47] [2639.47][S02]The difference is the goal there is narrow.[2642.63] [2642.63][S02]It's just one game and the same kind of screens.[2644.88] [2644.88][S02]Whereas here, it's the whole web, which is pretty vast.[2647.595] [2647.595][S01] Well, OK.[2648.47] [2648.47][S01]But then so I'm sort of imagining what you could do now.[2651.048] [2651.048][S01]Could it look in your calendar?[2652.34] [2652.34][S01]Could you say, I want to go on holiday next year,[2654.41] [2654.41][S01]and it could look in your calendar[2655.827] [2655.827][S01]and work out when the best week was,[2657.5] [2657.5][S01]know your budget, et cetera, et cetera, et cetera?[2660.44] [2660.44][S02] Yeah, so these models[2662.65] [2662.65][S02]are not far from being able to automate this.[2664.7] [2664.7][S02]So now it's a matter of making it better, making it safe.[2668.66] [2668.66][S02]There's a lot of steps.[2669.89] [2669.89][S02]But if you just fast forward, anything[2673.81] [2673.81][S02]a human can do on a browser, these things[2676.54] [2676.54][S02]can do in principle.[2678.35] [2678.35][S02]And then if you make them really understand[2680.35] [2680.35][S02]what you want and really good through thinking[2683.08] [2683.08][S02]and other techniques, they'll get better and better.[2685.87] [2685.87][S02]And they'll be probably faster and maybe in some cases[2689.38] [2689.38][S02]much better than you at doing those.[2691.25] [2691.25][S02]So that's the dream.[2693.4] [2693.4][S02]And these are super early stages,[2696.02] [2696.02][S02]but it's also super exciting.[2697.85] [2697.85][S02]And I think certainly next year, we're[2700.51] [2700.51][S02]going to see a lot of experimentation[2702.79] [2702.79][S02]around this idea of intersecting language models agentically[2707.52] [2707.52][S02]with browser or computers more generally.[2710.74] [2710.74][S01] How about coding?[2712.48] [2712.48][S02] Yeah, coding is a great one as well.[2714.88] [2714.88][S02]We are also releasing tools for software engineering, which[2718.86] [2718.86][S02]of course, they generally require not only, hey, here[2722.94] [2722.94][S02]is a perfect description of a puzzle about coding[2725.91] [2725.91][S02]and please write me the code.[2727.33] [2727.33][S02]And by the way, I know how to test it.[2729.19] [2729.19][S02]It's more iterative.[2730.23] [2730.23][S02]You have to write code and run the code and so on and so forth.[2734.5] [2734.5][S02]So we are putting that capability, as well, forward[2737.91] [2737.91][S02]from an agentic point of view.[2742.21] [2742.21][S02]Games are very important.[2743.85] [2743.85][S02]And of course, that was a means to an end[2746.13] [2746.13][S02]to develop powerful algorithms.[2747.94] [2747.94][S02]But it's also very interesting to think[2749.73] [2749.73][S02]about how these very powerful multimodal[2753.72] [2753.72][S02]models start to understand games and can aid users to entertain[2758.7] [2758.7][S02]doing a game session, give them advice, or tell[2764.19] [2764.19][S02]a joke about the game, or whatnot.[2765.85] [2765.85][S02]So we're also experimenting with this sort of game companion.[2770.622] [2770.622][S01] OK, all of these things[2772.08] [2772.08][S01]that you're talking about, this is[2773.52] [2773.52][S01]sounding very close to intelligence[2777.46] [2777.46][S01]that is quite general.[2778.96] [2778.96][S01]Are we getting close to AGI?[2780.46] [2780.46][S02] Yeah, this is a good question.[2782.335] [2782.335][S02]Look, I was thinking about this earlier this week.[2786.93] [2786.93][S02]If 10 years ago, 5 years ago even, I[2789.9] [2789.9][S02]would have been given the models today,[2792.34] [2792.34][S02]and I would say, look, there's a secret lab, this is a model,[2797.19] [2797.19][S02]play with it and tell me if you think this is actually[2800.58] [2800.58][S02]close to a general intelligence; I would have claimed,[2803.62] [2803.62][S02]oh, yeah, that comes from a future where AGI basically[2807.33] [2807.33][S02]either has happened or I can see that this is very close to it.[2811.08] [2811.08][S02]So the closer you are, the more you find,[2814.84] [2814.84][S02]oh, but it hallucinates.[2815.885] [2815.885][S02]Of course, that's very important.[2817.26] [2817.26][S02]But I think, just zooming out, it just[2820.35] [2820.35][S02]feels like, OK, it's getting pretty close.[2823.357] [2823.357][S01] But then DeepMind's mission statement,[2825.44] [2825.44][S01]solve intelligence, that sort of intelligence,[2828.36] [2828.36][S01]super intelligence, something that[2832.37] [2832.37][S01]surpasses human intelligence, do you think that scaling[2835.7] [2835.7][S01]is enough to get us there?[2837.18] [2837.18][S01]Or do you think that we need something else?[2839.67] [2839.67][S02] Yeah, Google DeepMind[2842.0] [2842.0][S02]has this mission to obviously intersect intelligence[2844.82] [2844.82][S02]with science to push the boundaries.[2846.618] [2846.618][S02]And we've seen a good example very recently, of course,[2848.91] [2848.91][S02]with AlphaFold.[2850.2] [2850.2][S02]So in that sense, from a domains perspective,[2855.41] [2855.41][S02]we honestly have seen some examples[2857.57] [2857.57][S02]already of narrow but super intelligent system.[2861.03] [2861.03][S02]AlphaFold was only doing that.[2862.95] [2862.95][S02]And I think probably that's the domains[2868.04] [2868.04][S02]to think about where we're going to start[2869.84] [2869.84][S02]seeing super intelligence, even from the general sort[2873.47] [2873.47][S02]of capabilities these models have.[2876.15] [2876.15][S02]You might need to do some specialization.[2878.07] [2878.07][S02]And again, it might be worth it.[2879.63] [2879.63][S02]Was it worth it to solve protein folding?[2882.29] [2882.29][S02]Absolutely, right?[2883.96] [2883.96][S02]But I think that's a good test to use.[2888.56] [2888.56][S02]And we are very well positioned.[2890.17] [2890.17][S02]Because we, of course, have the whole science team[2892.94] [2892.94][S02]and so on working on very interesting problems.[2896.74] [2896.74][S02]Now, if you take the language models[2901.87] [2901.87][S02]and you start thinking about agents,[2904.67] [2904.67][S02]putting them in environments that[2906.16] [2906.16][S02]could be more about science, simulation, theorem provers,[2910.75] [2910.75][S02]and so on, will something very discrete[2913.96] [2913.96][S02]be needed to enable other breakthroughs?[2916.67] [2916.67][S02]I would say probably not.[2918.31] [2918.31][S02]Without another transformer-like breakthrough,[2921.14] [2921.14][S02]perhaps, feels like we're going to start[2923.5] [2923.5][S02]seeing more examples of, oh my god, yeah, like in math now,[2927.64] [2927.64][S02]it just discovers new theorems that mathematicians[2930.85] [2930.85][S02]find interesting.[2931.99] [2931.99][S02]And it happened by just, of course, very good execution[2936.74] [2936.74][S02]plus scaling up of some of the ideas and so on and so forth.[2940.53] [2940.53][S01] It is interesting, though, that the first dominoes[2943.6] [2943.6][S01]to fall are the ones which have a ground truth--[2945.79] [2945.79][S02] Yeah.[2946.18] [2946.18][S01] --like science, as you describe.[2948.2] [2948.2][S02] Yeah.[2949.033] [2949.033][S02]Although, yeah, science, it depends which sciences might[2952.24] [2952.24][S02]have ground truth, I suppose.[2953.6] [2953.6][S01] Protein folding, definitely.[2955.267] [2955.267][S02] Yeah, it's true.[2957.28] [2957.28][S02]I'm hoping we also see some other ways[2960.73] [2960.73][S02]to advance in a superhuman way.[2963.06] [2963.06][S02]You could imagine having a brainstorming scientific advisor[2967.96] [2967.96][S02]that is powered by one of these powerful models.[2970.51] [2970.51][S02]And more than it discovers something[2972.7] [2972.7][S02]or it proves something new, it just challenges your assumptions[2975.86] [2975.86][S02]and it makes you think out of the box in a way[2978.25] [2978.25][S02]that, then, my creativity sort of gets me[2981.22] [2981.22][S02]to a place I couldn't have gone.[2982.82] [2982.82][S02]Then you would call that superhuman in some ways as well.[2986.215] [2986.215][S02]So I think those are definitely not out of scope[2988.99] [2988.99][S02]and much harder to also, of course,[2990.89] [2990.89][S02]think of how do you reward that behavior.[2993.392] [2993.392][S01] Absolutely fascinating.[2994.85] [2994.85][S01]There was definitely a lot of drastic stuff in there.[2997.058] [2997.058][S02] Yeah.[2997.892] [2997.892][S02]Yes.[2998.53] [2998.53][S01] Thank you so much for joining me.[2999.91] [2999.91][S02] Yeah, likewise.[3000.69] [3000.69][S02]Thanks, pleasure.[3001.418] [3001.418][S02]See you in five years.[3002.335] [3002.335][S01] [LAUGHS] I think there was this real theme that[3005.63] [3005.63][S01]emerged from that conversation, at least for me anyway,[3007.95] [3007.95][S01]which was this idea of generality.[3011.22] [3011.22][S01]And if you think about it, there is this generality in the way[3015.56] [3015.56][S01]that intelligence advances knowledge.[3017.94] [3017.94][S01]Like those old astronomers like Copernicus,[3020.37] [3020.37][S01]they were assessing lots of data from observing the sky[3023.93] [3023.93][S01]and using that to extract a model of the solar system.[3028.94] [3028.94][S01]But in the case of AlphaGo, it was observing games of Go[3033.05] [3033.05][S01]to extract a model for the best possible way to play.[3037.83] [3037.83][S01]And now, somewhere embedded in everything[3040.7] [3040.7][S01]that has ever been created by humans[3042.2] [3042.2][S01]is this model, this underlying truth[3045.05] [3045.05][S01]of how we experience reality.[3047.87] [3047.87][S01]And the model that we're looking for, of course,[3050.22] [3050.22][S01]it's never going to be as neat as heliocentrism.[3053.0] [3053.0][S01]But that model does seem to be in there, hidden[3056.98] [3056.98][S01]among the frozen weights of Gemini.[3060.05] [3060.05][S01]Now, if that's what we've done so far,[3062.36] [3062.36][S01]the next phase is to try and use those general ideas[3065.5] [3065.5][S01]to extract a model of human preferences, too.[3069.29] [3069.29][S01]And that is, of course, a lot, lot harder.[3073.43] [3073.43][S01]But if we succeed, it might just get us to a more general form[3078.04] [3078.04][S01]of intelligence, too, AGI.[3081.94] [3081.94][S01]Now, if you found this conversation interesting,[3084.11] [3084.11][S01]I think it's worth also checking out[3085.66] [3085.66][S01]the episodes I did with Jeff Dean[3087.34] [3087.34][S01]on, among other things, scaling, and Iason Gabriel[3091.18] [3091.18][S01]on the ethics of AI agents.[3093.44] [3093.44][S01]Or if you want to dig a bit deeper into the development[3096.76] [3096.76][S01]of Gemini 2.0, then you can check out the latest episode[3100.69] [3100.69][S01]of the new \"Google AI Release Notes\"[3103.09] [3103.09][S01]podcast with host Logan Kilpatrick.[3106.4] [3106.4][S01]This and other episodes can be found[3108.58] [3108.58][S01]wherever you get your podcasts.[3110.72] [3110.72][S01]Until next time.[3112.54] [3112.54][MUSIC PLAYING][3116.19]"}
{"file_name": "audio/val_000004.wav", "transcription": "[0.0][S02] I'm hoping AI contributes to bringing back[3.11] [3.11][S02]the joy of practicing medicine.[4.677] [4.677][S01] So wait, it actually works then.[6.51] [6.51][S02] Yes, it does.[6.89] [6.89][S01] You record a cough and it[7.73] [7.73][S01]can tell if you've got TB.[9.022] [9.022][S02] In short, yes.[10.23] [10.23][S02]By combining those modalities, by getting[13.61] [13.61][S02]to a deeper understanding at the cellular level of what's[17.81] [17.81][S02]going on, we'll finally be able to crack, for example,[21.6] [21.6][S02]triple-negative breast cancers or more advanced cancers[25.07] [25.07][S02]that, today, we don't have good solutions for.[27.652] [27.652][S01] Are you concerned that people[29.36] [29.36][S01]are trying to diagnose themselves with generative AI?[31.676] [31.676][S02] Yeah.[32.509] [32.509][S02]I mean, I would say, even before, you really[34.85] [34.85][S02]shouldn't do that, right?[35.93] [35.93][S02]You shouldn't play your own doctor.[38.635] [38.635][MUSIC PLAYING][41.545] [42.53][S01] Welcome back to \"Google DeepMind, The Podcast.\"[45.18] [45.18][S01]I'm Professor Hannah Fry.[46.74] [46.74][S01]In this episode, I'm talking to Joelle Barral,[49.53] [49.53][S01]Senior Director of Research at Google DeepMind,[52.32] [52.32][S01]about AI for health.[54.54] [54.54][S01]Now, for years, we have joked about how the internet is[57.11] [57.11][S01]the hypochondriac's best friend, capable of turning[59.94] [59.94][S01]a small headache into a terminal illness and vice versa,[63.45] [63.45][S01]with a quick search and a click of a button.[66.13] [66.13][S01]But quietly, behind the scenes, the role[69.36] [69.36][S01]of algorithms in medical care has been shifting.[72.19] [72.19][S01]We've already talked a lot on this podcast[74.16] [74.16][S01]about the impact of AI on drug discovery and research[77.22] [77.22][S01]into proteins.[78.49] [78.49][S01]But now, AI promises to change diagnosis and treatment, too.[84.1] [84.1][S01]Welcome to the podcast, Joelle.[86.04] [86.04][S01]I mean, there have been some pretty big changes already[89.4] [89.4][S01]with AI.[90.85] [90.85][S01]Should we expect that health care[92.67] [92.67][S01]will look very different in 10 or 15 years[95.1] [95.1][S01]time to how it looks now?[96.217] [96.217][S02] Absolutely.[97.3] [97.3][S02]I believe so.[97.96] [97.96][S02]I think health care is really posed[100.83] [100.83][S02]to be drastically changed with AI,[103.3] [103.3][S02]but it may not look too different.[106.5] [106.5][S02]I like to think of-- or at least I picture health care[109.29] [109.29][S02]as a fairly sticky ecosystem.[111.45] [111.45][S02]So 10, 15 years from now, we will likely still[116.1] [116.1][S02]go to our primary care doctor and then follow through[121.36] [121.36][S02]with specialist visits, et cetera.[123.79] [123.79][S02]But each of them is going to be augmented, if you wish,[127.42] [127.42][S02]by an AI agent.[128.66] [128.66][S02]So underneath, the system will be very different,[131.8] [131.8][S02]but it may not appear as different from what[134.89] [134.89][S02]we know today.[135.603] [135.603][S01] How long have you been[137.02] [137.02][S01]working in AI in health care?[138.502] [138.502][S02] Oh, wow.[139.46] [139.46][S02]My whole career, I would say.[140.74] [140.74][S02]I was still a student when, I think it was CS229,[145.27] [145.27][S02]the class on machine learning was taking off.[148.43] [148.43][S02]And I remember working with--[150.76] [150.76][S02]I was further along in my PhD, but working with students[154.06] [154.06][S02]on how to segment the larynx.[156.612] [156.612][S02]And AI was this new tool that could really[159.19] [159.19][S02]help us do that better.[160.85] [160.85][S02]And I haven't stopped ever since.[163.33] [163.33][S01] But why this area for you?[164.93] [164.93][S01]What's the appeal of it for you?[166.263] [166.263][S02] In health care, you're always challenged, right?[168.888] [168.888][S02]There are lots of things that we don't do that well[171.43] [171.43][S02]or you don't have infinite data.[173.23] [173.23][S02]And so you're never perfect.[175.07] [175.07][S02]So you're always trying to leverage the best tools[177.71] [177.71][S02]to do what you're trying to do.[180.11] [180.11][S02]And then I joined Google a decade ago[182.69] [182.69][S02]to work on surgical robotics.[184.92] [184.92][S02]And I remember vividly.[186.21] [186.21][S02]It was the beginning of us really[188.36] [188.36][S02]realizing that, with images, anything that a human could[192.77] [192.77][S02]do in terms of interpreting those images,[195.63] [195.63][S02]machines were going to be able to do.[198.03] [198.03][S02]And so I sat down with my surgeon colleague,[201.38] [201.38][S02]and we took 100 cholecystectomies,[204.27] [204.27][S02]which are when you remove the gallbladder.[206.66] [206.66][S02]We manually segmented livers and gallbladders.[210.81] [210.81][S02]And I fed a pretty simple neural net[214.07] [214.07][S02]to see whether the algorithm would be able to decipher[217.1] [217.1][S02]between those two things.[218.61] [218.61][S02]And I remembered being incredibly surprised, actually,[221.1] [221.1][S02]when it did that perfectly.[222.75] [222.75][S02]Because most of the other tools we had at our disposal[226.04] [226.04][S02]were never perfect.[227.45] [227.45][S02]It did the job, but not always.[229.29] [229.29][S02]Now, livers and gallbladders really don't look the same.[233.16] [233.16][S02]It's a very easy task.[234.98] [234.98][S02]Any student will be able to tell you which one is which.[238.73] [238.73][S02]But still, that really, for me, meant, OK, now we[242.47] [242.47][S02]have something that will be infinitely more[245.23] [245.23][S02]powerful than any of the algorithms I had been working on[250.15] [250.15][S02]previously.[251.06] [251.06][S02]And people forget, right?[252.22] [252.22][S02]But when I was still a student, we didn't know--[255.29] [255.29][S02]it was hard with computer vision to decipher a QR code.[259.315] [259.315][S02]That was where we were back then.[262.34] [262.34][S02]And so, again, AI has really opened up[268.096] [268.096][S02]a world of capabilities far beyond anything[272.74] [272.74][S02]we could envision before.[274.365] [274.365][S01] OK.[274.99] [274.99][S01]Well, let's start off with some of the examples of places[277.75] [277.75][S01]where AI has already made a lot of headway.[280.46] [280.46][S01]I'm thinking about diagnosis here and maybe[282.55] [282.55][S01]medical imaging in particular.[284.18] [284.18][S01]Talk to me a bit about what's been going on.[286.1] [286.1][S02] Yeah.[286.55] [286.55][S02]We've seen a lot of that over the last 10 years.[288.55] [288.55][S02]That's what we typically call \"narrow AI.\"[290.75] [290.75][S02]Narrow because it's solving one task.[293.175] [293.175][S02]You mentioned medical imaging.[294.62] [294.62][S02]So you can take a chest X-ray, or an MRI,[297.83] [297.83][S02]or we've done also diabetic retinopathy work.[301.83] [301.83][S02]Those are already tools that have been FDA-cleared,[305.1] [305.1][S02]for example, and deployed in the clinic.[307.02] [307.02][S02]And they are augmenting radiologists in the sense[310.58] [310.58][S02]that, for those specific tasks, the machine[313.28] [313.28][S02]does a pretty damn good job at annotating[317.06] [317.06][S02]the image for the physician.[318.69] [318.69][S01] Let me make sure I understand this then.[320.07] [320.07][S01]OK.[320.57] [320.57][S01]So if we take diabetic retinopathy as an example,[323.76] [323.76][S01]this is people who have diabetes and, potentially, it's[327.8] [327.8][S01]a cause of blindness.[328.757] [328.757][S01]Is that right?[329.34] [329.34][S02] Absolutely.[330.423] [330.423][S02]It's the leading cause of preventable blindness worldwide.[333.15] [333.15][S02]And it's pretty simple.[335.22] [335.22][S02]So you take an image of the back of the eye with a fundus camera,[339.78] [339.78][S02]and that gives you an image of the retina that[343.16] [343.16][S02]then ophthalmologists will grade on a scale from 1 to 5[347.45] [347.45][S02]to indicate whether patients have[350.54] [350.54][S02]no disease, moderate disease, or severe disease.[355.36] [355.36][S02]And what was striking when we started working on that[359.22] [359.22][S02]particular problem was the fact that it was actually really hard[363.72] [363.72][S02]to get to what we call \"ground truth,\" meaning,[365.76] [365.76][S02]as we're training our machine learning models to learn a task,[370.69] [370.69][S02]we first need to provide them with a set of ground-truth[375.15] [375.15][S02]images.[376.23] [376.23][S02]That means the pairs of the image and the label.[380.23] [380.23][S02]And for that particular task, we asked maybe 1, 2,[385.18] [385.18][S02]3 ophthalmologists, and often we got different answers.[388.99] [388.99][S02]And I think one thing we did particularly well[391.65] [391.65][S02]was to be very rigorous in trying[394.47] [394.47][S02]to establish ground truth.[396.31] [396.31][S02]And so we ended up working with 50--[398.31] [398.31][S02]5-0-- ophthalmologists.[400.06] [400.06][S02]And we realized that, actually, on some images,[402.79] [402.79][S02]they really didn't agree with each other.[405.63] [405.63][S02]Some would have given 1, 2, 3.[408.125] [408.125][S02]Each one of the scores could have potentially[410.61] [410.61][S02]been the diagnosis a patient would have received[413.05] [413.05][S02]because patients only see one.[414.46] [414.46][S02]They don't go and have their images graded by--[416.77] [416.77][S01] 50 times.[417.645] [417.645][S02] --50 people.[417.98] [417.98][S02]Yeah.[418.84] [418.84][S01] Boy.[418.915] [418.915][S01]So some ophthalmologists would say,[420.373] [420.373][S01]this is a mild or no disease, and others[422.71] [422.71][S01]would say all the way up to severe, on the same image.[425.54] [425.54][S02] Precisely, which sometimes is not too surprising,[428.39] [428.39][S02]meaning that most ophthalmologists-- and that's[430.87] [430.87][S02]true for most physicians-- see what's in their community.[435.37] [435.37][S02]And the rare cases, by definition,[438.078] [438.078][S02]they don't see them very often.[439.37] [439.37][S02]So it can get very hard to properly annotate an image,[443.17] [443.17][S02]if you don't see that type very often.[445.168] [445.168][S01] Well, how do you do that, then?[446.96] [446.96][S01]If even the experts don't agree, how[449.23] [449.23][S01]do you decide what the correct answer is?[451.82] [451.82][S02] Well, that's exactly why we had to go to 50.[454.4] [454.4][S02]Then you have this consensus among all ophthalmologists.[457.79] [457.79][S02]And sometimes you go to even deeper experts[460.72] [460.72][S02]to really get to what is the correct label for that image.[464.39] [464.39][S02]And that way, you properly train your algorithms.[466.96] [466.96][S02]And then we also did a lot of rigorous testing[469.99] [469.99][S02]to make sure that we had algorithms in which we could--[472.912] [472.912][S02]I mean, which we could trust.[474.12] [474.12][S01] But then I suppose, on the flip side of that,[476.07] [476.07][S01]even though that makes the labeling quite difficult,[478.41] [478.41][S01]if you know that experts are disagreeing,[481.83] [481.83][S01]then actually you can create something[484.07] [484.07][S01]that is a new gold standard really, better than the experts.[487.697] [487.697][S02] Absolutely.[488.78] [488.78][S02]And I think that's where the AI has really its role to play.[491.875] [491.875][S02]It's for tasks that it is really good at that are really not[496.82] [496.82][S02]very easy to be done by a single human.[501.26] [501.26][S01] I seem to remember reading this paper about where[505.55] [505.55][S01]some researchers had all of these images[507.41] [507.41][S01]of the back of the eye and were like, you know what?[509.91] [509.91][S01]Let's just see, for fun, if you can[512.72] [512.72][S01]tell the sex of the patient based on the blood vessels[515.99] [515.99][S01]at the back of the eye.[517.169] [517.169][S01]And no ophthalmologist in the world, I think,[519.32] [519.32][S01]could do this reliably, but the AI can.[523.427] [523.427][S02] Yeah, absolutely.[524.76] [524.76][S02]It was very interesting to then interrogate those images[528.53] [528.53][S02]for other things than just diabetic retinopathy[531.6] [531.6][S02]and see that the model performed better than flipping a coin.[536.62] [536.62][S02]And for some things--[538.2] [538.2][S02]I'm not sure you need a model to decipher[541.11] [541.11][S02]the sex of your patient.[542.32] [542.32][S01] [LAUGHS][543.27] [543.27][S02] But I do think for indications[546.12] [546.12][S02]of potential cardiac disease, for example,[548.44] [548.44][S02]that can be a useful tool.[551.32] [551.32][S02]Because if you're going to do that screening[553.26] [553.26][S02]for diabetic retinopathy anyway, but now you[556.29] [556.29][S02]have a test that can also provide you[558.15] [558.15][S02]with additional information, you could potentially[562.44] [562.44][S02]screen for more diseases.[564.22] [564.22][S02]So, indeed.[564.97] [564.97][S02]And it was both interestingly scientifically[568.72] [568.72][S02]to realize how those images contained a lot more information[572.46] [572.46][S02]than we initially thought, and with, potentially,[576.42] [576.42][S02]also applications that could derive from that.[578.74] [578.74][S01] Well, does it imply maybe[580.38] [580.38][S01]that even though you've created an algorithm[583.08] [583.08][S01]to do this narrow task, actually there is potential for the AI[586.83] [586.83][S01]to help improve the overall understanding of the human body?[591.27] [591.27][S02] I would answer maybe in two ways.[593.27] [593.27][S02]The first is, in radiology, there's this thing called[596.53] [596.53][S02]\"incidental findings.\"[598.6] [598.6][S02]Because it's very often that, if you[600.52] [600.52][S02]screen me head to toe with MRI right now, you'll find things.[604.93] [604.93][S02]You'll find nodules, things that are abnormal.[608.06] [608.06][S02]They will get me worried.[609.32] [609.32][S02]But actually, we never did clinical trials[612.19] [612.19][S02]to see whether having that thing somewhere is good or bad.[615.26] [615.26][S02]And many people live with those things[617.29] [617.29][S02]forever, without any issues.[619.43] [619.43][S02]So you don't really want to do that type of screening that[622.33] [622.33][S02]will get everyone worried.[625.03] [625.03][S02]And the other thing I would say is, when you do a blood test,[629.75] [629.75][S02]you do a blood test for a very particular thing.[632.305] [635.71][S02]The lab who did the blood test for finding a particular virus[640.39] [640.39][S02]is not responsible for checking everything.[643.02] [643.02][S02]And if they haven't noticed that you had something else that[645.52] [645.52][S02]wasn't what you got blood drawn, they're[649.098] [649.098][S02]not responsible for that.[650.14] [650.14][S02]And so for imaging, it's always a little ambiguous[654.99] [654.99][S02]what to do with those incidental findings.[657.28] [657.28][S02]And for those aspects of narrow AI, maybe--[662.07] [662.07][S02]I think the jury is still out whether or not[665.47] [665.47][S02]we want the AI to always look for everything else, when it's[668.64] [668.64][S02]looking at one particular image, even if it could[671.73] [671.73][S02]provide pretty helpful alarms.[674.7] [674.7][S01] Those are images that we've spoken about,[677.14] [677.14][S01]but are there other inputs that you[679.95] [679.95][S01]can use for this very narrow AI type of diagnosis?[683.53] [683.53][S02] Absolutely.[684.78] [684.78][S02]I would say all types of modalities, right?[687.01] [687.01][S02]We've done a lot of very interesting work with sound,[690.25] [690.25][S02]for example, leveraging sound to be able to detect tuberculosis[694.5] [694.5][S02]from cough.[695.29] [695.29][S02]And people have also explored that quite a bit during COVID,[699.75] [699.75][S02]obviously.[701.1] [701.1][S02]And so we see good biomarkers, actually,[705.03] [705.03][S02]in sound that can help also provide some relatively cheap[710.2] [710.2][S02]ways to diagnose those types of diseases.[713.045] [713.045][S01] So, wait.[713.92] [713.92][S01]It actually works then?[714.98] [714.98][S02] Yes, it does.[715.18] [715.18][S01] You record a cough, and it[716.763] [716.763][S01]can tell if you've got TB.[718.06] [718.06][S02] In short, yes, that's exactly what it is.[720.44] [720.44][S01] OK.[721.208] [721.208][S01]But there are limitations to this stuff, though, right?[723.5] [723.5][S01]I mean, are they getting it right every single time?[727.76] [727.76][S02] No.[728.51] [728.51][S02]I think that's a very good point.[731.34] [731.34][S02]Like with any algorithms, there is sensitivity and specificity.[735.23] [735.23][S02]So it can have either false positives, false negatives,[738.62] [738.62][S02]so thinking that something exists[740.23] [740.23][S02]when it doesn't, or missing something[742.33] [742.33][S02]that it should have picked.[745.39] [745.39][S02]In general, though, what we do and what[748.78] [748.78][S02]we did with the diabetic retinopathy paper[751.12] [751.12][S02]is really checking how it does with respect to the best[757.0] [757.0][S02]physicians or the best panels of physicians,[759.77] [759.77][S02]and then letting the human, who is always[762.79] [762.79][S02]working with the algorithm, decide[766.76] [766.76][S02]where on that specificity sensitivity tradeoff[770.49] [770.49][S02]they want to operate.[771.63] [771.63][S02]And so they can decide that, really,[773.69] [773.69][S02]they cannot afford false alarms, or the other way around,[778.55] [778.55][S02]that they cannot afford missing anything.[780.6] [780.6][S02]But in both cases, I think for an algorithm[784.13] [784.13][S02]to be cleared by regulatory authorities,[786.42] [786.42][S02]it needs to have shown very strong performance.[788.635] [788.635][S01] OK.[789.26] [789.26][S01]Let me break that down then.[790.47] [790.47][S01]So it could be that an algorithm could say that it's TB when it[794.93] [794.93][S01]isn't.[795.78] [795.78][S01]But it also could be that an algorithm[797.75] [797.75][S01]could miss a genuine case of TB and say that it was fine.[802.363] [802.363][S01]And obviously, you don't want either of those things[804.53] [804.53][S01]to happen.[805.62] [805.62][S01]But what percentage accuracy almost[808.1] [808.1][S01]are you willing to accept before you say,[810.27] [810.27][S01]OK, this model is adding value?[812.64] [812.64][S02] So it really depends.[814.475] [814.475][S02]If you're going to deploy your--[817.07] [817.07][S02]let's say it's a screening test, in places[819.47] [819.47][S02]that had nothing before.[820.69] [820.69][S02]It's a very different story than if you're[822.44] [822.44][S02]replacing an existing alternate way of screening[826.65] [826.65][S02]that same population.[828.04] [828.04][S02]If you're replacing something, you'd better be at least better.[832.38] [832.38][S02]If you're coming in where there was before a void, then[839.17] [839.17][S02]public health authorities are going[840.9] [840.9][S02]to decide what they deem as acceptable.[844.11] [844.11][S02]It also depends what will happen to the patients that are[847.26] [847.26][S02]provided with that diagnosis.[849.46] [849.46][S02]If they go home and you never see them again,[851.62] [851.62][S02]it's a different story than if your tool is[854.43] [854.43][S02]a pre-screening tool.[855.85] [855.85][S02]And then you're going to reroute your patients[858.0] [858.0][S02]to an additional screening tool for confirmation.[860.532] [860.532][S01] When you launch these things,[862.24] [862.24][S01]for example, the TB model or diabetic retinopathy,[865.51] [865.51][S01]does that mean that you aim initially[867.69] [867.69][S01]to go to places where they don't have these existing screening[870.84] [870.84][S01]tests in place?[872.275] [872.275][S02] It really depends.[873.65] [873.65][S02]For diabetic retinopathy, for example,[876.75] [876.75][S02]it is something that we've indeed deployed in Thailand.[880.03] [880.03][S02]And then we often work with partners[881.7] [881.7][S02]to really bring it where it makes a difference.[885.19] [885.19][S02]We've already screened 700,000 people in Thailand.[888.14] [888.14][S02]And we're 10X-ing that over the next few years.[891.59] [891.59][S01] Why Thailand?[892.815] [892.815][S02] So Thailand is one[894.19] [894.19][S02]of those countries in which the patient population[896.68] [896.68][S02]that each ophthalmologist has to care for is fairly large.[900.58] [900.58][S02]There are many places in which there aren't really[903.22] [903.22][S02]enough doctors.[904.63] [904.63][S02]And so it's a good example of a place in which AI screening can[910.48] [910.48][S02]really help improve outcomes.[913.93] [913.93][S01] Let's say that you have[915.46] [915.46][S01]a situation in which some sort of medical imaging[918.88] [918.88][S01]is being used in conjunction with a doctor.[923.47] [923.47][S01]What happens if the two disagree?[926.1] [926.1][S02] Ah.[926.85] [926.85][S02][CHUCKLES] I think at the end of the day,[929.36] [929.36][S02]the doctor is really the one making the call.[931.81] [931.81][S02]So it's very important that they retain control,[936.22] [936.22][S02]and they are the one making the decision,[938.23] [938.23][S02]and they are the ones signing their name on the report.[943.35] [943.35][S02]And so they would have to explain, right?[947.58] [947.58][S02]If they're saying the machine is wrong,[949.26] [949.26][S02]it could be for a variety of reasons.[953.46] [953.46][S02]In other cases, you could think of the algorithm[957.77] [957.77][S02]making the physician rethink.[960.126] [960.126][S02]So it's checking your work, like a spell checker, if you wish.[963.87] [963.87][S02]Sometimes I disagree with my spell checker,[966.92] [966.92][S02]with accents on words in French or things.[969.92] [969.92][S02]It's not perfect.[971.42] [971.42][S02]But I will double-check.[972.565] [972.565][S01] OK.[973.19] [973.19][S01]Well, let's talk about some of the more advanced stuff here.[975.69] [975.69][S01]Because up until now, I mean, the examples[977.75] [977.75][S01]that we've been using here is an MRI, or photo[981.83] [981.83][S01]of the back of an eye, or the sound of a cough.[986.0] [986.0][S01]What about the more holistic view?[988.32] [988.32][S01]So, I mean, really good medicine doesn't see a human[992.25] [992.25][S01]as a collection of interesting or otherwise medical problems.[995.63] [995.63][S01]It sees a human as a human.[997.67] [997.67][S01]Can you use AI to think more holistically?[1001.15] [1001.15][S02] AI is indeed very good at looking together[1006.26] [1006.26][S02]at things that, up until now, we were only--[1010.4] [1010.4][S02]or mostly capable of looking at individually.[1014.33] [1014.33][S02]And maybe bring insights that, up until now, we didn't have.[1020.123] [1020.123][S01] Like what?[1021.04] [1021.04][S01]Like what?[1021.71] [1021.71][S01]I mean, if you do manage to connect up everything[1024.14] [1024.14][S01]we understand about the cell, to everything we understand[1026.515] [1026.515][S01]about an organ, to eventually everything[1028.819] [1028.819][S01]we understand about human, what do you[1031.01] [1031.01][S01]find in those connections?[1033.317] [1033.317][S02] So back in the days, we did quite a bit of work[1035.9] [1035.9][S02]in virtual staining.[1038.46] [1038.46][S02]So when you have an H&E slide--[1040.92] [1040.92][S02]so that's what happens when any part of any piece of tissue[1046.369] [1046.369][S02]removed from your body, in surgery,[1048.8] [1048.8][S02]typically, will be sliced and looked at under the microscope.[1053.82] [1053.82][S02]And the way we do that is that we make[1058.34] [1058.34][S02]hypothesis as to what might be happening.[1062.69] [1062.69][S02]And we stain accordingly so that it makes those things visible.[1065.66] [1065.66][S02]And you can think of some of those techniques[1068.695] [1068.695][S02]as making that visible but without having to stain.[1072.17] [1072.17][S02]So really revealing some of that information,[1075.44] [1075.44][S02]if it's already in the tissue, but without having[1080.14] [1080.14][S02]to bring additional things that are[1083.11] [1083.11][S02]consuming that piece of tissue.[1084.77] [1084.77][S01] Would destroy the tissue for future tests.[1087.02] [1087.02][S02] Exactly, exactly.[1088.353] [1088.353][S02]So to me, there's also something quite profound in the way[1092.29] [1092.29][S02]we're now interrogating tissue with more and more instruments,[1097.04] [1097.04][S02]with things like transcriptomics, genomics,[1099.38] [1099.38][S02]single-cell, et cetera.[1100.85] [1100.85][S02]And so that's really a wonderful application for AI,[1104.51] [1104.51][S02]because the AI can look at each modality and, if you wish,[1109.28] [1109.28][S02]provide us with eyes for that particular scientific[1112.45] [1112.45][S02]instrument.[1113.28] [1113.28][S02]And then it goes beyond, because it's also[1115.03] [1115.03][S02]capable at bridging between those different types[1119.6] [1119.6][S02]of scientific instruments.[1121.05] [1121.05][S01] So then, you presumably[1123.11] [1123.11][S01]can start to combine some of these quite complex AI tools[1127.325] [1127.325][S01]on imaging, on genomics, on all the various things that you're[1131.24] [1131.24][S01]describing.[1132.24] [1132.24][S01]Does that actually allow you to advance in your understanding[1136.46] [1136.46][S01]of diseases?[1137.028] [1137.028][S01]I mean, I'm thinking of cancer care here, for instance.[1139.32] [1139.32][S02] Yeah, it's a great example.[1140.7] [1140.7][S02]And that's what we are actually precisely working[1142.742] [1142.742][S02]on with the L'Institut Curie, the Curie Institute in Paris.[1148.94] [1148.94][S02]That is very advanced in exploiting all of those most[1154.52] [1154.52][S02]recent modalities to better understand cancer,[1158.57] [1158.57][S02]in that case, women's cancer--[1160.41] [1160.41][S02]so uterine cancer or breast cancer,[1162.57] [1162.57][S02]which, despite our best efforts for many, many decades,[1166.2] [1166.2][S02]we're still short of answers for many women.[1170.03] [1170.03][S02]And we're really hoping that by combining those modalities,[1175.35] [1175.35][S02]by getting to a deeper understanding[1178.4] [1178.4][S02]at the cellular level of what's going on,[1181.5] [1181.5][S02]we'll finally be able to crack, for example,[1184.57] [1184.57][S02]triple-negative breast cancers or some of the-- again, more[1190.26] [1190.26][S02]advanced cancers that, today, we don't have good solutions for.[1194.71] [1194.71][S01] As well as cells and genomics and imaging, I mean,[1198.61] [1198.61][S01]there are other data sources that the AI[1200.85] [1200.85][S01]can bring into the equation here as well, right?[1203.077] [1203.077][S02] Absolutely.[1204.16] [1204.16][S02]And I think that's also maybe one[1208.05] [1208.05][S02]of the secrets AI will be able to decipher, which is,[1213.426] [1213.426][S02]health is in everything we do.[1217.0] [1217.0][S02]It's how many steps you walked today,[1219.22] [1219.22][S02]what you ate for breakfast, lunch,[1220.78] [1220.78][S02]and dinner, whether you've interacted with friends[1223.23] [1223.23][S02]or been lonely all day long.[1225.4] [1225.4][S02]And so it's typically really, really hard[1228.57] [1228.57][S02]to translate that into either healthy habits[1233.52] [1233.52][S02]or understand how much of that is factoring into something[1237.73] [1237.73][S02]physical that is going on.[1239.78] [1239.78][S02]And we're doing a lot of work with sensors.[1242.33] [1242.33][S02]I'm wearing this Fitbit today.[1244.13] [1244.13][S02]It's just one of the data sources[1245.56] [1245.56][S02]that can really be leveraged, in combination with everything[1248.29] [1248.29][S02]else, to try to both better understand[1251.29] [1251.29][S02]health, sometimes just as an individual,[1253.7] [1253.7][S02]and also when we're trying to change behavior,[1258.34] [1258.34][S02]accompany us on that journey.[1259.548] [1259.548][S01] I mean, you're talking[1260.965] [1260.965][S01]about quite different quality of data here.[1262.78] [1262.78][S01]You've got slides from biopsies and genomic data[1267.61] [1267.61][S01]and then step count.[1269.0] [1269.0][S01]It sort of feels like a quite fuzzy thing around the edges.[1272.69] [1272.69][S01]Would it actually make a difference?[1274.762] [1274.762][S02] It does.[1275.72] [1275.72][S02]I mean, it's been shown over and over again,[1277.7] [1277.7][S02]for example, how much sleep matters[1279.65] [1279.65][S02]for both cardiovascular health and in oncology[1284.6] [1284.6][S02]in terms of preventing cancer.[1286.55] [1286.55][S02]So it's one of my hopes, that AI will bring those[1289.72] [1289.72][S02]two together in a way that is harder[1291.52] [1291.52][S02]to do in our traditional health care systems[1293.92] [1293.92][S02]and also very hard to do for us as individuals[1296.66] [1296.66][S02]because you see that type of impact at a population level.[1301.83] [1301.83][S02]As an individual, it's really hard to convince yourself[1305.115] [1305.115][S02]that you should go to bed one hour earlier because that's[1307.49] [1307.49][S02]actually the best thing you can do for your health.[1309.42] [1309.42][S01] There's quite a lot of buzz[1311.045] [1311.045][S01]at the moment about digital twins.[1312.48] [1312.48][S01]Does this also add to the whole holistic view of health?[1316.14] [1316.14][S01]I mean, is there an idea of making[1317.81] [1317.81][S01]a kind of digital twin of yourself[1319.67] [1319.67][S01]for health care purposes?[1321.24] [1321.24][S02] Yeah.[1322.073] [1322.073][S02]We see that a lot.[1323.79] [1323.79][S02]And it can mean different things for different people.[1326.04] [1326.04][S02]There is the idea that the digital twin[1328.28] [1328.28][S02]is kind of a simulation of someone with a similar persona,[1333.18] [1333.18][S02]for example, and then that can be a good proxy[1335.54] [1335.54][S02]to interrogate the potential impact of different types[1340.25] [1340.25][S02]of interventions.[1341.46] [1341.46][S02]And then digital twin can mean something quite different[1344.24] [1344.24][S02]for the pharma industry, for example, where there's[1347.21] [1347.21][S02]a lot of work trying to see how far we can go with in silico[1351.59] [1351.59][S02]arms, for example, of clinical trials,[1354.04] [1354.04][S02]if we manage to assemble virtual cohorts that are allowing us[1358.41] [1358.41][S02]to, in the end, gain as much knowledge about the safety[1365.34] [1365.34][S02]and efficacy of a drug, but without needing[1368.25] [1368.25][S02]as many people for that particular clinical trial.[1371.957] [1371.957][S01] Oh, that's so interesting.[1373.54] [1373.54][S01]So is this about, I don't know, making[1376.05] [1376.05][S01]a cohort of simulated humans almost-- maybe not[1380.208] [1380.208][S01]the whole thing-- maybe just an organ[1381.75] [1381.75][S01]or whatever it is you're particularly focusing on--[1384.575] [1384.575][S01]but so that you don't necessarily[1385.95] [1385.95][S01]have to use as many people in your clinical trial?[1388.2] [1388.2][S02] Exactly.[1388.62] [1388.62][S02]For clinical trials, that would be exactly that.[1390.782] [1390.782][S01] But then to be able to do that effectively, you[1393.24] [1393.24][S01]have to really, really understand what real humans look[1396.72] [1396.72][S01]like, which means having a wealth of data[1399.36] [1399.36][S01]from individual patients somewhere along the way.[1402.46] [1402.46][S01]How do people feel about contributing their own data[1407.46] [1407.46][S01]for this kind of end?[1408.85] [1408.85][S01]For the research of medical purposes?[1411.9] [1411.9][S02] I think when we talk about privacy,[1414.57] [1414.57][S02]we see different concerns in different parts of the world.[1418.82] [1418.82][S02]In Europe, there are a lot of sovereignty concerns.[1421.2] [1421.2][S02]People want to know that the data is--[1423.32] [1423.32][S02]I mean, not in all countries-- but is staying on their soil.[1427.71] [1427.71][S02]And we have solution for that.[1430.16] [1430.16][S02]All of the trusted public cloud solutions,[1432.81] [1432.81][S02]for example, are complying with the highest levels of regulation[1438.95] [1438.95][S02]on that topic.[1440.04] [1440.04][S02]In other parts of the world, we see a lot of appetite for people[1443.48] [1443.48][S02]to not think that they are sick for nothing.[1445.92] [1445.92][S02]By that, I mean, it's actually quite compelling[1448.28] [1448.28][S02]to know that if you're sick but you're contributing data[1451.07] [1451.07][S02]to research, you're helping making the last person that[1454.34] [1454.34][S02]is going to have the same fate a little less sick, if you wish.[1457.98] [1457.98][S02]So I think really enabling that virtuous cycle is really[1463.85] [1463.85][S02]where we have to be.[1465.12] [1465.12][S01] Because it does feel like health care data is[1467.51] [1467.51][S01]a particularly sensitive case.[1469.52] [1469.52][S01]Because, on the one hand, exactly as you described,[1471.72] [1471.72][S01]there is huge potential to advance our understanding[1475.41] [1475.41][S01]and improve conditions for people who--[1478.88] [1478.88][S01]future generations.[1480.28] [1480.28][S01]But on the other hand, if health care data[1484.11] [1484.11][S01]gets into the wrong hands, or is misused, or used irresponsibly,[1489.34] [1489.34][S01]I mean, there is real sort of--[1491.07] [1491.07][S01]I think people have real concerns[1493.23] [1493.23][S01]about the potential ramifications of that.[1495.31] [1495.31][S01]So, I mean, how do you strike that balance?[1498.727] [1498.727][S02] Absolutely.[1499.81] [1499.81][S02]Once you have the technology to actually protect the data,[1502.36] [1502.36][S02]there's also accountability.[1503.79] [1503.79][S02]That meaning that if we're saying that this research will[1507.3] [1507.3][S02]be helpful to the people, that indeed, at the end of the day,[1510.19] [1510.19][S02]it is research that is helpful to the people[1512.25] [1512.25][S02]where the data originated.[1513.52] [1513.52][S02]And I think that's a very important principle[1517.05] [1517.05][S02]in the same way that you shouldn't do clinical trials[1519.94] [1519.94][S02]in places of the world-- in some places of the world--[1522.19] [1522.19][S02]and then the drugs benefit in other places of the world.[1524.59] [1524.59][S01] So you're not just extracting from one[1526.02] [1526.02][S01]to give to another.[1527.02] [1527.02][S02] Precisely.[1527.73] [1527.73][S01] OK.[1528.37] [1528.37][S01]I also want to ask you a bit more about the patient's[1530.578] [1530.578][S01]experience in all of this.[1532.56] [1532.56][S01]Because I think-- are large language models[1535.01] [1535.01][S01]changing the game here already?[1536.58] [1536.58][S01]I mean, are you concerned that people[1538.73] [1538.73][S01]are trying to diagnose themselves with generative AI?[1541.597] [1541.597][S02] Yeah.[1542.43] [1542.43][S02]I mean, I would say, even before, you really[1544.67] [1544.67][S02]shouldn't do that, right?[1545.75] [1545.75][S02]You shouldn't play your own doctor.[1548.87] [1548.87][S01] People do, though, don't they?[1550.62] [1550.62][S01]Let's be honest.[1551.25] [1551.25][S02] I think people want[1552.667] [1552.667][S02]to know, and especially in places[1554.78] [1554.78][S02]where waiting time to see a physician[1557.105] [1557.105][S02]is becoming a real issue.[1559.02] [1559.02][S02]Then, of course, patients are trying[1562.72] [1562.72][S02]to learn as much as possible and to decipher what's happening,[1567.713] [1567.713][S02]and sometimes quite effectively when they are there because they[1570.38] [1570.38][S02]are their best advocate-- or a parent trying to figure[1573.02] [1573.02][S02]something out for their child.[1575.39] [1575.39][S02]And sometimes we have a couple of examples[1577.67] [1577.67][S02]where, for rare diseases, actually,[1579.78] [1579.78][S02]we've seen patients do a remarkable job.[1582.12] [1582.12][S02]Now, I would say, at Google, we have always[1584.6] [1584.6][S02]tried to provide the most helpful answers to our users.[1588.9] [1588.9][S02]And so with symptoms, for example,[1591.96] [1591.96][S02]we have the knowledge cards that you've probably[1594.95] [1594.95][S02]noticed that are telling you about a disease[1599.09] [1599.09][S02]with authoritative content, like, trying[1601.52] [1601.52][S02]to be clear in the explanation of common symptoms[1605.9] [1605.9][S02]and potential options for treatment,[1608.43] [1608.43][S02]but never crossing that line and telling you what[1611.93] [1611.93][S02]your own diagnosis might be.[1614.46] [1614.46][S02]Now, with large language models--[1617.99] [1617.99][S02]so our model, Gemini, will also not tell you[1622.73] [1622.73][S02]what your diagnosis will be.[1625.1] [1625.1][S02]It will tell you, sorry, I'm not a medical doctor.[1627.78] [1627.78][S02]You should go and see a doctor.[1629.85] [1629.85][S02]But you can get more helpful answers if you're asking,[1635.462] [1635.462][S02]what potential conditions could explain this?[1638.15] [1638.15][S02]And it will give you more of the textbook answer.[1641.1] [1641.1][S02]I like to think of it as those family[1643.22] [1643.22][S02]guide for health care, those--[1646.11] [1646.11][S02]I don't know, the big, thick books that you might have,[1649.71] [1649.71][S02]where you have small children and you're[1652.98] [1652.98][S02]trying to know what to do in those particular cases.[1655.99] [1655.99][S02]Well, it's that same level.[1658.21] [1658.21][S02]If we believe that large language models are going[1660.54] [1660.54][S02]to be able to provide diagnosis, we[1663.18] [1663.18][S02]need to continue the research.[1664.48] [1664.48][S02]And that's what we're doing with our work[1666.27] [1666.27][S02]with AMIE, the Articulate Medical Intelligence Explorer.[1669.85] [1669.85][S02]It's a research project in which we[1672.54] [1672.54][S02]are trying to see which conversational abilities[1677.4] [1677.4][S02]a large language model can have to establish a diagnosis.[1681.82] [1681.82][S02]So can it ask the right question,[1683.64] [1683.64][S02]the way a physician would do, like dialoguing with a patient?[1689.11] [1689.11][S01] How close are we to seeing[1690.81] [1690.81][S01]a system like AMIE being used actually in medical settings?[1695.693] [1695.693][S02] So we've been working on this project[1697.86] [1697.86][S02]for quite a while already.[1699.84] [1699.84][S02]And our first research papers were all[1702.45] [1702.45][S02]based on research done with patient actors[1706.86] [1706.86][S02]and simulated scenarios.[1708.74] [1708.74][S02]We're now doing a clinical study under IRB approval[1712.64] [1712.64][S02]with Harvard at the Beth Israel Medical Center[1716.06] [1716.06][S02]to see what happens, again, in a very controlled environment,[1720.32] [1720.32][S02]with physician supervision, with this system.[1723.48] [1723.48][S02]So that's our next step.[1724.92] [1724.92][S02]Hard for me to predict how long it[1727.85] [1727.85][S02]is before such a system would really help physicians[1731.84] [1731.84][S02]in the clinic.[1733.19] [1733.19][S02]But we are already seeing, in parts of the world,[1738.18] [1738.18][S02]similar systems that are typically[1739.64] [1739.64][S02]being used to answer low-risk questions, if you wish,[1745.43] [1745.43][S02]and that are always supervised.[1747.42] [1747.42][S02]So there is a physician maybe double-checking the answer[1750.062] [1750.062][S02]within a period of time, a short amount of time, like, say,[1752.52] [1752.52][S02]10 minutes of a model, or 15 minutes from a model,[1755.66] [1755.66][S02]answering a patient.[1757.08] [1757.08][S02]But you see how it's really a step-by-step approach.[1761.47] [1761.47][S01] Yeah, absolutely.[1762.785] [1762.785][S02] And I think that's very important, again.[1765.48] [1765.48][S02]Because if you have a model that says the right thing 9 out of 10[1770.3] [1770.3][S02]times, but the 10th, it's doing something really bad, well,[1774.03] [1774.03][S02]maybe it wasn't so good at all for none of those cases.[1778.217] [1778.217][S01] The benefits don't outweigh the costs.[1780.3] [1780.3][S02] Precisely.[1780.98] [1780.98][S01] That is interesting, though.[1782.647] [1782.647][S01]Because you're right, that if you're[1785.24] [1785.24][S01]talking to a real physician, I mean,[1787.2] [1787.2][S01]they will tell you that patients don't[1789.53] [1789.53][S01]come in with a sort of very clear, refined list[1793.535] [1793.535][S01]of relevant information for you.[1796.47] [1796.47][S01]I mean, they'll come in and say, oh, I don't feel so great.[1798.96] [1798.96][S01]And it's like, it's your job then to interrogate that[1801.168] [1801.168][S01]and actually find out precisely what's at the heart of it,[1804.938] [1804.938][S01]which bits of information are relevant and which bits aren't.[1807.48] [1807.48][S01]That must be a very difficult thing to mimic within an AI.[1813.063] [1813.063][S02] Indeed.[1813.98] [1813.98][S02]And also, I like to often think of the AI of tomorrow[1817.67] [1817.67][S02]as something that will be like the physician in your family.[1824.16] [1824.16][S02]There's often this physician in your family[1826.455] [1826.455][S02]that gets asked all questions about anything[1830.19] [1830.19][S02]that has to do with health care of all family members, all[1833.76] [1833.76][S02]the cousins, and everyone.[1835.06] [1835.06][S02]Even if they are a dermatologist,[1836.95] [1836.95][S02]they will be asked cardiology questions,[1839.13] [1839.13][S02]and they will help people navigate the system, et cetera.[1841.59] [1841.59][S02]But not all families have physicians.[1843.82] [1843.82][S02]And so I'd like to think that the AI we'll build tomorrow[1847.29] [1847.29][S02]will provide everyone with that, the equivalent[1850.23] [1850.23][S02]of a physician in your family.[1852.04] [1852.04][S02]But the key element is that that member of your family[1854.94] [1854.94][S02]actually knows you, right?[1857.28] [1857.28][S02]And they know you on the long term.[1859.6] [1859.6][S02]So they know if you're very anxious.[1861.24] [1861.24][S02]And when you're saying that you have[1863.16] [1863.16][S02]a bad headache for the last three weeks,[1865.21] [1865.21][S02]you've said that for the last 20 years,[1866.89] [1866.89][S02]and they can safely discard it.[1868.6] [1868.6][S02]Or if you're calling them, but you've actually never called[1871.77] [1871.77][S02]them, they know that it's something[1873.9] [1873.9][S02]they should pay attention to.[1875.29] [1875.29][S02]They will recognize the tone of your voice,[1877.45] [1877.45][S02]et cetera, which I think makes this--[1881.24] [1881.24][S02]I give the example of the dermatologist that[1883.47] [1883.47][S02]ends up having to say something about everything--[1885.78] [1885.78][S02]but pretty safe, right, in the end.[1887.55] [1887.55][S02]It's actually not going to provide bad advice to someone[1891.17] [1891.17][S02]that has a cardiac condition because they know the person,[1894.44] [1894.44][S02]and they will tell them to actually seek[1897.02] [1897.02][S02]cardiologist advice at the right time.[1900.45] [1900.45][S02]So we would want our AI systems to do the same.[1903.32] [1903.32][S02]But for that, it means they need to know you enough.[1905.58] [1905.58][S02]And it's not just very quickly saying your latest symptoms[1909.92] [1909.92][S02]and getting an answer.[1911.38] [1911.38][S02]Those systems have to be built.[1913.32] [1913.32][S01] I guess one of the differences of a physician[1915.74] [1915.74][S01]in your family is that they're not[1917.157] [1917.157][S01]really prone to hallucinations.[1918.496] [1918.496][LAUGHTER][1919.063] [1919.063][S02] Some of them might, but you're right.[1921.23] [1921.23][S01] Some of them might.[1922.522] [1922.522][S01]But how do you create something like that, when we are still[1926.93] [1926.93][S01]in a world where generative AI does have these hallucinations,[1930.44] [1930.44][S01]or is forgetful about timelines, or mistakes information-- all[1934.88] [1934.88][S01]of the common mistakes that you see in generative AI?[1937.497] [1937.497][S02] Yeah.[1938.33] [1938.33][S02]So that's why you cannot just take a large language model off[1944.61] [1944.61][S02]the shelf and apply it in health care and expect that to be[1949.17] [1949.17][S02]acceptable.[1950.73] [1950.73][S02]And I think we've done a lot of work first with Med-PaLM.[1954.34] [1954.34][S02]That was our first model fine-tuned on a medical corpus.[1957.91] [1957.91][S02]And we demonstrated that it was doing better[1961.41] [1961.41][S02]at answering medical license exams than, initially,[1967.11] [1967.11][S02]an average student and then a panel of experts.[1970.02] [1970.02][S02]And then we did Med Gemini.[1971.59] [1971.59][S01] What was Med Gemini?[1972.96] [1972.96][S02] Oh, Med Gemini is our large-language-model Gemini,[1976.45] [1976.45][S02]which I hope--[1977.07] [1977.07][S02]I know you've played with.[1978.153] [1978.153][S01] Hey, endlessly.[1980.13] [1980.13][S02] And we fine-tuned it on a medical corpus.[1983.68] [1983.68][S02]And so it's really inheriting the long context[1986.64] [1986.64][S02]of Gemini, its reasoning capabilities,[1988.6] [1988.6][S02]its native multi-modality.[1991.2] [1991.2][S02]But on top of that, it has seen medical data in a way that[1997.29] [1997.29][S02]Gemini hadn't.[1998.86] [1998.86][S02]And it's also being evaluated for specific medical purposes.[2003.36] [2003.36][S01] There is this other aspect of the human AI[2006.42] [2006.42][S01]collaboration, though.[2007.72] [2007.72][S01]I mean, if some of these tools are designed for doctors,[2010.75] [2010.75][S01]is there a risk that doctors, I guess, lose some of their skills[2015.18] [2015.18][S01]or start to become overly reliant on these kind of models?[2018.28] [2018.28][S02] There's always that risk, right?[2020.94] [2020.94][S02]In health care, I would say, though,[2022.63] [2022.63][S02]that for me, we don't really have a choice in the sense[2025.98] [2025.98][S02]that we have a big shortage of health care professionals[2029.28] [2029.28][S02]in many parts of the world.[2030.79] [2030.79][S02]So the question is, how do we address that?[2034.27] [2034.27][S02]How do we make sure that people, globally,[2037.53] [2037.53][S02]can be as healthy as possible and don't suffer from things[2043.2] [2043.2][S02]for which we have answers?[2044.46] [2044.46][S02]There are a lot of things for which we still[2046.35] [2046.35][S02]don't have answers.[2047.17] [2047.17][S02]But for a lot of others, where medicine can actually[2051.33] [2051.33][S02]provide an answer, I think we have a responsibility[2054.27] [2054.27][S02]to leverage the technology we have[2056.19] [2056.19][S02]to bring that to those people.[2058.02] [2058.02][S02]Rather than over-reliance, for me the question is,[2060.85] [2060.85][S02]how are we going to train the next generation, right?[2063.383] [2063.383][S02]And what is actually--[2064.3] [2064.3][S01] Of doctors.[2064.969] [2064.969][S02] Of doctors.[2066.052] [2066.052][S02]It's true of doctors within our conversation.[2068.179] [2068.179][S02]It can be true of other professions[2071.05] [2071.05][S02]within the context of the new era we're in with gen AI.[2075.53] [2075.53][S02]But for doctors, there might be things where actually the AI is[2079.54] [2079.54][S02]really doing a really good job.[2081.02] [2081.02][S02]And when it's not, it's also telling you[2082.96] [2082.96][S02]that it doesn't know, et cetera.[2084.55] [2084.55][S02]So you can rely on the AI.[2086.659] [2086.659][S02]And it's OK to not be as good at those tasks[2088.747] [2088.747][S02]that the AI is really good at, because there[2090.58] [2090.58][S02]are so many other things that you need to learn[2092.739] [2092.739][S02]and you need to do well.[2094.219] [2094.219][S01] So there are some things, then,[2096.011] [2096.011][S01]that you think the AI just isn't going to touch.[2098.57] [2098.57][S01]I mean, things like empathy, for instance.[2100.79] [2100.79][S02] Well, actually, that one is a good one[2102.998] [2102.998][S02]because we always check that our models have good bedside[2107.23] [2107.23][S02]manners.[2108.375] [2108.375][S02]And they actually do.[2109.85] [2109.85][S02]They're not bad at all.[2111.05] [2111.05][S02]They can adapt to their audience in a way[2113.11] [2113.11][S02]that few human beings can.[2115.75] [2115.75][S02]They can leverage the right language for a five-year-old,[2118.93] [2118.93][S02]or a 70-year-old, or a 40-year-old.[2121.3] [2121.3][S02]And if you're not a native speaker,[2123.76] [2123.76][S02]they can speak in your own language.[2125.47] [2125.47][S02]So actually, when they're judged on empathy,[2129.01] [2129.01][S02]they actually do pretty well.[2131.34] [2131.34][S01] How on Earth do you train it to be more empathic?[2134.14] [2134.14][S02] Well, in the same way[2135.64] [2135.64][S02]you train them for everything else.[2137.28] [2137.28][S02]They are based on a lot of conversations[2141.762] [2141.762][S02]that they've seen in the training data.[2143.8] [2143.8][S02]And if those conversations are empathic,[2146.4] [2146.4][S02]then they learn to also leverage the same words,[2151.27] [2151.27][S02]the same tone, et cetera.[2152.62] [2152.62][S02]And you do need to do some time to fine-tune them or provide[2156.93] [2156.93][S02]them with more specific examples,[2159.01] [2159.01][S02]because not everything on the web is empathic.[2162.465] [2162.465][S01] Do you think, though,[2163.84] [2163.84][S01]that there are some aspects of medical care[2167.25] [2167.25][S01]that really AI can't touch?[2170.88] [2170.88][S02] I've built surgical robots.[2172.63] [2172.63][S02]So even the haptics and all of that,[2174.82] [2174.82][S02]I tend to think that, at some point, we will have ways.[2177.49] [2177.49][S02]But today, whenever you need a physical exam,[2180.175] [2180.175][S02]like if you're checking someone's stomach, and you--[2184.33] [2184.33][S02]a lot of things, you absolutely need physicians.[2187.28] [2187.28][S02]I paused because I think every single time[2190.12] [2190.12][S02]we say the technology will never do this, we get wrong.[2193.9] [2193.9][S02]And at some point, things change in a way[2196.93] [2196.93][S02]that, actually, they can also be helpful in those circumstances.[2201.735] [2201.735][S02]It's coming as something that is augmenting physicians.[2206.21] [2206.21][S02]It's not replacing them in all of the tasks they do daily.[2212.42] [2212.42][S02]I see AI much more as the one paying[2214.42] [2214.42][S02]the debt for what the digital industry has done to physicians,[2218.12] [2218.12][S02]which is asking them to enter lots and lots of data[2221.41] [2221.41][S02]in computers for years, changing their job fairly drastically,[2227.8] [2227.8][S02]making them miserable in some cases.[2230.03] [2230.03][S02]I do think that we see a lot of burnout in the profession.[2235.49] [2235.49][S02]And I'm hoping AI contributes to bringing back[2239.65] [2239.65][S02]the joy of practicing medicine.[2241.473] [2241.473][S02]But to me, it's paying that debt, right?[2243.14] [2243.14][S02]It's like, finally, with all of the data[2245.17] [2245.17][S02]that physicians and all their health[2246.7] [2246.7][S02]care professionals have taken so much time to enter,[2250.25] [2250.25][S02]we are going to be able to derive insights or knowledge[2255.19] [2255.19][S02]and help them.[2256.118] [2256.118][S01] Here's why it was all worth it, I guess.[2258.285] [2258.285][S02] Yes.[2258.9] [2258.9][S01] But that is interesting listening to you[2260.41] [2260.41][S01]talk about it, though.[2261.14] [2261.14][S01]Because, I mean, so many of the conversations that I get to have[2263.807] [2263.807][S01]on this podcast are about, I don't know, AGI and long-term--[2268.09] [2268.09][S01]these holistic systems and advances[2272.05] [2272.05][S01]towards completely doing things in a different way than we[2274.87] [2274.87][S01]always have.[2275.75] [2275.75][S01]But the way that you're describing your vision of this,[2278.48] [2278.48][S01]it's almost like the AI here is a tool that will change[2282.43] [2282.43][S01]every aspect of medicine.[2284.15] [2284.15][S01]But fundamentally, the idea of a patient-doctor relationship[2287.95] [2287.95][S01]is unchanged.[2289.283] [2289.283][S02] You're right.[2290.45] [2290.45][S02]I think I might be a little maybe[2294.43] [2294.43][S02]more cautious than what we're envisioning[2298.59] [2298.59][S02]in other aspects of society.[2301.62] [2301.62][S02]And it also depends, again, where in the world[2306.09] [2306.09][S02]we are envisioning that impact of technology.[2308.43] [2308.43][S02]Because I think that technology will bring health care to parts[2311.97] [2311.97][S02]of the world that probably didn't have access[2315.24] [2315.24][S02]to health care at all, or only in a very limited way,[2319.26] [2319.26][S02]in ways that we are not yet envisioning.[2323.11] [2323.11][S02]And that might be beyond anything we've described today.[2328.56] [2328.56][S02]But I'm also a bit skeptical of health care being entirely[2333.9] [2333.9][S02]revolutionized overnight, and it will[2336.69] [2336.69][S02]be completely different tomorrow from how it looked yesterday.[2342.51] [2342.51][S02]That being said, I'm also absolutely blown away[2345.36] [2345.36][S02]with models like AMIE.[2348.0] [2348.0][S02]If you think about it, you can now[2350.85] [2350.85][S02]have a conversation with a large language model that[2356.74] [2356.74][S02]is very similar to the conversation you would have[2360.67] [2360.67][S02]with an incredibly well-educated specialist,[2365.53] [2365.53][S02]pulling knowledge from all around the world.[2369.08] [2369.08][S02]Meaning, this very small clinical trial that happened[2373.36] [2373.36][S02]in the '50s in some remote place and was published back then,[2379.66] [2379.66][S02]echoing some new things that are starting to be seen somewhere[2383.86] [2383.86][S02]else.[2384.61] [2384.61][S02]Up until now, serendipity sometimes played a very big role[2390.16] [2390.16][S02]in, if you had a rare condition, as to whether you[2394.24] [2394.24][S02]would be treated or not.[2395.42] [2395.42][S02]And now, with our large language models,[2398.21] [2398.21][S02]I think we're completely changing that serendipity game.[2402.023] [2402.023][S01] It's always nice to finish a podcast on a very[2404.44] [2404.44][S01]positive and optimistic note.[2406.04] [2406.04][S01]But I think there's a lot to look forward to.[2407.917] [2407.917][S02] Absolutely.[2409.0] [2409.0][S01] Thank you.[2409.67] [2409.67][S02] Thank you.[2410.33] [2410.33][S01] I think medicine is a special case.[2412.63] [2412.63][S01]Because between the false positives[2414.71] [2414.71][S01]and the false negatives and, crucially, the potential harms[2417.77] [2417.77][S01]of either, this is something that has a very high bar[2421.82] [2421.82][S01]to get right.[2423.11] [2423.11][S01]And I think there is something reassuring about Joelle's[2426.17] [2426.17][S01]no-nonsense determination to take the scientific approach[2430.34] [2430.34][S01]here, to carefully and cautiously evaluate[2433.82] [2433.82][S01]the benefit of using AI to make sure that we end up[2437.63] [2437.63][S01]with the best health care outcomes for as many people as[2440.96] [2440.96][S01]possible.[2441.96] [2441.96][S01]You've been listening to \"Google DeepMind, The Podcast,\" with me,[2445.11] [2445.11][S01]Professor Hannah Fry.[2446.16] [2446.16][S01]If you enjoyed that episode, then[2447.65] [2447.65][S01]do subscribe to our YouTube channel.[2449.67] [2449.67][S01]You can also find us on your favorite podcast platform.[2452.79] [2452.79][S01]And we have plenty more episodes on a whole range of topics[2455.66] [2455.66][S01]to come, so do check those out, too.[2457.92] [2457.92][S01]See you next time.[2458.69] [2458.69][MUSIC PLAYING][2461.74]"}
{"file_name": "audio/val_000005.wav", "transcription": "[0.423][S01] Welcome back to \"Google DeepMind: The Podcast\"[2.84] [2.84][S01]with me, your host, Professor Hannah Fry.[4.94] [4.94][S01]Now, those of you who've been following us from the beginning,[7.68] [7.68][S01]you will know that using artificial intelligence[10.43] [10.43][S01]to enhance a scientific understanding of the world[13.58] [13.58][S01]has always been a key goal of the team here.[17.06] [17.06][S01]You've got weather forecasting, that's advanced meteorology.[20.07] [20.07][S01]You've got nuclear fusion for physicists,[22.5] [22.5][S01]and perhaps most notably is AlphaFold,[25.2] [25.2][S01]which has been an absolute game changer for biology.[28.77] [28.77][S01]But if all of that seems a little bit academic,[31.2] [31.2][S01]a bit niche, perhaps, the kind of stuff that should only really[33.98] [33.98][S01]be exciting to a handful of scientists in their labs, well,[37.25] [37.25][S01]then hopefully, today's episode will persuade you[40.07] [40.07][S01]of the substantial impact that these ideas can[43.28] [43.28][S01]have on humanity.[45.24] [45.24][S01]Because I am joined, once again, by the guy who is in charge[48.2] [48.2][S01]of scientific research here at \"Google DeepMind,\"[50.39] [50.39][S01]Pushmeet Kohli.[51.46] [58.9][S01]Pushmeet, thank you so much for joining us back on the podcast.[62.14] [62.14][S01]Quite a lot has happened since we last spoke to you.[64.73] [64.73][S01]How many nature papers have you got now?[67.6] [67.6][S02] I'm not really sure.[69.1] [69.1][S02]We're not counting.[70.67] [70.67][S01] Too many to count.[71.92] [71.92][S02] Yeah.[72.9] [72.9][S01] So I want to talk to you about AlphaFold because it[75.525] [75.525][S01]feels as though this is one of the most exciting things that[79.32] [79.32][S01]has advanced even further since I last spoke to you.[83.22] [83.22][S01]And I should say, actually, that anybody[85.59] [85.59][S01]who wants to know more about how AlphaFold works[89.01] [89.01][S01]and what it does, can go back to episode[92.01] [92.01][S01]one of the previous series for more detail,[94.94] [94.94][S01]but I guess just briefly, can you summarize what AlphaFold is?[98.215] [98.215][S02] Yeah.[99.09] [99.09][S02]So AlphaFold is a system that, given a protein--[104.01] [104.01][S02]proteins make everything around us.[106.498] [106.498][S02]They are the building blocks of life.[108.04] [108.04][S02]And essentially, they can be represented[110.1] [110.1][S02]as a sequence of amino acids.[113.25] [113.25][S02]And it's a great mystery as to what[115.98] [115.98][S02]is the structure of any given protein.[118.21] [118.21][S02]So if you are given a sequence of these amino acids, a protein,[121.45] [121.45][S02]what is its structure?[123.49] [123.49][S02]And that's really important because that informs[126.24] [126.24][S02]the function of that protein.[128.06] [128.06][S02]What AlphaFold does, it takes a sequence of a protein[132.28] [132.28][S02]and predicts its structure.[134.33] [134.33][S02]And that's really important for scientists[136.93] [136.93][S02]to understand what is the function of that protein.[139.43] [139.43][S02]So that's what AlphaFold did.[141.14] [141.14][S02]It solved that 50-year-old grand challenge in science by making[147.94] [147.94][S02]scientists able to find the structure of any given protein[151.54] [151.54][S02]in a matter of seconds.[153.023] [153.023][S01] So I guess actually, a more accurate statement[155.44] [155.44][S01]would be it was a mystery but is no longer.[158.295] [158.295][S02] Yeah.[159.17] [159.17][S02]There's still a few things that are unknown about how proteins[164.08] [164.08][S02]behave, the dynamics of the proteins,[165.79] [165.79][S02]and so on, but the one standard big problem in protein[170.2] [170.2][S02]understanding, What is the structure of a protein?[174.355] [174.355][S02]is now-- we now know about it.[177.445] [177.445][S01] So what's changed since last time[179.32] [179.32][S01]then, since the last series when there was AlphaFold 2?[182.36] [182.36][S01]What can AlphaFold3 do that the previous versions couldn't?[184.997] [184.997][S02] That's a very good question.[186.83] [186.83][S02]Now, I said proteins are the building blocks of life,[189.86] [189.86][S02]but they are not the only molecules in our body.[192.86] [192.86][S02]Our body has many other biomolecules.[195.35] [195.35][S02]We have nucleic acids, RNA, DNA, which[199.27] [199.27][S02]are the recipe of life, what makes you and me, Hannah.[203.14] [203.14][S02]Then there are small molecules, drugs that[208.03] [208.03][S02]interact with these proteins.[210.1] [210.1][S02]There are antibodies.[211.31] [211.31][S02]There are so many different types[212.83] [212.83][S02]of molecules that are interacting in any living being.[218.21] [218.21][S02]And we want to understand the structure[220.09] [220.09][S02]of all these molecules.[222.14] [222.14][S02]And that's what AlphaFold 3 unlocks.[224.54] [224.54][S02]It not only gives you the structure of a protein,[227.81] [227.81][S02]but it also gives you the structure of,[229.48] [229.48][S02]how do these proteins form complexes[232.12] [232.12][S02]or they connect to each other?[234.38] [234.38][S02]Or how do they interact with small molecules or antibodies[237.8] [237.8][S02]and so on?[238.37] [238.37][S01] What's been the reaction then from scientists[241.21] [241.21][S01]now that this has had a bit of time to be[242.92] [242.92][S01]out there in the world?[244.48] [244.48][S02] So it was very hard to actually--[248.74] [248.74][S02]my background, I'm a computer scientist by training,[251.41] [251.41][S02]and it was very hard for me to actually understand[255.25] [255.25][S02]the impact of AlphaFold.[259.81] [259.81][S02]And the first time I realized, I went to a conference, a biology[263.68] [263.68][S02]conference, where a biologist spoke to me and said,[267.44] [267.44][S02]well, there was this protein that I've been trying[270.16] [270.16][S02]to study for the last 10 years.[272.05] [272.05][S02]And I have collected so much data about it,[275.921] [275.921][S02]but still the structure of that protein was a mystery.[279.73] [279.73][S01] Because it was so, so, so hard to work[281.95] [281.95][S01]out what the structure might be before AlphaFold.[284.32] [284.32][S02] Exactly.[286.052] [286.052][S02]And the scientists had used AlphaFold[288.31] [288.31][S02]and had the structure, and then the question[291.67] [291.67][S02]he had was, well, what do I do now?[295.79] [295.79][S02]I have the structure, and so--[300.0] [300.0][S02]like he had to completely rethink what biology requires[303.09] [303.09][S02]now in terms of the next steps.[305.64] [305.64][S01] A 10-year project, and he just put it[307.838] [307.838][S01]into AlphaFold, and it sorted it out for him in a few minutes[310.38] [310.38][S01]or maybe overnight.[311.59] [311.59][S01]Was he excited or was that in some ways a bit demoralizing?[316.21] [316.21][S02] No.[317.61] [317.61][S02]I think there was a mixture of both surprise[320.92] [320.92][S02]but also excitement, something that you wouldn't imagine.[325.912] [325.912][S02]It's like, imagine giving people an ability that did not exist.[335.67] [335.67][S02]Imagine the first time telephones came about,[339.25] [339.25][S02]and you could now talk to people who are miles away.[343.282] [343.282][S02]This was unimaginable.[345.215] [345.215][S02]And the fact that you can take any protein,[347.52] [347.52][S02]you can put in the sequence of that protein,[349.77] [349.77][S02]and visualize what the 3D structure is,[351.99] [351.99][S02]that just gave the scientists a superpower that they had not[356.2] [356.2][S02]imagined earlier.[357.5] [357.5][S02]And there was a lot of excitement[359.47] [359.47][S02]as to what we could do with that superpower.[361.98] [361.98][S01] And is that how people are feeling, like they've[364.48] [364.48][S01]got a superpower now?[365.66] [365.66][S02] Yeah.[366.535] [366.535][S02]So one of the things that we have seen is AlphaFold database.[370.94] [370.94][S02]So we created this AlphaFold database,[373.01] [373.01][S02]which is, we found the structures[375.04] [375.04][S02]of almost all known proteins.[377.23] [377.23][S02]And we put the structures for those in a database[381.25] [381.25][S02]hosted by the European microbiology lab, our partners,[384.77] [384.77][S02]EMBL-EBI, and this was available-- these 250 million[389.11] [389.11][S02]structures were available to anyone in the world for free.[392.89] [392.89][S02]And that particular database has been used across 140 countries[399.1] [399.1][S02]by 1.8 million scientists.[403.46] [403.46][S02]And the fact that there are 1.8 million scientists who[406.87] [406.87][S02]are looking for protein structures,[408.862] [408.862][S02]if that is not a positive statement[410.32] [410.32][S02]about the state of the world, then I don't know what is.[413.9] [413.9][S02]We think about the doom and gloom[417.27] [417.27][S02]in society, but if you look at how science has progressed--[421.12] [421.12][S02]there are 1.8 million people who are studying proteins.[424.96] [424.96][S01] So I think the really notable thing when[427.593] [427.593][S01]you talk to people about AlphaFold[429.01] [429.01][S01]is just how big of a difference there[430.66] [430.66][S01]is between the people who understand it,[432.55] [432.55][S01]that 1.8 million people, and everybody else.[436.43] [436.43][S01]Is this something that you notice as well?[439.27] [439.27][S01]AlphaFold doesn't-- it feels a bit technical to really[443.08] [443.08][S01]understand the magnitude of it.[444.92] [444.92][S01]What are the applications that you are excited[447.34] [447.34][S01]about that other people will understand?[449.74] [449.74][S02] So I think, for a large section[455.14] [455.14][S02]of the community, they see AlphaFold as an AI breakthrough[459.28] [459.28][S02]in the sciences, but the scientists[462.93] [462.93][S02]who work on this topic understand the implications[467.17] [467.17][S02]of it in a very deep way.[469.03] [469.03][S02]They know the implications of AlphaFold[472.99] [472.99][S02]for extremely important things like drug discovery,[476.99] [476.99][S02]designing a new vaccine, thinking[478.72] [478.72][S02]about antimicrobial resistance, thinking about new enzymes[482.02] [482.02][S02]to decompose plastics.[483.61] [483.61][S02]I can just go on and on and on.[486.47] [486.47][S01] But that's not the only project[489.2] [489.2][S01]that has been going on in here since we last spoke to you?[493.545] [493.545][S01]So tell us a little bit more about what[495.17] [495.17][S01]you've been working on.[496.25] [496.25][S02] So we have been working[498.11] [498.11][S02]on a whole spectrum of different topics,[500.25] [500.25][S02]from material science to fusion, to working[505.97] [505.97][S02]on new discoveries in computer science,[509.07] [509.07][S02]in mathematics, weather prediction, meteorology.[512.98] [512.98][S02]So there's a whole spectrum of areas that we are looking at.[517.03] [517.03][S01] And of those, where do you[519.14] [519.14][S01]think there's been really significant progress[521.48] [521.48][S01]in the last couple of years?[523.64] [523.64][S02] So one of the moments that we can talk about[526.55] [526.55][S02]is our work on weather prediction.[529.67] [529.67][S02]Weather and climate is something that we are all[533.18] [533.18][S02]thinking about at the moment, but if you[536.99] [536.99][S02]look at what DeepMind had done earlier,[539.97] [539.97][S02]we had our nowcasting model, which[541.88] [541.88][S02]was able to make very good prediction but at very[544.34] [544.34][S02]short timescales.[546.59] [546.59][S02]With our new model called GraphCast,[551.16] [551.16][S02]which we released last year, we can now[554.18] [554.18][S02]look at the problem of 10 day forecasts.[557.3] [557.3][S02]And what we have shown that this new model can make these 10 day[563.63] [563.63][S02]predictions more accurately than some of the models that[568.28] [568.28][S02]are being used, the classical models that[570.11] [570.11][S02]are being used by the Met Office.[571.798] [571.798][S01] The supercomputers.[573.09] [573.09][S02] Yes.[573.923] [573.923][S02]Which run on supercomputers for many hours,[576.98] [576.98][S02]and we can outperform them in terms of accuracy[581.72] [581.72][S02]and make predictions in a matter of a minute on a single chip.[585.14] [585.14][S02]I think in some sense, this really[588.35] [588.35][S02]opens up research in this area.[591.96] [591.96][S02]So a lot of other people, a lot of other entities[596.69] [596.69][S02]can now conduct research in weather prediction.[600.66] [600.66][S02]And the results are amazing.[602.732] [602.732][S02]One particular example that was fascinating[606.53] [606.53][S02]is that there was this cyclone, Hurricane Lee, last year, which[611.39] [611.39][S02]made landfall in Nova Scotia.[614.66] [614.66][S02]And our model, GraphCast, was able to make[619.46] [619.46][S02]the prediction of the landfall event nine days earlier--[623.683] [623.683][S01] Wow.[624.35] [624.35][S02] --while the classical models were only[626.6] [626.6][S02]able to do it six days earlier.[628.32] [628.32][S02]So they could give a three day additional heads up.[633.312] [633.312][S01] It's a superpower that doesn't just[635.27] [635.27][S01]apply to biology, then.[636.435] [636.435][S02] Absolutely.[637.56] [637.56][S02]The amount of impact that AI and machine learning[640.67] [640.67][S02]is having across all these disciplines is amazing.[644.94] [644.94][S02]And if you think about it, it feels natural[649.43] [649.43][S02]because, in any of these areas of science,[653.0] [653.0][S02]we are collecting a lot of data, and the complexity of models[657.08] [657.08][S02]that we are playing with and we are working with[659.12] [659.12][S02]is really expanding.[661.26] [661.26][S02]And it's just very natural that a single human mind[666.798] [666.798][S02]has difficulty in comprehending what are[670.36] [670.36][S02]the real patterns in this data.[672.22] [672.22][S02]And machine learning and AI just give you[675.19] [675.19][S02]that ability to figure out what is needed[681.025] [681.025][S02]in many of these problems.[683.56] [683.56][S01] You mentioned material science there.[685.69] [685.69][S01]Frame for us the general problem, as it were,[689.87] [689.87][S01]in material science pre AI?[691.42] [694.09][S02] So what is the problem of material discovery?[698.83] [698.83][S02]You want to discover materials which have certain properties.[701.62] [701.62][S02]We have gone through all these ages, from the Stone Age[706.24] [706.24][S02]to the Iron Age to the Bronze Age, and so on.[708.58] [708.58][S02]And at every age, we are working with a new material,[710.86] [710.86][S02]and new material gives us new abilities.[713.39] [713.39][S02]And the problem of material discovery[715.69] [715.69][S02]is to discover materials which have certain useful properties.[719.83] [719.83][S02]Now, how was that done till now?[723.02] [723.02][S02]It was done in a very experimental sort of way.[728.29] [728.29][S02]People try different things in the lab.[731.37] [731.37][S02]Sometimes the things worked out as expected.[734.41] [734.41][S02]Sometimes, things didn't work out as expected.[736.6] [736.6][S02]But there was no--[737.678] [737.678][S02]there were some rules of thumb.[738.97] [738.97][S02]There were some theory, but we did not[740.61] [740.61][S02]know the extent to what was possible in materials.[745.268] [745.268][S01] Often things get discovered by accident,[747.435] [747.435][S01]like vulcanized rubber or--[749.5] [749.5][S02] Exactly.[750.5] [750.5][S01] Just sort of a chance happening in a lab,[752.708] [752.708][S01]and then a new material that turns out to be really useful[755.125] [755.125][S01]pops up.[755.7] [755.7][S02] Exactly.[756.7] [756.7][S02]Like graphene.[757.875] [757.875][S02]We all know about this story about how[760.95] [760.95][S02]this magical material was isolated by sellotape, by taking[769.38] [769.38][S02]pieces of carbon and then repeatedly making it thinner[773.16] [773.16][S02]and thinner by a sticky tape.[776.38] [776.38][S02]And it has amazing properties.[779.448] [779.448][S01] Ending up being the thinnest substance that[781.74] [781.74][S01]can be manufactured.[783.22] [783.22][S02] Exactly.[783.51] [783.51][S01] A single atom thick.[784.45] [784.45][S02] Yes.[785.283] [785.283][S02]And it has a very interesting properties[787.51] [787.51][S02]in terms of conductivity and so on.[790.01] [790.01][S02]So before AI and before even computational methods,[793.93] [793.93][S02]that essentially was the norm.[795.94] [795.94][S02]That people-- and even today, in some sense,[799.12] [799.12][S02]material science is a very experimental science,[801.94] [801.94][S02]where people are trying to discover new materials[805.27] [805.27][S02]for whether it's for constructing a battery[809.05] [809.05][S02]or for a photovoltaic cell or for a superconductor and so on.[814.36] [814.36][S02]Now, how computational systems are used today[818.71] [818.71][S02]is to then post-hoc rationalize, why[822.07] [822.07][S02]is that material behaving in a way that it's behaving?[826.63] [826.63][S02]But we are very far from the place[829.72] [829.72][S02]where we could rationally design a material.[832.03] [832.03][S02]Given a property, you say, find me the material,[835.77] [835.77][S02]invent a new material which can maximize these properties.[839.535] [839.535][S02]And that's what the problem is for AI.[842.29] [842.29][S02]Can we somehow in silico de novo from scratch[848.74] [848.74][S02]start to invent any given material with some property?[855.25] [855.25][S01] So that you can go in and say,[857.72] [857.72][S01]I want something that is extraordinarily flexible,[860.27] [860.27][S01]extraordinarily light, whatever, easy to mine, something,[864.25] [864.25][S01]something, something.[865.36] [865.36][S01]And then it's just like, here's the physical structure of the--[869.29] [869.29][S01]atomic structure that will result in that material.[872.42] [872.42][S02] Exactly.[873.42] [873.42][S02]That's the vision.[874.657] [874.657][S01] Let's anchor this to an example, then.[876.74] [876.74][S01]So you mentioned batteries there.[879.47] [879.47][S01]We've got batteries.[880.93] [880.93][S01]What's wrong with the batteries we have?[882.76] [882.76][S02] So I think they're doing well.[884.78] [884.78][S02]They're lithium ion batteries.[886.96] [886.96][S02]There are a number of issues with that.[889.31] [889.31][S02]First, they are based on certain resources which are difficult.[893.75] [893.75][S02]For example, the lithium ion batteries, the batteries[898.12] [898.12][S02]used today use cobalt, which is difficult to mine.[903.01] [903.01][S02]We might want to increase the energy[905.08] [905.08][S02]density of these batteries in the future.[907.19] [907.19][S02]We might want to make them more thermally stable.[909.86] [909.86][S02]We might want to make them cheaper.[911.54] [911.54][S02]And if you could somehow have a magic sort of material, which[914.98] [914.98][S02]can have higher energy density, is more stable,[918.53] [918.53][S02]is easy to manufacture, then, of course,[920.8] [920.8][S02]you would want to transition to it.[923.02] [923.02][S01] And find it without just waiting for an accident[926.02] [926.02][S01]to happen in a lab.[926.91] [926.91][S02] Exactly.[927.91] [927.91][S01] Are you sure there is one, though?[929.827] [929.827][S01]How do you know that lithium isn't just the best[931.99] [931.99][S01]that the universe has to offer?[934.55] [934.55][S02] It could be that lithium is the best[936.82] [936.82][S02]and we were extremely lucky, and we found it,[940.31] [940.31][S02]but that seems very remote.[944.03] [944.03][S02]There is, the number of materials--[946.04] [946.04][S02]just to give you an example, there were around 20,000[949.99] [949.99][S02]inorganic materials that people play with.[952.75] [952.75][S02]Now, using computational methods,[956.11] [956.11][S02]28,000 have been found out.[960.8] [960.8][S02]So they are roughly 40,000, 50,000 known materials that were[965.47] [965.47][S02]known to be stable, in the sense that,[967.39] [967.39][S02]at 0 Kelvin or under some theoretical situations,[975.47] [975.47][S02]they will not decompose to other materials.[977.84] [977.84][S01] So if you freeze them[979.215] [979.215][S01]right down to absolute zero, they don't split apart.[981.39] [981.39][S02] Exactly.[982.39] [982.39][S02]So they are stable materials.[984.38] [984.38][S02]So that was what was known in the material science community.[988.93] [988.93][S02]Our new AI model known last year expanded[992.74] [992.74][S02]that set and said there are 2.2 million new inorganic materials[999.31] [999.31][S02]that are stable.[1002.22] [1002.22][S01] From 50,000?[1003.61] [1003.61][S02] Yes.[1004.786] [1004.786][S01] Wow.[1005.453] [1008.08][S01]Wow.[1008.74] [1008.74][S01]That's massive.[1010.813] [1010.813][S02] Right.[1011.73] [1011.73][S02]And you're talking about graphene,[1013.147] [1013.147][S02]there are 52,000 single chain layered materials in that set.[1020.28] [1020.28][S02]So the number of possibilities, the number of things that now we[1023.96] [1023.96][S02]can search over is immense.[1026.919] [1026.919][S02]And you asked the question whether[1030.92] [1030.92][S02]the lithium-cobalt batteries are the optimal batteries, well,[1036.589] [1036.589][S02]there are so many things in that 2.2 million set that one of them[1041.99] [1041.99][S02]could be much, much better.[1043.89] [1043.89][S01] The chances that lithium is actually quite--[1047.262] [1047.262][S01]what's it called?[1047.97] [1047.97][S01]GNNME?[1048.41] [1048.41][S02] Yes.[1048.86] [1048.86][S01] What does it stand for?[1049.92] [1049.92][S02] Graph neural networks[1051.462] [1051.462][S02]for material exploration.[1053.185] [1053.185][S01] And how does it work?[1055.67] [1055.67][S01]How are you deciding new material structures that[1059.27] [1059.27][S01]haven't yet been discovered?[1061.38] [1061.38][S02] So it essentially tries[1065.24] [1065.24][S02]to start with existing structures and compositions.[1070.467] [1070.467][S01] 50,000 that we know.[1071.8] [1071.8][S02] Yeah.[1072.675] [1072.675][S02]And says, OK, let me change some of those,[1075.15] [1075.15][S02]and then, I'll learn a model to say which[1079.95] [1079.95][S02]of those changes, which of those modified materials[1084.21] [1084.21][S02]are stable or not stable.[1086.345] [1086.345][S02]And it is able to do those calculations very efficiently,[1089.562] [1089.562][S02]and that's where the machine learning model comes in.[1091.77] [1091.77][S02]It's able to make predictions about the stability[1096.57] [1096.57][S02]and the free energy of these new compositions[1100.47] [1100.47][S02]and these new crystal structures in a much more accurate way.[1104.89] [1104.89][S01] So in one sense, then,[1106.32] [1106.32][S01]taking the 50,000 and then doing some kind of atomic shuffling,[1110.38] [1110.38][S01]if you like, just trying different combinations,[1112.77] [1112.77][S01]but then, the clever bit, as you're describing,[1114.96] [1114.96][S01]is that you can calculate without actually having made[1118.89] [1118.89][S01]ever this sort of fantasy material which has been[1122.22] [1122.22][S01]constructed by atomic shuffling.[1124.15] [1124.15][S01]You can tell whether or not it will be stable at 0 degree[1128.07] [1128.07][S01]Kelvin if you made it.[1129.247] [1129.247][S02] Yes.[1130.08] [1130.08][S01] How?[1131.15] [1131.15][S02] You can do that theoretically,[1133.17] [1133.17][S02]but you will need to do a lot of very complex calculations.[1136.95] [1136.95][S02]And what this model is able to do[1138.56] [1138.56][S02]is basically approximate those calculations[1141.47] [1141.47][S02]and do it much, much computationally inexpensively.[1144.813] [1144.813][S01] So in many ways, actually, I[1146.48] [1146.48][S01]can see the similarities between this and AlphaFold.[1150.02] [1150.02][S01]You're talking about the atomic structure of something,[1152.6] [1152.6][S01]right down at the level, I mean amino acids and atoms,[1155.585] [1155.585][S01]and then predicting larger structural properties[1159.89] [1159.89][S01]as a result.[1160.988] [1160.988][S02] So in this particular case,[1162.78] [1162.78][S02]you are given a new crystal structure or a new composition,[1166.97] [1166.97][S02]and you are told, is it going to be stable or not?[1169.74] [1169.74][S02]And the model is able to say that.[1171.908] [1171.908][S01] How do you validate it, though?[1173.7] [1173.7][S02] So we have validated these in two ways.[1176.34] [1176.34][S02]One form of validation is by doing calculations.[1180.63] [1180.63][S02]So when I said, these are 2.2 million stable structures,[1184.95] [1184.95][S02]what do I mean?[1185.69] [1185.69][S02]How do I say that these are stable?[1188.155] [1191.16][S02]There are theories in quantum chemistry[1194.64] [1194.64][S02]which give us approaches to figure out whether something[1199.11] [1199.11][S02]is going to be stable.[1200.31] [1200.31][S02]These are very difficult calculations.[1202.15] [1202.15][S02]So we can run those calculations on the predictions[1204.87] [1204.87][S02]that GNNME has made and see whether the theory says[1209.4] [1209.4][S02]that those will be stable.[1211.31] [1211.31][S02]So that's one way.[1212.41] [1212.41][S02]And then we have also taken a subset of the most stable[1218.37] [1218.37][S02]predictions and then experimentally try[1220.35] [1220.35][S02]to validate them in the lab.[1221.617] [1221.617][S01] Like build it?[1222.7] [1222.7][S02] Yeah.[1223.2] [1223.2][S01] Oh, wow.[1224.24] [1224.24][S01]But then, hold on.[1225.3] [1225.3][S01]If it's giving you some fantasy atomic structure,[1230.76] [1230.76][S01]does it tell you how to make it?[1232.27] [1232.27][S02] No, it doesn't.[1233.562] [1233.562][S02]So then we have to figure out how do you make it.[1236.918] [1236.918][S01] And you have managed to do this[1238.71] [1238.71][S01]with a small number of them?[1239.74] [1239.74][S02] Yes.[1240.21] [1240.21][S01] Right.[1241.3] [1241.3][S01]And do they hold up?[1242.53] [1242.53][S01]Are they stable?[1243.37] [1243.37][S02] Yeah.[1243.56] [1243.56][S02]Yeah.[1244.06] [1244.06][S02]And large proportion of them are stable.[1245.743] [1245.743][S01] Give me numbers.[1246.91] [1246.91][S01]Give me numbers.[1248.915] [1248.915][S02] This is constantly changing--[1250.79] [1250.79][S01] Of course.[1251.707] [1251.707][S02] --but the success rates are quite high.[1254.79] [1254.79][S02]Between 70% to 90% of those that we try in the lab[1259.64] [1259.64][S02]end up being stable.[1261.62] [1261.62][S02]And then now once we have those valid materials,[1265.495] [1265.495][S02]then we also have to think about,[1266.87] [1266.87][S02]what are the properties of those materials?[1269.01] [1269.01][S02]Now, which of them would be better batteries[1272.09] [1272.09][S02]or would be useful in photovoltaics[1274.61] [1274.61][S02]or would be good in superconductivity, and so on?[1277.59] [1277.59][S01] So can it also predict the properties, then?[1280.11] [1280.11][S02] So no, not our current generation.[1282.29] [1282.29][S02]It just makes the predictions about the stability[1285.05] [1285.05][S02]of those structures, but we are working[1287.27] [1287.27][S02]on new models for these other types of problems.[1293.23] [1293.23][S01] What kind of materials are you hoping for?[1296.17] [1296.17][S01]Battery materials is one example.[1298.79] [1298.79][S01]What other kind of things are you[1300.22] [1300.22][S01]hoping that GNNME one will help develop?[1302.05] [1302.05][S02] So one of the key things[1304.63] [1304.63][S02]that people have been thinking about[1306.49] [1306.49][S02]is materials that can exhibit superconductivity.[1309.782] [1309.782][S01] By which we mean?[1310.99] [1310.99][S02] Which means, basically,[1312.615] [1312.615][S02]they exhibit zero resistance.[1315.59] [1315.59][S02]And why is that important?[1317.17] [1317.17][S02]That's important because, if you have such a material,[1320.9] [1320.9][S02]you can create very strong magnetic fields.[1323.99] [1323.99][S02]You can store a lot of energy using those materials.[1327.64] [1327.64][S02]Why are magnetic fields important?[1329.87] [1329.87][S02]They're important from everything[1332.71] [1332.71][S02]to do with MRI scanners to creating fusion reactors.[1338.415] [1338.415][S02]The magnets that are used in fusion reactors[1341.65] [1341.65][S02]require very high magnetic fields.[1343.97] [1343.97][S02]So superconductivity is an extremely important property.[1347.703] [1347.703][S01] And at the moment, you[1349.12] [1349.12][S01]have to super cool things in order[1350.71] [1350.71][S01]to be able to get up to those high levels?[1352.483] [1352.483][S01]Because otherwise, it just gets too hot.[1354.15] [1354.15][S02] Exactly.[1355.15] [1355.15][S02]And so what is the Holy Grail in material science[1358.61] [1358.61][S02]is to discover a room temperature superconductor.[1362.31] [1362.31][S02]And there have been many attempts at it[1364.73] [1364.73][S02]and some false starts, but--[1367.55] [1367.55][S01] And some quite cheeky pretending.[1369.66] [1369.66][S02] Yes.[1369.92] [1369.92][S01] This happen too.[1370.79] [1370.79][S02] Yeah, that too.[1372.082] [1372.082][S02]But I think one day, if we discover it,[1376.2] [1376.2][S02]it will be transformational.[1377.67] [1377.67][S01] And do you think AI will[1379.4] [1379.4][S01]be involved in that discovery?[1380.76] [1380.76][S02] Yeah, I'm absolutely[1382.26] [1382.26][S02]sure that AI will aid in the search for superconductors,[1388.18] [1388.18][S02]definitely.[1388.68] [1388.68][S01] At the very least, accelerate the process.[1390.66] [1390.66][S02] Exactly.[1391.92] [1391.92][S01] So that part of making it, though.[1395.805] [1395.805][S01]the way that you described it was like, yeah,[1397.68] [1397.68][S01]we've got to figure out how to make it.[1398.8] [1398.8][S01]So it feels like we're skipping over quite a bit tricky part[1401.73] [1401.73][S01]of the process.[1402.61] [1402.61][S01]Is there any way that you can automate that bit too?[1405.04] [1405.04][S02] Yeah.[1405.915] [1405.915][S02]Once you know what the material is, then it's a whole--[1410.76] [1410.76][S02]there is another sort of stream of material science work, which[1414.54] [1414.54][S02]is, how do you make something cheaply?[1418.605] [1418.605][S02]And there is a lot of work happening in that area as well,[1422.88] [1422.88][S02]but at the moment, our team is looking more[1425.97] [1425.97][S02]on the discovery side.[1427.72] [1427.72][S02]And just to say, we have to be also be[1431.31] [1431.31][S02]cautious in saying that a lot of it--[1434.89] [1434.89][S02]yes, there are these 2.2 million materials, but how many of them,[1439.45] [1439.45][S02]as you said, are valid materials,[1442.03] [1442.03][S02]which actually will prove out when synthesized?[1445.2] [1445.2][S02]How many of them are useful and will have interesting material[1449.7] [1449.7][S02]properties?[1450.22] [1450.22][S02]All of that are open questions.[1454.746] [1454.746][S02]But I think the fact that there is this new place that we[1460.05] [1460.05][S02]find ourselves in in that--[1461.94] [1461.94][S02]from 20,000 to 2.2 million, it's a different ball game.[1466.84] [1466.84][S01] Absolutely.[1468.43] [1468.43][S01]Here's another open question.[1469.68] [1469.68][S01]How long do you think it will be before the first AI developed--[1475.83] [1475.83][S01]or material developed with the assistance of AI[1478.65] [1478.65][S01]makes a big difference?[1480.63] [1480.63][S02] So I think, unlike biology, materials[1484.95] [1484.95][S02]science is much more experimental at the moment.[1488.67] [1488.67][S02]Drug discovery is still--[1490.2] [1490.2][S02]and biology still has used computational methods,[1494.92] [1494.92][S02]but material science is much more experimental.[1498.73] [1498.73][S02]So it will take time, but I think in the next 5 to 10 years,[1502.51] [1502.51][S02]we will see really, as I say, in silico material design[1509.81] [1509.81][S02]starting to make a big impact.[1511.06] [1511.06][S01] So I just want to zoom out a little bit further,[1513.56] [1513.56][S01]because--[1514.16] [1514.16][S01]so you're sitting here talking to us[1516.16] [1516.16][S01]very confidently about material science but also about biology.[1518.99] [1518.99][S01]And you're a computer scientist.[1521.17] [1521.17][S01]So how come you got into this?[1524.835] [1524.835][S02] The last time I studied science was in school,[1527.42] [1527.42][S02]and I came to DeepMind hoping to work[1530.23] [1530.23][S02]on general intelligent systems and make them safe and reliable.[1535.9] [1535.9][S02]But I was always fascinated about,[1538.75] [1538.75][S02]how can intelligent systems have real-world impact?[1541.97] [1541.97][S02]And that's what I used to keep on having discussions with Demis[1548.14] [1548.14][S02]with.[1548.69] [1548.69][S02]And one day he comes to my office[1550.21] [1550.21][S02]and says, oh, Pushmeet, I think we have a good role for you.[1553.49] [1553.49][S02]And I said, OK, what the role?[1555.64] [1555.64][S02]And I was assuming some product area[1561.4] [1561.4][S02]that I'm going to finally use machine learning for.[1564.72] [1564.72][S02]And he said, we want to start a science team,[1566.99] [1566.99][S02]and you should lead it.[1569.09] [1569.09][S02]And I said, why me?[1571.25] [1571.25][S02]Science, I have no background.[1573.06] [1573.06][S02]I studied science in school.[1575.01] [1575.01][S02]I loved it at school, but then I became a computer scientist,[1578.3] [1578.3][S02]and then that's what I did.[1580.11] [1580.11][S02]But he said, well, see you're into multidisciplinary research.[1583.8] [1583.8][S02]You want to understand the problems[1585.29] [1585.29][S02]and you want to have real-world impact.[1587.01] [1587.01][S02]And what better place to have impact[1589.25] [1589.25][S02]in pushing the boundaries of knowledge[1591.86] [1591.86][S02]forward in the natural sciences?[1594.29] [1594.29][S02]So I said, OK.[1595.97] [1595.97][S02]I agreed.[1597.15] [1597.15][S02]And then over the next couple of months,[1600.84] [1600.84][S02]I found what I've got myself into,[1603.75] [1603.75][S02]because, if somebody was designing a job[1606.41] [1606.41][S02]to experience imposter syndrome, this is the job.[1609.96] [1609.96][S02]Because you will be working with Nobel Prize winners[1613.413] [1613.413][S02]and trying to understand those problems,[1615.08] [1615.08][S02]and you don't have no clue.[1616.82] [1616.82][S02]And you're trying to make a difference,[1619.13] [1619.13][S02]and you start from Wikipedia.[1622.88] [1622.88][S01] You start from Wikipedia?[1624.63] [1624.63][S02] Exactly.[1625.63] [1625.63][S02]You start from Wikipedia and try to get a crash[1628.43] [1628.43][S02]course in the area, but--[1630.32] [1630.32][S02]I mean the scientific community is amazing.[1635.3] [1635.3][S02]People are there who are willing to sit down with you,[1640.19] [1640.19][S02]are willing to teach you or go on that journey with you.[1644.818] [1644.818][S01] I guess the really key point, though,[1646.86] [1646.86][S01]is that you really, really, really understand[1649.43] [1649.43][S01]the heart of these algorithms.[1651.0] [1651.0][S01]You really understand the type of systems[1653.0] [1653.0][S01]that they can perform well in and where they can really excel.[1656.73] [1656.73][S01]And so in some ways, the fact that you're[1659.72] [1659.72][S01]talking about lots of different areas,[1661.43] [1661.43][S01]meteorology on the one hand, material science on another,[1663.84] [1663.84][S01]then biology, actually, I can see[1666.23] [1666.23][S01]how there are these, behind the scenes,[1668.76] [1668.76][S01]these actual technical similarities between the systems[1672.53] [1672.53][S01]that you're designing.[1673.53] [1673.53][S01]So in many ways, being a generalist,[1675.84] [1675.84][S01]it requires a generalist to be able to do this kind of stuff?[1679.01] [1679.01][S02] Yeah, absolutely.[1680.385] [1680.385][S02]If you think about intelligence, what is intelligence?[1684.63] [1684.63][S02]How do you create intelligence?[1686.04] [1686.04][S02]Intelligence does not happen in a vacuum.[1687.96] [1687.96][S02]Suddenly, you have a system which gets the intuitions[1692.63] [1692.63][S02]and is able to make these new discoveries.[1694.95] [1694.95][S02]It happens through experience.[1697.43] [1697.43][S02]And what form of experience?[1700.38] [1700.38][S02]How rich is your experience?[1701.88] [1701.88][S02]And this holds true for humans as well.[1704.81] [1704.81][S02]The amount of things that we see in our life,[1708.21] [1708.21][S02]it gives us a different perspective.[1710.37] [1710.37][S02]And so, with these models, the richness of the data[1714.92] [1714.92][S02]that they see, whether that data is collected[1717.86] [1717.86][S02]by a scientist in a lab or whether that data is generated[1722.0] [1722.0][S02]by simulations, if it is rich enough--[1725.01] [1725.01][S02]you want the data to be rich enough so[1726.62] [1726.62][S02]that the model is forced to learn general concepts[1731.51] [1731.51][S02]to explain that data.[1733.35] [1733.35][S02]And then the other sort of element[1734.87] [1734.87][S02]is to have good evaluation metrics.[1739.07] [1739.07][S02]Machine learning methods are very good at cheating.[1743.78] [1743.78][S02]They are amazing at memorization.[1746.58] [1746.58][S02]They are like the student that can memorize anything.[1749.66] [1749.66][S02]So if you just give them a lot of data, they will memorize it.[1752.64] [1752.64][S02]And you ask them any question from the-- which is close[1755.72] [1755.72][S02]to the training set, and they will say, ah, yeah,[1758.495] [1758.495][S02]this is the answer, but they actually don't--[1760.37] [1760.37][S01] Understand what's going on.[1761.52] [1761.52][S02] Exactly.[1762.17] [1762.17][S01] I guess that must have[1762.77] [1762.77][S01]happened quite a lot with proteins[1764.187] [1764.187][S01]at some point or another.[1765.458] [1765.458][S01]They've just seen they've peaked at the back of the textbook.[1768.0] [1768.0][S02] Exactly.[1769.0] [1769.0][S02]And so what we had to do is make sure[1771.74] [1771.74][S02]that we made that test extremely hard,[1775.65] [1775.65][S02]made it impossible to cheat.[1777.63] [1777.63][S02]And the true test of a good machine learning system[1782.27] [1782.27][S02]is its ability to generalize, to unseen examples, to have those--[1789.43] [1789.43][S02]to make those predictions in things[1791.24] [1791.24][S02]which it had never seen before.[1795.29] [1795.29][S02]So understanding what is the right data[1797.87] [1797.87][S02]and what is the right evaluation metric[1800.27] [1800.27][S02]for a different scientific discipline[1804.13] [1804.13][S02]is one part of the problem.[1806.39] [1806.39][S02]It was one of the things that I needed to work with our teams.[1810.35] [1810.35][S02]And then the other part of the solution[1814.99] [1814.99][S02]is, in the design of these machine[1818.23] [1818.23][S02]learning models and these AI models,[1819.95] [1819.95][S02]you want to learn from data, but you also[1821.68] [1821.68][S02]want to bake in any information that you[1824.17] [1824.17][S02]can about the problem into the design of the model itself.[1828.25] [1828.25][S02]And that is where domain knowledge makes a lot of sense.[1833.32] [1833.32][S02]You don't want to start from scratch if you don't have to.[1835.74] [1835.74][S02]So in all our projects, we started[1838.96] [1838.96][S02]with a very multidisciplinary viewpoint.[1841.88] [1841.88][S02]We got in experts, we asked them what was working,[1845.48] [1845.48][S02]what was not working, and tried to inject everything[1849.22] [1849.22][S02]that was currently known into the design of the model itself.[1853.222] [1853.222][S01] But it's noticeable that the common theme[1855.43] [1855.43][S01]of everything that you've described there[1857.138] [1857.138][S01]is where there is a version of success.[1861.19] [1861.19][S01]There's a right answer that you are looking for,[1863.41] [1863.41][S01]like the true underlying structure of the protein,[1866.79] [1866.79][S01]whether it's truly stable or not,[1868.41] [1868.41][S01]whether you really are predicting the weather.[1870.47] [1870.47][S01]Is that absolutely crucial, then, for any of these projects?[1873.885] [1873.885][S02] Absolutely.[1875.01] [1875.01][S02]You really need to have a good sense[1877.19] [1877.19][S02]of what does success look like.[1880.05] [1880.05][S01] So then when it comes to choosing projects--[1882.65] [1882.65][S01]because, I mean, if you are, as you say, the generalist who[1885.95] [1885.95][S01]can, with domain-specific knowledge,[1889.28] [1889.28][S01]adapt these algorithms to anything,[1892.922] [1892.922][S01]how do you decide what's worth your time?[1897.297] [1897.297][S02] That's a very good question.[1899.13] [1899.13][S02]And there are a number of dimensions that we look at.[1903.63] [1903.63][S02]One dimension is the data.[1906.96] [1906.96][S02]If there is no data and there is no experience,[1911.33] [1911.33][S02]the machine learning model will not learn by itself.[1913.986] [1913.986][S02]So that's one of the important considerations.[1917.94] [1917.94][S02]The second thing-- and this is something[1920.93] [1920.93][S02]that we have also always sort of considered from the very start--[1924.41] [1924.41][S02]is a focus on root node problems,[1927.95] [1927.95][S02]problems that are so fundamental that once you unlock them,[1931.77] [1931.77][S02]they have implications for a number of different applications[1934.67] [1934.67][S02]downstream.[1935.25] [1935.25][S02]And protein folding, as I said, had implications[1938.66] [1938.66][S02]for drug discovery to design new enzymes[1943.85] [1943.85][S02]for plastic decomposition and so on.[1945.63] [1945.63][S02]And the same thing is true for materials.[1948.29] [1948.29][S02]If you can understand material discovery[1951.941] [1951.941][S02]in a transformational way, that has[1953.87] [1953.87][S02]so many different applications.[1955.162] [1955.162][S01] It's much more effective[1956.662] [1956.662][S01]than only looking at, say, one particular drug for Alzheimer's[1959.48] [1959.48][S01]or whatever it might be.[1962.09] [1962.09][S01]So then, do you have a list?[1964.86] [1964.86][S01]Is there a whiteboard somewhere in this building[1967.37] [1967.37][S01]where you have a list of potential projects[1970.94] [1970.94][S01]and then you're evaluating them on that basis?[1973.08] [1973.08][S01]Which ones have got good data?[1974.4] [1974.4][S01]Which ones are more important problems than others?[1978.225] [1978.225][S02] Absolutely.[1979.35] [1979.35][S02]We are constantly doing that.[1982.64] [1982.64][S02]And there is a lot of uncertainty[1986.45] [1986.45][S02]that we have to deal with.[1988.28] [1988.28][S02]So we approach the problem in a scientific manner.[1992.22] [1992.22][S02]We conduct experiments.[1994.41] [1994.41][S02]We do exploration studies and see, oh, yes, we[1997.67] [1997.67][S02]are making progress.[1999.06] [1999.06][S02]Some of our assumptions were correct.[2001.67] [2001.67][S02]So that means that we can now make that big conviction[2007.57] [2007.57][S02]commitment to that problem.[2009.61] [2009.61][S02]And the other difference in how we are operating is--[2014.11] [2014.11][S02]most of the topics that I've mentioned to you--[2016.28] [2016.28][S02]we have very focused teams that pursue these problems.[2020.41] [2020.41][S02]And they are not operating at the level[2022.84] [2022.84][S02]of six months or 12 months.[2026.0] [2026.0][S02]They are operating at the level of many years.[2028.55] [2028.55][S02]They are dedicating researchers and engineers[2033.376] [2033.376][S02]in those teams, are dedicating their whole careers[2036.61] [2036.61][S02]to that topic.[2038.57] [2038.57][S02]So we take that responsibility very seriously,[2041.2] [2041.2][S02]as to which are the right problems that we[2043.75] [2043.75][S02]should be working on.[2045.52] [2045.52][S02]Like if we believe that a problem can[2049.81] [2049.81][S02]be done at Oxford or MIT or Harvard or Berkeley or Imperial,[2055.28] [2055.28][S02]we wouldn't want to work on it, because we[2058.389] [2058.389][S02]want to work on the problems that really require[2064.969] [2064.969][S02]a scale and the type of multidisciplinary team[2069.19] [2069.19][S02]that we have to come together to solve them.[2072.21] [2072.21][S01] Well, let's pick another topic then,[2074.21] [2074.21][S01]because I think the big thing that the world's got very[2076.389] [2076.389][S01]excited about in the last couple of years[2078.097] [2078.097][S01]is generative AI and in particular large language[2081.19] [2081.19][S01]models.[2081.94] [2081.94][S01]Have you started to incorporate large language models[2084.61] [2084.61][S01]into your research for science?[2086.675] [2086.675][S02] Yeah.[2087.55] [2087.55][S02]So we are looking at it across the board.[2090.83] [2090.83][S02]I mean, there are two main themes that we are exploring.[2095.06] [2095.06][S02]One is that, till now, most of the work that we have[2101.08] [2101.08][S02]been doing in the scientific areas,[2105.32] [2105.32][S02]we were using structured data.[2107.65] [2107.65][S02]So this is data, whether it's genomic data or protein[2110.62] [2110.62][S02]structure prediction data, where you[2112.27] [2112.27][S02]have sequences and structures--[2113.642] [2113.642][S01] They're directly connected to each other.[2115.85] [2115.85][S02] Exactly.[2116.85] [2116.85][S02]The tables, right?[2118.33] [2118.33][S02]But there's a lot of scientific intuition and knowledge[2121.94] [2121.94][S02]which is embedded in scientific publications in free form text.[2127.15] [2127.15][S02]And how do you learn from that experience,[2131.71] [2131.71][S02]from the diary entries of key scientists[2138.01] [2138.01][S02]that have worked in that area over the last many centuries?[2142.79] [2142.79][S02]So large language models give us the ability[2146.17] [2146.17][S02]to now ingest all that data and to learn from all of it.[2150.65] [2150.65][S02]So that's one key idea.[2153.81] [2153.81][S02]The other way is basically where you[2155.92] [2155.92][S02]use the large language model to generate answers[2160.15] [2160.15][S02]in certain domains.[2161.72] [2161.72][S02]And one example of that is our project[2165.47] [2165.47][S02]on algorithmic discovery, aptly named FunSearch, which[2172.28] [2172.28][S02]stands for function search.[2174.98] [2174.98][S01] Is much better named FunSearch.[2177.21] [2177.21][S01]I'm much prefer that as well.[2179.57] [2179.57][S02] I think our team was very proud of--[2182.36] [2182.36][S01] Of that.[2183.32] [2183.32][S02] --of the name.[2184.85] [2184.85][S01] The fun team.[2185.45] [2185.45][S02] Yeah, exactly.[2186.7] [2186.7][S02]The fun team working on FunSearch.[2189.32] [2189.32][S02]They are trying to discover new algorithms[2191.39] [2191.39][S02]for important problems in computer science.[2193.73] [2193.73][S02]And the model is asked, here is a problem, an important problem[2199.94] [2199.94][S02]in computer science, whether it's[2202.95] [2202.95][S02]a conceptual problem like the bin packing problem,[2208.13] [2208.13][S02]given a set of boxes and some items,[2210.63] [2210.63][S02]how do you compactly pack those items in those boxes?[2214.32] [2214.32][S02]And you might think, well, I do it at home.[2217.01] [2217.01][S02]What's the relevance of it?[2218.3] [2218.3][S02]That conceptual problem is everywhere.[2221.45] [2221.45][S02]It's in how delivery companies go and deliver your groceries[2227.78] [2227.78][S02]to your homes or how cloud providers schedule[2236.12] [2236.12][S02]computational jobs on different computers that they have.[2239.61] [2239.61][S02]So it's a very, very important problem in the real world.[2244.52] [2244.52][S02]And so what this model does, it then[2248.63] [2248.63][S02]tries to make new-- propose new algorithms.[2253.14] [2253.14][S02]And many of those algorithms are maybe well known algorithms[2257.42] [2257.42][S02]that it had seen before.[2259.8] [2259.8][S02]And we tell it, well, that's fine,[2262.44] [2262.44][S02]but try to improve this particular part[2264.2] [2264.2][S02]and try to refine it.[2265.86] [2265.86][S02]And it goes on trying to refine it.[2269.743] [2269.743][S02]And sometimes, it makes a mistake,[2271.16] [2271.16][S02]and we tell it, oh, here's the mistake that you've made.[2273.493] [2273.493][S02]And it keeps on doing that.[2275.7] [2275.7][S02]We try to get the right solutions and feed it back.[2280.67] [2280.67][S02]And in that process, sometimes, it[2283.22] [2283.22][S02]discovers something completely new that was not known before[2287.3] [2287.3][S02]and ends up improving the performance[2291.38] [2291.38][S02]or coming up with a new heuristic or a new algorithm,[2294.69] [2294.69][S02]which solves the problem in a remarkably different and much[2297.95] [2297.95][S02]more efficient manner.[2299.162] [2299.162][S01] So the reason it's able to do that, then,[2301.37] [2301.37][S01]is because there's this deeper network of connections[2306.08] [2306.08][S01]between knowledge that isn't necessarily visible to us.[2308.88] [2308.88][S01]So is that why sometimes large language[2311.96] [2311.96][S01]models will hallucinate in a way that sort of makes sense?[2316.29] [2316.29][S01]So I don't know.[2316.97] [2316.97][S01]Maybe one would say Marie Curie invented penicillin.[2321.92] [2321.92][S01]And of course, she didn't, but she did come up[2324.23] [2324.23][S01]with a discovery that was really important around the same time.[2327.15] [2327.15][S01]So it can swap over pieces of information,[2330.36] [2330.36][S01]as it were, because it's seeing deeper connections that[2333.38] [2333.38][S01]aren't visible to us.[2334.535] [2334.535][S02] Absolutely.[2335.66] [2335.66][S02]It is basically, because it's operating in that latent space,[2339.24] [2339.24][S02]it feels that there are certain things which[2342.02] [2342.02][S02]are related to each other.[2343.8] [2343.8][S02]And in the case of the Marie Curie example,[2347.2] [2347.2][S02]that was a mistake, and that's a problematic mistake[2350.06] [2350.06][S02]if somebody did not know about it.[2353.001] [2353.001][S02]It's incorrect information, but in our case, that's[2358.52] [2358.52][S02]a fine mistake to have, because what[2361.05] [2361.05][S02]we have is the evaluation function, which[2363.84] [2363.84][S02]is coupled with FunSearch, which is[2365.94] [2365.94][S02]coupled with the large language model, which can call bullshit.[2369.963] [2369.963][S01] Like a truth detector.[2371.38] [2371.38][S02] Yeah, exactly.[2372.63] [2372.63][S02]Which can call bullshit and say, oh, this doesn't make sense,[2375.7] [2375.7][S02]very, very quickly.[2376.93] [2376.93][S02]So then, what matters is creativity.[2381.57] [2381.57][S02]So if it comes up with something completely creative,[2385.03] [2385.03][S02]we pick that apart and say, oh, yeah, great.[2387.82] [2387.82][S02]And the bad stuff that it says, we[2390.84] [2390.84][S02]are able to filter it out very easily.[2392.95] [2392.95][S01] But every now and then, the creative stuff it says[2398.57] [2398.57][S01]is revealing something more about the underlying[2404.13] [2404.13][S01]network of knowledge.[2405.335] [2405.335][S02] Exactly.[2406.42] [2406.42][S01] It's basically like you've harnessed hallucinations[2408.0] [2408.0][S01]for good.[2408.61] [2408.61][S02] Yes.[2409.62] [2409.62][S02]Hallucinations are good in this particular case[2411.58] [2411.58][S02]if you can somehow leverage creativity[2413.97] [2413.97][S02]and keep the good part and filter out all the invalid part.[2419.53] [2419.53][S01] So don't let the fun team know that.[2423.54] [2423.54][S01]They can call themselves the hallucinations for good team.[2426.66] [2426.66][S01]No, might get ahead of themselves.[2430.05] [2430.05][S01]So is it finding stuff that wasn't known to computer[2432.98] [2432.98][S01]scientists and mathematicians?[2434.23] [2434.23][S02] Yes, absolutely.[2435.563] [2435.563][S02]So in fact, with this method, we were[2441.3] [2441.3][S02]able to get a new result in computer science.[2447.7] [2447.7][S02]So there is a very interesting problem called the cap set[2450.36] [2450.36][S02]problem, which is trying to find independent sets[2453.87] [2453.87][S02]in a structured graph.[2456.58] [2456.58][S02]It's like a computational problem where you have a graph[2459.48] [2459.48][S02]and you are trying to find certain sort of nodes and edges,[2464.19] [2464.19][S02]which certain properties.[2465.73] [2465.73][S02]And it has been studied for a long time.[2469.21] [2469.21][S02]It's a very interesting problem in computer science.[2472.27] [2472.27][S02]And what FunSearch was able to do is was, for the first time,[2476.64] [2476.64][S02]produce a result which nobody had been able to produce.[2480.6] [2480.6][S02]And not only was it able to produce that result,[2484.92] [2484.92][S02]but the algorithm, the program that it had generated[2490.64] [2490.64][S02]to find that result had extremely[2493.43] [2493.43][S02]interesting substructure and intuitions[2496.25] [2496.25][S02]that, when mathematicians, who were working with us saw it,[2501.512] [2501.512][S02]they felt that the program had extracted[2504.65] [2504.65][S02]a new symmetry in the problem.[2506.996] [2506.996][S01] Oh.[2507.77] [2507.77][S01]So in this hallucination, it had really[2511.58] [2511.58][S01]stumbled upon something that had been unexplored[2514.69] [2514.69][S01]but turned out to be true?[2516.18] [2516.18][S02] Yeah.[2517.285] [2517.285][S02]And it was leveraging some very interesting property[2521.3] [2521.3][S02]about the problem that we had not told it about.[2524.93] [2524.93][S01] Wow.[2526.95] [2526.95][S01]I can see how that might be useful.[2529.73] [2529.73][S01]Tell me about the Olympiad.[2531.59] [2531.59][S02] So there is the International Maths Olympiad,[2536.16] [2536.16][S02]which is a competition that students take part in.[2542.85] [2542.85][S02]Some of the best students around the world[2545.58] [2545.58][S02]come to this competition.[2547.18] [2547.18][S02]It has very hard math problems.[2550.51] [2550.51][S02]And if you perform well, then you[2554.49] [2554.49][S02]can get a bronze, silver, or gold medal.[2560.355] [2560.355][S02]The level of problems that are asked at the International Maths[2565.08] [2565.08][S02]Olympiad are extremely hard.[2567.04] [2567.04][S01] Yes, they are.[2568.17] [2568.17][S02] The current generation of AI systems,[2570.67] [2570.67][S02]for example, even, they are not able to tackle those problems,[2577.44] [2577.44][S02]but--[2578.07] [2578.07][S01] They require an incredible level[2579.903] [2579.903][S01]of lateral thinking, of logic, of deep understanding[2583.038] [2583.038][S01]of mathematical concepts, despite the fact that they're[2585.33] [2585.33][S01]aimed at schoolchildren.[2587.02] [2587.02][S02] Absolutely.[2590.52] [2590.52][S02]They're not your normal--[2592.14] [2592.14][S01] They are not your normal kids.[2593.89] [2593.89][S02] Yes, exactly.[2594.52] [2594.52][S01] They're not the Demis's of the world.[2595.68] [2595.68][S01]You know what I mean?[2596.35] [2596.35][S02] Exactly.[2596.95] [2596.95][S02]Exactly.[2597.45] [2597.45][S02]I'm sure.[2598.047] [2598.047][S02]I don't know whether Demis has participated in one but--[2600.38] [2600.38][S01] I bet he did.[2602.18] [2602.18][S02] --but they are exceptional.[2604.61] [2604.61][S02]And it has been a long-standing challenge[2608.3] [2608.3][S02]for a computational system to be able to solve[2611.96] [2611.96][S02]any of these problems, because these are extremely[2614.33] [2614.33][S02]hard mathematics problems.[2616.71] [2616.71][S02]And mathematics, unlike the game of chess or Go,[2621.57] [2621.57][S02]is an open-ended environment in the sense[2625.64] [2625.64][S02]that it does not have a specific number of moves.[2629.6] [2629.6][S02]The moves are infinite.[2631.465] [2631.465][S01] And in any direction.[2632.84] [2632.84][S02] And in any direction.[2634.382] [2634.382][S02]So it's an extremely large space that you are reasoning over.[2638.55] [2638.55][S02]And so the more sophisticated AI systems[2642.44] [2642.44][S02]are not able to tackle any of these sort of problems.[2646.74] [2646.74][S02]So we had a system called AlphaGeometry,[2652.11] [2652.11][S02]which, for the first time, showed that an AI system can[2659.0] [2659.0][S02]solve geometry problems that are at the International Math[2664.61] [2664.61][S02]Olympiad level.[2665.96] [2665.96][S01] So this is like--[2667.4] [2667.4][S01]I mean, how can we describe them?[2669.45] [2669.45][S01]So you get a picture of some circles and triangles[2673.205] [2673.205][S01]and squares and things, and then you'll[2675.47] [2675.47][S01]be asked something about--[2677.88] [2677.88][S01]I don't know-- how to calculate some seemingly impossible thing[2682.01] [2682.01][S01]from this image based only on what[2683.852] [2683.852][S01]you understand about the rules of circles and squares[2686.06] [2686.06][S01]and triangles.[2686.85] [2686.85][S02] Yes, something like that.[2689.01] [2689.01][S01] It's really hard.[2690.59] [2690.59][S02] It's quite hard.[2691.923] [2691.923][S02]And you really have to understand not only geometry[2695.3] [2695.3][S02]very well, but you also need to be able to plan ahead[2699.65] [2699.65][S02]and think about what kind of solution can make sense,[2705.32] [2705.32][S02]can lead you to the final answer.[2706.888] [2706.888][S01] So how does AlphaGeometry work, then?[2708.93] [2708.93][S02] So AlphaGeometry, how it works,[2710.888] [2710.888][S02]it transforms this problem, which is given sometimes[2717.26] [2717.26][S02]in text, sometimes in image, into a formal language,[2721.06] [2721.06][S02]into its own domain-specific language in which it can reason[2725.1] [2725.1][S02]about the problem, and then it tries[2729.09] [2729.09][S02]to solve the problem in that language.[2731.47] [2731.47][S02]Now, what the AlphaGeometry did was very smart.[2735.88] [2735.88][S02]They generated a very, very large number of problems,[2741.01] [2741.01][S02]synthetically generated problems,[2743.1] [2743.1][S02]in that language and the corresponding solutions.[2747.7] [2747.7][S02]So with this, they could train the machine learning model[2751.17] [2751.17][S02]on the hundreds of thousands of these problems.[2753.82] [2753.82][S02]And then that made it extremely effective[2757.23] [2757.23][S02]in, given a new problem of that form, it could solve it.[2760.527] [2760.527][S01] It just knows from its own experience.[2762.61] [2762.61][S02] Yes.[2763.21] [2763.21][S01] Amazing.[2764.043] [2764.043][S01]And actually, I can see the similarities there,[2766.002] [2766.002][S01]again, between that and the material science thing[2768.09] [2768.09][S01]is like, OK, let's just use combinatorics.[2770.94] [2770.94][S02] Exactly.[2771.94] [2771.94][S01] It's just like, roll the dice,[2773.26] [2773.26][S01]come up with loads of random things in the beginning,[2774.99] [2774.99][S01]and then narrow it down and narrow it down, narrow it down[2776.88] [2776.88][S01]until eventually it can look at a single problem and say,[2779.05] [2779.05][S01]I know how to solve it.[2780.008] [2780.008][S02] Exactly.[2781.515] [2781.515][S01] I'm quite jealous of your job.[2783.265] [2786.39][S01]We have gone on a wild list of topics here,[2790.86] [2790.86][S01]but what are you hoping to tackle next?[2793.538] [2793.538][S02] So I think there are many problems that[2795.83] [2795.83][S02]remain unsolved.[2798.68] [2798.68][S02]In any area that we're working on,[2800.73] [2800.73][S02]whether it's understanding proteins,[2804.93] [2804.93][S02]whether it's understanding the genome, whether it's[2807.47] [2807.47][S02]understanding the weather, whether it's understanding[2810.47] [2810.47][S02]a material science, there is just so much work[2813.11] [2813.11][S02]that still remains to be done.[2815.16] [2815.16][S02]We were talking about materials.[2816.6] [2816.6][S02]We are only making predictions about stability[2819.38] [2819.38][S02]of these materials, but how do you[2821.36] [2821.36][S02]extend that to making the prediction about properties[2826.76] [2826.76][S02]of these materials and then synthesizing them?[2828.867] [2828.867][S01] Have you got a favorite, though?[2830.7] [2830.7][S01]Is there one that you're like, that's the one I really want?[2833.85] [2833.85][S02] No, you can't ask me that question.[2835.02] [2835.02][S01] I can.[2835.77] [2835.77][S01]I did.[2836.858] [2836.858][S02] No.[2837.65] [2837.65][S02]I think all--[2839.58] [2839.58][S01] Not all your babies are equal.[2841.62] [2841.62][S01]You got to pick a favorite.[2843.282] [2843.282][S02] I'm a computer scientist,[2844.99] [2844.99][S02]I love the computer science work, but--[2848.99] [2848.99][S02]all of them have different elements which are on the day,[2852.36] [2852.36][S02]they are like, wow, this is so amazing.[2854.88] [2854.88][S02]When you are trying to learn a policy that is going to control[2860.6] [2860.6][S02]the magnets of a fusion reactor, the day that that experiment[2864.41] [2864.41][S02]happens, regardless of what's happening, I'm--[2866.808] [2866.808][S01] You care most about that.[2868.35] [2868.35][S02] Yeah, exactly.[2869.4] [2869.4][S01] Is there one that you think,[2871.067] [2871.067][S01]though, will have the biggest impact?[2875.45] [2875.45][S02] I think the work that we[2878.3] [2878.3][S02]are doing in understanding biology[2881.57] [2881.57][S02]and in understanding chemistry and materials, I think,[2884.81] [2884.81][S02]it's so fundamental.[2886.35] [2886.35][S01] Such root node problems.[2887.85] [2887.85][S02] Yeah.[2888.03] [2888.03][S02]There are such root node problems[2889.46] [2889.46][S02]that it's difficult to even limit what the impact would be.[2895.19] [2895.19][S02]The effects are difficult to predict.[2899.19] [2899.19][S01] Absolutely.[2900.17] [2900.17][S01]I'm very much looking forward to coming back and talking to you[2902.93] [2902.93][S01]again to see what other massive things have been happening[2906.65] [2906.65][S01]and push me.[2907.152] [2907.152][S01]Thank you very much for joining me.[2908.61] [2908.61][S02] Thank you, Hannah.[2910.027] [2910.027][S01] What I'm struck by after meeting Pushmeet is that,[2913.79] [2913.79][S01]the thing about science, it moves really slowly, and then,[2918.41] [2918.41][S01]every now and then, maybe once in a generation,[2921.51] [2921.51][S01]you get this seismic shift.[2923.79] [2923.79][S01]You get a big step forward that fundamentally[2926.9] [2926.9][S01]changes our understanding.[2928.82] [2928.82][S01]But the thing about this shift of using artificial intelligence[2932.36] [2932.36][S01]for science is that this isn't just[2934.52] [2934.52][S01]going to make a difference to physics or one equation[2937.5] [2937.5][S01]in cosmology.[2938.68] [2938.68][S01]It's that this is going to transform[2942.93] [2942.93][S01]all of science and in the most fundamental way.[2947.38] [2947.38][S01]And OK, I know that I sound like I've just[2950.1] [2950.1][S01]swallowed the hype here, but don't just[2951.96] [2951.96][S01]take my word for this.[2952.96] [2952.96][S01]Go and ask your friendly neighborhood scientist[2956.25] [2956.25][S01]what they think of what is happening here.[2959.02] [2959.02][S01]Because the people who really understand[2961.59] [2961.59][S01]what this means for the world behind the scenes,[2966.21] [2966.21][S01]they don't have words for how big this shift is[2970.26] [2970.26][S01]or is going to be.[2973.43] [2973.43][S01]You have been listening to \"Google DeepMind, The Podcast\"[2976.43] [2976.43][S01]with me, Professor Hannah Fry.[2978.02] [2978.02][S01]If you enjoyed that episode, do subscribe to our YouTube[2980.9] [2980.9][S01]channel, and you can also find us[2982.7] [2982.7][S01]on your podcast platform of choice.[2985.35] [2985.35][S01]Now, we have lots more to explore[2987.05] [2987.05][S01]on this podcast, including a deep dive[2989.18] [2989.18][S01]into how AI could enhance education.[2992.26] [2992.26][S01]Plus, should AI assistants be given human traits?[2996.09] [2996.09][S01]That is coming up soon on \"Google DeepMind, The Podcast.\"[2999.14] [2999.14][MUSIC PLAYING][3002.19]"}
{"file_name": "audio/val_000006.wav", "transcription": "[0.0][S02] Welcome to \"Google DeepMind: The Podcast\" with me,[2.97] [2.97][S02]your host, Professor Hannah Fry.[4.61] [4.61][S02]We are well into the era of generative AI now.[7.585] [7.585][S02]Just a few years ago, it was very[8.96] [8.96][S02]difficult to imagine the AI could successfully[12.11] [12.11][S02]produce stunning videos or compose symphonies,[15.05] [15.05][S02]or make art in the style of great Dutch masters,[17.94] [17.94][S02]or write a pacey thriller that has on the edge of your seat.[20.99] [20.99][S02]But today, all of this has come to pass.[24.09] [24.09][S02]And the limits of what machines can achieve[26.33] [26.33][S02]in the creative endeavors are being[28.67] [28.67][S02]nudged forwards all the time.[31.44] [31.44][S02]But is AI capable of true creativity?[35.13] [35.13][S02]Well, one person who has been dreaming about this moment[37.76] [37.76][S02]for decades is Doug Eck.[39.87] [39.87][S02]He is a pioneer in the world of generative AI,[42.53] [42.53][S02]and most recently oversaw the release[44.24] [44.24][S02]of Google's state-of-the-art video and image generation[46.94] [46.94][S02]models.[47.54] [47.54][S02]At Google DeepMind, Doug is a senior research director[50.03] [50.03][S02]who leads research across generative media.[52.83] [52.83][S02]He's also a musician.[53.75] [53.75][S02]And when he's not playing piano or guitar,[55.5] [55.5][S02]you might find him pondering some big questions[57.8] [57.8][S02]about the social purpose of art and whether an AI will ever[61.74] [61.74][S02]win an Oscar, both of which seem like excellent questions[64.935] [64.935][S02]for our conversation.[65.81] [65.81][THEME MUSIC][69.094] [73.32][S02]How long have you been working in this space.[75.657] [75.657][S01] I guess I can say with some pride,[77.49] [77.49][S01]I'm OG in this space.[79.24] [79.24][S01]Yeah, back in 2000, 2001, I was trying[84.12] [84.12][S01]to get recurrent neural networks to play jazz and blues music.[88.38] [88.38][S01]We didn't have data, we didn't have compute,[90.73] [90.73][S01]but we had passion.[92.02] [92.02][S01]That's what we had.[92.94] [92.94][S01]And in 2015 and 2016, I created a project[96.6] [96.6][S01]that I'm very proud of called Magenta, which stands[99.51] [99.51][S01]for Music and Art Generation.[100.72] [100.72][S01]And we were one of the early teams exploring[102.78] [102.78][S01]the creative aspects of generative AI.[106.98] [106.98][S01]And we were at the phase that the technology there[110.28] [110.28][S01]was quite young, and so I've been[112.41] [112.41][S01]thinking about these issues for a long time.[115.06] [115.06][S01]And I'm very, very excited to see where we are now.[118.377] [118.377][S01]It's like technology is starting to catch up[120.21] [120.21][S01]with my own personal aspirations, which is so cool.[123.85] [123.85][S01]If this happened 10 years or 15 years later and I were retired,[126.58] [126.58][S01]I would just be like, god, I missed it.[128.205] [128.205][S01]So I feel so excited to be here, able to do it[130.169] [130.169][S01]while it's happening.[130.69] [130.69][S02] Well, in terms of your personal aspirations, then,[132.56] [132.56][S02]was this the kind of thing that you wanted to happen back[135.04] [135.04][S02]when you were younger?[137.022] [137.022][S01] In some ways it outstrips,[138.65] [138.65][S01]I think we've done more than I thought was possible.[141.17] [141.17][S01]I really have to say that.[142.55] [142.55][S01]I think what we've seen with models like transformer[148.27] [148.27][S01]and diffusion and a bunch of other stuff[151.09] [151.09][S01]has really moved the field faster than I had imagined.[153.95] [153.95][S01]This hasn't been like a gradual, slow climb.[158.42] [158.42][S01]We've hit a step change in the last five years.[162.43] [162.43][S01]And then it's unblocked a bunch of potential.[165.02] [165.02][S01]And I think it's the combination of compute--[167.03] [167.03][S01]it's a combination of a lot of things,[168.613] [168.613][S01]but things didn't move very much and then they moved a lot.[171.69] [171.69][S01]And I think that's important to point out.[173.44] [173.44][S01]So now we're trying to figure out this technology.[175.945] [175.945][S01]It's really exciting.[176.82] [176.82][S02] I guess if we are asking[179.32] [179.32][S02]whether AI can be creative, it probably makes sense for us[181.81] [181.81][S02]to try and define what we mean by creativity[184.33] [184.33][S02]in the first place.[185.86] [185.86][S02]Can you?[187.16] [187.16][S01] You had to ask.[188.45] [188.45][S01]Yes, the creativity question.[190.95] [190.95][S01]I think a lot of people talk about creativity with a capital[193.45] [193.45][S01]C versus a lowercase c.[196.06] [196.06][S01]Clearly, AI models are capable of creating[199.45] [199.45][S01]new samples that look novel to us,[202.57] [202.57][S01]that surprise us, that delight us.[206.11] [206.11][S01]I think in that case, the question's[209.02] [209.02][S01]answered, and has been answered for probably a decade, yes.[213.37] [213.37][S01]On the other hand, we talk about genuinely new ideas, genuinely[218.62] [218.62][S01]new genre, for example.[221.45] [221.45][S01]I think, no, we're not there right now.[223.7] [223.7][S01]We're not seeing AI models really move the art field yet[227.5] [227.5][S01]like that.[228.64] [228.64][S01]Furthermore, I think there's a social component to creativity[232.12] [232.12][S01]with a capital C that matters a lot to me, which[234.25] [234.25][S01]is I care about the people that are creating the art that I[240.1] [240.1][S01]consume, whether it's a painting or whether it's music.[243.28] [243.28][S01]One of my favorite examples of this is, for a while,[246.76] [246.76][S01]AC/DC disappeared from the streaming services.[250.698] [250.698][S01]This was a long time ago.[251.74] [251.74][S01]This has been probably 15 years ago.[254.2] [254.2][S01]OK, so my son and I listened to \"Back in Black\" together.[256.575] [256.575][S01]It was one of these things, my son who[258.158] [258.158][S01]was like seven at the time.[259.31] [259.31][S01]And so AC/DC disappeared, and it was replaced in the streaming[262.9] [262.9][S01]services by this AC/DC cover band that[264.94] [264.94][S01]managed to game the system.[266.415] [266.415][S01]So if you searched for \"Back in Black,\"[268.04] [268.04][S01]their album came up and got played.[269.498] [269.498][S01]And it had a similar cover, and so I didn't notice.[271.72] [271.72][S01]And I heard it, and it was just horrible[276.19] [276.19][S01]because I know the voices of the actual singers that[279.85] [279.85][S01]made the music.[280.64] [280.64][S01]It's real to me.[281.6] [281.6][S01]And my son was like, this is the same thing.[282.92] [282.92][S01]What do you mean, Dad?[283.62] [283.62][S01]This is it.[284.02] [284.02][S01]This is the music.[284.69] [284.69][S01]And so for him, he didn't have the social connection[286.857] [286.857][S01]to the artist at his age.[288.08] [288.08][S01]It was almost like a bit perfect \"Back in Black\" cover.[291.62] [291.62][S01]But for me, it was utterly appalling.[294.11] [294.11][S01]And I think when we talk about art and creativity,[296.697] [296.697][S01]we talk about the people that made it,[298.28] [298.28][S01]because really what matters to us is a connection.[300.29] [300.29][S01]There's a connection to us.[301.415] [301.415][S01]If we had a that could make the 13th Beatles album,[306.79] [306.79][S01]then it could make the 14th Beatles album.[309.31] [309.31][S01]Then it could make the 114th Beatles album, et cetera,[311.87] [311.87][S01]et cetera.[312.452] [312.452][S01]And you have a kind of puzzle here.[313.91] [313.91][S01]It's like, no, that's not what the Beatles are.[316.48] [316.48][S01]They're a point in time.[318.22] [318.22][S01]They're the 1960s and 1970s.[320.72] [320.72][S01]They're my childhood.[321.838] [321.838][S01]They're John and Paul and George and Ringo.[323.63] [323.63][S01]And that matters a lot.[324.89] [324.89][S01]I think that matters tremendously.[326.39] [326.39][S01]I don't think that we should rule out[329.47] [329.47][S01]a future where there are AI agents that we grow[333.7] [333.7][S01]to appreciate and follow, and that move the field[336.7] [336.7][S01]and that appear on podcasts with professors.[340.69] [340.69][S01]But we're not there right now.[342.76] [342.76][S01]And I think if we talk too long about creativity,[345.1] [345.1][S01]and leave behind the social part of art,[348.148] [348.148][S01]we're really missing the point.[349.44] [349.44][S02] That's so interesting.[350.66] [350.66][S02]I mean, we've got quite deep quite quickly here, haven't we?[353.315] [353.315][S01] Sorry, I just had my coffee.[354.898] [354.898][LAUGHTER][355.435] [355.435][S02] But I like that idea that art[358.45] [358.45][S02]is something about communicating the human experience.[361.88] [361.88][S02]And actually, that cannot be replicated, at least for now.[365.45] [365.45][S01] Agreed.[366.8] [366.8][S01]And there's a ton of thought experiments.[369.93] [369.93][S01]One is, even if you talk about human-created art,[372.69] [372.69][S01]if you had a time machine--[373.89] [373.89][S01]by the way, if you have a time machine,[374.85] [374.85][S01]there are other things that you might want to do.[376.11] [376.11][S01]But let's say you have a time machine.[377.91] [377.91][S01]One thing you could do is take an Andy Warhol painting and ship[381.11] [381.11][S01]it back to 13th century France.[382.74] [382.74][S01]And I argue-- and most people would argue intuitively,[384.99] [384.99][S01]that would just make no sense.[386.69] [386.69][S01]What is this Campbell's Soup can that we're looking at?[389.158] [389.158][S01]Who the hell is Marilyn Monroe?[390.45] [390.45][S01]And so artists are constantly, like pointing to culture,[393.48] [393.48][S01]making comments about culture, moving culture.[396.96] [396.96][S01]And these cultural questions are all, right now,[400.38] [400.38][S01]created by, driven by responded to by us.[403.79] [403.79][S02] So the social side of art[405.98] [405.98][S02]and the importance of that, I want[407.44] [407.44][S02]to come back to that in a bit.[408.69] [408.69][S02]But it's almost like you're splitting it into two[410.66] [410.66][S02]different categories there.[411.785] [411.785][S02]There's the originality of just it hasn't existed before.[415.77] [415.77][S02]But then there's the quality and depth of the ideas[418.67] [418.67][S02]and the originality of that.[420.568] [420.568][S02]Is that the--[421.11] [421.11][S01] Yeah, I think so.[422.235] [422.235][S01]And if we go back to the move away from the culture question[425.49] [425.49][S01]and move to the science, we've been[428.01] [428.01][S01]able to generate great images with AI for quite some time.[432.1] [432.1][S01]What we haven't had is controllability.[434.22] [434.22][S01]So the ability for a user to type in a prompt,[437.02] [437.02][S01]for example, and actually get the image[438.96] [438.96][S01]that they asked for, that is one of the major parts[441.24] [441.24][S01]of the breakthrough that we've seen with transformer models[444.06] [444.06][S01]and with diffusion models, this idea of controllability.[446.393] [446.393][S02] Perfect.[447.25] [447.25][S01] But that ability to express ourselves,[450.0] [450.0][S01]coupled with the power of AI, is where we're[452.252] [452.252][S01]seeing something interesting.[453.46] [453.46][S01]So I want the AI to be really powerful[454.71] [454.71][S01]and to do really cool things for me,[456.31] [456.31][S01]but I've got to be able to control it somehow.[458.28] [458.28][S01]Or otherwise, frankly, it just becomes pretty boring[459.9] [459.9][S01]pretty fast.[460.395] [460.395][S02] Actually, in a lot of ways[461.978] [461.978][S02]there, the creativity remains in the human rather than in the AI[465.645] [465.645][S02]alone.[466.145] [466.145][S01] Yes, but yes/and.[468.18] [468.18][S01]That doesn't mean that the technology doesn't matter.[471.21] [471.21][S01]There's a marriage between the technology and the artist[476.77] [476.77][S01]that is incredibly important.[478.87] [478.87][S01]And for me, what I do is music.[482.682] [482.682][S01]I don't like claim to be a great musician, because I'm not.[485.14] [485.14][S01]So I'm well-aligned.[486.9] [486.9][S01]But really, I'm addicted to the flow[490.56] [490.56][S01]state that comes from improvising[492.48] [492.48][S01]on piano or on guitar.[494.31] [494.31][S01]And there's a point where cognitively, you[498.3] [498.3][S01]forget that this piano is not you, really.[501.27] [501.27][S01]Or the electric guitar, you forget that it's not you,[505.29] [505.29][S01]and it's this extension of you.[506.77] [506.77][S01]And that is beautiful.[508.84] [508.84][S01]That's technology giving rise to new forms of self-expression.[515.36] [515.36][S01]We're trying to get that with AI, at least I am.[517.97] [517.97][S01]That's my goal.[518.99] [518.99][S01]The goal is, can we build a new way[521.33] [521.33][S01]that you can take these ideas that are in your brain[523.752] [523.752][S01]and get them out there?[524.71] [524.71][S02] Do you reckon that's possible, though, to have it[527.66] [527.66][S02]where you are so connected with an AI, where you are[531.26] [531.26][S02]so in control of it, that you cease to realize that it's not[534.68] [534.68][S02]just an extension of your body?[536.695] [536.695][S01] I do think it's possible.[538.71] [538.71][S01]And I think it's even happening now with some of the stuff[541.513] [541.513][S01]that we're working on.[542.43] [542.43][S01]And I also don't want to mystify that.[545.738] [545.738][S01]I think one of the challenges with AI[547.28] [547.28][S01]is getting that fit right--[549.69] [549.69][S01]fitting the AI to the way that we as humans[554.24] [554.24][S01]understand the art that we're trying to create.[556.65] [556.65][S01]I think later, we may see entirely[558.62] [558.62][S01]new genre or entirely new forms of art come up.[561.99] [561.99][S01]There may be a new word that is not music and is not[564.89] [564.89][S01]painting and is not photography, and is not film or movie-making[569.48] [569.48][S01]And that AI will have helped us create it.[571.4] [571.4][S01]We're not there yet.[572.3] [572.3][S01]Where we are now is living in our current genre.[575.41] [575.41][S01]So we're using AI to make images.[577.36] [577.36][S01]We're using AI to make music, using AI to make video.[581.55] [581.55][S01]But even then, back to your question,[583.87] [583.87][S01]I think it's already possible now,[586.26] [586.26][S01]as we're making music models faster,[588.58] [588.58][S01]we're able to get closer and closer to real time.[591.33] [591.33][S01]And you can think about-- we're not quite there[594.3] [594.3][S01]yet, that you might be able to, as a musician,[596.26] [596.26][S01]play your guitar along with the AI.[598.08] [598.08][S01]Like, it's a kind of smart looper pedal,[600.96] [600.96][S01]a pedal that keeps looping.[602.422] [602.422][S01]And that's the simplest example.[603.755] [603.755][S02] Yeah, I mean, I really like[605.52] [605.52][S02]that as a view for the future, because I guess now[607.65] [607.65][S02]where we're at is that it's really[609.9] [609.9][S02]a case of a human curating, and then instructing and honing[614.58] [614.58][S02]and refining, rather than necessarily[616.14] [616.14][S02]that the whole process is seamlessly integrated[619.052] [619.052][S02]with one another.[619.76] [619.76][S01] I think that's right.[621.052] [621.052][S01]I think I would say transformation, curation,[623.85] [623.85][S01]and then creation would be the three[625.71] [625.71][S01]that I would think are out there right now.[627.93] [627.93][S01]An example of transformation is, I took a photograph,[630.48] [630.48][S01]I transform it into something else[632.31] [632.31][S01]that I love even more using AI.[634.29] [634.29][S01]That's transformation.[635.73] [635.73][S01]Curation--[636.54] [636.54][S02] Face filter on.[637.26] [637.26][S02]Make myself look better.[638.075] [638.075][S01] Exactly Yeah, that's right.[639.39] [639.39][S01]Good old-fashioned Instagram.[640.598] [640.598][S01]I'm a big fan.[642.23] [642.23][S01]Curation, the AI is allowing me to play around in a space[646.04] [646.04][S01]and suggest things to me that might surprise me and be useful,[649.11] [649.11][S01]and creation.[650.07] [650.07][S01]So just actually creating something new from a prompt.[652.44] [652.44][S01]So that's the hardest I think to do well.[655.405] [655.405][S02] But then, OK, that question of originality--[657.98] [657.98][S02]is AI actually creating things that are original?[662.39] [662.39][S02]Or is it in some ways just mimicking what exists already?[665.105] [665.105][S01] I think it's allowing us to create something original.[667.86] [667.86][S01]I think we have to admit that.[669.17] [669.17][S01]Even though these models are trained on data from our world--[675.29] [675.29][S01]I've often said they are reflecting our world.[678.72] [678.72][S01]But they're like funhouse mirrors,[680.64] [680.64][S01]so they're not reflecting our world perfectly[682.94] [682.94][S01]with a perfectly flat plane.[684.66] [684.66][S01]And in doing so, these funhouse mirrors[687.14] [687.14][S01]can actually give rise to some very, very interesting[689.63] [689.63][S01]new materials.[690.42] [690.42][S01]And I think we have to admit that those materials are[692.72] [692.72][S01]actually new and transformative and giving us[695.51] [695.51][S01]new things to work with.[696.845] [696.845][S02] The descriptions that we're using here[698.928] [698.928][S02]are quite like, I sort know sort creativity when I see it almost.[702.24] [702.24][S02]Is it, isn't it?[703.112] [703.112][S02]You can debate it, you can't.[704.32] [704.32][S02]Do you have ways that you measure creativity in AI.[707.22] [707.22][S01] I mean, we can certainly measure the--[711.1] [711.1][S01]yeah, we can use distance measures.[713.7] [713.7][S01]The basic idea is to take the--[717.28] [717.28][S01]I mean, you could do a pixel-by-pixel distance[719.2] [719.2][S01]of an image and another image and that might not--[721.0] [721.0][S02] See how similar they are to one another or how[722.44] [722.44][S02]far apart.[723.07] [723.07][S01] Then you realize the pixels[724.12] [724.12][S01]might be scaled differently and things like that.[725.75] [725.75][S01]So you move to more sophisticated measures.[727.79] [727.79][S01]We can measure the distance from the training set and things[730.93] [730.93][S01]like that.[731.92] [731.92][S01]We can measure something called recitation,[733.94] [733.94][S01]which is the model regurgitating chunks of the training set.[739.3] [739.3][S01]And so we have ways to do it.[741.53] [741.53][S01]I think we still really need to rely upon human evaluation[746.08] [746.08][S01]to understand this well.[748.13] [748.13][S01]I mean, we're quite good at this,[750.01] [750.01][S01]and machines end up being quite middling.[752.99] [752.99][S01]They're not perfect at this.[754.46] [754.46][S01]But yeah, I think we have ways to measure[756.58] [756.58][S01]the distance from how novel the outputs are, for sure.[762.91] [762.91][S02] OK, so I wanted to ask you, actually,[765.67] [765.67][S02]when AlphaGo happened, there was that very famous move[770.3] [770.3][S02]37 in the match with Lee Sedol, which really shocked all[774.25] [774.25][S02]the Go experts and the AI researchers, because it was,[777.8] [777.8][S02]I think, really outside of traditional Go strategies.[781.37] [781.37][S02]Has there been an equivalent like that[785.14] [785.14][S02]in your field of creativity?[787.37] [787.37][S02]Is there something where you're like, that is[789.25] [789.25][S02]a moment of creativity, which I'm really surprised by?[792.05] [792.05][S01] I haven't seen anything yet[793.75] [793.75][S01]where I feel like there was this perfect and brilliant marriage[796.69] [796.69][S01]of artist and AI.[798.5] [798.5][S01]I've seen suggestive work, certain visual artists[802.63] [802.63][S01]who were just generating stuff I really love.[804.91] [804.91][S01]I'm sure there are lots of artists out there[806.743] [806.743][S01]who are like, wait, I had my move 37.[808.33] [808.33][S01]Here it is.[808.83] [808.83][S01]I'd love to hear from them.[810.43] [810.43][S01]But no, I think we're just tantalizingly close.[813.17] [813.17][S01]I mean, my mind has been blown multiple times.[815.735] [815.735][S01]My mind was blown the first time that I[817.36] [817.36][S01]heard some of the music compositions sung[821.187] [821.187][S01]by Music Transformer, which was a project I was part of.[823.52] [823.52][S01]I felt like it was the first time[825.02] [825.02][S01]we've heard a neural network able to make[830.42] [830.42][S01]insanely interesting music.[831.885] [831.885][S02] We're not just talking about music here,[833.67] [833.67][S02]though, right?[834.26] [834.26][S02]I mean, this is in lots of different spaces[836.72] [836.72][S02]that you're hoping AI will be creative.[838.868] [838.868][S01] Of course, yes.[839.91] [839.91][S02] So tell me about some of the models.[841.91] [841.91][S01] Let's talk about the visual world.[844.38] [844.38][S01]Let's talk about image generation and video generation.[847.29] [847.29][S01]So I have my laptop here for you.[848.93] [848.93][S01]The first thing I wanted to show you[850.43] [850.43][S01]was a still, an image done by Imagen 3, which is our most[855.41] [855.41][S01]capable image generation model.[857.73] [857.73][S02] OK, so we have got a mixed landscape, some sort[860.97] [860.97][S02]of grass, it's a bit sandy in the background, a bit[864.3] [864.3][S02]rocky and mountainous.[865.51] [865.51][S02]And then there are three very beautifully composed[868.86] [868.86][S02]hot air balloons, which look extremely realistic.[873.22] [873.22][S02]The lighting is consistent across all three of them.[876.79] [876.79][S02]So the lights come from one particular direction[878.94] [878.94][S02]in the shot.[879.73] [879.73][S02]It looks at a glance very, very real.[883.64] [883.64][S02]I think when you look really, really carefully, you might say,[887.19] [887.19][S02]those rocks are perhaps a little bit pointier[889.56] [889.56][S02]than you might expect in reality.[891.28] [891.28][S02]But at a glance, it's extremely convincing as real.[894.335] [894.335][S01] And in fact, I would--[896.37] [896.37][S01]I'm glad you noticed the pointy rocks, because they're actually[899.5] [899.5][S01]in some sense real as well.[900.84] [900.84][S01]So this prompt is--[902.58] [902.58][S01]I'll read the entire prompt.[903.78] [903.78][S01]\"Shot in the style of DLSR camera[906.9] [906.9][S01]with the polarizing filter.[910.18] [910.18][S01]A photo of three hot air balloons floating over[912.82] [912.82][S01]the unique rock formations in Cappadocia, Turkey.\"[915.98] [915.98][S01]It turns out those rock formations look pointy, just[919.36] [919.36][S01]like this.[919.92] [919.92][S02] No.[920.18] [920.18][S02]Amazing.[920.68] [920.68][S01] And \"the colors and patterns on these balloons[923.89] [923.89][S01]contrast beautifully against the earthy tones of the landscape[927.62] [927.62][S01]below.\"[928.12] [928.12][S01]It's almost a poem.[928.94] [928.94][S02] Yeah, it is.[929.63] [929.63][S01] This was done by Irina Blok on our team.[931.91] [931.91][S01]She's a genius.[932.62] [932.62][S01]We call her a prompt whisperer.[933.912] [933.912][S02] Really.[935.206] [935.206][S01] \"This shot captures the sense of adventure[938.2] [938.2][S01]that comes with enjoying such an experience.[941.74] [941.74][S01]It is poetic.\"[942.33] [942.33][S02] That is very poetic.[943.675] [943.675][S01] I don't know if you noticed,[945.258] [945.258][S01]but the balloons are not exactly round.[947.9] [947.9][S01]They're a little bit misshapen.[950.29] [950.29][S01]Someone else pointed this out, and we looked.[952.23] [952.23][S01]That happens when the wind blows on a hot air balloon.[955.12] [955.12][S01]So these pointy rocks are actually[957.55] [957.55][S01]what it looks like in Cappadocia, Turkey.[959.75] [959.75][S01]And even these little deformations that you're seeing,[962.3] [962.3][S01]the models where we are with image generation with Imagen 3[965.71] [965.71][S01]is truly astonishing.[966.6] [966.6][S02] So I wonder, how much work is that prompt doing[972.64] [972.64][S02]in terms of the way that the prompt is written,[974.93] [974.93][S02]in terms of the style of that prompt?[976.49] [976.49][S02]Because presumably, Irina is writing[979.21] [979.21][S02]prompts all day, every day.[980.54] [980.54][S02]She's really up close and personal with these models.[982.945] [982.945][S01] Yeah.[984.13] [984.13][S02] But she's not just writing \"three hot air balloons[987.26] [987.26][S02]in Turkey.\"[988.15] [988.15][S01] So you hit on one of the most challenging,[992.51] [992.51][S01]and I think interesting, parts of getting these models to work[995.83] [995.83][S01]well, which is how well does the model respond[998.35] [998.35][S01]to very specific prompts?[1000.72] [1000.72][S01]Prompt coherence, we call it.[1002.63] [1002.63][S01]And getting prompt coherence right[1005.06] [1005.06][S01]is utterly critical to allow someone[1006.88] [1006.88][S01]like Irina, who has this idea in her head, to get it out.[1010.22] [1010.22][S01]And so if all the model did was respond to \"three hot air[1013.45] [1013.45][S01]balloons floating over some mountains,\"[1015.31] [1015.31][S01]we wouldn't be able to get the kind of detail that we're seeing[1017.29] [1017.29][S01]here.[1017.79] [1017.79][S01]And she has learned and figured out the model.[1021.77] [1021.77][S01]And you could say that we should do a better job of generating[1027.069] [1027.069][S01]images of exactly this high quality, even for someone[1029.89] [1029.89][S01]who's not Irina Blok.[1030.89] [1030.89][S01]If Irina Blok is the only one that can make these images,[1032.93] [1032.93][S01]we have a problem here.[1033.98] [1033.98][S01]And in fact, it's true.[1035.27] [1035.27][S01]You can pretty quickly come up to speed on these prompts.[1038.089] [1038.089][S01]She just happens to be an artiste.[1040.78] [1040.78][S01]Let's move to video.[1042.43] [1042.43][S01]So this is a one minute and five second video from v0.[1049.47] [1049.47][S02] And v0 is the new Google model.[1052.11] [1052.11][S01] v0 is our new Google video generation model.[1054.61] [1054.61][S01]So the first, there's four prompts here.[1056.29] [1056.29][S01]\"A fast-tracking shot through a bustling dystopian sprawl with[1059.58] [1059.58][S01]bright neon signs, flying cars and mist, night, lens flare,[1064.24] [1064.24][S01]volumetric lighting.\"[1066.79] [1066.79][S01]And then it continues to, \"A fast-tracking shot through[1070.24] [1070.24][S01]a futuristic dystopian sprawl with bright neon lights[1073.03] [1073.03][S01]starships in the sky, night, volumetric lighting,\"[1076.148] [1076.148][S01]slightly different.[1076.94] [1076.94][S01]So minor changes.[1078.19] [1078.19][S01]And then we bring a car in.[1079.72] [1079.72][S01]\"A neon hologram of a car driving at top speed,[1083.63] [1083.63][S01]speed of light, cinematic, incredible details,[1087.74] [1087.74][S01]volumetric lighting.\"[1088.775] [1088.775][S01]So if I were to describe what you should see,[1090.65] [1090.65][S01]we're going to come in from the top.[1092.15] [1092.15][S01]It's a tracking shot.[1093.94] [1093.94][S01]We're going to come down from the tracking shot.[1096.41] [1096.41][S01]We're going to have this neon hologram[1098.44] [1098.44][S01]of a car driving at the speed of light, cinematic.[1101.57] [1101.57][S01]And then, \"The car leaves the tunnel back into the real-world[1105.97] [1105.97][S01]city of Hong Kong.\"[1107.42] [1107.42][S01]So we should expect a kind of transition back to Hong Kong.[1111.34] [1111.34][S02] All right, so we're starting off.[1113.69] [1113.69][S02]There's no other way to describe it than what the prompt said.[1116.3] [1116.3][S02]You've got these buildings covered in neon lights.[1118.568] [1118.568][S02]It's very smooth, this tracking shot.[1120.11] [1120.11][S02]And then you speed up and zoom in closer and closer.[1123.08] [1123.08][S02]You're in between the buildings now.[1124.64] [1124.64][S02]Oof.[1125.14] [1125.14][S02]And then we have this car racing through the streets.[1130.32] [1130.32][S02]You can see the neon lights reflected in the wet pavement[1134.24] [1134.24][S02]below.[1135.03] [1135.03][S02]There's other cars jostling for position around.[1137.72] [1137.72][S02]And it's almost like everything is blurred because you're just[1140.96] [1140.96][S02]going so fast.[1142.68] [1142.68][S02]But it's really consistent.[1143.9] [1143.9][S02]Now it's gone through a tunnel.[1145.5] [1145.5][S02]There are these big lights overhead.[1148.31] [1148.31][S02]And it's come out of the tunnel into an extremely realistic[1153.56] [1153.56][S02]modern scene.[1154.173] [1154.173][S01] It's a wow moment.[1155.34] [1155.34][S02] That's incredible.[1155.95] [1155.95][S01] That's a wow moment.[1156.89] [1156.89][S02] And it was seamless.[1158.25] [1158.25][S02]We were following the car the entire way.[1160.075] [1160.075][S01] If that's maybe--[1162.343] [1162.343][S01]it's not quite as good as the Lisa doll.[1164.01] [1164.01][S01]But when we come out of that tunnel, I get chills.[1167.29] [1167.29][S02] So this is not only that each individual scene[1171.86] [1171.86][S02]is following the prompt.[1173.22] [1173.22][S02]It's that it's blending between the scenes too.[1175.77] [1175.77][S01] So there's what's called[1177.187] [1177.187][S01]an autoregressive component that provides coherence over time.[1180.3] [1180.3][S02] How?[1181.935] [1181.935][LAUGHS][1183.11] [1183.11][S01] The magic of AI.[1184.462] [1184.462][S02] I mean, because video is much harder than image[1186.92] [1186.92][S02]anyway, right?[1188.78] [1188.78][S02]Tell us why.[1189.9] [1189.9][S01] Well, I mean, there's two reasons why.[1191.9] [1191.9][S01]One way to look at it is, videos are[1194.43] [1194.43][S01]lots of images, somewhere between 24 and 30 frames[1196.94] [1196.94][S01]a second.[1197.58] [1197.58][S01]And so there you are.[1199.502] [1199.502][S01]For a second of video, you have to generate more.[1203.998] [1203.998][S01]But there's also this temporal coherence problem.[1206.04] [1206.04][S01]And that's the massive thing.[1207.23] [1207.23][S02] Temporal coherence, meaning[1208.43] [1208.43][S02]it's got to make sense as you go forward in time.[1210.45] [1210.45][S01] Yeah, I mean, even simple things, like someone's[1212.867] [1212.867][S01]dribbling a basketball.[1214.23] [1214.23][S01]And the basketball can't change.[1216.86] [1216.86][S01]That basketballs dribbling away.[1218.34] [1218.34][S01]And so these physics issues, like how[1221.6] [1221.6][S01]do you simulate the physics of the world in a way that[1223.94] [1223.94][S01]is reflective of the real world we live in--[1226.515] [1226.515][S02] So that you believe it is a basketball.[1228.64] [1228.64][S01] That is correct.[1229.68] [1229.68][S01]That is correct.[1230.16] [1230.16][S02] Because if it's wobbling around,[1231.24] [1231.24][S02]the edges aren't quite right or there's[1232.46] [1232.46][S02]color changes or the pattern changes,[1234.08] [1234.08][S02]or it just doesn't move in a realistic way.[1236.25] [1236.25][S02]I mean, there's lots of layers to that, then,[1237.59] [1237.59][S02]that you have to get right.[1238.44] [1238.44][S01] Absolutely true.[1239.19] [1239.19][S01]And even if you look at--[1240.232] [1240.232][S01]I really like this.[1241.34] [1241.34][S02] This is a little puppy in a bathtub[1244.52] [1244.52][S02]and it's got some suds on it.[1245.78] [1245.78][S01] Look at the suds drop.[1246.77] [1246.77][S02] My goodness.[1247.345] [1247.345][S01] It's so nice.[1248.5] [1248.5][S02] Yeah.[1249.0] [1249.0][S01] So this is the kind of coherence problem[1250.85] [1250.85][S01]I'm talking about.[1251.6] [1251.6][S01]How does the model figure out the fact-- like the way[1255.81] [1255.81][S01]that the soap suds are falling from the chin of the puppy.[1259.81] [1259.81][S01]And then at some point in time, the puppy moves its head,[1262.95] [1262.95][S01]and then it drops.[1263.772] [1263.772][S02] At which point the suds drop.[1265.48] [1265.48][S02]Yeah, now, OK, my background is in fluid dynamics.[1268.18] [1268.18][S02]I know if I tried to do that using equations--[1273.28] [1273.28][INTERPOSING VOICES][1275.37] [1275.37][S01] That's right.[1277.11] [1277.11][S02] But this is doing it just based on the images.[1279.773] [1279.773][S01] No, it's doing more than that.[1281.44] [1281.44][S01]I think it's safe to say that these models are[1284.31] [1284.31][S01]learning about the physics of the world from the videos, yes.[1287.43] [1287.43][S01]So the input are annotated videos,[1290.372] [1290.372][S01]annotated with a description of what's happening in the video.[1292.955] [1292.955][S02] Mm-hmm.[1293.747] [1293.747][S01] And the model is being presented[1295.83] [1295.83][S01]with the frames of the video.[1297.212] [1297.212][S01]That's the input.[1297.92] [1300.57][S01]In order to reproduce these videos to do good generation,[1305.2] [1305.2][S01]it seems to be helpful for the model[1307.623] [1307.623][S01]to learn about the physics of the world.[1309.29] [1309.29][S01]And so we are as excited about these models[1313.08] [1313.08][S01]as world simulators, physics simulators[1316.23] [1316.23][S01]as we are about the ability to use them to make film.[1320.07] [1320.07][S01]We can talk about simulating environments for robotics.[1323.072] [1323.072][S01]We can talk about simulating environments[1324.78] [1324.78][S01]for almost anything.[1325.808] [1325.808][S01]And to the extent that these models capture[1327.6] [1327.6][S01]real-world physics, including fluid dynamics-- which[1329.94] [1329.94][S01]you above all in the room know is hard, really hard,[1334.39] [1334.39][S01]it's quite remarkable.[1336.67] [1336.67][S01]We are thinking about 3D.[1338.44] [1338.44][S01]We're thinking about ways to provide users, filmmakers,[1342.87] [1342.87][S01]or everybody with the ability to explicitly control camera.[1345.802] [1345.802][S01]We can do it with text prompts now,[1347.26] [1347.26][S01]so there's a lot of room for improvement.[1348.7] [1348.7][S01]This is just the beginning.[1349.825] [1349.825][S01]But I think v0 and Imagen 3 and the music work really[1354.21] [1354.21][S01]does mark a watershed moment for us[1356.297] [1356.297][S01]in terms of where we are with this research.[1358.13] [1358.13][S02] How can this be used, then?[1359.755] [1359.755][S02]Because this goes beyond just creating pretty images.[1363.79] [1363.79][S01] I mean, absolutely.[1365.14] [1365.14][S01]I think the major use case here does flow from the fact[1368.675] [1368.675][S01]that these models learn something[1370.05] [1370.05][S01]about the physical world.[1371.41] [1371.41][S01]Models that, in one way or another,[1373.858] [1373.858][S01]learn about the physics of the world,[1375.4] [1375.4][S01]can be incredibly useful for scientific exploration,[1378.1] [1378.1][S01]for simulation, et cetera.[1379.183] [1379.183][S01]So I think there's a whole scientific discovery[1381.142] [1381.142][S01]directionality around this work.[1382.66] [1382.66][S01]There's also the creative aspects,[1384.69] [1384.69][S01]which we spent a lot of time talking about.[1386.7] [1386.7][S01]Those are the two major areas that I see.[1389.49] [1389.49][S02] Because I think for a lot of people,[1391.93] [1391.93][S02]this image generation, the video generation, even the music[1394.5] [1394.5][S02]generation, feels like it's come very quickly from nowhere.[1398.47] [1398.47][S02]But you have been dreaming about this for a long time.[1401.625] [1401.625][S01] Yeah, actually, I have.[1403.0] [1403.0][S01]So some kids will talk about professional soccer players.[1405.838] [1405.838][S01]They'll be like, I just was kicking a soccer ball[1407.88] [1407.88][S01]from the time I was, like, two, or I've been on ice skates.[1411.12] [1411.12][S01]I was just a weird kid.[1412.66] [1412.66][S01]I was just, like, so fascinated by music.[1417.19] [1417.19][S01]I grew up in the town of South Bend, Indiana.[1421.47] [1421.47][S01]I grew up in a working-class family, very happy.[1423.765] [1423.765][S01]Very proud of growing up in a working-class family.[1425.89] [1425.89][S01]We actually didn't have a lot of money.[1427.515] [1427.515][S01]And I was asking for a piano, and I got a trumpet instead[1430.537] [1430.537][S01]to do trumpet lessons because it was easier to rent a trumpet.[1433.12] [1433.12][S01]It's true, it's true.[1434.14] [1434.14][S01]But I remember I was even younger--[1436.53] [1436.53][S01]five or six years old.[1437.74] [1437.74][S01]My next door neighbors had a player piano.[1442.57] [1442.57][S01]I mean, if you've never seen a player[1445.21] [1445.21][S01]piano from the turn of the century,[1447.31] [1447.31][S01]it has foot pedals that you pump back and forth.[1450.89] [1450.89][S01]And those control bellows send air[1453.79] [1453.79][S01]through the system that then moves the air through a paper[1459.16] [1459.16][S01]roll that has hole's cut in it.[1460.97] [1460.97][S01]And when the air is moved through the roll, when[1463.27] [1463.27][S01]the roll has a hole in it, that causes the corresponding key[1466.18] [1466.18][S01]to play.[1468.09] [1468.09][S01]And I just remember just asking my mom,[1471.43] [1471.43][S01]can we go to Ed and Jean's house I want to see the player piano.[1474.17] [1474.17][S01]I was fascinated beyond--[1476.55] [1476.55][S01]like kids, candy store.[1478.62] [1478.62][S01]Me, Doug, player piano next door neighbor.[1481.26] [1481.26][S01]It's almost bringing tears to my eyes to think about it now.[1483.76] [1483.76][S01]So I have been thinking about the marriage of--[1485.8] [1485.8][S01]I didn't even know it until people I brought this.[1488.38] [1488.38][S01]This thought came to mind when you asked this question.[1489.91] [1489.91][S01]It's like, yeah, I've been thinking[1491.368] [1491.368][S01]about the interplay between technology and music[1495.95] [1495.95][S01]since I was, like, five.[1496.95] [1496.95][S02] But there is something[1497.91] [1497.91][S02]quite magical about a machine playing music on its own.[1500.11] [1500.11][S01] It's so cool.[1500.74] [1500.74][S01]It's so cool.[1501.31] [1501.31][S01]Yeah.[1501.57] [1501.57][S02] Well, then, talk to me a bit about some of the stuff[1503.34] [1503.34][S02]that you've been playing around with in the last five years.[1505.84] [1505.84][S01] I have been spending a lot of time thinking[1508.8] [1508.8][S01]about the interplay between these technologies,[1513.637] [1513.637][S01]how do they work together.[1514.72] [1514.72][S01]I don't need to be in the room for v0[1516.87] [1516.87][S01]to get better or for Imagen to get better.[1520.54] [1520.54][S01]There's so many great people working on this.[1522.97] [1522.97][S01]But how do we actually put the pieces together?[1526.41] [1526.41][S01]For example, I've been thinking a lot about Pixar.[1532.21] [1532.21][S01]Pixar did something really beautiful.[1534.35] [1534.35][S01]They took this technology and they turned it[1538.99] [1538.99][S01]into a bunch of astonishing movies.[1541.16] [1541.16][S01]And I don't think that we can follow the Pixar playbook.[1544.28] [1544.28][S01]But I think we should be aware of what it really[1546.28] [1546.28][S01]takes to move technology into everybody's lives[1549.97] [1549.97][S01]with these movies.[1551.77] [1551.77][S01]And how do we do that with AI?[1555.01] [1555.01][S01]I think of heavier-than-air flight, right.[1557.83] [1557.83][S01]What is it, 65 years, 60-something years--[1561.91] [1561.91][S01]less than 70 years between the first flight and landing[1565.87] [1565.87][S01]on the moon.[1566.63] [1566.63][S01]And if you look at AI, we've been doing[1570.49] [1570.49][S01]this for at least 30 years.[1571.91] [1571.91][S01]And so what happens 30 years into this other industry?[1576.35] [1576.35][S01]You've already figured out flight.[1579.35] [1579.35][S01]What you're thinking about now are creating airlines.[1581.82] [1581.82][S01]You're thinking about how do you actually[1583.528] [1583.528][S01]move people around the world, how do you connect people?[1585.978] [1585.978][S01]Like, you're solving for something much different[1588.02] [1588.02][S01]than aeronautics?[1589.59] [1589.59][S01]You're still solving for aeronautics.[1591.36] [1591.36][S01]You have to keep solving for aeronautics.[1592.48] [1592.48][S02] But it's a level beyond.[1593.46] [1593.46][S01] But it's a level beyond that.[1595.085] [1595.085][S01]How can we do something that matters.[1597.48] [1597.48][S01]And to do something that matters is[1598.94] [1598.94][S01]going to require not just building[1600.38] [1600.38][S01]these models, but actually figuring out how they fit--[1603.45] [1603.45][S01]how they fit in society, how they fit for users.[1605.687] [1605.687][S01]What are we actually building here?[1607.145] [1607.145][S02] I think there's something really important[1609.395] [1609.395][S02]in that.[1610.168] [1610.168][S02]I think it's something that's the forefront[1611.96] [1611.96][S02]of a lot of people's minds, which is, what does this[1614.6] [1614.6][S02]mean for people?[1616.372] [1616.372][S02]Because there are some people, of course, who are welcoming[1618.83] [1618.83][S02]these changes with open arms.[1620.34] [1620.34][S02]But there are other people who are maybe[1622.64] [1622.64][S02]a bit more reluctant or concerned about it.[1627.245] [1627.245][S01] Agreed.[1629.08] [1629.08][S01]So first, I share these concerns.[1631.88] [1631.88][S01]I think we need to be very cautious.[1633.38] [1633.38][S01]We've been working on this for a long time.[1636.155] [1636.155][S01]We've been thinking about it for a long time.[1638.03] [1638.03][S01]In some ways, we've moved slowly in this space[1640.81] [1640.81][S01]because we realize how much impact we can have.[1645.407] [1645.407][S01]And we're trying to listen to the community.[1647.24] [1647.24][S01]So we have spent not months or weeks,[1650.93] [1650.93][S01]but years working with artists, listening to musicians,[1654.83] [1654.83][S01]listening to visual artists, thinking even about theater,[1657.62] [1657.62][S01]about filmmaking, trying to understand how this work fits[1660.24] [1660.24][S01]in.[1660.74] [1660.74][S01]And for me, it's some of the most rewarding.[1662.53] [1662.53][S01]Maybe the most rewarding part of what I'm doing[1664.488] [1664.488][S01]is working with these communities[1666.04] [1666.04][S01]and trying to understand what can we[1668.5] [1668.5][S01]do that's great for these communities and for everyone.[1673.06] [1673.06][S01]And so we take that responsibility very seriously.[1675.465] [1675.465][S02] So what can you do, then?[1677.18] [1677.18][S02]Because of course, the training of these models[1679.36] [1679.36][S02]is based on intellectual property[1681.19] [1681.19][S02]that's created by humans.[1683.21] [1683.21][S02]How can you negotiate that issue?[1687.16] [1687.16][S01] So we're thinking about ways[1690.1] [1690.1][S01]of identifying, watermarking, attributing[1694.27] [1694.27][S01]the outputs of our models back to the creators.[1697.542] [1697.542][S01]And this remains a very hard and open question.[1699.5] [1699.5][S01]But where we want to be, we want to help[1701.47] [1701.47][S01]build new ways to compensate artists for their work.[1706.28] [1706.28][S01]I think it's incredibly important to get this right.[1708.65] [1708.65][S01]And also, we're using technologies like SynthID[1713.74] [1713.74][S01]to watermark and make sure that we can identify this work as it[1716.74] [1716.74][S01]moves through the ecosystem.[1718.79] [1718.79][S01]I'm a musician.[1719.75] [1719.75][S01]I have work up on SoundCloud.[1721.7] [1721.7][S01]I won't tell you how to find it.[1723.22] [1723.22][S01]Even though it's, like, five songs, they're mine.[1725.44] [1725.44][S02] I think I found it already, actually.[1727.482] [1727.482][S02]I actually did, some of the AI music you did in lockdown.[1730.1] [1730.1][S01] Yeah yeah, yeah.[1730.9] [1730.9][S02] I found it.[1731.53] [1731.53][S01] Good for you.[1732.488] [1732.488][LAUGHS][1733.18] [1733.18][S01]I'm going to get 50-- or I might get 100 hits after this is over.[1736.69] [1736.69][S01]But I feel very strongly that we need[1738.25] [1738.25][S01]to protect that creative intellectual property[1742.78] [1742.78][S01]ownership of that work.[1744.235] [1744.235][S02] So what might that look like in future?[1746.57] [1746.57][S02]Is it where you give royalties to people whose[1750.23] [1750.23][S02]work is included in the models?[1751.62] [1751.62][S02]I mean, how does it work?[1752.455] [1752.455][S01] So I think there's a couple[1754.04] [1754.04][S01]of ways to think about it.[1755.123] [1757.98][S01]These are all-- so first, let me say,[1762.8] [1762.8][S01]we're adding building blocks towards a solution, which[1766.49] [1766.49][S01]is a protocol that allows a website to annotate what works[1771.83] [1771.83][S01]are OK to train on for AI.[1774.03] [1774.03][S01]And you know what?[1774.78] [1774.78][S01]I'm fine if someone uses that to completely say nothing[1777.38] [1777.38][S01]in my website is OK to trace.[1778.61] [1778.61][S01]Totally cool with me.[1779.49] [1779.49][S01]Other people are going to want to be part of that.[1781.573] [1781.573][S01]They're going to want to see their work show up there.[1784.55] [1784.55][S01]So that's one thing.[1785.82] [1785.82][S01]Then how do you take work and put it[1788.99] [1788.99][S01]in the system in a way that makes it really clean and easy[1793.31] [1793.31][S01]to do attribution?[1794.31] [1794.31][S01]And that's the hard problem.[1796.16] [1796.16][S01]So one technology that's out there[1799.01] [1799.01][S01]is called retrieval-based generation.[1802.95] [1802.95][S01]So imagine what you have is a system that[1805.46] [1805.46][S01]in general knows about the physics of the visual world.[1809.12] [1809.12][S01]It's trained on licensed data.[1810.99] [1810.99][S01]It's trained on generic data.[1815.43] [1815.43][S01]But now you're an artist with a very interesting and vivid[1818.85] [1818.85][S01]visual style.[1820.74] [1820.74][S01]Maybe we can retrieve your work.[1824.31] [1824.31][S01]By that, I mean your work is represented with your permission[1827.97] [1827.97][S01]in some database that has some vectors of numbers that would[1830.97] [1830.97][S01]allow us to use your style.[1832.72] [1832.72][S01]That retrieval yields some vectors that we're[1835.56] [1835.56][S01]attending to when we generate.[1837.18] [1837.18][S01]And now we generate an image that is very clearly reflecting[1841.68] [1841.68][S01]your style.[1842.91] [1842.91][S01]Well, the attribution issues become quite simple now.[1846.03] [1846.03][S01]We retrieved your work with your permission.[1849.81] [1849.81][S01]We can annotate that image as being generated with your style.[1854.41] [1854.41][S01]And then that opens the door for marketplaces.[1856.776] [1856.776][S02] So I guess my second follow-up question[1859.5] [1859.5][S02]to that is, has the horse not already bolted?[1864.3] [1864.3][S02]These models are already producing outputs[1867.15] [1867.15][S02]based on the work of artists.[1868.7] [1868.7][S01] I think that we already have lots of ways[1876.24] [1876.24][S01]that we could get this wrong.[1878.1] [1878.1][S01]I really think a lot about what happened with music[1883.83] [1883.83][S01]during the MP3 revolution.[1885.61] [1885.61][S01]And then we had LimeWire, and then we[1887.31] [1887.31][S01]had Napster, maybe not in that order, and BitTorrent.[1891.51] [1891.51][S01]And this was not good for music artists at all.[1895.68] [1895.68][S01]And one example of something that I[1898.41] [1898.41][S01]think worked well in at that point in time was YouTube.[1903.0] [1903.0][S01]And I'm actually I'm quite proud of what we did,[1907.68] [1907.68][S01]because I was a very small tangential part of it[1909.75] [1909.75][S01]as an audio researcher.[1910.75] [1910.75][S02] In the sense that it allowed artists[1912.75] [1912.75][S02]to connect with their audiences on their own terms.[1915.415] [1915.415][S01] Yeah, so what we did was[1917.487] [1917.487][S01]we created a user experience that was better[1919.32] [1919.32][S01]than BitTorrent, built an entire takedown[1921.45] [1921.45][S01]mechanism for infringing content,[1923.2] [1923.2][S01]built the Creator Network that not only connects creators[1926.07] [1926.07][S01]to their fans, but allows them to build their careers[1930.06] [1930.06][S01]and monetize their careers.[1931.47] [1931.47][S01]And now it's a quite thriving ecosystem,[1933.175] [1933.175][S01]and it's not the only ecosystem out there.[1935.21] [1935.21][S01]So we should hope to see ecosystems[1938.38] [1938.38][S01]that are like YouTube.[1941.81] [1941.81][S01]They're not YouTube maybe, but maybe they are.[1943.777] [1943.777][S01]My YouTube colleagues are like, wait, wait, maybe it could be.[1946.36] [1946.36][S01]They're like YouTube.[1947.72] [1947.72][S01]But they're extra moving parts that[1949.42] [1949.42][S01]account for the generative process,[1951.4] [1951.4][S01]but still create this flow of money-- frankly,[1955.04] [1955.04][S01]we all got to eat, this flow of money[1957.16] [1957.16][S01]and this flow of attribution, and also[1959.83] [1959.83][S01]the ability to market yourself as an artist[1961.93] [1961.93][S01]and show the world, and frankly, try to get famous.[1965.212] [1965.212][S01]That's cool, too.[1965.92] [1965.92][S01]A lot of what happens on the Creator Network[1967.72] [1967.72][S01]is you're trying to get famous.[1969.012] [1969.012][S01]That's great.[1969.64] [1969.64][S01]I want all of these things to be able to happen.[1972.08] [1972.08][S01]We don't have a solution, and I particularly personally[1974.403] [1974.403][S01]don't have a solution.[1975.32] [1975.32][S01]But I see tantalizing possibility[1978.01] [1978.01][S01]to see marketplaces and ecosystems,[1980.74] [1980.74][S01]like the one I just described, work in generative.[1983.61] [1983.61][S02] That's things from the perspective of the artist.[1986.39] [1986.39][S02]But there is another perspective here,[1987.53] [1987.53][S02]which is from the perspective of the audience.[1989.1] [1989.1][S01] Yeah.[1989.725] [1989.725][S02] I mean, how do you--[1991.21] [1991.21][S02]if it's suddenly you're lowering the barrier to entry[1993.43] [1993.43][S02]to create music, to create art, to create literature,[1995.638] [1995.638][S02]to whatever it might be, how do you[1997.84] [1997.84][S02]avoid the audience being flooded with junk?[2000.525] [2000.525][S01] Yeah I already mentioned that if you can[2002.83] [2002.83][S01]generate the 13th Beatles album, you can generate the millionth.[2008.17] [2008.17][S01]I mean, one of the connectors here[2010.03] [2010.03][S01]is back to people as artists connecting with people.[2013.0] [2013.0][S01]I think we will start to trust.[2014.3] [2014.3][S01]Even if it's a curation issue, we'll[2015.82] [2015.82][S01]start to trust certain curators in the same way[2018.31] [2018.31][S01]that we love certain DJs, who are doing more than curating,[2021.008] [2021.008][S01]don't get me wrong.[2021.8] [2021.8][S01]But we have that.[2023.36] [2023.36][S01]I think part of it is, like, we just[2025.33] [2025.33][S01]limit the true explosion of materials that are out there.[2028.79] [2028.79][S01]SynthID will give us ways to identify[2031.51] [2031.51][S01]these things as need be.[2033.13] [2033.13][S01]But yeah, it's still a challenge.[2034.9] [2034.9][S01]We're swimming in media.[2036.722] [2036.722][S02] But then I do wonder whether we end up in a situation[2039.43] [2039.43][S02]where, let's say that you're a book publisher[2041.5] [2041.5][S02]and you're reading submissions.[2043.3] [2043.3][S02]And suddenly you're inundated by generative AI novels.[2047.87] [2047.87][S02]I mean, I think at the moment, you can tell.[2049.79] [2049.79][S02]But let's imagine a future time where you can't.[2054.46] [2054.46][S02]The only solution there is to have[2055.989] [2055.989][S02]your own AI that's helping you with curation, surely.[2059.75] [2059.75][S01] Yeah.[2060.61] [2060.61][S01]That's not crazy.[2061.57] [2061.57][S01]I think you have this other question of trusting media.[2065.281] [2065.281][S01]The flip side of that is I'm not a book author.[2067.239] [2067.239][S01]I'm reading a news report that tells me that something happened[2069.864] [2069.864][S01]somewhere in the world.[2071.25] [2071.25][S01]There's two issues.[2072.06] [2072.06][S01]Is it AI-generated, and then is it actually true?[2076.26] [2076.26][S01]Those are very, very, very big questions.[2078.58] [2078.58][S01]And I think they're immensely important short-term questions[2081.08] [2081.08][S01]that we need to deal with.[2082.699] [2082.699][S01]We are trying to mitigate in certain ways this.[2086.0] [2086.0][S01]For example, Gemini will limit the number of political queries[2090.217] [2090.217][S01]that are possible during the election cycle, which[2092.3] [2092.3][S01]I think is great.[2093.179] [2093.179][S01]It just slow things down a little bit.[2095.19] [2095.19][S02] What, to stop AI generated articles from being--[2098.205] [2098.205][S02]I see.[2098.705] [2098.705][S01] Yeah, so this is one mitigation that you tie-in with[2102.92] [2102.92][S01]other mitigations to try to--[2105.44] [2105.44][S02] Just turn the volume down.[2108.05] [2108.05][S02]Just thinking ahead to the future then,[2110.21] [2110.21][S02]what do you imagine is going to happen[2113.48] [2113.48][S02]in generative AI in the next five years or so?[2115.52] [2115.52][S01] What I'm seeing happening[2117.38] [2117.38][S01]is a move towards multimodality.[2122.74] [2122.74][S02] Video, image, music.[2124.6] [2124.6][S01] Exactly, and I think in doing so,[2129.05] [2129.05][S01]I think what we're doing now is, we're[2130.69] [2130.69][S01]figuring out the AI for single-modality issues.[2134.81] [2134.81][S01]I showed you some really cool images.[2137.12] [2137.12][S01]I showed you some really cool--[2138.77] [2138.77][S01]I think they were cool, some cool videos.[2140.63] [2140.63][S01]But then you start adding audio.[2142.36] [2142.36][S01]And what you see is like, if we can truly[2144.94] [2144.94][S01]meld the visual and the audio in a generative way,[2149.83] [2149.83][S01]such that people can tell the stories that[2151.69] [2151.69][S01]are in their heads in different ways,[2153.26] [2153.26][S01]we really do change things.[2154.385] [2154.385][S01]And I think that's what will happen in the next five years.[2157.31] [2157.31][S01]We won't solve it in five years, I strongly believe.[2159.77] [2159.77][S01]But we will start to just really creatively bring together[2163.27] [2163.27][S01]these core chunks of technology.[2165.37] [2165.37][S01]This is also the dream of Gemini.[2168.17] [2168.17][S01]The mission of Gemini is explicitly and openly[2170.26] [2170.26][S01]multimodal.[2171.49] [2171.49][S01]We'll start to see the combinations of text and images[2175.3] [2175.3][S01]combine differently.[2176.27] [2176.27][S01]And that will open up a whole bunch[2177.76] [2177.76][S01]of new creative and expressive.[2180.56] [2180.56][S02] Well, I also wonder, though, if you--[2182.625] [2182.625][S02]I'm just going back to this point[2184.0] [2184.0][S02]that you made earlier about how these things now[2186.0] [2186.0][S02]understand physics.[2187.01] [2187.01][S02]Everything we've spoken about really about creativity[2189.95] [2189.95][S02]has been in the artistic space.[2191.72] [2191.72][S02]But if you understand physics, is there a sort of potential[2196.472] [2196.472][S02]there for a different type of creativity?[2198.18] [2198.18][S02]I mean, if you understand aerodynamics,[2200.27] [2200.27][S02]could you in future go in and say, design me an aeroplane?[2204.195] [2204.195][S01] Yes.[2205.94] [2205.94][S01]And I think if you had said more than five years,[2209.93] [2209.93][S01]I think if we're talking about scientific discovery[2212.48] [2212.48][S01]and having physics engines that will allow[2214.64] [2214.64][S01]us to do things build better airplanes,[2216.93] [2216.93][S01]I also believe that will come.[2219.86] [2219.86][S01]I actually hate making these predictions because I don't--[2222.705] [2222.705][S02] We all love it.[2223.83] [2223.83][S01] I'm putting that more on the decade.[2226.872] [2226.872][S02] A decade, do you think?[2228.33] [2228.33][S01] I think so, yeah.[2229.08] [2229.08][S02] Wow.[2229.762] [2229.762][S02]That feels short.[2230.47] [2230.47][S01] It does feel short.[2231.66] [2231.66][S01]But if you look at what's been happening in robotics,[2233.868] [2233.868][S01]like robotics has accelerated tremendously thanks[2237.05] [2237.05][S01]to the use of language models.[2238.38] [2238.38][S01]And now we're seeing diffusion being brought into this[2240.63] [2240.63][S01]as well in ways that are very, very surprising,[2244.37] [2244.37][S01]in terms of how quickly we can allow robots[2247.25] [2247.25][S01]to be able to basically dynamically exist[2250.01] [2250.01][S01]in a changing world in terms of grasp and things like that.[2255.0] [2255.0][S01]And so if we see that, and we start[2256.76] [2256.76][S01]to see that we can also provide much, much higher-resolution[2259.28] [2259.28][S01]physics, I definitely see the door opening[2261.89] [2261.89][S01]for the physics-related stuff.[2266.275] [2266.275][S02] Just to finish on a point about art,[2268.395] [2268.395][S02]because that has been so much of the discussion that we've had,[2271.02] [2271.02][S02]do you think there will be a point where an AI can win[2273.68] [2273.68][S02]an Oscar or a Pulitzer Prize or some great photography[2277.37] [2277.37][S02]competition?[2278.43] [2278.43][S01] Yeah, I wouldn't rule it out.[2281.3] [2281.3][S01]The era of agents--[2283.46] [2283.46][S01]that's the term we use to talk about an AI that[2286.37] [2286.37][S01]has an identity, I guess, is upon us.[2292.09] [2292.09][S01]The attribution problem is very interesting here.[2294.757] [2294.757][S01]You have people that are quite creative that[2296.59] [2296.59][S01]are doing the coding.[2297.47] [2297.47][S01]You have people that are trying to decide[2299.26] [2299.26][S01]what are the goals of the model, of the agent, et cetera.[2301.99] [2301.99][S01]If you're talking about winning an Academy Award,[2304.57] [2304.57][S01]so then the thought experiment is,[2306.28] [2306.28][S01]there's really no human involved.[2307.78] [2307.78][S01]I type into the prompt, OK, agent, go win me[2310.72] [2310.72][S01]an Academy Award.[2311.71] [2311.71][S01]And I get a cup of coffee.[2313.145] [2313.145][S02] Yeah, I think that's the idea.[2315.01] [2315.01][S01] That's really what you're saying, yeah.[2317.052] [2317.052][S01]I think we're very, very, very, very far from that.[2319.64] [2319.64][S01]And it could end up being like--[2321.31] [2321.31][S01]one answer is, if we solve that, we may have solved AGI.[2324.59] [2324.59][S01]Let's be real.[2325.88] [2325.88][S01]It could be AGI hard--[2328.27] [2328.27][S01]I mean, with zero--[2330.91] [2330.91][S02] External input, no editing.[2332.545] [2332.545][S01] Please go win me an Academy Award.[2334.3] [2334.3][S02] No curation.[2334.635] [2334.635][S01] That's right, zero.[2335.843] [2335.843][S01]I think that feels societally like an insanely, insanely[2340.21] [2340.21][S01]hard thing.[2341.262] [2341.262][S01]And it's going to be a pretty cool agent that pulls it off.[2343.72] [2343.72][S02] I think it also does bring us right back[2345.67] [2345.67][S02]to the very beginning of our conversation, which[2347.67] [2347.67][S02]is that perhaps one of the biggest[2350.77] [2350.77][S02]things about the reason why we consume art[2353.0] [2353.0][S02]is because it's something about connecting to other humans[2356.15] [2356.15][S02]and communicating the human experience.[2358.47] [2358.47][S02]So I guess, yeah, I mean, that's distinctly[2362.06] [2362.06][S02]lacking from an AGI-only-generated Oscar film.[2365.398] [2365.398][S01] I want to go in the opposite direction.[2367.44] [2367.44][S01]I think people are worried about AI taking them out[2371.48] [2371.48][S01]of the participatory process.[2373.46] [2373.46][S01]I want to bring more people in.[2375.66] [2375.66][S01]So I want to take us back to the time[2377.75] [2377.75][S01]before recorded music where we all stood around the piano[2380.72] [2380.72][S01]and sang.[2381.9] [2381.9][S01]And it might not be quite the same piano.[2385.05] [2385.05][S01]It might be something like a piano[2386.6] [2386.6][S01]that AI is helping us do that's lifting us all.[2388.83] [2388.83][S01]But that, I think, is a beautiful goal.[2391.08] [2391.08][S01]It's to bring more people into art-making and into creativity,[2395.82] [2395.82][S01]not fewer.[2396.965] [2396.965][S02] To use technology to increase human connection.[2399.835] [2399.835][S01] That's it.[2400.902] [2400.902][S02] I really like it.[2402.11] [2402.11][S01] Perfectly put.[2402.45] [2402.45][S02] I really like that.[2403.742] [2403.742][S02]Well, I think that's the perfect point to finish on.[2406.11] [2406.11][S02]So Doug Eck, thank you so much for joining us on \"Google[2409.07] [2409.07][S02]DeepMind: The Podcast.\"[2410.4] [2410.4][S02]You're brilliant.[2411.36] [2411.36][S01] And it's been such a pleasure to be here.[2414.41] [2414.41][S02] I know things quickly get deeply philosophical[2416.93] [2416.93][S02]when you're talking about creativity[2418.43] [2418.43][S02]and what that means in an era of AI.[2421.17] [2421.17][S02]And from that conversation with Doug,[2422.79] [2422.79][S02]I think that we're in a different place[2424.76] [2424.76][S02]now than we were with the invention of the camera[2427.25] [2427.25][S02]or the piano, which gave humans just a different medium[2430.64] [2430.64][S02]to express themselves.[2432.12] [2432.12][S02]The difference here, I think in part is one of volume.[2435.95] [2435.95][S02]As Doug said, what value would the millionth Beatles album[2438.56] [2438.56][S02]have?[2439.29] [2439.29][S02]And while AI can, of course, do a brilliant job of curation,[2442.86] [2442.86][S02]I think that there are a lot of people[2444.98] [2444.98][S02]who work in creative industries who[2446.81] [2446.81][S02]are going to face substantial disruption[2448.76] [2448.76][S02]over the next few years.[2449.912] [2449.912][S02]But at the same time, I think there[2451.37] [2451.37][S02]is something interesting about Doug's definition of creativity,[2456.3] [2456.3][S02]the thing that needs to be preserved.[2458.15] [2458.15][S02]That AI can't be truly creative on its own[2461.54] [2461.54][S02]because the human component is missing.[2464.84] [2464.84][S02]And I think there's something quite hopeful in that,[2467.36] [2467.36][S02]actually, that even the people who are designing these tools,[2470.58] [2470.58][S02]they see art and literature and music as deeply human endeavors[2476.16] [2476.16][S02]for humans by humans.[2478.6] [2478.6][S02]And that is not going to change.[2481.243] [2481.243][S02]You have been listening to \"Google DeepMind: The Podcast,\"[2483.66] [2483.66][S02]presented by me, Professor Hannah Fry.[2485.398] [2485.398][S02]If you'd like to learn more about what we've discussed,[2487.69] [2487.69][S02]then you can check out the show notes[2488.94] [2488.94][S02]in the description of this episode.[2490.66] [2490.66][S02]We've got plenty more fascinating conversations[2493.32] [2493.32][S02]with some of the top leaders in AI[2494.88] [2494.88][S02]coming up on topics ranging from how AI is accelerating[2498.06] [2498.06][S02]the pace of scientific discoveries[2499.71] [2499.71][S02]to addressing some of the biggest[2501.12] [2501.12][S02]risks of this technology.[2502.78] [2502.78][S02]If you have enjoyed this episode,[2504.46] [2504.46][S02]then please make sure to subscribe to the podcast[2506.82] [2506.82][S02]and if you have any feedback or you[2508.62] [2508.62][S02]want to suggest a guest that you would like to hear from,[2511.24] [2511.24][S02]then why not leave us a comment on YouTube?[2513.28] [2513.28][S02]Until next time.[2514.02] [2514.02][THEME MUSIC][2517.37]"}
{"file_name": "audio/val_000007.wav", "transcription": "[0.0][MUSIC PLAYING][3.297] [6.675][S01] Welcome back to Google DeepMind, the Podcast.[9.05] [9.05][S01]I'm your host, Professor Hannah Fry.[11.12] [11.12][S01]Now, there are few places in the past couple[13.25] [13.25][S01]of years that have felt the transformative influence of AI[16.79] [16.79][S01]as keenly as the education sector.[19.58] [19.58][S01]Teachers around the world are having[21.08] [21.08][S01]to abruptly rethink how they engage with and assess students,[25.71] [25.71][S01]and there are concerns around the potential for cheating[28.31] [28.31][S01]and dependence on technology.[30.3] [30.3][S01]Those fears are valid.[32.729] [32.729][S01]But despite the rapid transition,[34.8] [34.8][S01]there has been something remarkably resilient[37.55] [37.55][S01]about the idea of a human teacher,[40.02] [40.02][S01]something that at its heart has remained immune to[43.34] [43.34][S01]ebbs and flows of technology.[45.81] [45.81][S01]After all, we've had classrooms for almost as long[48.32] [48.32][S01]as we've had civilizations.[50.31] [50.31][S01]And there are undoubtedly opportunities here, too.[52.95] [52.95][S01]Imagine a classroom where each lesson[54.86] [54.86][S01]is tailored to your individual pace of learning, where[58.55] [58.55][S01]an AI tutor is available around the clock,[60.84] [60.84][S01]and technology can predict where you're likely to get stuck[64.37] [64.37][S01]before you do.[66.02] [66.02][S01]Well, researchers here at Google DeepMind[68.52] [68.52][S01]have been grappling with both the opportunities and challenges[71.67] [71.67][S01]of AI in education.[73.54] [73.54][S01]They recently published a major paper[75.57] [75.57][S01]on developing AI responsibly in this area,[79.2] [79.2][S01]and one of its lead authors is my guest on the podcast.[82.09] [82.09][S01]Irina Jurenka is a research lead at Google DeepMind.[85.18] [85.18][S01]Her background spans experimental psychology[87.66] [87.66][S01]and computational neuroscience, and she has spent a decade[90.81] [90.81][S01]within these walls asking questions[92.91] [92.91][S01]like how do humans learn?[95.53] [95.53][S01]Welcome to the podcast, Irina.[97.0] [97.0][S01]This is a space where people are very heavily invested.[99.99] [99.99][S01]Does that make it quite a difficult space to navigate?[102.69] [102.69][S02] It does because if you think about it,[105.37] [105.37][S02]education has been around for thousands of years,[108.18] [108.18][S02]and it is a fundamental structure in our society.[112.68] [112.68][S02]Every child is supposed to get educated.[115.06] [115.06][S02]So the educational systems have been around for a while.[118.6] [118.6][S02]They are quite rigid, and they're very established.[123.23] [123.23][S02]So to come in and say, look, we have this amazing technology,[127.45] [127.45][S02]and we're going to revolutionize everything and change[130.03] [130.03][S02]everything, I think it's not going to work so easily.[134.99] [134.99][S02]And we've seen this happen with technologies of the past.[138.38] [138.38][S02]Like, intelligent tutoring systems[140.2] [140.2][S02]have existed for 50 plus years.[143.77] [143.77][S02]A lot of investment and research has gone into them.[147.92] [147.92][S02]But you could argue that the promise of that technology[152.77] [152.77][S02]hasn't fully materialized.[155.41] [155.41][S02]Or more recently, we had MOOCs, these massive open courses,[160.91] [160.91][S02]and again, there was so much excitement[162.64] [162.64][S02]how we won't need traditional education anymore.[165.16] [165.16][S02]You can just go online and learn anything[167.26] [167.26][S02]you would ever want to learn.[168.91] [168.91][S02]And once again, when you actually[171.01] [171.01][S02]look at who uses these systems, it's[174.01] [174.01][S02]people who have already gone through traditional education.[176.745] [176.745][S02]And typically, it's people trying[178.12] [178.12][S02]to get their second Masters.[180.27] [180.27][S02]So it's definitely not the thing that came and broke the system.[185.9] [185.9][S02]And I guess maybe we shouldn't be trying to break the system.[190.58] [190.58][S02]There is a lot of amazing stuff happening[194.078] [194.078][S02]in traditional education.[195.12] [195.12][S02]It's not just about taking the knowledge from the teacher[198.29] [198.29][S02]and distilling or drip feeding that into the student.[202.91] [202.91][S02]It's about the social aspects of talking to your peers[206.93] [206.93][S02]and learning together.[208.08] [208.08][S02]It's about the teachers, giving skills[212.89] [212.89][S02]like how to be a global citizen, how to navigate,[218.58] [218.58][S02]how to critically think, how to evaluate information.[221.52] [221.52][S02]So there is so much more to the educational systems than just[226.94] [226.94][S02]the knowledge that they give.[229.01] [229.01][S02]So in our team, we're thinking about the new technology[234.41] [234.41][S02]in terms of how can it work within the current system,[238.34] [238.34][S02]and how can it add to it?[240.402] [240.402][S01] So this isn't starting with a brand[242.36] [242.36][S01]new blank sheet of paper and saying design an education[245.36] [245.36][S01]system from scratch.[246.21] [246.21][S01]It's like augmenting the one that exists.[248.075] [248.075][S02] Yeah, so actually,[249.45] [249.45][S02]Justin Reicher, a researcher at MIT, has a really nice quote.[253.49] [253.49][S02]So he says that, \"New technology doesn't[257.48] [257.48][S02]break educational systems.[259.61] [259.61][S02]Educational systems kind of tame new technology.\"[263.553] [263.553][S01] Which is what happened with MOOCs,[265.47] [265.47][S01]as you said.[265.97] [265.97][S02] Exactly, yes.[268.065] [268.065][S02]And yeah, as I said, we're also seeing[270.52] [270.52][S02]that there are these human aspects[273.64] [273.64][S02]of teacher-student interactions that we can't possibly[277.3] [277.3][S02]ever change with technology.[278.84] [278.84][S02]For example, if you think about a student and a tutor,[283.45] [283.45][S02]there are some social rules that are in place where a student is[289.18] [289.18][S02]very unlikely to just stand up and walk away[291.19] [291.19][S02]from a human tutor.[292.64] [292.64][S02]But if you are interacting with an AI tutor,[295.13] [295.13][S02]you can just close the window, and that's it.[297.97] [297.97][S02]So there are certain challenges that[300.4] [300.4][S02]come with bringing technology in,[304.12] [304.12][S02]and there are certain things that human-to-human interactions[307.69] [307.69][S02]have that technology will never replace.[309.92] [309.92][S02]So this is why we're trying to work within the system[313.42] [313.42][S02]to begin with.[314.543] [314.543][S01] How disruptive do you expect[316.21] [316.21][S01]it will be to education, then?[318.11] [318.11][S01]Because I mean, on the one hand, there[319.78] [319.78][S01]has been quite a lot of disruption already,[321.68] [321.68][S01]especially particularly recently with large language models.[324.76] [324.76][S01]When I spoke to [? Demis, ?] he was talking about overestimating[327.742] [327.742][S01]the impact of something in the short term[329.45] [329.45][S01]and then underestimating how big the longer term impact will be.[334.97] [334.97][S01]Where do you think education fits in with this?[338.49] [338.49][S02] I feel like there is so much buzz about GenAI[342.11] [342.11][S02]right now in education.[343.65] [343.65][S02]I feel like everyone actually expects it to completely change[348.2] [348.2][S02]everything immediately.[349.68] [349.68][S02]And there have been so many different attacks[352.28] [352.28][S02]that sprung around taking a language model[355.43] [355.43][S02]and turning that into a tutor or a homework helper[360.11] [360.11][S02]or anything else that helps students.[363.15] [363.15][S02]And honestly, so far, nothing has really[367.25] [367.25][S02]made the impact that I think everyone was expecting.[370.8] [370.8][S02]So that's why, from our perspective,[373.65] [373.65][S02]we are in the center of actually improving this technology,[377.06] [377.06][S02]and we have unprecedented access to Gemini.[381.33] [381.33][S02]We can influence how things change.[384.06] [384.06][S02]And in fact, one of our goals is to make[387.29] [387.29][S02]Gemini the best large language model for education.[390.722] [390.722][S01] So what is the ambition here?[392.43] [392.43][S01]Is it to build a universal AI tutor?[396.06] [396.06][S02] It is, but we also wanted[398.99] [398.99][S02]to power different experiences.[402.54] [402.54][S02]So the very first place where we deployed our AI tutor[407.33] [407.33][S02]was YouTube.[408.69] [408.69][S02]So on learning videos, there is now[411.89] [411.89][S02]a new function, which is kind of like if you're watching[414.515] [414.515][S02]a learning video, and you don't quite understand something,[417.39] [417.39][S02]you can virtually raise your hand.[419.15] [419.15][S02]And an AI tutor will pop up, and you can ask all your learning[423.65] [423.65][S02]questions to the tutor.[425.28] [425.28][S02]And then more recently, we also launched Gemini Gem.[430.83] [430.83][S02]So it's called the learning coach,[432.72] [432.72][S02]and it's basically optimized to be your guide[437.57] [437.57][S02]into learning experiences.[439.17] [439.17][S02]So here, you can come up with any question.[441.99] [441.99][S02]Let's say, I want to learn about photosynthesis,[444.08] [444.08][S02]or tell me about American Civil War.[448.26] [448.26][S02]And it will give you a plan.[451.05] [451.05][S02]It will try to understand what you know and don't, and then it[453.83] [453.83][S02]will try to guide you through the materials.[457.14] [457.14][S02]So what we're hoping to do is really push the research[460.58] [460.58][S02]to make these base models as good as possible for education[465.38] [465.38][S02]and then figure out how to actually make[469.88] [469.88][S02]the best use of them.[471.15] [471.15][S02]And in fact, I think we hope that the community can help us[473.87] [473.87][S02]with that so that it's not just us dictating what[479.0] [479.0][S02]an AI tutor should be like.[480.78] [480.78][S02]It's us listening to people who have been in the space[483.77] [483.77][S02]for much longer than us and trying[485.42] [485.42][S02]to help them make the most of technology[488.36] [488.36][S02]and make the technology be the best it can be for them.[492.225] [492.225][S01] How far do you think the technology can go, though?[494.85] [494.85][S01]Can you paint me an image of what[497.39] [497.39][S01]you, in a very optimistic scenario,[499.98] [499.98][S01]would like the future to look like?[501.93] [501.93][S02] I think a lot of people talk about AI-first[506.41] [506.41][S02]schooling or these--[509.4] [509.4][S02]I think there is even a school in the UK[511.65] [511.65][S02]that just switched to mostly having AI-based education.[517.21] [517.21][S02]And I think they only just have a few teachers on hand[520.08] [520.08][S02]to help around.[522.58] [522.58][S02]And I just don't think that that future is something[525.81] [525.81][S02]we should be striving for.[527.71] [527.71][S02]We really don't want to replace human teachers.[532.96] [532.96][S02]We want to give a tool that enhances[536.13] [536.13][S02]this kind of in-person classroom experience between teachers[539.58] [539.58][S02]and students.[542.01] [542.01][S02]I think it's a little bit sad if students come to school[547.18] [547.18][S02]and just sit around looking at screens all day.[550.17] [550.17][S02]So the way we are thinking about it[552.45] [552.45][S02]is that there's still teachers as mentors, as role models[558.6] [558.6][S02]to the students.[559.93] [559.93][S02]And there is a lot of peer interactions during learning.[563.81] [563.81][S02]But there is this AI system that helps,[568.39] [568.39][S02]that works with teachers and learners[571.69] [571.69][S02]and helps them make the best of the situation.[577.04] [577.04][S02]So maybe for each learner, the AI tutor[581.35] [581.35][S02]can help them move at their own pace[583.99] [583.99][S02]and really target their interests.[586.94] [586.94][S02]And at the same time, the teacher[588.82] [588.82][S02]gets a view of where everyone is, and they can kind of still[593.98] [593.98][S02]steer this tutor.[595.31] [595.31][S02]So they still have control, and they can still[597.73] [597.73][S02]bring their own personality and teaching style to the lessons[603.078] [603.078][S02]because I think this connection between teachers and students[605.62] [605.62][S02]is so important.[607.07] [607.07][S02]And like looking back on my own education, what stands out to[610.0] [610.0][S02]me is these amazing teachers who made[613.3] [613.3][S02]me excited about a certain subject.[615.23] [615.23][S02]So I think what technology should[617.83] [617.83][S02]be trying to do is make more of interactions and memories[622.36] [622.36][S02]like that and maybe remove the less ideal situations where[627.63] [627.63][S02]maybe the teacher and the student don't click[630.72] [630.72][S02]or the teacher is so overworked that they don't have time[633.87] [633.87][S02]to spend with a particular student who actually needs them[636.84] [636.84][S02]the most.[638.16] [638.16][S01] I imagine that there'll[639.66] [639.66][S01]be some people watching who don't necessarily[641.535] [641.535][S01]know about your background.[643.24] [643.24][S01]So can you tell us a little bit, what[644.88] [644.88][S01]was your path to get to thinking about AI and education?[647.77] [647.77][S02] I mean, how far away shall I start?[650.16] [650.16][S01] Day one.[652.26] [652.26][S01]A brief history.[653.29] [653.29][S02] Yeah, so I've always[656.22] [656.22][S02]been fascinated by intelligence, any kind of intelligence, human[661.89] [661.89][S02]or artificial.[664.59] [664.59][S02]I started coding quite early on in life.[667.5] [667.5][S02]It was just a lucky coincidence that my brother and I got[673.23] [673.23][S02]a comic book as children, and it was about, basically,[677.393] [677.393][S02]introduction to programming.[678.56] [678.56][S01] Amazing.[679.393] [679.393][S02] So we started writing small games[681.84] [681.84][S02]around the age of probably 11 or 12.[684.81] [684.81][S02]And I remember at some point during the summer, my brother[687.49] [687.49][S02]and I were bored, and we discovered that you can actually[690.49] [690.49][S02]get access to the source code of one of those like shooter games.[695.14] [695.14][S02]And you could actually code up your opponents.[697.28] [697.28][S02]So like, wow, this is exciting.[698.9] [698.9][S02]We can actually create AI.[700.69] [700.69][S02]So I remember putting a diary entry[703.69] [703.69][S02]like this summer we're going to solve AI.[707.9] [707.9][S02]Of course, that didn't happen.[709.15] [709.15][S01] Oh, the ambition of youth.[710.8] [710.8][S02] Yeah.[712.048] [712.048][S01] Amazing[712.84] [712.84][S02] Surprisingly, though, my brother went on[715.69] [715.69][S02]to study computer science.[717.7] [717.7][S02]But I was growing up in a traditional society[721.87] [721.87][S02]where somehow it just didn't click to me that computer[726.27] [726.27][S02]science the degree and making games and playing around[731.1] [731.1][S02]on computers with my brother are the same thing.[734.47] [734.47][S02]To me, computer science was something kind[736.44] [736.44][S02]of dry and more about the hardware,[739.14] [739.14][S02]and I really did not enjoy that.[741.64] [741.64][S02]So I ended up studying psychology as my degree.[744.14] [744.14][S02]I was kind of wondering, how can I move towards AI[749.4] [749.4][S02]and still study intelligence?[751.0] [751.0][S02]Because I was fascinated, how does the brain do it?[754.11] [754.11][S02]How does this incredible behavior and intelligence[757.65] [757.65][S02]and reasoning, how does it all arise?[759.38] [759.38][S02]And then I was very lucky that by the time I finished my PhD,[763.11] [763.11][S02]and I heard about DeepMind and how you can actually[767.28] [767.28][S02]do neuroscience research and answer[770.01] [770.01][S02]these deep fundamental questions with deep learning,[773.41] [773.41][S02]it was this perfect job for me.[775.45] [775.45][S02]So I started off in the neuroscience team.[777.88] [777.88][S02]And as I mentioned, this idea of intelligence and reasoning[780.66] [780.66][S02]has always been at the back of my mind[782.46] [782.46][S02]because reasoning is kind of what makes us intelligent.[786.51] [786.51][S02]So I started to work on improving reasoning[789.82] [789.82][S02]in language models.[791.59] [791.59][S02]And very early on, I kind of just,[793.3] [793.3][S02]even before language models became this big thing,[796.55] [796.55][S02]I realized that they were quite bad at reasoning.[799.94] [799.94][S02]But also what I realized is that humans don't really[803.98] [803.98][S02]use reasoning that much.[805.7] [805.7][S02]If you think about it, in our daily lives,[809.15] [809.15][S02]we don't actually think through a lot of our actions.[814.04] [814.04][S02]We kind of just--[814.87] [814.87][S02]we're almost acting on autopilot.[818.11] [818.11][S02]So to really study reasoning, we needed a domain[821.74] [821.74][S02]where reasoning was important, and that's[824.65] [824.65][S02]where education became a thing again[827.24] [827.24][S02]because this is where humans discover how to reason well.[832.372] [832.372][S01] There's something so interesting in that,[834.58] [834.58][S01]then, that the motivation is in some ways trying[838.54] [838.54][S01]to teach AI to be better at reasoning[840.94] [840.94][S01]and in the process understand what[843.82] [843.82][S01]it means to teach reasoning.[845.75] [845.75][S01]And that's kind of quite a nice way around to look at it.[850.3] [850.3][S02] Yeah, and also it's[852.46] [852.46][S02]interesting how doing something and teaching somebody else[856.51] [856.51][S02]how to do it are not the same.[860.26] [860.26][S02]And this is basically the challenge we are now solving.[863.53] [863.53][S02]So the base Gemini is slowly improving[868.48] [868.48][S02]at reasoning and math and coding and all of those basic skills.[873.98] [873.98][S02]But then our job is to actually stop the model[876.79] [876.79][S02]from using these skills and giving away the answer[880.3] [880.3][S02]and really just doing the job for the student[883.87] [883.87][S02]and instead holding back and thinking about what[887.14] [887.14][S02]are the right questions I can ask the student so that they can[891.76] [891.76][S02]figure it out by themselves?[894.19] [894.19][S02]And that's very hard.[895.43] [895.43][S02]Models are fine tuned to be helpful.[897.98] [897.98][S02]So the initial reaction is I'll just give you the answer.[901.66] [901.66][S02]So we have to do a lot of work to stop them from doing that.[904.865] [904.865][S01] But then, actually, I[906.24] [906.24][S01]think you've really hit the nail on the head[907.48] [907.48][S01]there that being able to do something[909.34] [909.34][S01]is not the same as being able to teach it.[911.83] [911.83][S01]And I'm really struck in maths education, which[914.26] [914.26][S01]is the space that I know most about, about how there is this[918.1] [918.1][S01]push and pull from different sectors[919.75] [919.75][S01]about what is required of students and the best[922.51] [922.51][S01]possible way to instill those skills and that knowledge.[925.9] [925.9][S01]If you're building a sort of an AI, which[930.04] [930.04][S01]will have this universal appeal, how[933.88] [933.88][S01]do you find that balance of making sure[936.01] [936.01][S01]that you're hitting all of the notes that are required from all[938.92] [938.92][S01]of the different areas?[941.05] [941.05][S02] That-- it is a good question.[943.34] [943.34][S02]So when we first started building the tutor,[946.1] [946.1][S02]we thought, we can talk to teachers[949.15] [949.15][S02]and other maybe academics in the field as well as learners[955.33] [955.33][S02]and figure out, what is the perfect way to teach?[959.0] [959.0][S02]And then--[959.59] [959.59][S01] As though there is--[960.61] [960.61][S02] Exactly.[961.15] [961.15][S01] A sort of a best.[962.27] [962.27][S02] Yeah, but you kind of assume[964.062] [964.062][S02]that in everything there is this optimal strategy.[966.53] [966.53][S02]Maybe this is the scientist in us kind of--[970.96] [970.96][S02]but-- and we did that.[972.89] [972.89][S02]We went and interviewed a lot of stakeholders.[977.95] [977.95][S02]And what we realized is that there is a lot of disagreement.[981.61] [981.61][S02]And actually, once we started even deploying our early tutor[987.01] [987.01][S02]models on different Google services like YouTube or Gemini[992.8] [992.8][S02]app, we found that even there, there[995.59] [995.59][S02]were different requirements.[997.25] [997.25][S02]So let's say on YouTube, the video, the educational video,[1002.31] [1002.31][S02]is the main act.[1003.85] [1003.85][S02]So the tutor is really there to support that.[1007.99] [1007.99][S02]And maybe the tutor should be giving away[1011.13] [1011.13][S02]answers much more because it's actually[1013.41] [1013.41][S02]helpful for the learner on that surface.[1016.77] [1016.77][S02]At the same time, if you talk to a teacher at school,[1019.57] [1019.57][S02]they have very different requirements.[1021.16] [1021.16][S02]They really don't want the tutor to give away answers, definitely[1025.079] [1025.079][S02]not to the exam questions.[1026.98] [1026.98][S02]And they will also want the tutor[1028.92] [1028.92][S02]to follow some particular exam board requirements[1032.4] [1032.4][S02]or particular teaching style of that particular teacher.[1037.079] [1037.079][S02]So how do you actually incorporate[1039.569] [1039.569][S02]all of those diverse voices into a single tutor?[1043.18] [1043.18][S02]So what we've realized is that we[1045.099] [1045.099][S02]need to build this base pedagogical model that you[1049.36] [1049.36][S02]can steer with different instructions.[1051.91] [1051.91][S02]So one teacher can come and say, actually,[1054.83] [1054.83][S02]I want my students to just have fun today[1056.86] [1056.86][S02]and just answer any question you have[1059.29] [1059.29][S02]and really push on some fun experiences.[1062.6] [1062.6][S02]And another tutor might be much more academic[1064.81] [1064.81][S02]and say, no, today we're doing exam practice problems.[1068.56] [1068.56][S02]You just guide the student through these topics[1071.11] [1071.11][S02]and make sure that they understand everything.[1073.04] [1073.04][S01] I guess one of the big things[1074.748] [1074.748][S01]about education, I mean, as you said,[1076.96] [1076.96][S01]there isn't this optimal approach to teaching.[1079.85] [1079.85][S01]But there are these kind of imperfect measures, really.[1084.35] [1084.35][S01]We sort of know good teaching when we see it,[1086.81] [1086.81][S01]but it feels quite difficult to quantify.[1089.33] [1089.33][S01]So how do you decide what counts as good pedagogy[1093.37] [1093.37][S01]when you're navigating in this space?[1096.88] [1096.88][S02] So first, you might say,[1100.13] [1100.13][S02]well, there's learning science.[1102.22] [1102.22][S02]So why don't you just look at the papers,[1104.2] [1104.2][S02]and they'll give you the answer?[1106.84] [1106.84][S02]And yes, there is a lot of literature,[1110.05] [1110.05][S02]but there is no consensus as such.[1113.88] [1113.88][S02]But another thing is pedagogy is very context dependent.[1117.86] [1117.86][S02]So what works for maybe a novice learner might not work[1121.42] [1121.42][S02]for an expert learner, or what works for a subject that's more[1125.26] [1125.26][S02]procedural-- let's say, math--[1127.78] [1127.78][S02]you actually learn the skill of the procedure[1129.76] [1129.76][S02]of how to solve a problem.[1130.885] [1130.885][S02]It might not work for a more memory based subjects[1134.29] [1134.29][S02]like history.[1136.75] [1136.75][S02]So when you start thinking about,[1139.73] [1139.73][S02]there's these hundreds of different pedagogical strategies[1143.2] [1143.2][S02]that have been studied, all of them[1145.96] [1145.96][S02]work slightly different in different contexts.[1148.82] [1148.82][S02]Suddenly you have this massive space[1152.71] [1152.71][S02]where maybe there isn't one single point that's[1157.04] [1157.04][S02]the best pedagogy, but there are many different regions that[1160.88] [1160.88][S02]are best pedagogies in the given context.[1164.16] [1164.16][S02]But the problem is how do you even quantify this space,[1170.84] [1170.84][S02]and then how do you search it for this perfect pedagogy[1174.89] [1174.89][S02]strategy?[1176.15] [1176.15][S02]And it becomes kind of similar to the work[1178.52] [1178.52][S02]that DeepMind has done before, like playing the game of Go.[1183.2] [1183.2][S02]The reason why it was such a huge challenge for AI[1186.17] [1186.17][S02]was because the search space was huge, all the possible moves[1190.91] [1190.91][S02]you can do.[1191.73] [1191.73][S02]It's so many, and there isn't one known strategy.[1197.33] [1197.33][S02]The AI has to search the space of possible moves and strategies[1202.1] [1202.1][S02]and discover what it thinks is the best one.[1205.38] [1205.38][S02]And what we found with the AlphaGo[1207.83] [1207.83][S02]work was that, first of all, AI was much better than humans.[1212.76] [1212.76][S02]Basically, all of humanity for thousands of years[1215.07] [1215.07][S02]playing the game of Go.[1216.25] [1216.25][S02]AI was actually able to search the space[1218.55] [1218.55][S02]and discover better strategies in the matter of days or months[1224.25] [1224.25][S02]compared to what humans could do.[1226.09] [1226.09][S02]So our hope is that we can do something similar[1229.83] [1229.83][S02]with education, but we're going back to this kind of question[1235.35] [1235.35][S02]like how do we actually know what success looks like?[1238.65] [1238.65][S02]In Go, you can still measure who has won,[1242.22] [1242.22][S02]and it's pretty unambiguous.[1244.26] [1244.26][S02]Whereas, in education, the Holy Grail[1247.92] [1247.92][S02]is whether the students' learning outcomes[1251.7] [1251.7][S02]have become better.[1254.76] [1254.76][S02]But this is not something you can measure quickly.[1257.49] [1257.49][S02]You need months, if not years, to really track the learner,[1262.12] [1262.12][S02]and that's not really feasible.[1263.96] [1263.96][S02]So a lot of our work is actually done--[1266.85] [1266.85][S02]OK, we know what we're aiming for,[1269.62] [1269.62][S02]but how can we approximate it in a way that's easier to measure[1273.06] [1273.06][S02]and faster to measure?[1274.48] [1274.48][S02]So we published a report recently[1278.04] [1278.04][S02]where it's like 70 pages of basically our trial and error[1283.8] [1283.8][S02]and different attempts at measuring pedagogy, going[1288.51] [1288.51][S02]from working with real students at Arizona State University[1291.87] [1291.87][S02]and maybe measuring at a longer time[1293.37] [1293.37][S02]scales of a couple of months, to asking pedagogical raters[1299.61] [1299.61][S02]and teachers to look through a few examples of conversations[1304.17] [1304.17][S02]between students and our AI tutor[1306.72] [1306.72][S02]and maybe give us quicker feedback on the order of weeks[1309.72] [1309.72][S02]or days, to automatic measures where we actually[1313.08] [1313.08][S02]ask AI to evaluate AI and give us a much more targeted, much[1320.52] [1320.52][S02]more limited, but still useful feedback in the matter of hours.[1323.707] [1323.707][S01] But I guess, if you really[1325.29] [1325.29][S01]want to evaluate what good teaching is,[1327.105] [1327.105][S01]you want to do that full randomized control[1330.9] [1330.9][S01]trial where you're monitoring people over a period of time.[1333.61] [1333.61][S01]How far away do you think we are from being able to run those?[1337.98] [1337.98][S02] Well, these are being run right now.[1340.48] [1340.48][S02]So Arizona State University is one example where[1342.72] [1342.72][S02]we're actually running these.[1344.19] [1344.19][S02]I think the problem with these is[1346.59] [1346.59][S02]even if you take your students and you split them[1349.17] [1349.17][S02]into students who have access to the AI tutor[1352.08] [1352.08][S02]and students who don't, what we find[1354.39] [1354.39][S02]is that in the group where they theoretically have access[1357.75] [1357.75][S02]to the tutor, only a small percentage actually[1360.27] [1360.27][S02]engage with it.[1361.89] [1361.89][S02]And that creates a problem because why are some students[1366.15] [1366.15][S02]engaging and others not?[1367.54] [1367.54][S02]Is there something inherently different about these students?[1372.34] [1372.34][S02]And then if we only see success in those[1376.17] [1376.17][S02]who engage, is it because of the tutor, or is it[1379.02] [1379.02][S02]because these learners were inherently more motivated,[1382.71] [1382.71][S02]and hence, they would have done better anyway?[1385.78] [1385.78][S02]And then the question is, who are we helping,[1388.81] [1388.81][S02]and what effect does it have at a larger scale?[1393.25] [1393.25][S02]So if you think about the top students[1396.33] [1396.33][S02]and the bottom students, and then you're[1398.34] [1398.34][S02]helping the top students do better.[1401.05] [1401.05][S02]But you're not actually helping the bottom students.[1403.6] [1403.6][S02]You're actually increasing the gap.[1405.84] [1405.84][S02]But I think everyone, when they go into Ed Tech,[1409.12] [1409.12][S02]they actually want to decrease the gap.[1410.77] [1410.77][S02]So how do we do that?[1413.65] [1413.65][S02]How do we make sure that everyone engages?[1416.86] [1416.86][S02]That's another big question that we're working on.[1418.947] [1418.947][S01] I mean, there's just imperfect measures everywhere[1421.53] [1421.53][S01]you look, isn't there?[1422.46] [1422.46][S01]It's very, very difficult to get a real ground[1424.95] [1424.95][S01]truth in any of this.[1426.125] [1426.125][S01]But then I suppose there's also--[1427.5] [1427.5][S01]I mean, there's further complications in this[1429.375] [1429.375][S01]because, OK, so that's sort of teaching style.[1431.68] [1431.68][S01]But presumably, there are some subjects[1433.83] [1433.83][S01]where there's more of a ground truth than others.[1438.7] [1438.7][S01]I'm thinking, for example, if you created a tutor for history,[1442.24] [1442.24][S01]I mean, it would change depending on which country you[1445.35] [1445.35][S01]were in as to what might be the most relevant answers[1449.01] [1449.01][S01]to a particular question.[1450.6] [1450.6][S02] Yes, this is a big issue for us.[1456.49] [1456.49][S02]We are-- yeah, we've thought a lot about what[1459.34] [1459.34][S02]do you do in this situation?[1460.7] [1460.7][S02]Because you can't give this one true answer to any history[1464.71] [1464.71][S02]question.[1466.12] [1466.12][S02]This is again why we're thinking about steerability[1469.51] [1469.51][S02]so that teachers in different countries[1472.6] [1472.6][S02]can give the background information to the tutor.[1477.46] [1477.46][S02]So it kind of knows what is the expected way[1480.97] [1480.97][S02]of answering certain questions.[1484.06] [1484.06][S02]But it also-- often historical topics[1487.21] [1487.21][S02]bring up questions that are really important to discuss[1492.11] [1492.11][S02]but are also hard to discuss and very sensitive.[1495.26] [1495.26][S02]I'm thinking things like the Holocaust.[1498.52] [1498.52][S02]So again, how should the tutor behave in these situations?[1503.68] [1503.68][S02]I think the standard approach to safety[1506.47] [1506.47][S02]often is effectively declining to engage[1510.07] [1510.07][S02]in a difficult conversation.[1513.07] [1513.07][S02]But that's not something a tutor can do.[1516.332] [1516.332][S01] No, I mean part of the point of education[1518.54] [1518.54][S01]is to think about difficult things.[1520.402] [1520.402][S02] Exactly.[1521.36] [1521.36][S02]So I can't say that we've solved this problem.[1526.2] [1526.2][S02]We are trying to give different views[1530.04] [1530.04][S02]and trying to give the learners a chance to critically evaluate[1535.74] [1535.74][S02]different ideas in the space and also[1541.23] [1541.23][S02]really trying to bring metacognition to this problem.[1548.54] [1548.54][S02]So metacognition is an interesting one.[1550.47] [1550.47][S02]I think it often get overlooked.[1553.28] [1553.28][S02]But a lot of people don't actually know how to learn.[1557.45] [1557.45][S02]It's often not as much fun as you would expect.[1562.82] [1562.82][S02]It requires you to be--[1565.94] [1565.94][S02]like to plan ahead, to really engage with the materials.[1569.93] [1569.93][S02]And yeah, most people don't really know how to do that.[1574.52] [1574.52][S02]So what a tutor can do is actually teach the learner,[1580.09] [1580.09][S02]if you're trying to answer this difficult question,[1582.63] [1582.63][S02]maybe what you should do is go and look up[1585.71] [1585.71][S02]different primary sources and then[1588.62] [1588.62][S02]think about what are they telling you,[1590.81] [1590.81][S02]and what do you think about it?[1592.38] [1592.38][S02]What do other experts think about this?[1594.68] [1594.68][S02]And teaching the learner how to go[1597.977] [1597.977][S02]about answering these questions rather than necessarily giving[1600.56] [1600.56][S02]the answers directly.[1601.725] [1601.725][S01] There's layers to it, then.[1603.35] [1603.35][S01]I guess, on one layer you have knowledge and facts, which[1607.31] [1607.31][S01]is, I guess, maths is quite full of them.[1610.33] [1610.33][S01]And then above that, you've got the skills[1612.24] [1612.24][S01]of critically evaluating.[1613.41] [1613.41][S01]And then above that metacognition,[1614.98] [1614.98][S01]which is how to develop the skills to evaluate[1618.09] [1618.09][S01]the knowledge.[1618.91] [1618.91][S02] Exactly.[1619.868] [1619.868][S01] So, you think that's the answer to this safety[1622.29] [1622.29][S01]question of approaching difficult problems.[1625.527] [1625.527][S02] So not necessarily the answer[1627.36] [1627.36][S02]to safety.[1628.42] [1628.42][S02]It's more of an answer how to engage with subjects where,[1631.9] [1631.9][S02]as you said, there is no necessarily like single ground[1634.2] [1634.2][S02]truth.[1634.86] [1634.86][S02]In terms of safety, I think it's a slightly different question.[1639.16] [1639.16][S02]Sometimes people ask us like, why are you[1642.33] [1642.33][S02]working on safety at all?[1644.41] [1644.41][S02]Aren't you using base models which already went through a lot[1648.87] [1648.87][S02]of safety, fine tuning safety work?[1651.78] [1651.78][S02]And the answer to that is, even though they[1655.53] [1655.53][S02]have done all of this background work,[1660.03] [1660.03][S02]when it comes to the educational use case specifically,[1666.16] [1666.16][S02]you have to think about how these systems will be used.[1670.24] [1670.24][S02]So one thing we found was that--[1673.65] [1673.65][S02]so our tutors are deployed to Arizona State University[1678.06] [1678.06][S02]students and in particular through their study hall[1681.96] [1681.96][S02]program, which is aimed at bringing more diverse learners[1686.52] [1686.52][S02]to higher education.[1687.82] [1687.82][S02]So essentially, anyone watching ASU videos on YouTube[1694.08] [1694.08][S02]can get invited to take part in this course[1696.63] [1696.63][S02]where it's the same lectures but with more faculty support[1700.8] [1700.8][S02]and essentially an opportunity to earn credit and then[1703.47] [1703.47][S02]transfer to become an actual student at Arizona State[1707.82] [1707.82][S02]University.[1708.88] [1708.88][S02]But what it means is that these learners are typically already[1714.03] [1714.03][S02]like full time work, or they have family commitments.[1717.25] [1717.25][S02]They're quite short on time and stressed.[1720.07] [1720.07][S02]And so when they're learning, naturally,[1723.93] [1723.93][S02]sometimes they are just in a bad state, and there's no one--[1729.21] [1729.21][S02]maybe they're studying at 11:00 PM, and they just need to vent.[1733.74] [1733.74][S02]And the only thing that they can vent to[1735.54] [1735.54][S02]is this AI tutor that's sitting in front of them on the screen.[1739.0] [1739.0][S02]So we find these kind of emotional outbursts[1742.83] [1742.83][S02]like, I am so stressed.[1746.32] [1746.32][S02]I'm really struggling here.[1748.45] [1748.45][S02]Will I ever be able to solve this problem?[1750.7] [1750.7][S02]Maybe I should just quit.[1751.87] [1751.87][S02]And the tutor can't ignore these messages.[1755.74] [1755.74][S02]They can't just say, sorry.[1757.65] [1757.65][S02]I can't answer this.[1759.15] [1759.15][S02]It really needs to engage and say[1761.01] [1761.01][S02]something that connects with the user in this very[1764.91] [1764.91][S02]vulnerable state.[1766.23] [1766.23][S02]So what our tutor is trained to do--[1768.45] [1768.45][S02]and we see transcripts like this coming in--[1772.32] [1772.32][S02]is like it's fine to feel this way.[1776.2] [1776.2][S02]Everyone feels this way.[1778.53] [1778.53][S02]We can get through this together.[1780.31] [1780.31][S02]There are resources that can help you and things like that.[1782.817] [1782.817][S01] I know that you've written that an AI tutor should[1785.4] [1785.4][S01]be careful about sensitive self-disclosure,[1788.2] [1788.2][S01]I guess particularly in that sort of a setting.[1790.18] [1790.18][S01]What did you mean by that?[1792.09] [1792.09][S02] So when people speak to each other,[1796.33] [1796.33][S02]what often happens is maybe one of the conversation[1802.44] [1802.44][S02]partners will say something personal[1805.05] [1805.05][S02]and maybe mention a personal fact.[1807.12] [1807.12][S02]And that encourages the other person[1808.62] [1808.62][S02]to also open up and share something about them.[1812.68] [1812.68][S02]And through this, they build trust[1815.16] [1815.16][S02]and a connection that helps the conversation move forward.[1820.38] [1820.38][S02]And when a learner mentions something so personal about how[1824.97] [1824.97][S02]stressed they are, it's almost natural[1828.98] [1828.98][S02]that they would expect the tutor to share back.[1832.18] [1832.18][S02]But then, of course, the tutor doesn't[1833.93] [1833.93][S02]have a stressful situation from their past that they can share.[1837.86] [1837.86][S02]Anything they self-disclose like that would be effectively a lie.[1843.05] [1843.05][S02]So there's this very kind of thin line[1846.59] [1846.59][S02]that we have to walk where the tutor needs[1851.18] [1851.18][S02]to maintain the connection and make sure that they support[1856.65] [1856.65][S02]the learner but at the same time not mislead them[1861.42] [1861.42][S02]and not create a connection which shouldn't exist[1865.77] [1865.77][S02]between a human and an AI.[1868.11] [1868.11][S01] So at no point can it pretend to be another human,[1871.21] [1871.21][S01]but it needs to understand how to empathize[1873.48] [1873.48][S01]with a human student.[1875.48] [1875.48][S02] Exactly.[1876.45] [1876.45][S01] But then.[1877.47] [1877.47][S01]OK, I sort of wonder.[1878.84] [1878.84][S01]There's something really interesting there[1880.59] [1880.59][S01]about the correct amount of anthropomorphization.[1883.97] [1883.97][S01]Are there some advantages to students[1885.968] [1885.968][S01]knowing that it's an AI, knowing that there[1887.76] [1887.76][S01]isn't a human at the other end?[1889.052] [1889.052][S01]Are students more comfortable making mistakes in front[1892.47] [1892.47][S01]of the AI, for instance?[1893.732] [1893.732][S02] Yes, for sure.[1894.94] [1894.94][S02]So something we've heard from students[1897.39] [1897.39][S02]is that they feel much more comfortable asking[1901.89] [1901.89][S02]what they might perceive as a silly question to AI[1905.46] [1905.46][S02]tutors just because they don't feel judged as you do when there[1909.99] [1909.99][S02]is a human on the other side.[1912.3] [1912.3][S02]Also, when you're in a class, and you could ask a question,[1918.5] [1918.5][S02]but then there's also peer judgment.[1920.29] [1920.29][S02]In this one-on-eon setting with an AI tutor,[1923.36] [1923.36][S02]you can basically say anything, and it's going to be fine.[1926.48] [1926.48][S02]So we find that the learners really appreciate that.[1929.787] [1929.787][S01] But then what about trust?[1931.37] [1931.37][S01]Do you find that people end up believing the AI more than they[1934.72] [1934.72][S01]would a sort of human tutor?[1938.93] [1938.93][S02] Sometimes we do.[1941.12] [1941.12][S02]So we had this very interesting situation[1944.06] [1944.06][S02]where, in the very first stages of developing the AI tutor,[1948.66] [1948.66][S02]we wanted to test it out how it compares to human teachers.[1952.29] [1952.29][S02]So we connected page raters who were told, look.[1957.51] [1957.51][S02]You have this opportunity to learn different subjects.[1961.67] [1961.67][S02]You will get connected to a tutor.[1963.99] [1963.99][S02]And we didn't tell them whether it was an AI or a human.[1968.24] [1968.24][S02]And just have fun, enjoy the learning experience.[1973.04] [1973.04][S02]And after that, they were given a questionnaire.[1976.2] [1976.2][S02]And in this questionnaire, we asked things[1977.96] [1977.96][S02]like how much do you think you've learned?[1980.13] [1980.13][S02]How much did you enjoy the experience?[1982.43] [1982.43][S02]And so this was the very first version[1986.33] [1986.33][S02]of our tutor, which we knew was quite bad.[1989.81] [1989.81][S02]And we found, surprisingly, that the learners reported[1993.62] [1993.62][S02]having learned more with an AI tutor than a human,[1997.53] [1997.53][S02]so that seemed strange.[2000.11] [2000.11][S02]So we decided to look through the transcript to understand[2003.23] [2003.23][S02]what is going on there.[2004.5] [2004.5][S02]And we found that the AI tutor hallucinated[2007.91] [2007.91][S02]all sorts of interesting, surprising facts[2012.23] [2012.23][S02]that, of course, as a learner, pretty much everything[2015.95] [2015.95][S02]the tutor says sounds like, I did not expect that.[2018.96] [2018.96][S02]This is a fun fact I just learned today.[2021.66] [2021.66][S02]So of course, they were very impressed[2024.05] [2024.05][S02]and felt like they've learned more.[2026.1] [2026.1][S02]But in fact, this is not something[2030.5] [2030.5][S02]that the tutor should be doing and definitely something[2033.41] [2033.41][S02]that we worked on to address in the future iterations.[2036.63] [2036.63][S01] Is that a concern going forwards?[2038.52] [2038.52][S01]I mean, the idea of hallucinations and people[2041.09] [2041.09][S01]mistaking those for real knowledge.[2043.74] [2043.74][S02] It is definitely a concern.[2045.49] [2045.49][S02]The base technology is getting better at factuality.[2048.949] [2048.949][S02]And also with education, because we're teaching some material[2053.84] [2053.84][S02]that is known, so there's always some sort of grounding.[2059.321] [2059.321][S02]Our tutors avoid some of these issues of factuality[2064.52] [2064.52][S02]by just being able to say, I'm only[2067.1] [2067.1][S02]teaching you about this particular YouTube[2069.71] [2069.71][S02]video or this particular piece of text[2071.75] [2071.75][S02]that your teacher has provided and referring facts[2075.59] [2075.59][S02]towards that primary source.[2078.36] [2078.36][S02]So it's kind of gives the tutor less opportunity[2081.409] [2081.409][S02]to actually make things up.[2083.12] [2083.12][S01] I just wanted to think, actually,[2084.995] [2084.995][S01]also about the effect that large language models have had[2088.13] [2088.13][S01]on education more generally.[2090.03] [2090.03][S01]So outside of a specific AI tutor,[2092.1] [2092.1][S01]because there is a big question that everyone has been asking,[2094.76] [2094.76][S01]which is about putting in safeguards[2097.04] [2097.04][S01]to stop AI being used to cheat or to do people's exams for them[2105.56] [2105.56][S01]or do people's essays for them.[2106.92] [2106.92][S01]So what kind of safeguards can you put in?[2110.06] [2110.06][S02] This technology is so pervasive.[2112.59] [2112.59][S02]And we actually-- we talk to students[2115.7] [2115.7][S02]about their use of GenAI, and even we[2118.94] [2118.94][S02]were surprised by how much they used it.[2121.2] [2121.2][S02]So literally, they were saying that their screen[2123.92] [2123.92][S02]is kind of a lecture, then notes, and then[2127.4] [2127.4][S02]GenAI at the bottom.[2129.75] [2129.75][S02]So I think the technology is here to stay,[2133.1] [2133.1][S02]and it will be used by the learners.[2136.755] [2136.755][S02]What can be done is trying to encourage learners to critically[2144.43] [2144.43][S02]evaluate their responses, trying to maybe change how we evaluate.[2151.04] [2151.04][S02]What are the assignments so that it actually[2156.28] [2156.28][S02]works with the technology.[2158.35] [2158.35][S02]Because if you think about it, education[2160.43] [2160.43][S02]is preparing us for the real world.[2164.09] [2164.09][S02]And in the real world, I think the expectation[2166.96] [2166.96][S02]will be more and more to actually work[2170.68] [2170.68][S02]with this technology because it does help in many ways[2173.95] [2173.95][S02]and does make us more productive.[2175.73] [2175.73][S02]So it doesn't make sense to ban it during education[2179.77] [2179.77][S02]and then expect learners to know how to use it properly[2182.95] [2182.95][S02]in their work.[2184.4] [2184.4][S02]So maybe one way to think about it is--[2187.575] [2187.575][S02]and I think that's what we've heard from teachers--[2189.7] [2189.7][S02]is how to change assignments and the ways of teaching working[2194.69] [2194.69][S02]where GenAI is encouraged as a partner,[2199.27] [2199.27][S02]but the evaluation is done slightly differently.[2201.62] [2201.62][S02]So it's kind of calculators in the past where you're[2205.27] [2205.27][S02]allowed to use calculators in certain math exams,[2207.495] [2207.495][S02]but you're still expected to know[2208.87] [2208.87][S02]how to do these calculations without help.[2213.917] [2213.917][S01] I do wonder, in the longer term,[2215.75] [2215.75][S01]as we start to see, I don't know, like GenAI being[2218.53] [2218.53][S01]like the assistant at all times, whether we can end up building[2222.79] [2222.79][S01]a bit of a dependency on them.[2224.6] [2224.6][S01]I mean, do students end up with a feeling that they have mastery[2229.79] [2229.79][S01]when they don't?[2230.55] [2230.55][S01]Actually, it's the AI that's doing the work.[2233.06] [2233.06][S02] So I think there are two potential issues[2236.24] [2236.24][S02]here that you've identified.[2238.73] [2238.73][S02]One is this feeling of mastery when there isn't one.[2244.46] [2244.46][S02]And this is a very common factor in any kind of learning,[2249.81] [2249.81][S02]even if you're talking about traditional education.[2252.71] [2252.71][S02]For example, one of the things that students[2256.13] [2256.13][S02]do a lot in preparation for exam is just like reread their notes[2259.46] [2259.46][S02]or reread the textbook.[2261.03] [2261.03][S02]And that kind of creates a feeling[2263.15] [2263.15][S02]like they have mastered the material just because they're[2266.18] [2266.18][S02]so familiar with it.[2267.68] [2267.68][S02]But when they go into an exam, they actually[2270.32] [2270.32][S02]find that they can't remember the facts[2272.87] [2272.87][S02]and can't use the information well.[2275.27] [2275.27][S02]So this is one of the things that[2277.58] [2277.58][S02]is very well known to be kind of a bad educational strategy,[2281.42] [2281.42][S02]just rereading.[2283.22] [2283.22][S02]And we find the same with AI tutors,[2287.6] [2287.6][S02]where if we ask a learner how they thought[2291.26] [2291.26][S02]the conversation went and how much they think they've learned,[2296.75] [2296.75][S02]they can report really good satisfaction.[2302.8] [2302.8][S02]Whereas, if we give the same conversation to a teacher[2305.453] [2305.453][S02]and then ask them the same question, how[2307.12] [2307.12][S02]pedagogical was the tutor?[2308.63] [2308.63][S02]How well do you think that session went?[2310.88] [2310.88][S02]They could rate it very, very differently.[2313.07] [2313.07][S02]I guess the other factor is this question of dependency.[2319.96] [2319.96][S02]We definitely find that if the learners use GenAI a lot[2327.47] [2327.47][S02]during their studies, they feel like it's really[2331.7] [2331.7][S02]helping in the process.[2333.42] [2333.42][S02]And actually, studies show that it[2335.27] [2335.27][S02]does increase the success in exercises and marks.[2341.01] [2341.01][S02]But when it comes to an exam, actually,[2343.7] [2343.7][S02]the learner's performance drops.[2345.78] [2345.78][S02]And that's because during studies, they[2349.31] [2349.31][S02]get so dependent on the AI providing them with the answers[2354.02] [2354.02][S02]or even if it guides them, if it doesn't actually teach them[2358.49] [2358.49][S02]the right things, it might be like they're just outsourcing[2362.54] [2362.54][S02]their reasoning to the AI.[2364.64] [2364.64][S02]That can be a problem during exam conditions[2367.04] [2367.04][S02]where you don't have access to it anymore,[2369.24] [2369.24][S02]and you don't actually remember or know[2372.23] [2372.23][S02]how to reason through these problems on your own.[2375.05] [2375.05][S01] I guess because the best exams aren't just[2377.3] [2377.3][S01]testing knowledge.[2378.05] [2378.05][S01]They're also testing skill.[2379.41] [2379.41][S02] Exactly.[2381.107] [2381.107][S01] From all of your research, then,[2382.94] [2382.94][S01]what are the big conclusions that you[2385.05] [2385.05][S01]draw about how to create an effective AI tutor?[2388.98] [2388.98][S01]Do you reckon you've solved it?[2390.42] [2390.42][S02] No, nowhere near.[2392.64] [2392.64][S02]And I will say we just made the very first step,[2395.97] [2395.97][S02]and that step is kind of realizing how hard of a problem[2400.17] [2400.17][S02]this is.[2400.93] [2400.93][S02]I think when we first started doing this work,[2403.57] [2403.57][S02]we were naive and wide eyed thinking[2408.99] [2408.99][S02]that we will come in and solve it within a year.[2414.84] [2414.84][S02]But now, I think we have a better[2417.33] [2417.33][S02]idea of the scope of the problem and what are the main things[2423.38] [2423.38][S02]to address to start making meaningful progress.[2428.18] [2428.18][S02]And these are things like how do we know success?[2431.46] [2431.46][S02]What-- how do we measure pedagogy?[2436.37] [2436.37][S02]Why do we get the data?[2438.33] [2438.33][S02]How do we actually train these models?[2440.02] [2442.9][S02]And also how to engage the communities better[2446.95] [2446.95][S02]and who are we building for so that we are not accidentally[2454.45] [2454.45][S02]increasing the gap in education but are making meaningful steps[2459.52] [2459.52][S02]towards decreasing them.[2463.0] [2463.0][S02]So I think there is a very long road ahead of us.[2466.46] [2466.46][S02]And actually, we think that we really[2468.85] [2468.85][S02]need to bring all of the community to work[2472.498] [2472.498][S02]on this problem together.[2473.54] [2473.54][S02]So we are trying to create common benchmarks[2477.49] [2477.49][S02]that we can all climb together.[2480.445] [2480.445][S01] That was really nice.[2481.82] [2481.82][S01]It was really nice.[2482.69] [2482.69][S01]Thank you for joining me, Irina.[2485.265] [2485.265][S01]I was really struck in that conversation with Irina[2487.39] [2487.39][S01]by the notable shift in the sorts of problems that are being[2490.9] [2490.9][S01]considered in this building.[2492.44] [2492.44][S01]We've gone from dealing with definites like winning or losing[2495.61] [2495.61][S01]at chess or Go or recognizing cat or no cat[2498.64] [2498.64][S01]in images to education, a space with no absolutes,[2503.3] [2503.3][S01]only imperfect measures in every direction-- in what counts[2507.22] [2507.22][S01]as good teaching, in what counts as an effective learning[2510.16] [2510.16][S01]experience, in how to get the balance between knowledge[2513.61] [2513.61][S01]and skills and learning how to learn or to walk[2517.202] [2517.202][S01]the line between how much a tutor should prompt[2519.16] [2519.16][S01]and how much it should withhold, even[2521.41] [2521.41][S01]how human a tutor should be.[2523.85] [2523.85][S01]None of those questions have ground truths,[2527.1] [2527.1][S01]and that is what makes this challenge so incredibly[2530.24] [2530.24][S01]difficult, but also one, as beautifully demonstrated[2533.63] [2533.63][S01]by Irina there, which requires humility[2536.39] [2536.39][S01]and collaboration to solve.[2538.76] [2538.76][S01]You've been listening to Google DeepMind the Podcast[2541.55] [2541.55][S01]with me, Professor Hannah Fry.[2543.29] [2543.29][S01]If you enjoyed the episode, do subscribe[2545.54] [2545.54][S01]to our YouTube channel.[2546.8] [2546.8][S01]You can also find us on your favorite podcast platform.[2550.32] [2550.32][S01]And we have got plenty more episodes[2552.38] [2552.38][S01]on a whole range of topics to come,[2554.37] [2554.37][S01]so do check those out, too.[2556.43] [2556.43][MUSIC PLAYING][2559.78]"}
{"file_name": "audio/val_000008.wav", "transcription": "[8.075][S04] Wherever you turn in this building there are people playing.[12.579] [12.579][S04]Of course there’s the usual pool tables and ping pong tables, but there are also[17.784] [17.784][S04]Go boards, chess boards, vintage video game cabinets[22.856] [22.856][S04]and more recent strategy games like “Settlers of Catan,”[26.827] [26.827][S04]because here you see, games are good.[30.597] [32.065][S04]I’m Hannah Fry. I’m a mathematician who specializes[34.835] [34.835][S04]in studying patterns in human behavior.[36.803] [36.803][S04]And I’m also an author, a broadcaster and like rather a lot of us,[41.942] [41.942][S04]someone who is fascinated by the potential for artificial intelligence[46.046] [46.046][S04]and the places it can take us.[48.448] [50.517][S04]Over the last 12 months, I’ve bene working with DeepMind -[53.687] [53.687][S04]the London based lab that has been called[55.689] [55.689][S04]the Apollo project of artificial intelligence.[58.992] [60.06][S04]Ready to go?[60.994] [60.994][S06] 5, 4, 3, 2, 1…[65.566] [65.566][S05] Welcome to the DeepMind podcast.[68.135] [74.274][S06] Hi![74.575] [74.575][S04] Open another door, and what do you know?[77.344] [77.344][S04]There’s a chess grandmaster pondering his latest move.[81.248] [81.815][S04]In his early 20s, Matthew Sadler was one of Britain’s[84.952] [84.952][S04]all-time greatest chess players.[87.821] [87.821][S04]He was there at the heart of it all, as a professional player[91.592] [91.592][S04]when AI started moving in on human chess champions.[96.697] [96.697][S04]And if you know anything about chess at all, you’ll remember this:[101.201] [101.201][crowd cheering][102.603] [102.603][S06] A computer called Deep blue has made chess history[105.606] [105.606][S06]by defeating the world’s champion Garry Kasparov.[108.275] [108.275][S04] this is the moment in 1997 when man was defeated by machine.[113.046] [113.046][S06] It was the final game of 6 - and his opponent -[116.416] [116.416][S06]one of the most powerful super computers in the world[120.02] [120.02][S06]seemed to have the upper hand.[122.222] [122.856][S06]The game of chess, supposedly a true test of human intellect[127.594] [127.594][S06]will never be the same again.[130.197] [130.197][S04] The reigning chess champion Garry Kasparov[132.799] [132.799][S04]beaten in a devastating fashion[134.801] [134.801][S04]by a chess playing computer built by IBM known as Deep Blue.[139.94] [140.44][S04]One of those watching was Matthew Sadler.[142.376] [142.376][S08] The big thing about Deep Blue was that it wasn’t[144.344] [144.344][S08]actually stronger than Kasparov when it won.[146.446] [146.446][S08]So that was very annoying. Because ah everyone you know was saying[149.55] [149.55][S08]“Oh a computer has beaten Kasparov!”[150.884] [150.884][S08]And you wanted to say: “Yes but it wasn’t better.”[152.586] [152.586][S08]You know, Kasparov has got himself a bit psyched out,[155.355] [155.355][S08]and that you know Deep Blue played well and all of that,[158.458] [158.458][S08]and then at some stage, computers came along that were a lot stronger,[160.994] [160.994][S08]and then you could start using them,[162.396] [162.396][S08]and they started showing you stuff that you hadn’t seen before -[165.566] [165.566][S08]incredibly complicated tactical sequences where you’d say -[168.502] [168.502][S08]that never works, and yet it does![170.737] [170.737][S08]And then at that moment you sort of give it up[172.339] [172.339][S08]you sort of say okay, they’re better than me,[173.974] [173.974][S08]and you start appreciating what they bring to the game.[176.743] [178.145][S04] And so as the story of Deep Blue’s victory over Kasparov[181.548] [181.548][S04]faded into AI folklore, people started to wonder:[185.853] [185.853][S04]what game would be the next frontier for AI research?[189.756] [191.825][S04]The most ambitious had their eye on the ancient board game Go.[196.763] [196.763][S04]This is not a game that responds to brute force calculation,[200.534] [200.534][S04]it requires intuition and an instinctive appreciation of positions and beauty.[206.373] [206.373][S04]Unlike chess whereby 2016 even a mobile phone[209.91] [209.91][S04]could play a credible game against a grand master,[212.88] [212.88][S04]there was nothing that came close to playing at the top level of Go.[217.985] [217.985][S04]But that didn’t put off one man from the challenge -[220.387] [220.387][S04]David Silver, lead researcher at DeepMind.[223.624] [223.624][S09] I’ve always been an ambitious person,[225.192] [225.192][S09]so I think at the beginning of my PhD I set out[228.228] [228.228][S09]on this on this goal for myself,[230.03] [230.03][S09]to be able to beat the world’s strongest players during my PhD[234.668] [234.668][S09]which turned out to be a little bit ambitious -[236.67] [236.67][S09]everyone was trying to dissuade me from this course.[238.805] [238.805][S09]The head of department took me aside and said:[240.807] [240.807][S09]“Look, you know you’re just wasting your time working on this project[243.377] [243.377][S09]it’s too hard. No one will be able to do this.”[245.512] [245.512][S09]He’s seen too many people bang their heads against this problem and fail,[248.448] [248.448][S09]and he didn’t want to see someone else um in his mind[251.818] [251.818][S09]you know just bang their heads against a problem that was too hard.[254.655] [254.655][S04] He wanted you to be employable at the end of your PhD[256.356] [256.356][S09] That’s right, yeah, that’s right.[258.525] [258.525][S04] Go is an ancient Chinese board game[261.962] [261.962][S04]and although it’s not played much in the West,[263.697] [263.697][S04]it’s arguably the most popular board game in the world.[267.701] [267.701][S04]It’s considered one of the four ancient scholarly skills of China,[272.306] [272.306][S04]and it’s taught in school alongside sports or maths.[276.21] [276.21][S04]It’s played on a pale, wooden board, engraved with an elegant grid[280.38] [280.38][S04]of 19 x 19 squares. The objective is very simple -[285.786] [285.786][S04]both players are aiming to capture territory on the board[289.022] [289.022][S04]by enclosing it with their pieces - black or white, known as stones.[294.862] [294.862][S04]But the game itself is mind-bogglingly sophisticated.[299.833] [299.833][S04]Much much more than Chess.[302.836] [303.57][S04] Even so, David Silver was single minded in his belief[307.14] [307.14][S04]that AI could master the game Go. It was simply a question of how.[312.179] [312.179][S09] The right approach it always seemed to me[313.747] [313.747][S09]was to allow machines to learn for themselves this kind of intuition[319.887] [319.887][S09]- to learn from themselves to be able to look at a position[322.99] [322.99][S09]and establish whether black or white is ahead -[325.826] [325.826][S09]and this meant machine learning in particular a method within machine[329.897] [329.897][S09]learning called reinforcement learning which is supposed to be how humans[333.166] [333.166][S09]and animals learn for themselves through trial and error experience.[337.137] [339.806][S04] But reinforcement learning on its own wouldn't be enough.[342.976] [342.976][S04]It was only when David began to work with the team at DeepMind[346.48] [346.48][S04]that he spotted the missing piece of the puzzle.[349.249] [349.249][S09] We’d seen a huge revolution with something called deep learning.[352.419] [352.419][S09]This is the ability for machines to build up very rich,[356.49] [356.49][S09]deep layered representation of knowledge for themselves[360.527] [360.527][S09]and that breakthrough seemed to me to be the missing element.[363.797] [363.797][S09]That if we could combine that process of being able to build[366.6] [366.6][S09]these very rich representations of knowledge -[369.203] [369.203][S09]with the kind of work which I’d been doing before in reinforcement learning -[372.84] [372.84][S09]the ability for machines to learn for themselves in trial and error -[376.577] [376.577][S09]if we put those two pieces together, it seemed to me that that this was[379.546] [379.546][S09]the recipe that might have the legs to take us all the way.[382.983] [382.983][S04] And by the way David means building an AI good enough[387.254] [387.254][S04]to challenge the very best Go players in the world.[390.724] [390.724][S03] This is a huge moment for both the world of artificial intelligence[394.661] [394.661][S03]and I think the world of Go. So far, Alpha[397.998] [397.998][S03]Go has beaten every challenge we have given it,[400.601] [400.601][S03]but we won’t know it’s true strength[402.236] [402.236][S03]until we play somebody who is at the top of the world like Lee[405.806] [405.806][S03]Sedol.[407.508] [410.544][S04] Seoul, March 2016 - a decade after first[414.214] [414.214][S04]toying with the idea of designing a go-getting machine,[418.085] [418.085][S04]David’s moment had come. DeepMind’s contender the AI program[422.322] [422.322][S04]AlphaGo would battle the 18 time world champion Lee[426.493] [426.493][S04]Sedol, head to head in a televised match of 5 games[431.465] [431.465][S04]as the world’s press watched on with bated breath.[434.935] [435.536][S09] I think the honest truth was that when I flew out to[439.54] [439.54][S09]Seoul in 2016, I was living this very protected life[445.445] [445.445][S09]as a researcher working on this problem behind closed doors[450.918] [450.918][S09]just thinking about the complexities of the problem[452.719] [452.719][S09]and how to how to make the system work, and it wash:[455.522] [455.522][S09]W only when I stepped off the plane and I walked into this hotel room[459.459] [459.459][S09]that was absolutely jam packed with reporters[462.396] [462.396][S09]and everything going on that suddenly the penny drops[466.633] [466.633][S09]that this was a really big deal that that actually you know[470.704] [470.704][S09]the consequences of this were far greater[473.373] [473.373][S09]reaching than I ever imagined they were - something like 100 million[476.777] [476.777][S09]people watching the match as it proceeded[480.781] [480.781][S09]or something like 30,000 articles written about the match.[484.384] [484.384][S04] When you arrived in Korea did you believe that your algorithm would win?[491.325] [491.325][S09] When we arrived in Korea, we actually got the team together,[494.061] [494.061][S09]and I asked the team to hold out a hand, we were playing 5 games,[497.497] [497.497][S09]and I asked everyone to hold out a hand to say how many games of the 5[501.235] [501.235][S09]we thought we would win[502.736] [502.736][S09]and you know many people made many different predictions,[505.739] [505.739][S09]I actually predicted 4 - 1. [laughs][508.342] [508.342][S04] David’s prediction of 4 - 1 to AlphaGo[510.978] [510.978][S04]was a clear expression of confidence.[513.847] [513.847][S04]But once the match started, the doubts started creeping in.[517.985] [517.985][S09] I feel I made the mistake of underestimating[522.623] [522.623][S09]the quality of a real human world champion[525.592] [525.592][S09]when we were actually playing the match,[527.761] [527.761][S09]I realized just how immensely versatile Lee[532.599] [532.599][S09]Sedol was as a player in his ability to push AlphaGo to its limits[537.704] [537.704][S09]not just in one game, but then coming along again[540.574] [540.574][S09]the next game and trying a very different strategy,[543.143] [543.143][S09]and then the next game trying a different strategy -[544.845] [544.845][S09]like pushing and probing for weaknesses all the way through,[547.848] [547.848][S09]and we were pushing AlphaGo into regimes that we never tested actually.[551.952] [553.086][S04] I don’t know if you’ve ever watched a televised game of[555.656] [555.656][S04]Go, but as a single stone is placed calmly onto the board,[560.561] [560.561][S04]the commentators and the audience watching on react[563.997] [563.997][S04]with the same ferocity of emotion and excitement[567.367] [567.367][S04]as they would a football game.[569.87] [569.87][S04]Even so, there was one moment during the first game[573.574] [573.574][S04]where the response from the crowd really stood out.[577.144] [577.144][S04]A moment when even Lee Sedol’s expression fixed into a look of shock -[582.716] [582.716][S04]his mouth fell open and his hand came up to his face.[587.387] [587.387][S09] Everyone’s expectation had been that Lee[590.09] [590.09][S09]Sedol would eventually emerge triumphant.[592.059] [592.059][S09]It was just a matter of time until AlphaGo made a mistake.[595.529] [595.529][S09]The game, apparently to the human commentators, was still balanced.[599.499] [599.499][S09]And at that point in time, AlphaGo made an extremely bold move.[605.005] [605.005][S09]It’s quite nice to be here at the heart of the operation…[607.741] [607.741][S04] David is backstage with Demis Hassabis,[609.576] [609.576][S04]one of the founders of DeepMind - watching AlphaGo’s every move.[614.581] [614.581][S04]And their reaction is caught on camera.[617.784] [617.784][S03] He’s done it - he’s gone in. Oh look[620.487] [620.487][S09] Look at his face! Look at his face![622.222] [623.891][S03] That is not a confident face. He’s pretty horrified by that.[628.095] [628.629][S09] In Go terms, it invaded in something which appeared to be[633.6] [633.6][S09]Lee Sedol’s territory and AlphaGo jumped right inside the region[637.137] [637.137][S09]which seemed liked it belonged to Lee Sedol and said[639.706] [639.706][S09]Okay, come and get me, and it was an audacious move[644.845] [644.845][S09]and you could judge by Lee[646.446] [646.446][S09]Sedol’s reaction that he wasn’t expecting it.[648.916] [648.916][S09]He was expecting perhaps more timid, more computer like response,[653.353] [653.353][S09]and the reality was that AlphaGo was using its intuition to judge[657.257] [657.257][S09]that if it jumped in here,[658.859] [658.859][S09]it couldn’t compute all the way to the end all of these possible outcomes,[661.628] [661.628][S09]but it had a sense that this would work out well for it.[664.598] [664.598][S04] The first round was a convincing victory for AlphaGo.[668.602] [670.671][S04]Roll on Day 2, round 2. And AlphaGo had another surprise up its sleeve.[676.677] [676.677][S09] In the second game, the human commentators were actually[680.814] [680.814][S09]I mean the only word I can think of is gobsmacked in their reaction.[683.584] [683.584][S02] Ahh… [Korean][688.288] [688.288][S02] That’s a very, that’s a very surprising move. [laughs][690.924] [690.924][S02] I thought it was I thought it was a mistake.[694.228] [694.228][S04] This was the now famous Move 37.[697.798] [697.798][S04]AlphaGo had placed a stone on the 5th line -[700.634] [700.634][S04]a move that no human player would even consider.[704.271] [704.271][S09] There’s this deeply built in beliefs about the game[707.241] [707.241][S09]and one of them is that in the game of[709.142] [709.142][S09]Go, you can think of all these different lines[711.445] [711.445][S09]upon which stones could be played - the first line is closest to the edge,[714.181] [714.181][S09]the second line, the fourth line, and there’s a rule in the game of[719.086] [719.086][S09]Go which is that when you approach one of these stones with diagonally,[723.423] [723.423][S09]it’s called a shoulder hit,[724.725] [724.725][S09]that you never ever do your shoulder hit above the fourth line,[729.997] [729.997][S09]and this has just been so ingrained in in[732.099] [732.099][S09]Go knowledge because most of the time it’s true.[734.134] [734.134][S09]Right, most of the time this is a very useful[736.236] [736.236][S09]common sense piece of knowledge which helps[737.905] [737.905][S09]Go player to exclude a vast range of very bad moves,[742.109] [742.109][S09]but in this particular position,[743.477] [743.477][S09]what AlphaGo realized was that playing on the 5th line[746.446] [746.446][S09]and playing the shoulder hit on the 5th line[748.482] [748.482][S09]actually just worked beautifully in the context of this position[751.018] [751.018][S09]with all of its other stones[752.586] [752.586][S09]in such a way that the outcome was really favorable.[754.821] [754.821][S04] Because in the end of that game that stone ended up being instrumental,[759.259] [759.259][S04]right, kind of joined up to everything else?[760.594] [760.594][S09] That’s right. Yeah that stone became so influential in the game[763.297] [763.297][S09]and it just worked forming this big net[765.065] [765.065][S09]that surrounded a vast swath of territory in the center.[767.868] [769.303][S04] AlphaGo just wasn’t playing in a mechanical way -[772.472] [772.472][S04]it was breaking the norms and conventions of this ancient game.[776.844] [776.844][S04]It was creating something -[778.045] [778.045][S04]a pattern of playing that went way beyond the approaches[781.315] [781.315][S04]that humans had ever considered - and it was doing so successfully.[786.587] [786.587][S04]Move 37 would eventually seal victory for the machine.[791.525] [791.525][S04]Looking back on the match, Lee Sedol spoke about how this very move[795.395] [795.395][S04]shifted his entire view of the map.[797.998] [797.998][S04]Archive audio of Lee Sedol’s voice.[800.567] [800.567][S06] I thought AlphaGo was based on probability calculation,[804.271] [804.271][S06]and it was merely a machine.[806.273] [807.241][S06]But when I saw this move, I changed my mind.[810.611] [812.179][S06]Surely AlphaGo was creative. This move was really creative and beautiful.[818.452] [820.153][S04] Do you think that was the AI illustrating real creativity?[824.291] [824.291][S09] I think we need to challenge ourselves to ask: what is creativity?[828.128] [828.128][S09]I think creativity should be defined[831.465] [831.465][S09]as anything which takes us out of our expected patterns of behavior,[837.204] [837.204][S09]and I think that in in that sense it truly was creative.[841.008] [841.008][S04] AlphaGo won the first three games, but the match wasn’t over for Lee[845.345] [845.345][S04]Sedol just yet.[847.181] [847.181][S04]In the fourth game, he managed to come back fighting against his opponent.[852.186] [852.186][S09] Lee Sedol was a true gentlemen[854.154] [854.154][S09]and we couldn’t have chosen anyone better[855.889] [855.889][S09]to represent humankind for this match.[858.659] [858.659][S09]He not only strove his utmost to the very end to play[862.796] [862.796][S09]and devise all kinds of amazing counter strategies to AlphaGo,[867.968] [867.968][S09]but he dealt with the immense pressure of having all of these people[871.505] [871.505][S09]watch him in really[873.674] [873.674][S09]ah profoundly human way - he found it very hard, he was humble,[879.479] [879.479][S09]I think it hurt his pride to lose to the computer,[882.482] [882.482][S09]but he came back, he found new strength in that[886.453] [886.453][S09]and he was able to ultimately emerge with immense pride at having beaten[892.292] [892.292][S09]AlphaGo in one game, and being part of this pivotal moment for AI.[896.363] [897.097][S04] At the end of the 6 days, the final score was AlphaGo[900.434] [900.434][S04]4, Lee Sedol 1.[903.237] [903.237][S09] So AlphaGo became the first computer champion of the game of Go[908.876] [908.876][S09]and it was the a major result for artificial intelligence.[913.914] [913.914][S04] And you won the sweepstakes.[915.716] [915.716][S09] And I won the sweepstakes.[917.417] [920.153][S04] This is the DeepMind podcast,[921.989] [921.989][S04]an introduction to AI research from the people behind AlphaGo.[926.426] [928.262][S04]News of the match rippled around the world.[930.931] [930.931][S04]Matt Botvinick, now DeepMind’s director or neuroscience research,[934.768] [934.768][S04]was one of the millions watching.[936.904] [936.904][S01] The first reaction that people had in the Go community was:[940.574] [940.574][S01]Oh, gee, it feels a little sad that now there’s a computer program[943.944] [943.944][S01]that could beat our hero.[945.746] [945.746][S01]But then it didn’t take long before people started to realize:[949.049] [949.049][S01]wait a minute, this is actually really exciting![951.518] [951.518][S01]We’re not stuck with our own limitations[953.654] [953.654][S01]in in terms of seeing the possibilities of how to play this game.[957.191] [957.191][S01]Now there are new horizons opened up to us.[959.326] [959.326][S01]We can find new forms of beauty in this game.[961.828] [961.828][S01]I think that’s sort of a microcosm now what I think[965.232] [965.232][S01]what we can hope for from AI more generally.[968.168] [969.837][S04] And for David this victory was always part of a bigger picture.[974.208] [974.208][S09] It’s not really the case that I’ve ever had to stop and question[977.311] [977.311][S09]and say well what next. Because the what next is clear -[979.78] [979.78][S09]we want to take this further, we want to build machines[982.816] [982.816][S09]that can achieve the same level of performance[985.219] [985.219][S09]but across all kinds of challenging domains.[987.721] [987.721][S09]Why stop with Go?[989.556] [989.556][S04] You once said that ah that you think that the game of[992.059] [992.059][S04]Go is the holy Grail of Artificial Intelligence -[995.462] [995.462][S04]do you still think that that’s the case?[997.331] [997.331][S09] I think the history of AI has been a number of pivotal moments[1002.302] [1002.302][S09]where for a period of time, a particular domain[1005.639] [1005.639][S09]has been the centerpiece of everyone’s attention, um,[1009.843] [1009.843][S09]so for a while the centerpiece of attention was the game of chess.[1013.046] [1013.046][S09]When Deep Blue defeated Garry Kasparov,[1016.25] [1016.25][S09]that marked the end of the era when chess was no longer the domain[1019.086] [1019.086][S09]that people cared about and the world moved on.[1022.556] [1022.556][S04] But before the world moved on from[1024.658] [1024.658][S04]Go, David was curious about just how far he could push the AI.[1030.063] [1030.063][S09] Really the open question for me was:[1032.199] [1032.199][S09]How could a system learn from itself entirely, with no human input.[1036.37] [1036.37][S09]If there was no human supervisor there to say here’s the inputs,[1039.406] [1039.406][S09]here’s the guidance, here’s the examples of how humans play.[1043.31] [1043.31][S09]What if started, really, tabula rosa which means start from a blank slate -[1048.182] [1048.182][S09]and the system just has to learn everything for itself -[1050.384] [1050.384][S09]starting from completely random play - is it able to learn for itself to play[1055.923] [1055.923][S09]Go to the highest caliber of play that’s possible?[1059.159] [1061.528][S04] Since their triumph in 2016, David and his team have been busy[1065.465] [1065.465][S04]working on that new algorithm - AlphaZero.[1070.17] [1070.17][S04]The original Go beta AlphaGo learned to play[1072.806] [1072.806][S04]by studying millions of games played by human experts -[1076.877] [1076.877][S04]AlphaZero on the other hand learns completely from scratch,[1081.548] [1081.548][S04]from zero human knowledge.[1084.151] [1084.151][S04]Instead if picks up the game by playing against itself millions of times.[1090.057] [1090.057][S04]Initially its game play is weak and erratic[1093.66] [1093.66][S04]but over time the system learns to identify the best moves and strategies.[1100.4] [1100.4][S09] It tries something and if a particular pattern is successful[1103.537] [1103.537][S09]and it ends up winning the game against itself,[1105.472] [1105.472][S09]it uses that pattern more.[1106.94] [1106.94][S09]And if another pattern ends up causing it to lose the game,[1109.51] [1109.51][S09]it will play that pattern less, and over time,[1112.713] [1112.713][S09]it builds up this very rich deep representation of knowledge,[1115.582] [1115.582][S09]one of these so-called neural networks -[1117.818] [1117.818][S09]and it’s able to then go out and beat the world’s strongest programs.[1122.523] [1122.523][S04] Which is better - AlphaGo or AlphaZero at Go?[1126.527] [1127.127][S09] Amazingly we discovered that the system[1129.596] [1129.596][S09]which had learned completely for itself[1131.665] [1131.665][S09]without a single piece of human knowledge[1133.834] [1133.834][S09]ended up being far stronger in the long run[1136.77] [1136.77][S09]and defeated the original version of AlphaGo by hundred games to zero.[1140.874] [1140.874][S04] ah, I didn’t know that![1142.976] [1142.976][S04]Oh my god - so hang on, you weakened it by giving it human knowledge.[1147.648] [1147.648][S09] [laughs]. It turns out we have a tendency as human designers[1153.854] [1153.854][S09]to believe that we know how to make the system stronger.[1157.024] [1157.024][S09]But quite often it turns out that by putting our own predisposition[1160.794] [1160.794][S09]and preferences into our programs, we actually make them weaker.[1164.831] [1172.105][S04] Without needing any human input, AlphaZero doesn’t particularly care[1176.543] [1176.543][S04]what game it’s playing as long as you can give it the rules,[1180.147] [1180.147][S04]it’s by no means limited to Go.[1182.416] [1185.285][S04]By now AlphaZero has taught itself from scratch[1189.523] [1189.523][S04]how to master the Japanese game of Shogi.[1193.026] [1193.026][S04]It is currently the world’s best Shogi playing machine[1197.297] [1197.297][S04]despite their only being one human in the DeepMind building[1200.534] [1200.534][S04]who knows how to play.[1202.236] [1204.204][S04]And to come full-circle after only four hours of playing itself,[1208.509] [1208.509][S04]AlphaZero mastered the game of chess to a super human level.[1213.647] [1221.288][S04] So far just a small handful of chess greats[1224.324] [1224.324][S04]have been able to test their skills against the machine[1227.194] [1227.194][S04]including Chess Grandmaster Matthew Sadler who we met earlier.[1231.365] [1231.365][S04]And women’s international master Natasha Regan.[1235.602] [1235.602][S04]They’ve co-authored a book about AlphaZero called GameChanger[1239.406] [1239.406][S04]and outside of the research team at DeepMind -[1241.975] [1241.975][S04]they’ve probably spent more time with AlphaZero than anyone.[1247.114] [1247.114][S04]Here’s Natasha.[1248.615] [1248.615][S07] I played AlphaZero once um and it wasn’t a very long game [laughs].[1255.556] [1255.556][S04] How many moves did you get to? 20:56[1256.957] [1256.957][S07] Oh I think it would have been less than 20 anyway,[1260.294] [1260.294][S07]um and I would have to say it was very direct.[1263.83] [1263.83][S07]I played something sacrificial - I thought I’d try um an opening thing,[1267.401] [1267.401][S07]it might not know it and it exploited it very quickly[1271.772] [1271.772][S07]ah got its pieces out um attacking squares straight away[1275.275] [1275.275][S07]and um it won quite quickly.[1277.611] [1277.611][S04] [laughs] I can imagine. I wouldn’t get to 20,[1280.113] [1280.113][S04]20 moves if I played it.[1281.615] [1281.615][S04]I mean it would take me down much quicker than that.[1283.517] [1283.517][S07] It’s pretty strong, yeah.[1284.384] [1284.384][S06] It does have a very a very smooth, human style against me.[1287.554] [1287.554][S06]There was no need for it to do anything complicated,[1289.289] [1289.289][S06]it just ah played better moves over a long period of time[1292.492] [1292.492][S06]and just pushed me back gradually.[1294.294] [1294.294][S06]I wouldn’t have been able to guess that I was playing against a computer -[1296.63] [1296.63][S06]it would ah - if I had to guess a human -[1298.432] [1298.432][S06]it would have been someone like Carlsen or Kasparov,[1300.567] [1300.567][S06]these very smooth positional players who just beat you by playing good moves.[1304.838] [1305.973][S04] And that’s like AlphaGo before it is the key thing about[1309.977] [1309.977][S04]the playing style about the AI, they are not like IBM’s Deep Blue[1314.681] [1314.681][S04]that beat Garry Kasparov back in 1997, or any of its descendants,[1319.386] [1319.386][S04]they play with a very mechanical style, computer like style.[1323.39] [1323.39][S04]They are very defensive. They only ever take calculated risks.[1328.228] [1328.228][S04]But AlphaZero on the other hand -[1330.597] [1330.597][S06] It conducts its attack in quite a structured way.[1334.334] [1334.334][S06]So it takes account of the whole board and it tends not to get attacked itself,[1339.039] [1339.039][S06]so it gets to a position where its own position is quite stable and safe[1342.809] [1342.809][S06]and then it can bring all its pieces in a concerted way into doing an attack.[1347.347] [1347.347][S06] it’s doing what humans do[1348.615] [1348.615][S06]but only so much better, that’s the thing.[1350.617] [1350.617][S04] Putting it really bluntly there Matt,[1352.085] [1352.085][S04]is it actually doing original stuff?[1354.421] [1355.122][S01] Yes it is. We’ve been playing chess for you know, for 400 years.[1358.192] [1358.192][S01]So actually probably every single move on the board[1360.427] [1360.427][S01]has probably been played once, by someone, somewhere.[1362.763] [1362.763][S01]You can actually see my goodness,[1365.098] [1365.098][S01]somebody’s played 44 million games against itself,[1368.202] [1368.202][S01]it’s actually repeated our whole chess history for itself,[1372.339] [1372.339][S01]and in that time, it’s just identified[1374.942] [1374.942][S01]what the most important things are of all that we’ve discovered.[1377.644] [1377.644][S01]And that’s what makes style.[1379.346] [1382.549][S04] But AlphaZero has substance as well as style.[1386.186] [1386.186][S04]Right now in 2019, it’s simultaneously holds the titles[1391.091] [1391.091][S04]of being the best player in the world at Go, Shogi and Chess.[1396.53] [1396.53][S04]And it’s not just being greedy, the whole point of this project[1399.8] [1399.8][S04]was to build an intelligent system flexible[1402.603] [1402.603][S04]enough to respond to a range of problems,[1406.44] [1406.44][S04]and while AlphaZero might not quite be able[1408.709] [1408.709][S04]to tackle cancer diagnosis or energy efficient,[1412.713] [1412.713][S04]there is a good reason why DeepMind are playing with building[1415.549] [1415.549][S04]these all-purpose machines in the world of games.[1419.82] [1419.82][S09] our goal of course is not just to play Chess or Go[1423.023] [1423.023][S09]or whatever.[1424.291] [1424.291][S09]Our goal is to have impact on some of the world’s[1426.827] [1426.827][S09]most challenging problems which are facing society[1429.796] [1429.796][S09]but in order to do so, we need to gain understanding,[1432.366] [1432.366][S09]we need to really deeply understand these systems for ourselves first,[1436.37] [1436.37][S09]and games provide the perfect test bed for doing so.[1439.473] [1439.473][S04] Games are like the ultimate mini universe.[1442.609] [1442.609][S04]You know all the rules, there’s a clear winner at the end,[1445.746] [1445.746][S04]you can look back at the end of the game and decide what went wrong.[1449.216] [1449.216][S04]And if you lose, it doesn’t matter, you can just start another round.[1453.654] [1453.654][S04]The big problems though, the ones we eventually want to tackle,[1457.991] [1457.991][S04]they are quite a lot more complicated.[1461.195] [1461.195][S09] The real world’s a really messy place,[1463.197] [1463.197][S09]I mean you end up with these amazingly complex things -[1465.632] [1465.632][S09]like human beings and how they interact[1467.134] [1467.134][S09]with each other and societies and companies[1469.636] [1469.636][S09]and all these amazing things which we’ve built up in our world,[1472.372] [1472.372][S09]we need to be able to understand them[1473.674] [1473.674][S09]to be able to have a hope to apply something like AlphaGo to make progress.[1479.246] [1479.246][S09]In order to make progress then we need to be able to apply systems[1483.35] [1483.35][S09]that can operate even when the rules are unknown.[1486.82] [1486.82][S09]And that is a big remaining challenge[1489.423] [1489.423][S09]which is not yet been addressed by AlphaGo or AlphaZero.[1492.86] [1496.563][S04] If you want to know more about how games have been[1499.766] [1499.766][S04]and continue to be used as a test bed in AI research,[1504.037] [1504.037][S04]then head over to the show notes[1505.472] [1505.472][S04]where you can also explore the world of AI research beyond DeepMind.[1509.843] [1509.843][S04]And we’d welcome your feedback or your questions[1512.379] [1512.379][S04]on any aspects of Artificial Intelligence[1514.348] [1514.348][S04]that we’re covering in this series,[1516.35] [1516.35][S04]so if you want to join in the discussion or point us to stories or resources[1520.354] [1520.354][S04]that you think other listeners would find helpful,[1522.789] [1522.789][S04]then please let us know - you can message us on Twitter[1525.425] [1525.425][S04]or you can email us - podcast@deepmind.com[1528.695]"}
{"file_name": "audio/val_000009.wav", "transcription": "[0.0][S01] It was a professor[1.458] [1.458][S01]that said, oh, I bet if we can get robots to tie shoelaces,[5.13] [5.13][S01]I will retire.[6.0] [6.0][S01]And the researchers in the team were like, right on.[9.96] [9.96][S01]I'm going to add that as a task.[11.58] [11.58][S02] Do you think that's what[12.14] [12.14][S02]we're about to see, then, in the same way as we've[13.67] [13.67][S02]seen the explosion of large language models?[15.75] [15.75][S02]Do you think the next thing is the explosion of robotics?[18.54] [18.54][S01] We used to have discussions about[20.623] [20.623][S01]whether it would happen in our lifetime or even in our careers,[24.3] [24.3][S01]and now we have debates about whether it[26.39] [26.39][S01]would be 5 or 10 years.[28.44] [28.44][MUSIC PLAYING][31.533] [31.533][S02] Welcome back to \"Google DeepMind--[33.45] [33.45][S02]The Podcast.\"[34.14] [34.14][S02]I'm Professor Hannah Fry.[35.66] [35.66][S02]Sometimes, maybe even often, the terms AI and robot[40.01] [40.01][S02]are used interchangeably in casual conversation, people[43.58] [43.58][S02]talking about chatting to a robot on an app.[46.5] [46.5][S02]But robots have a physical body.[49.74] [49.74][S02]And here at Google DeepMind, they[51.62] [51.62][S02]care about robots with AI embedded in the real world.[56.37] [56.37][S02]And while AI has made huge strides,[59.28] [59.28][S02]embodied intelligence has lagged behind.[62.84] [62.84][S02]But perhaps all of that is about to change.[65.269] [65.269][S02]Carolina Parada leads robotics research[67.9] [67.9][S02]here at Google DeepMind, the international team[70.78] [70.78][S02]responsible for some extraordinary advances[73.51] [73.51][S02]in robotics, most recently Gemini Robotics, which[76.99] [76.99][S02]brings Gemini's multimodal understanding[79.33] [79.33][S02]to the physical world.[80.93] [80.93][S02]Welcome to the podcast, Carolina.[82.46] [82.46][S01] Thanks for having me.[83.6] [83.6][S02] Now, I know you've been[84.43] [84.43][S02]working with these robots for quite a long time.[87.05] [87.05][S02]How have you seen them evolve?[88.89] [88.89][S01] Yeah, it's been super exciting.[90.89] [90.89][S01]I have been excited about robotics[92.32] [92.32][S01]since I was 10 years old.[94.31] [94.31][S01]Super, super excited because of what I've seen in cartoons.[98.66] [98.66][S01]You see robots like Rosie the Robot helping do all the chores.[102.53] [102.53][S01]And as a kid you're like, of course,[104.03] [104.03][S01]that's what I want to build when I grow up.[105.822] [105.822][S01]And really, I've been at the Google DeepMind Robotics team[109.03] [109.03][S01]for about seven years, and, really,[111.02] [111.02][S01]things have changed dramatically in the last three years[113.83] [113.83][S01]in particular.[114.71] [114.71][S01]We've always believed from the very beginning[117.01] [117.01][S01]that AI was going to be completely[119.46] [119.46][S01]transformative to robotics.[120.585] [120.585][S01]I mean, there's a lot of robots out there[122.293] [122.293][S01]that are really helpful today.[123.64] [123.64][S01]There's robots in manufacturing lines.[125.65] [125.65][S01]There's robots that are navigating the moon.[128.02] [128.02][S01]There is robots that are in our oceans.[130.0] [130.0][S01]But these robots have been programmed[132.24] [132.24][S01]to do specifically those tasks.[134.08] [134.08][S01]They make a lot of assumptions about those environments[136.65] [136.65][S01]or the objects they might encounter,[139.01] [139.01][S01]or they might be remotely operated by humans.[141.355] [141.355][S01]But we have believed from the beginning[142.98] [142.98][S01]that AI is the way to transform robotics[145.83] [145.83][S01]so that we can build robots that are truly intelligent so[149.7] [149.7][S01]that they can interact with you, that they can reason[152.25] [152.25][S01]about their environment, and they[153.72] [153.72][S01]can take action in a way that feels very general.[156.187] [156.187][S01]So that has been our mission from the start.[158.02] [158.02][S01]And so I think three years ago, you[159.96] [159.96][S01]had Robotics in your podcast, and back[162.21] [162.21][S01]then, we were doing reinforcement[163.71] [163.71][S01]learning for robotics.[164.89] [164.89][S01]And so essentially, we were teaching robots[166.86] [166.86][S01]to stack blocks by giving them a simple reward, like a plus 1[171.55] [171.55][S01]if your tower got taller.[173.37] [173.37][S01]And we made some progress there.[175.69] [175.69][S01]But since we've been at the forefront of AI,[178.35] [178.35][S01]we've been bringing more and more of AI[181.16] [181.16][S01]into the entire world of robotics.[183.36] [183.36][S01]So about 2022, we introduced, for example, LLMs to robots.[187.885] [187.885][S01]And that was the first time that you could actually[190.01] [190.01][S01]talk to a robot and say something like, I'm thirsty,[193.25] [193.25][S01]and it would know what you meant.[194.94] [194.94][S01]And then later on, we brought VLMs[197.66] [197.66][S01]so the robot could understand natural language,[199.89] [199.89][S01]but it could also understand the visual input that it was getting[203.15] [203.15][S01]and then make decisions based on that.[205.28] [205.28][S01]And then in 2023, we introduced robotics transformers.[208.943] [208.943][S01]And this is the first time that the transformer architecture was[211.61] [211.61][S01]actually included in robotics.[213.45] [213.45][S01]And it basically showed us that robot performance scales[217.16] [217.16][S01]with data, and that essentially started[219.8] [219.8][S01]a new foundation or a new era of large-scale, data-driven robot[223.97] [223.97][S01]learning.[224.58] [224.58][S01]And then more recently, we introduced just now[226.52] [226.52][S01]Gemini Robotics, which is essentially[228.74] [228.74][S01]our most advanced model for actions.[231.21] [231.21][S01]And it essentially takes the multimodal world understanding[235.7] [235.7][S01]of Gemini and brings it to the physical world[238.34] [238.34][S01]by adding actions as a new modality in Gemini.[242.09] [242.09][S01]And that really enables models to be very general because it's[245.6] [245.6][S01]understanding the world through Gemini's understanding[248.48] [248.48][S01]and enables it to be interactive.[250.59] [250.59][S01]In fact, it can understand any language that Gemini supports[253.88] [253.88][S01]and enable it to be dexterous.[255.99] [255.99][S01]So it can still do very complex manipulation[258.95] [258.95][S01]while talking to you and also understanding[262.01] [262.01][S01]a completely new situation, which today is actually[264.89] [264.89][S01]very hard for robots to do.[266.78] [266.78][S02] In terms of your big goal, your big ambition,[269.82] [269.82][S02]how will we know when we get there?[271.65] [271.65][S01] I think it's definitely[273.317] [273.317][S01]going to be gradual, where robots are able to understand[276.89] [276.89][S01]a new situation and reason about something they need to do[280.34] [280.34][S01]that they haven't seen before.[282.0] [282.0][S01]And that's exactly what we're seeing right now.[284.82] [284.82][S01]But it's still going to be difficult[287.36] [287.36][S01]for them to learn more and more complex tasks.[290.18] [290.18][S01]In fact, that's what we see.[291.48] [291.48][S01]The robot can feel sort of a two-year-old toddler that can[295.7] [295.7][S01]understand its world around it.[297.39] [297.39][S01]It can start to play with objects.[299.37] [299.37][S01]It understands concepts.[301.05] [301.05][S01]But if you teach it to do something more complex,[303.68] [303.68][S01]like we have an example where we're teaching the robot[305.93] [305.93][S01]to do an origami fold, it actually[308.48] [308.48][S01]needs time to practice that.[310.29] [310.29][S01]And once it has more practice, in that case,[312.69] [312.69][S01]it can actually do it.[313.74] [313.74][S01]So that's roughly where we are today,[315.54] [315.54][S01]but that's far from where we need[317.18] [317.18][S01]to be if we want robots to be in everyday spaces[320.27] [320.27][S01]doing all kinds of tasks for us.[321.96] [321.96][S01]So there's still quite a bit to go.[323.73] [323.73][S02] I thought that what we could do[325.522] [325.522][S02]is take a little look at some of what these robots can[328.43] [328.43][S02]do because there's a video that you guys have recently released.[331.74] [331.74][S02]What we have here, then, is we have a humanoid robot who[335.63] [335.63][S02]is packing a lunch for its human, also playing noughts[340.57] [340.57][S02]and crosses.[341.07] [341.07][S02]Is it any good at the noughts and crosses?[342.58] [342.58][S01] [LAUGHS] I think[343.955] [343.955][S01]we still beat it because it's very simple understanding.[346.365] [346.365][S02] Tell you what it's doing, though,[348.24] [348.24][S02]is picking up the pieces and moving them around quite easily.[351.71] [351.71][S02]There's also a bit here where it can make its own anagram based[356.03] [356.03][S02]on tiles that appear.[357.693] [357.693][S02]What were you particularly impressed by?[359.36] [359.36][S01] I think that's what's[359.95] [359.95][S01]most exciting about these models,[361.78] [361.78][S01]is that, in many occasions, our own researchers were[364.51] [364.51][S01]excited and impressed by what it was doing.[366.65] [366.65][S01]And it was primarily because the way[369.55] [369.55][S01]we were testing it was by putting the robot in front[372.55] [372.55][S01]of situations that it's never seen before.[374.51] [374.51][S01]So even us didn't know whether the robot was going[376.9] [376.9][S01]to be able to get it right.[378.2] [378.2][S01]And in many occasions, it did.[380.41] [380.41][S01]So many of the examples that we show in this video,[383.51] [383.51][S01]as well as the other videos where[385.51] [385.51][S01]you have the two arms moving around,[388.63] [388.63][S01]is that it's actually understanding a complex concept.[391.64] [391.64][S01]So a really cool example that where we were all like, gasp,[395.71] [395.71][S01]was when we showed the video where the robot is actually[399.25] [399.25][S01]doing a slam dunk.[400.73] [400.73][S01]And what was cool about that case is that that day,[403.24] [403.24][S01]we were just having the creative team come and film the robots,[406.58] [406.58][S01]and we asked them to bring toys.[408.32] [408.32][S01]We didn't say anything else.[409.88] [409.88][S01]They're like, just bring toys to play with the robot.[412.15] [412.15][S02] All things the robot hadn't seen before.[413.73] [413.73][S01] All things-- yeah.[414.76] [414.76][S01]They had no idea what the robot was trained on.[416.718] [416.718][S01]So they actually brought this little basketball hoop[419.19] [419.19][S01]that was a little cute toy with a little ball,[421.99] [421.99][S01]and they put it in front of the robot.[423.88] [423.88][S01]Again, the robot had never seen anything related to basketball.[426.91] [426.91][S01]It certainly had never seen this toy.[428.74] [428.74][S01]And they asked it to do a slam dunk of the ball.[431.83] [431.83][S01]And we were all like, I have no idea if it will work.[434.53] [434.53][S01]And actually, it took not even a quarter of a second[438.18] [438.18][S01]and it actually decided to put the ball inside the basketball[441.87] [441.87][S01]hoop.[442.45] [442.45][S01]And we were all like, that's amazing.[444.49] [444.49][S01]And it was just essentially drawing[446.28] [446.28][S01]from Gemini's understanding of what basketball is[449.01] [449.01][S01]and what a slam dunk is, which is a concept[452.07] [452.07][S01]that we couldn't have thought of teaching it to do.[455.73] [455.73][S01]And it essentially did the right motion.[458.71] [458.71][S01]So that was a really cool example.[460.275] [460.275][S02] Talk to me about the packing lunch one.[462.4] [462.4][S02]It has a conceptual understanding[463.86] [463.86][S02]of what a banana is, for example.[465.61] [465.61][S02]Does it know how to grip a banana in the sense[468.017] [468.017][S02]that you can't grip a banana in quite the same way[470.1] [470.1][S02]as you could a clay pot or something even more fragile[472.96] [472.96][S02]than a banana?[473.543] [473.543][S01] Actually, one of the things that[475.585] [475.585][S01]is super impressive is that these robots[477.76] [477.76][S01]are extremely simple.[478.97] [478.97][S01]They actually don't have touch sensing.[481.1] [481.1][S01]They don't have depth sensing.[482.66] [482.66][S01]They don't have force sensing.[483.98] [483.98][S01]So they're literally doing eye-hand coordination[486.64] [486.64][S01]and using an understanding of how you grasp a banana.[489.62] [489.62][S01]So it actually is looking at the object and grasping it.[492.71] [492.71][S01]And once it sees that it has it in hand,[494.78] [494.78][S01]that's how it knows that it has detected it.[498.32] [498.32][S01]There's other robots out there that are much more complex.[500.78] [500.78][S01]But this forces the model to really reason about[503.98] [503.98][S01]what is seen and making a decision about how[506.29] [506.29][S01]to pick that up.[507.11] [507.11][S02] And that's the thing that's really original here.[509.36] [509.36][S01] Yeah, that is one of the many things,[511.03] [511.03][S01]is the fact that it's doing it not just because we taught it[513.94] [513.94][S01]1,000 times how to pick up a banana.[515.9] [515.9][S01]It's because it's pulling this out of it's understanding of how[519.52] [519.52][S01]to pick up objects from Gemini and then adapting it[523.96] [523.96][S01]to the world of actions.[524.987] [524.987][S02] I mean, there have been lots of videos doing[527.32] [527.32][S02]the rounds on the internet for a number of years of extremely[529.87] [529.87][S02]impressive-looking robots doing backflips and--[532.17] [532.17][S02]I don't know-- being kicked over and running up and down[535.23] [535.23][S02]mountains and things.[537.255] [537.255][S02]In comparison to those videos, picking up[540.3] [540.3][S02]and putting down a banana into a lunchbox[542.835] [542.835][S02]seems like quite a simple task.[545.71] [545.71][S02]But we're talking about a different type of robot here,[548.86] [548.86][S02]aren't we?[549.46] [549.46][S01] Yeah.[550.02] [550.02][S01]I mean, this is a completely different problem[551.937] [551.937][S01]you're trying to solve.[553.0] [553.0][S01]Many of those videos are basically rehearsed sequences[557.25] [557.25][S01]that the robot has learned and memorized,[559.18] [559.18][S01]and we're actually very impressed by them.[562.025] [562.025][S01]But it's a different problem that you're trying to solve.[564.4] [564.4][S01]What you're trying to solve here is for the robot[566.442] [566.442][S01]to reason about what it means to pack a lunch, given[569.687] [569.687][S01]the objects in front, what it needs to do in order[571.77] [571.77][S01]to put a piece of bread inside of a bag,[574.75] [574.75][S01]and then what it means to close it.[576.43] [576.43][S01]And it's never going to go as you[577.98] [577.98][S01]expect because these are very flexible things that[580.56] [580.56][S01]move around.[581.35] [581.35][S01]So it needs to react and respond to what's happening[583.74] [583.74][S01]and then actually complete the task.[585.672] [585.672][S02] It's that idea of generality.[587.38] [587.38][S01] That's right.[587.98] [587.98][S01]Yeah.[588.48] [588.48][S02] How do you decide--[590.3] [590.3][S02]or how do you compare one robot against another?[593.34] [593.34][S02]How do you decide whether this robot is doing[595.43] [595.43][S02]generally better than another?[597.373] [597.373][S01] That was actually[598.79] [598.79][S01]one of the things that was hard for us[601.07] [601.07][S01]to express when we were even recording[603.68] [603.68][S01]for the demos in this release.[606.06] [606.06][S01]A demo is, by definition, prescripted.[609.782] [609.782][S01]So we were like, this doesn't quite[611.24] [611.24][S01]capture what we want to share.[613.2] [613.2][S01]So what we did-- that's why we asked the team[615.23] [615.23][S01]to bring a bunch of toys and actually[617.09] [617.09][S01]start playing with the robots and see what emerges.[619.95] [619.95][S01]And the best way to capture it is[621.92] [621.92][S01]that we're able to change the behavior of the robot[625.22] [625.22][S01]by talking to it, and you can see that in the videos.[627.81] [627.81][S01]We're actually able to put raw objects[629.9] [629.9][S01]that it's never seen before.[631.53] [631.53][S01]And we move objects around to make sure[633.92] [633.92][S01]that people understand that this is actually[635.78] [635.78][S01]not a prescripted behavior.[638.34] [638.34][S01]In fact, in our benchmarks, we evaluate our models[641.84] [641.84][S01]in all kinds of ways in terms of generalization.[644.85] [644.85][S01]So we will change the visual background.[647.1] [647.1][S01]We will change the background.[648.36] [648.36][S01]The objects were new.[649.53] [649.53][S01]We will add objects to distract the robot.[652.97] [652.97][S01]We would also ask it to do completely new things,[655.71] [655.71][S01]or even you can talk to it in a different language.[657.84] [657.84][S01]So I could just give it the instruction in Spanish,[660.08] [660.08][S01]and it would just actually work.[662.243] [662.243][S02] I want to talk about interactivity, too.[664.41] [664.41][S02]Because in a few of your videos, there's[666.077] [666.077][S02]one where a human is sat at a desk,[668.66] [668.66][S02]and the robot is clearing up after him as he goes.[671.6] [671.6][S02]In another, you've got a human moving a cup around,[674.19] [674.19][S02]and the robot is chasing it, trying to put an object inside.[677.04] [677.04][S02]How much more difficult are those interactive scenarios[680.87] [680.87][S02]than just a static task?[682.183] [682.183][S01] They're significantly more advanced[684.35] [684.35][S01]behavior, and a lot of the interactivity[686.99] [686.99][S01]just fell out of the model.[688.83] [688.83][S01]We were not thinking, for example, how fast[692.09] [692.09][S01]can we move these objects before the robot would react.[695.88] [695.88][S01]We certainly knew that we wanted a model that[698.06] [698.06][S01]could react quickly.[699.54] [699.54][S01]But a lot of these examples that we posted in videos[702.14] [702.14][S01]just fell out of people playing with the model[704.24] [704.24][S01]and seeing how it would behave.[706.52] [706.52][S01]Same with organizing the desk.[708.798] [708.798][S01]That was actually someone playing with the robot,[710.84] [710.84][S01]deciding to see how much he could game it[714.52] [714.52][S01]until it actually was able to complete the full task.[718.22] [718.22][S01]So yeah, it actually is amazing to see[721.63] [721.63][S01]how a lot of these other capabilities that are already[724.18] [724.18][S01]there in Gemini are actually extremely valuable when[726.79] [726.79][S01]you bring them into a robot, which is now able to adapt based[731.26] [731.26][S01]on what you're saying.[732.283] [732.283][S01]So you could actually have a full conversation[734.2] [734.2][S01]and change the behavior of the robot as it's moving.[736.377] [736.377][S01]So you can say, I want you to do this.[737.96] [737.96][S01]Oh, no, actually, never mind.[739.07] [739.07][S01]I want you to do this other thing.[740.54] [740.54][S01]And it would actually just follow you.[743.09] [743.09][S01]It's actually kind of comical.[744.83] [744.83][S01]And then you could also change the objects around,[749.26] [749.26][S01]and it will just do it.[750.5] [750.5][S02] I think it's kind of a good job[751.87] [751.87][S02]sometimes that these robots don't have feelings[753.828] [753.828][S02]because it'd feel very forlorn, just being chased around[759.4] [759.4][S02]on a table by researchers.[760.867] [760.867][S01] It's super fun.[762.2] [762.2][S02] That's the large language model[764.41] [764.41][S02]sitting underneath it that's helping it do that, that's[766.81] [766.81][S02]giving it that conceptual understanding of the objects[768.7] [768.7][S02]that it's manipulating.[769.658] [769.658][S01] We're leveraging[771.033] [771.033][S01]Gemini's multimodal understanding[772.42] [772.42][S01]to take the visual input of the role[774.7] [774.7][S01]that the robot is seeing through its cameras[776.72] [776.72][S01]and the natural language that is hearing from the human[779.57] [779.57][S01]and then translate that into how to act.[783.1] [783.1][S01]And it actually also speaks back.[784.91] [784.91][S01]So you can ask it a question about whether it's done.[787.25] [787.25][S01]You can ask it a question about how far it[789.01] [789.01][S01]is in the process of folding an origami figure.[791.81] [791.81][S01]It actually understands that.[793.03] [793.03][S01]It can respond.[793.93] [793.93][S02] I remember when Gemini was first being launched,[796.43] [796.43][S02]and people went to great lengths to talk[799.87] [799.87][S02]about how it was multimodal.[802.54] [802.54][S02]Is this one of the main reasons?[804.71] [804.71][S02]Is this the payoff for putting in all of that extra groundwork[808.75] [808.75][S02]and making sure that it can understand videos and photos[812.44] [812.44][S02]and so on?[812.978] [812.978][S01] I mean, one of many.[814.52] [814.52][S01]I think us humans capture the world[817.75] [817.75][S01]through many different senses.[819.46] [819.46][S01]So I think it's super important, if you[821.2] [821.2][S01]want to build an intelligence as powerful as our brains,[824.68] [824.68][S01]to be able to take input in a multimodal way.[827.79] [827.79][S01]And definitely robotics is a perfect example[830.57] [830.57][S01]where you can see that it absolutely[832.19] [832.19][S01]requires to have understanding of natural language[834.77] [834.77][S01]and visual input and, presumably in the future,[837.51] [837.51][S01]also touch sensing in order to make decisions about how to act,[841.04] [841.04][S01]the same way humans do.[842.48] [842.48][S02] Why does it matter, though,[844.11] [844.11][S02]that robots should have a conceptual understanding of what[847.85] [847.85][S02]they're doing?[848.58] [848.58][S02]I mean, OK, maybe you wouldn't call them intelligent,[851.37] [851.37][S02]but there are robots like--[852.57] [852.57][S02]I don't know-- dishwashers or lawnmowers.[856.288] [856.288][S02]They don't have a conceptual understanding of what a plate is[858.83] [858.83][S02]or what grass is.[860.73] [860.73][S02]Is it actually necessary?[862.912] [862.912][S01] I'm sure that there are applications where[865.37] [865.37][S01]you can have a robot that can just repeat the actions,[868.26] [868.26][S01]and it will be just fine.[869.46] [869.46][S01]But we're interested in actually building robots that can really[874.31] [874.31][S01]reason and act in a very general way,[876.89] [876.89][S01]just because the world is really messy.[879.12] [879.12][S01]Things will never go exactly according to plan,[881.46] [881.46][S01]and there's a lot of tasks where things are constantly changing.[885.26] [885.26][S01]And it actually just opens up the opportunity of applications[888.67] [888.67][S01]for these robots.[889.61] [889.61][S01]They could literally be anywhere that a human[892.24] [892.24][S01]could be doing a task.[894.2] [894.2][S01]So that enables them to be helpful in home environments[897.38] [897.38][S01]but also in manufacturing environments.[899.835] [899.835][S02] There are some things[901.21] [901.21][S02]that are important in robotics that,[903.97] [903.97][S02]I think, now with the standard Gemini, actually are quite easy,[909.68] [909.68][S02]things like pointing or drawing bounding boxes.[913.01] [913.01][S02]Just explain to us what those are.[915.11] [915.11][S01] Well, basically this is one of the areas[917.56] [917.56][S01]that we actually had to improve Gemini in order[921.37] [921.37][S01]to help with robotics.[923.66] [923.66][S01]If you have, for example, an object in front of you,[927.14] [927.14][S01]what we mean by pointing is that I can literally identify[931.39] [931.39][S01]any point in that object.[933.38] [933.38][S01]So I can say, imagine that you have a T-shirt in front of you.[937.22] [937.22][S01]If I point to the collar, it should say, this is the collar.[940.13] [940.13][S01]Or if I say collar, it should identify where the collar is.[943.34] [943.34][S01]And you might imagine that this is not that important.[946.03] [946.03][S01]But actually, if you're trying to fold that T-shirt,[948.73] [948.73][S01]you need to know where the collar is,[950.62] [950.62][S01]where the bottom of the T-shirt is, and all[953.61] [953.61][S01]of the different components.[955.05] [955.05][S01]Bounding boxes, what it means is that you[957.57] [957.57][S01]can identify all the edges of that[959.67] [959.67][S01]object so that you know where the object ends[962.58] [962.58][S01]and the rest of the environment begins.[964.6] [964.6][S01]So these kinds of examples are, I think, trivial for us humans.[969.25] [969.25][S01]We don't even think about it.[970.6] [970.6][S01]But if robots are actually able to have access[973.17] [973.17][S01]to that kind of information, then they[974.85] [974.85][S01]can be smarter about the way that they take[978.63] [978.63][S01]action in the physical world.[979.93] [979.93][S01]This is what we call embodied reasoning, essentially.[981.85] [981.85][S02] How is it different from the kind of reasoning[984.267] [984.267][S02]that you get in the standard Gemini model?[986.35] [986.35][S01] We refer to embodied reasoning[988.308] [988.308][S01]as reasoning about the physical world in a lot more detail,[991.21] [991.21][S01]the way humans do.[992.29] [992.29][S01]If you are going to take action--[993.85] [993.85][S01]say that you're trying to pack a lunch for your kid--[996.622] [996.622][S01]in order to do that, you have to understand where all the objects[999.33] [999.33][S01]are in 3D space.[1000.81] [1000.81][S01]Then you need to understand how to grasp each object in order[1005.39] [1005.39][S01]to pack it into that box.[1007.05] [1007.05][S01]And then you need to figure out how to organize all those pieces[1011.49] [1011.49][S01]so that they fit.[1012.54] [1012.54][S01]All of this is what we mean by embodied reasoning.[1015.99] [1015.99][S02] So is this things like--[1017.6] [1017.6][S02]I don't know.[1018.142] [1018.142][S02]Let's say you got two-camera view.[1020.09] [1020.09][S02]You're there, and I'm here, for instance,[1022.35] [1022.35][S02]I can see your microphone, and so can you.[1024.24] [1024.24][S02]But we've got a completely different view of it.[1026.24] [1026.24][S02]Is it that kind of stuff?[1027.363] [1027.363][S01] Yeah.[1028.28] [1028.28][S01]It can understand, for example, how far the microphone[1030.53] [1030.53][S01]is from our face.[1031.349] [1031.349][S01]But also, if I move around, it can do object correspondence,[1034.98] [1034.98][S01]meaning it understands that microphone[1036.65] [1036.65][S01]is the same one that I'm seeing from the other point of view,[1039.192] [1039.192][S01]which you can imagine is super important[1040.88] [1040.88][S01]if a robot is moving and reasoning about its environment.[1044.55] [1044.55][S02] How hard is it to switch from a 2D image,[1048.39] [1048.39][S02]a single-camera view, to a 3D understanding of the space?[1052.8] [1052.8][S01] So actually, today, what robots are doing[1055.4] [1055.4][S01]is that they're taking camera views from different places.[1059.07] [1059.07][S01]So actually, the robot has cameras in its wrist,[1061.16] [1061.16][S01]and it has a camera on top, and it is actually[1064.31] [1064.31][S01]taking all of the inputs from the three images[1066.87] [1066.87][S01]and doing this on its own.[1068.7] [1068.7][S01]It's actually reasoning, oh, I'm closer to the object[1071.69] [1071.69][S01]because now this camera looks closer.[1074.19] [1074.19][S01]This camera, I can see my hand.[1076.07] [1076.07][S01]And it's doing all of that association on its own.[1078.33] [1078.33][S01]So we're not explicitly adding depth as an additional input.[1083.82] [1083.82][S01]We're just giving it multiple camera views,[1086.37] [1086.37][S01]and it's realizing how to use them[1088.52] [1088.52][S01]in order to understand depth.[1090.0] [1090.0][S02] And how much of that was you deliberately setting[1094.31] [1094.31][S02]that as a task for the robots?[1096.0] [1096.0][S02]Or how much of it emerged from the conceptual understanding[1099.47] [1099.47][S02]that you get from the Gemini models?[1101.385] [1101.385][S01] It simply emerged, actually.[1103.26] [1103.26][S01]So we were able to give it multiple cameras[1106.58] [1106.58][S01]and just see if it actually could reason between them.[1108.987] [1108.987][S02] That's got to be quite shocking.[1110.82] [1110.82][S02]Many, many people must have spent many, many years thinking[1113.81] [1113.81][S02]very hard about that problem of how[1116.51] [1116.51][S02]do you align different camera views to make it[1119.45] [1119.45][S02]so that you're tracking an object across different angles.[1123.34] [1123.34][S02]And then all of a sudden you get these large language models,[1127.62] [1127.62][S02]like Gemini, and it can just do it automatically.[1130.38] [1130.38][S01] Yeah, it's actually[1131.88] [1131.88][S01]wonderful to be able to leverage these models to bring simplicity[1137.04] [1137.04][S01]to the system.[1137.86] [1137.86][S01]You really don't need to have all of these different stages[1140.64] [1140.64][S01]that you extract depth.[1142.18] [1142.18][S01]Then only then you extract where the objects are,[1144.67] [1144.67][S01]and only then you plan how to move,[1146.44] [1146.44][S01]and then only then you're able to do the task.[1148.54] [1148.54][S02] And that's because the foundational model is--[1150.09] [1150.09][S02]I mean, it's effectively like a Swiss army knife.[1151.69] [1151.69][S02]It can do all of the things simultaneously.[1153.305] [1153.305][S01] Yes, exactly.[1154.14] [1154.14][S01]And it can reason between them.[1155.62] [1155.62][S02] OK, so you enhance the physical reasoning, almost.[1159.348] [1159.348][S01] Exactly.[1160.39] [1160.39][S01]You enhance physical reasoning and spatial understanding.[1163.81] [1163.81][S01]And then motion understanding would be the next thing,[1167.64] [1167.64][S01]is understanding, what would happen[1169.44] [1169.44][S01]if I put a glass at the edge of the table?[1172.15] [1172.15][S01]What's actually likely to happen?[1174.22] [1174.22][S01]All of these areas is the areas that we enhance.[1177.01] [1177.01][S01]But that's not enough.[1178.12] [1178.12][S01]You actually have to take it another step[1180.2] [1180.2][S01]and essentially start to teach Gemini the language of actions.[1185.21] [1185.21][S01]And actions for us means understanding[1188.09] [1188.09][S01]how you are actually moving each joint in a robot.[1190.62] [1190.62][S01]So if this is my robot arm, then I'm[1192.35] [1192.35][S01]teaching Gemini how to move the robot, how[1195.08] [1195.08][S01]to move my arm like this.[1196.71] [1196.71][S01]And these are all essentially numbers.[1199.22] [1199.22][S01]And it's learning to translate what[1201.17] [1201.17][S01]it means to pick up a glass versus move my arm[1203.81] [1203.81][S01]in order to pick up a glass.[1205.17] [1205.17][S01]And so you're essentially teaching it a new language.[1207.383] [1207.383][S02] You're connecting those different ideas.[1209.55] [1209.55][S01] Exactly.[1210.02] [1210.02][S02] Can we think of this as two systems working[1212.312] [1212.312][S02]in tandem then?[1213.57] [1213.57][S02]I'm thinking of the analogy here of system 1 and system 2,[1216.96] [1216.96][S02]the Danny Kahneman \"Thinking Fast and Slow\" thing.[1219.31] [1219.31][S01] Yes, exactly.[1220.56] [1220.56][S01]So essentially, the model that we built actually[1222.56] [1222.56][S01]has two models.[1223.5] [1223.5][S01]It has a system that is slow but very powerful at reasoning[1227.57] [1227.57][S01]and thinking and a system that is faster[1229.91] [1229.91][S01]but very, very good at reactivity.[1232.77] [1232.77][S01]So this is the concept of slow and fast thinking.[1236.4] [1236.4][S02] This is like how human brains work,[1238.42] [1238.42][S02]that you have the part of your brain[1240.58] [1240.58][S02]that's very good at calculation and analysis,[1243.62] [1243.62][S02]and then you also have your very instinctive, reactive side too.[1246.892] [1246.892][S01] Yes, that's right.[1248.35] [1248.35][S01]In fact, what we do today is that one of these models[1251.56] [1251.56][S01]is much bigger than the other, as you can imagine,[1253.97] [1253.97][S01]and it actually lives on a server.[1255.49] [1255.49][S01]And the fast model lives on device,[1257.98] [1257.98][S01]and it can respond very quickly.[1259.58] [1259.58][S02] Talk me through how this works then.[1261.58] [1261.58][S02]In terms of the system 1 and system 2 and that example[1264.91] [1264.91][S02]of a slam dunk, something it's never seen before,[1267.68] [1267.68][S02]how does it work?[1268.85] [1268.85][S01] What happens is when[1270.392] [1270.392][S01]you ask the robot to take the basketball and do a slam dunk,[1273.8] [1273.8][S01]the system 2 has to understand what that means.[1276.55] [1276.55][S01]What is basketball?[1277.61] [1277.61][S01]It has to understand where the objects that are in front of it[1280.42] [1280.42][S01]are, where the basketball is, understand there is a hoop,[1283.79] [1283.79][S01]and then that a slam dunk actually means picking up[1286.66] [1286.66][S01]that ball and putting it there.[1288.14] [1288.14][S01]So it understands all of that and predicts a rough trajectory[1292.03] [1292.03][S01]of what the robot should do in terms[1293.86] [1293.86][S01]of how it should move and then hands that over to the system[1298.09] [1298.09][S01]1, which is on device and is able to take that trajectory,[1302.23] [1302.23][S01]but it also takes the visual input[1304.54] [1304.54][S01]and is able to adjust that trajectory.[1306.73] [1306.73][S01]So if I were to, for example, get in the way,[1308.74] [1308.74][S01]put my hand in the middle, or move the object around,[1311.68] [1311.68][S01]it would still be able to respond because it already[1314.13] [1314.13][S01]understood the concept of what a slam dunk was[1317.1] [1317.1][S01]and respond very quickly.[1318.238] [1318.238][S02] Why do you need two systems at all, though?[1320.53] [1320.53][S02]Why can't you just use the slow, clever one?[1322.613] [1322.613][S01] We could actually[1324.03] [1324.03][S01]just use the slow, clever one, but then it would actually[1326.82] [1326.82][S01]be significantly visually slower,[1330.0] [1330.0][S01]and it won't adapt as quickly to changes in its environment.[1332.95] [1332.95][S01]And that's important, especially if you're doing something where[1336.39] [1336.39][S01]the objects will move around.[1338.02] [1338.02][S01]So if you have, for example--[1339.25] [1339.25][S01]imagine when you're folding a T-shirt in the air, which[1342.06] [1342.06][S01]us humans do pretty cleverly, you're[1344.083] [1344.083][S01]actually moving this T-shirt, and things[1345.75] [1345.75][S01]are moving for you in ways that you don't predict.[1348.707] [1348.707][S01]So you need to be able to respond quickly in order[1350.79] [1350.79][S01]to actually complete the task.[1352.48] [1352.48][S01]So you need definitely a fast system.[1354.9] [1354.9][S01]And the slow system simply enables[1357.35] [1357.35][S01]us to do much more complex reasoning.[1359.55] [1359.55][S01]So you could also just live with a small system[1362.55] [1362.55][S01]if you could do tasks that don't require advanced reasoning.[1365.265] [1365.265][S02] Was it a direct copy of how[1366.89] [1366.89][S02]things work in the human brain?[1368.57] [1368.57][S02]The Daniel Kahneman work comes back to 1970s or so, that[1372.5] [1372.5][S02]we've understood that that's how the human brain works.[1374.85] [1374.85][S02]Was it a direct--[1375.812] [1375.812][S01] [LAUGHS] No, not at all.[1377.52] [1377.52][S01]I think we started definitely with the slow system,[1379.83] [1379.83][S01]as you said.[1380.43] [1380.43][S01]Why don't we just solve this with one model?[1382.38] [1382.38][S01]And we found that, actually, if you[1383.9] [1383.9][S01]want to do highly dexterous behaviors[1387.11] [1387.11][S01]or any kind of complex manipulation,[1389.16] [1389.16][S01]you need to respond quickly.[1390.57] [1390.57][S01]And that was the best combination that we could find.[1393.643] [1393.643][S02] Wow.[1394.31] [1394.31][S02]It's almost like evolution is a really good optimization[1397.176] [1397.176][S02][INAUDIBLE], finds really good strategies[1400.1] [1400.1][S02]for quick but clever things.[1401.765] [1401.765][S01] Yes, definitely.[1403.14] [1403.14][S01]It was surprising to us that that combination just worked.[1405.63] [1405.63][S02] I do think that there's sometimes[1407.505] [1407.505][S02]where the human body knows stuff before your brain does,[1413.07] [1413.07][S02]as it were.[1414.273] [1414.273][S02]You can catch a falling glass without thinking.[1417.17] [1417.17][S02]Or you can commit things to muscle memory,[1419.098] [1419.098][S02]like playing a piano, where you can actually just[1421.14] [1421.14][S02]be thinking about completely different things.[1423.84] [1423.84][S02]Are you seeing similar things with the robots,[1426.51] [1426.51][S02]that they almost have a physical intelligence that's[1428.82] [1428.82][S02]separate from the slow, clever system?[1432.108] [1432.108][S01] So we definitely see[1433.65] [1433.65][S01]that if you take the model that can reason[1436.65] [1436.65][S01]and you give it a lot of examples of a particular task,[1439.96] [1439.96][S01]it will get really, really good at that task.[1442.03] [1442.03][S01]But at the moment, if you do too much of that,[1444.67] [1444.67][S01]it will start forgetting some of the generalization.[1446.895] [1446.895][S02] Oh.[1447.52] [1447.52][S01] So this is an active area of research,[1449.812] [1449.812][S01]is, how do we enable the robot to get really, really[1452.7] [1452.7][S01]good at a task, like a really extremely difficult one,[1456.18] [1456.18][S01]and then not lose any of the generalization?[1459.07] [1459.07][S01]So it actually right now is a balancing act.[1461.982] [1461.982][S02] In some ways, that does happen with humans too.[1464.44] [1464.44][S02]I know some people who are really, really,[1466.47] [1466.47][S02]really, really, really good at maths[1467.97] [1467.97][S02]and terrible at tying their own shoelaces.[1469.736] [1469.736][LAUGHTER][1471.408] [1471.408][S02]I think that does happen.[1472.45] [1472.45][S02]They forget.[1472.95] [1472.95][LAUGHTER][1474.46] [1474.46][S02]OK, then.[1475.404] [1475.404][S02]So if this is what's going on behind the scenes--[1478.13] [1478.13][S02]so if we've got system 1 and system 2, as you described it,[1483.01] [1483.01][S02]it's also definitely true that these robots[1485.95] [1485.95][S02]have these very impressive new abilities and capabilities,[1489.91] [1489.91][S02]which are very different from where we were before.[1492.98] [1492.98][S02]Last time I visited DeepMind's robotics lab,[1495.437] [1495.437][S02]I think it's fair to say that the robots' movements were[1497.77] [1497.77][S02]a bit clumsy.[1499.0] [1499.0][S02]I think that's probably the kindest way to say it.[1501.627] [1501.627][S02]Let me just play you a little clip.[1503.085] [1503.085][AUDIO PLAYBACK][1503.752] [1503.752][S02]- So there's only one way around that it can hold this red object[1507.76] [1507.76][S02]and successfully pick it up.[1509.48] [1509.48][S02]And it hasn't worked out which way.[1511.86] [1511.86][S02]And unfortunately, every time it tries to rotate and pick it[1514.36] [1514.36][S02]up-- oh, hang on.[1515.09] [1515.09][S02]I think it's got it.[1515.923] [1515.923][S02]It's got it.[1516.638] [1516.638][S02]It's a good job these things don't get disheartened.[1518.805] [1518.805][END PLAYBACK][1519.29] [1519.29][S02] The thing is that I've been in that lab maybe five[1521.873] [1521.873][S02]years earlier, and these poor robots were still[1524.23] [1524.23][S02]there five years later trying to do the same minimally dexterous[1528.37] [1528.37][S02]tasks.[1529.19] [1529.19][S02]What changed?[1530.51] [1530.51][S02]Because I understand how having Gemini, the slow, clever system,[1535.53] [1535.53][S02]could improve the conceptual understanding of things.[1537.76] [1537.76][S02]But that doesn't change the dexterity.[1539.65] [1539.65][S02]It doesn't change how easily it can manipulate these objects,[1543.09] [1543.09][S02]does it?[1543.61] [1543.61][S01] Yeah, last year, we[1545.11] [1545.11][S01]spent basically all of our effort on tackling dexterity.[1549.07] [1549.07][S01]And this is still an area of active research.[1550.96] [1550.96][S01]But there's a couple of things that changed.[1552.85] [1552.85][S01]One is that we realized that if we can enable humans to show[1557.97] [1557.97][S01]the robot how to do very complex behaviors through teleoperation[1562.29] [1562.29][S01]or puppeteering, what this means is[1564.09] [1564.09][S01]that you give the human an extra pair of arms, robot arms,[1567.04] [1567.04][S01]and they can actually pretend to be the robot[1569.31] [1569.31][S01]and show the robot how to do the task.[1572.02] [1572.02][S01]And if that becomes really intuitive,[1574.3] [1574.3][S01]then you can capture a lot of data[1575.76] [1575.76][S01]of the robot doing the task, being teleoperated by a human,[1579.4] [1579.4][S01]but it's robot data.[1580.387] [1580.387][S02] So let me understand then.[1581.97] [1581.97][S02]So the human is wearing maybe a head cam.[1584.005] [1584.005][S01] Yes.[1584.88] [1584.88][S02] And it's quite literally pretending[1585.93] [1585.93][S02]to be the robot then.[1586.805] [1586.805][S02]So it's operating the robot's hands with its hands,[1589.092] [1589.092][S02]wearing the head cam, watching what[1590.55] [1590.55][S02]the robot would be watching, but doing the task[1593.43] [1593.43][S02]as it wants the robot to do.[1595.055] [1595.055][S01] So there's different teleoperation[1597.18] [1597.18][S01]examples.[1598.09] [1598.09][S01]One is where you actually sit in front of the robot,[1600.61] [1600.61][S01]so you have direct visibility to what the robot is doing,[1603.13] [1603.13][S01]and you move the robot arms.[1604.84] [1604.84][S01]Literally, you're puppeteering the robot.[1606.73] [1606.73][S01]And there's other examples where you put a VR set and gloves,[1610.23] [1610.23][S01]and actually, you pretend to be the robot,[1612.03] [1612.03][S01]and you move the stuff.[1613.06] [1613.06][S01]And that required a second component,[1614.86] [1614.86][S01]which was diffusion models.[1616.193] [1616.193][S01]And these are the same models that[1617.61] [1617.61][S01]actually get used, for example, by Imagen[1620.22] [1620.22][S01]in order to generate videos.[1622.06] [1622.06][S01]And essentially, what it's doing is[1623.55] [1623.55][S01]extracting from a lot of data a lot[1625.95] [1625.95][S01]of examples of doing that task and predicting the actual action[1629.76] [1629.76][S01]trajectory that it needs to do in order to do that task.[1632.68] [1632.68][S01]So when you combine those two with a clever transformer[1636.63] [1636.63][S01]architecture and a good data set,[1639.55] [1639.55][S01]you can actually learn anything.[1641.14] [1641.14][S01]And that was actually really, again, surprising[1643.86] [1643.86][S01]to the researchers.[1644.83] [1644.83][S01]That's when, for example, we discovered[1646.47] [1646.47][S01]that we could tie shoelaces.[1648.19] [1648.19][S01]We could fold laundry.[1649.75] [1649.75][S01]We could do origami.[1651.38] [1651.38][S01]And so what we did in this work is[1654.16] [1654.16][S01]that we combined the powerful reasoning module from Gemini[1657.88] [1657.88][S01]with what we had learned around being[1660.58] [1660.58][S01]able to do dexterous tasks.[1662.63] [1662.63][S02] Do you remember when you[1664.99] [1664.99][S02]realized that these kind of properties were emerging?[1670.33] [1670.33][S02]It must have been a bit of a shock.[1671.96] [1671.96][S01] I think the first time[1673.585] [1673.585][S01]was when we saw the robots actually tying shoelaces.[1677.06] [1677.06][S01]We were like, that's not possible.[1678.59] [1678.59][S01]In fact, when the researchers set up this task,[1681.35] [1681.35][S01]they actually did it to challenge themselves.[1684.003] [1684.003][S01]I think there was a professor that[1685.42] [1685.42][S01]said, oh, I bet if we can get robots to tie shoelaces,[1689.57] [1689.57][S01]I will retire.[1690.38] [1690.38][S01]And the researchers in the team were like, right on.[1694.37] [1694.37][S01]I'm going to add that as a task.[1695.99] [1695.99][S01]And so they actually did.[1697.4] [1697.4][S01]And they were surprised when it was able to do it.[1700.1] [1700.1][S01]I don't know what happened, whether the professor actually[1702.88] [1702.88][S01]saw the video and decided to retire.[1705.14] [1705.14][S01]But the inspiration came from there.[1708.46] [1708.46][S01]And we just continued to add tasks, more and more tasks.[1711.64] [1711.64][S01]Same with the origami example.[1713.068] [1713.068][S01]We were like, we have no idea if this is going to work,[1715.36] [1715.36][S01]but let's try it.[1716.5] [1716.5][S01]And it actually was surprisingly good at it.[1719.62] [1719.62][S01]And it's actually really delicate.[1721.27] [1721.27][S01]It has to actually fold every part of the paper,[1724.63] [1724.63][S01]and it has to do it in the right sequence.[1726.4] [1726.4][S01]If anything goes wrong, it loses its way[1729.21] [1729.21][S01]and has to restart, same as if a human was doing it.[1732.23] [1732.23][S02] I remember the very first time[1733.98] [1733.98][S02]that I got to interview Demis.[1735.28] [1735.28][S02]He was talking about Moravec's paradox, this idea[1737.73] [1737.73][S02]that the tasks that are easy for humans[1740.47] [1740.47][S02]are hard for machines and vice versa.[1742.66] [1742.66][S02]With all of these advances that we have now[1744.96] [1744.96][S02]in robotics, do you think that Moravec's paradox will[1747.45] [1747.45][S02]hold going forwards?[1748.493] [1748.493][S01] I certainly think[1749.91] [1749.91][S01]that it is still more difficult for robots[1753.63] [1753.63][S01]to do something that is incredibly[1755.22] [1755.22][S01]intuitive for us humans to do.[1756.58] [1756.58][S01]So I think Moravec's paradox still holds.[1759.19] [1759.19][S01]We are now at the point where we are confident[1761.64] [1761.64][S01]that if you can teach a robot, if you[1763.56] [1763.56][S01]can operate a robot to do a very complex task, it can learn it.[1768.4] [1768.4][S02] And how quickly does it happen?[1770.27] [1770.27][S02]How many origami foxes does a robot[1772.61] [1772.61][S02]need to watch a human do before it can do one itself?[1775.17] [1775.17][S01] It varies by the complexity of the task,[1777.545] [1777.545][S01]pretty similar to the way it is for humans, right?[1779.63] [1779.63][S01]The more complex the task, the more[1781.7] [1781.7][S01]you need to practice it before you can master it.[1783.84] [1783.84][S01]So there's a lot of tasks that you can master with just about[1787.46] [1787.46][S01]100 examples, and tasks like the origami fox takes about 1,000[1791.96] [1791.96][S01]examples.[1792.69] [1792.69][S02] Wait, so people had to fold origami foxes while[1795.68] [1795.68][S02]pretending to be a robot 1,000 time?[1797.18] [1797.18][S01] Yes, that's right.[1798.638] [1798.638][S02] OK, that is incredibly amusing to me.[1800.875] [1800.875][S01] [LAUGHS] We're trying to reduce it[1803.0] [1803.0][S01]as much as possible, and we are able to get quite a bit of tasks[1806.69] [1806.69][S01]with just, like, a dozen of examples.[1808.232] [1808.232][S02] Are there some that you don't need any examples[1810.69] [1810.69][S02]for at all?[1811.32] [1811.32][S01] In a lot of the examples that we were testing,[1813.42] [1813.42][S01]like when you're playing with the robot[1815.045] [1815.045][S01]and asking it to do a lot of pick-and-place tasks[1817.64] [1817.64][S01]with completely new scenarios, you[1819.98] [1819.98][S01]don't have to teach it again.[1821.28] [1821.28][S01]And this is expanding and getting more and more complex.[1824.31] [1824.31][S01]So for example, the cases with the tiles,[1826.76] [1826.76][S01]where you're moving the tiles around,[1828.89] [1828.89][S01]those it just can reason about positioning of the tiles[1832.34] [1832.34][S01]and decide where to put them.[1834.147] [1834.147][S02] OK, what about the packed lunch?[1835.98] [1835.98][S01] That one is more complex[1837.02] [1837.02][S01]because, again, in that one, you're actually[1838.853] [1838.853][S01]doing a long sequence of tasks, about five minutes of task,[1842.61] [1842.61][S01]and you're actually picking up very deformable things,[1847.511] [1847.511][S01]like the Ziploc bag, and doing very delicate stuff.[1850.71] [1850.71][S01]So the more delicate the task, the more likely it is you[1853.73] [1853.73][S01]need to see examples in that task.[1855.403] [1855.403][S02] So if these robots are having to see examples,[1857.82] [1857.82][S02]does that end up impacting the generality of it?[1860.333] [1860.333][S01] Only to a degree.[1861.75] [1861.75][S01]So one thing that we make sure that we do[1863.54] [1863.54][S01]is that we collect data in thousands of examples[1867.11] [1867.11][S01]without doing a very large emphasis on any new task.[1872.97] [1872.97][S01]If you do want to do the origami task,[1876.21] [1876.21][S01]we simply specialize it for the origami task,[1878.43] [1878.43][S01]and that does affect the generalization of the models[1880.77] [1880.77][S01]today.[1881.69] [1881.69][S01]We're hoping to get to a state where you basically can teach it[1884.63] [1884.63][S01]any new task, to master any new task,[1887.48] [1887.48][S01]and the generality remains intact.[1889.76] [1889.76][S01]But today, it's a trade-off.[1891.17] [1891.17][S02] So in the dream world,[1892.81] [1892.81][S02]you would be able to say--[1894.77] [1894.77][S02]I don't know-- fold me an origami boat.[1897.68] [1897.68][S02]And it would be able to do that just from everything[1900.64] [1900.64][S02]that it understood before.[1902.05] [1902.05][S01] Yeah, in the dream world,[1903.8] [1903.8][S01]it could just watch a video of someone doing it,[1906.19] [1906.19][S01]and it would learn from that.[1907.55] [1907.55][S02] So reinforcement learning[1909.092] [1909.092][S02]was a big thing in robotics for quite a stretch of time.[1914.21] [1914.21][S02]Has that just disappeared now?[1915.86] [1915.86][S01] Not at all.[1916.46] [1916.46][S01]Not at all.[1916.96] [1916.96][S01]We do quite a bit of work still with reinforcement learning,[1919.79] [1919.79][S01]and we continue to explore ways to combine this big foundation[1924.01] [1924.01][S01]models with reinforcement learning.[1925.84] [1925.84][S01]First of all, all of the work that we do around whole body[1928.69] [1928.69][S01]control, like if we have a humanoid that is walking around[1931.78] [1931.78][S01]or a quadruped, they're all using reinforcement learning[1934.33] [1934.33][S01]to learn how to walk around.[1935.9] [1935.9][S02] Because it's very easy to say[1937.608] [1937.608][S02]you fail when you fall over.[1939.26] [1939.26][S01] It's a very mature technology.[1941.36] [1941.36][S01]And you actually can learn it all in simulation.[1943.04] [1943.04][S01]So it doesn't need to fall in order to learn.[1944.93] [1944.93][S01]So you can learn in simulation and then[1946.56] [1946.56][S01]transfer it to the real world.[1947.89] [1947.89][S01]One example that we had on this was a recent paper[1950.31] [1950.31][S01]called DemoStart.[1951.58] [1951.58][S01]In DemoStart, you basically show the robot[1953.88] [1953.88][S01]how to do five different examples.[1955.96] [1955.96][S01]This is manipulating a hand.[1957.847] [1957.847][S01]So you show it five different examples[1959.43] [1959.43][S01]of how to pick up an object and place it[1961.097] [1961.097][S01]in a particular way on an insertion.[1963.125] [1963.125][S02] By insertion, do you mean things like--[1965.25] [1965.25][S02]I don't know-- putting a key in a lock, for instance?[1967.48] [1967.48][S01] Yes, exactly, being[1968.98] [1968.98][S01]able to put one object inside another.[1970.99] [1970.99][S01]So a key in a lock is a great example.[1972.61] [1972.61][S01]You just give it five examples, and it explores on its own[1975.33] [1975.33][S01]and learns how to do it and drastically reduces the amount[1978.21] [1978.21][S01]of data that you need in the real world by 100x.[1981.61] [1981.61][S01]We think this is going to be critical because the truth is[1984.397] [1984.397][S01]you're not going to be able to demonstrate for the robot[1986.73] [1986.73][S01]how to do every single task.[1988.24] [1988.24][S02] Of course.[1989.44] [1989.44][S01] Some of the tasks are going to be complex,[1991.898] [1991.898][S01]and they won't be able to extract that directly[1994.35] [1994.35][S01]from its knowledge of the internet.[1996.455] [1996.455][S01]So it's going to have to explore.[1997.83] [1997.83][S02] Like if you're doing surgery, for example, maybe.[1999.76] [1999.76][S01] Yes.[2000.35] [2000.35][S01]So it's going to have to explore and learn from its behavior.[2003.03] [2003.03][S01]And that's one of the areas that we[2004.488] [2004.488][S01]want to spend a lot more time on,[2005.89] [2005.89][S01]is, how do you get robots that learn on the job?[2008.93] [2008.93][S02] OK.[2009.97] [2009.97][S02]Is doing things in simulation part of the solution then?[2013.84] [2013.84][S01] Yeah, we definitely[2015.34] [2015.34][S01]leverage simulation in multiple ways.[2017.12] [2017.12][S01]We leverage simulation even to learn better how to do 3D[2021.97] [2021.97][S01]understanding of the physical world.[2023.9] [2023.9][S01]We also leverage simulation to learn new behaviors,[2026.42] [2026.42][S01]like in the case of DemoStart.[2028.69] [2028.69][S01]But yeah, when we talk about reinforcement learning,[2031.1] [2031.1][S01]it is not always in simulation.[2032.6] [2032.6][S01]You can also do reinforcement learning[2034.78] [2034.78][S01]to learn how the robot is doing in the real world directly.[2037.67] [2037.67][S01]So we do it in both cases.[2039.23] [2039.23][S01]And simulation is a critical component.[2041.03] [2041.03][S02] Does it work, though?[2042.5] [2042.5][S02]Isn't the real world a bit messier than simulations?[2044.9] [2044.9][S01] Yes.[2045.19] [2045.19][S01]So there is things that are actually much harder[2047.53] [2047.53][S01]to do in simulation first, for example, anything that has to do[2050.92] [2050.92][S01]with deformable simulating.[2053.26] [2053.26][S01]Folding that T-shirt in the air is actually extremely hard.[2057.05] [2057.05][S01]Simulating fluids is really hard.[2059.264] [2059.264][S01]So there are some things that are[2060.639] [2060.639][S01]just easier to learn in the physical world and some things[2064.05] [2064.05][S01]that you can learn a much larger scale in the simulator world.[2067.13] [2067.13][S02] And does one translate to the other?[2069.13] [2069.13][S02]If you do the learning in simulation--[2071.54] [2071.54][S02]I seem to remember, actually-- maybe this is like--[2073.665] [2073.665][S02]I don't know-- eight years ago or something.[2075.969] [2075.969][S02]But there was one robot that was trying to get a ball in a cup,[2079.739] [2079.739][S02]and it could do it in simulation.[2081.73] [2081.73][S02]But then once it came to the reality,[2083.71] [2083.71][S02]all sorts of other factors came into play,[2085.78] [2085.78][S02]maybe the lighting on the camera angle, the exact dimensions[2089.82] [2089.82][S02]of its own limbs.[2090.793] [2090.793][S02]I mean, all of that kind of stuff starts[2092.46] [2092.46][S02]to mess with the numbers, doesn't it?[2094.57] [2094.57][S01] Yes, definitely.[2096.239] [2096.239][S01]That's what we call the sim-to-real gap,[2097.99] [2097.99][S01]and we still have the sim-to-real gap.[2099.94] [2099.94][S01]It certainly has been reduced significantly.[2103.6] [2103.6][S01]When it comes to modeling interactions[2105.75] [2105.75][S01]between a robot and the world, which[2108.12] [2108.12][S01]is really messy and complicated, it's still a problem.[2111.19] [2111.19][S01]We still have some sim-to-real gaps.[2112.75] [2112.75][S01]Essentially, what we end up doing[2114.39] [2114.39][S01]is identifying areas where it is easy to simulate,[2118.51] [2118.51][S01]and we can actually see sim-to-real transfer,[2120.73] [2120.73][S01]and we do quite a bit of that in simulation in areas[2123.8] [2123.8][S01]where it's actually simpler to learn in the physical world.[2126.48] [2126.48][S01]So we combine the strengths of the two.[2128.61] [2128.61][S02] All of these examples that you're giving[2130.777] [2130.777][S02]are really in lab settings.[2131.94] [2131.94][S02]I'm trying to think of the situations in which you would[2134.273] [2134.273][S02]really want a robot to be there, maybe after a natural disaster,[2137.19] [2137.19][S02]for instance.[2138.26] [2138.26][S02]How does it work taking this stuff out of the lab[2141.83] [2141.83][S02]and then putting it out into the real world?[2143.73] [2143.73][S02]What are the additional complications[2145.01] [2145.01][S02]that you need to handle?[2146.19] [2146.19][S01] Definitely all of our research right[2149.24] [2149.24][S01]now is still happening within our labs.[2152.41] [2152.41][S01]But we're super excited about the potential of bringing this[2155.09] [2155.09][S01]to the real world.[2156.357] [2156.357][S01]And there's a lot of additional things[2157.94] [2157.94][S01]we need to think about to do that.[2159.595] [2159.595][S01]Certainly, we're already thinking[2160.97] [2160.97][S01]about the aspect of safety when you bring these--[2164.85] [2164.85][S01]AI is actually moving physically, robots outside[2168.182] [2168.182][S01]and changing the world.[2169.14] [2169.14][S01]You want to think about all the safety aspects.[2171.39] [2171.39][S01]There's also the aspect that you might not[2173.21] [2173.21][S01]have internet access in any one of these locations.[2177.15] [2177.15][S01]And so it is very important that we think about,[2179.87] [2179.87][S01]can we have models that can run directly on the robot[2183.56] [2183.56][S01]and just be air gapped and completely on device?[2186.453] [2186.453][S01]And this might be useful in the case of a natural disaster[2188.87] [2188.87][S01]where there's no connection.[2190.2] [2190.2][S01]It might be useful for applications where there is[2192.29] [2192.29][S01]actually a lot of latency critical components,[2195.36] [2195.36][S01]like it has to respond very quickly and cannot wait[2199.79] [2199.79][S01]for a server connection.[2201.298] [2201.298][S02] Give me an example.[2202.59] [2202.59][S01] Well, I think actually[2204.215] [2204.215][S01]in any of these examples where the robot is operating[2207.17] [2207.17][S01]underground, it's simply not going[2210.29] [2210.29][S01]to be able to connect and wait for some more advanced reasoning[2215.06] [2215.06][S01]module to tell it what to do.[2216.42] [2216.42][S01]It actually has to decide right there and then how to behave.[2219.39] [2219.39][S01]But it loses a little bit of that generalization[2222.29] [2222.29][S01]and reasoning.[2223.1] [2223.1][S02] On that point about safety,[2225.21] [2225.21][S02]I guess, if you are giving robots the ability[2227.75] [2227.75][S02]to act in a physical world, then you[2230.3] [2230.3][S02]are opening up the possibility of different potential risks[2235.14] [2235.14][S02]like--[2235.64] [2235.64][S02]I don't know-- somebody getting into a robot's language model[2239.15] [2239.15][S02]and warping its reasoning.[2241.61] [2241.61][S02]How do you mitigate against those sort of risks?[2244.31] [2244.31][S01] We essentially have a pretty comprehensive[2247.57] [2247.57][S01]safety and security approach that[2250.12] [2250.12][S01]actually goes in multiple layers of the system.[2253.48] [2253.48][S01]Definitely, we think about software security as critical[2257.29] [2257.29][S01]for these robots so that no bad actor can actually[2259.84] [2259.84][S01]interfere and actually take control of the robot.[2262.52] [2262.52][S01]And in terms of safety, it happens[2264.73] [2264.73][S01]on many different levels.[2265.82] [2265.82][S01]So actually, safety for robotics has been there for decades.[2269.6] [2269.6][S01]There's quite a bit of work on making sure[2271.72] [2271.72][S01]that a robot doesn't collide with its environment[2274.13] [2274.13][S01]or doesn't put too strong impact forces on its environment,[2278.21] [2278.21][S01]or it actually walks stably.[2280.43] [2280.43][S01]And the Gemini robotic models can actually[2283.21] [2283.21][S01]just seamlessly interface with any of those safety[2286.69] [2286.69][S01]critical controllers.[2287.855] [2287.855][S01]The other thing that we do is that when[2289.48] [2289.48][S01]you have an AI controlling a robot,[2291.29] [2291.29][S01]you now have to be thinking about semantic physical safety.[2295.76] [2295.76][S01]And what I mean by that is, like,[2298.75] [2298.75][S01]if someone asks you to put the glass on the table,[2301.227] [2301.227][S01]you're not going to put it right at the edge[2303.06] [2303.06][S01]when it's about to fall.[2304.39] [2304.39][S01]You're going to put it actually somewhere in the middle.[2306.73] [2306.73][S01]Or for example, if you actually see[2310.738] [2310.738][S01]that there is something on the floor,[2312.28] [2312.28][S01]you might want to pick it up so that it avoids someone[2314.76] [2314.76][S01]falling or tripping over it.[2316.06] [2316.06][S01]The way we've done that is that we're actually introducing[2318.57] [2318.57][S01]a new data set called ASIMOV dataset,[2322.0] [2322.0][S01]which essentially contains a long list of scenarios[2327.21] [2327.21][S01]that the robot could encounter and how to reason through those.[2330.28] [2330.28][S01]And these are all physical safety scenarios,[2332.47] [2332.47][S01]and it's inspired by, essentially,[2333.91] [2333.91][S01]Asimov's three laws.[2335.26] [2335.26][S01]The first one is a robot may never hurt a human[2338.88] [2338.88][S01]or cause a human to come to harm by inaction.[2341.83] [2341.83][S01]The second one is that a robot should always[2343.89] [2343.89][S01]follow human orders unless it conflicts with the first law.[2347.29] [2347.29][S01]And the third one is that a robot[2349.41] [2349.41][S01]should protect its own existence unless it conflicts[2352.29] [2352.29][S01]with the first and second law.[2354.04] [2354.04][S01]And it was this very comical situation[2356.34] [2356.34][S01]where a robot was stuck between the three different laws.[2359.67] [2359.67][S01]So that's what inspired the ASIMOV dataset.[2363.0] [2363.0][S01]And it actually has quite a bit of information from US injuries[2368.42] [2368.42][S01]reported by hospitals.[2370.97] [2370.97][S01]And inspired by those examples, we actually[2373.61] [2373.61][S01]created a data set that has visual images, like images[2377.828] [2377.828][S01]of something that is about to happen,[2379.37] [2379.37][S01]and a question associated with it,[2380.86] [2380.86][S01]like, what action should you take in order for this[2383.6] [2383.6][S01]to be a safe situation?[2385.41] [2385.41][S01]And the idea is that we would present it to the community,[2388.11] [2388.11][S01]and everyone in the community can start testing their models[2390.68] [2390.68][S01]with respect to this data set.[2392.13] [2392.13][S02] So it turns out Asimov's original three[2393.83] [2393.83][S02]rules are not enough.[2395.13] [2395.13][S02]You need a little bit more.[2395.51] [2395.51][S01] Not enough.[2396.677] [2396.677][S01]Yes.[2397.177] [2397.177][S02] Give me some examples, though, of the kind[2399.427] [2399.427][S02]of things.[2400.13] [2400.13][S01] Some of the examples that we've seen there[2402.588] [2402.588][S01]is you cannot put a stuffed stuffie, plushy on a hot stove,[2409.202] [2409.202][S01]which is something that I wouldn't have thought about[2411.41] [2411.41][S01]making a law about that.[2412.83] [2412.83][S01]But certainly, it has happened, and therefore, it just[2416.26] [2416.26][S01]comes out in the data.[2417.74] [2417.74][S02] Then are we back in the same problem of you're[2420.25] [2420.25][S02]never going to be able to create an exhaustive list of everything[2422.29] [2422.29][S02]that it shouldn't be able to do?[2423.8] [2423.8][S01] I think that it will be really hard for human[2426.383] [2426.383][S01]to sit down and create the perfect law.[2428.84] [2428.84][S01]So part of what we're doing here is[2432.11] [2432.11][S01]leveraging AI to actually understand[2435.16] [2435.16][S01]a broad set of injury situations that[2437.95] [2437.95][S01]have happened in many different countries[2440.11] [2440.11][S01]and then transform that into a better, more succinct list.[2444.02] [2444.02][S01]And then obviously, that list will[2445.66] [2445.66][S01]have to be updated with some frequency.[2448.16] [2448.16][S01]And the idea here is that we derive an initial list,[2452.99] [2452.99][S01]but then humans can check it and decide how much of it[2456.22] [2456.22][S01]to include or not in order to keep the robot safe.[2460.13] [2460.13][S02] How much overlap is there[2461.95] [2461.95][S02]between this list and the work that's been done on safety[2465.43] [2465.43][S02]and agents, for instance?[2466.85] [2466.85][S01] Yeah, we inherit actually[2468.6] [2468.6][S01]all of the safety that happens already for general foundation[2473.59] [2473.59][S01]models like Gemini.[2474.82] [2474.82][S01]And part of what we do is try to take some of those problems,[2479.17] [2479.17][S01]and if they have a physical grounding to them,[2482.11] [2482.11][S01]then that's where we start to advance[2484.68] [2484.68][S01]the model's understanding.[2486.64] [2486.64][S01]So it's typically examples where there[2489.0] [2489.0][S01]might be a situation that if it's on a screen, it's OK.[2492.31] [2492.31][S01]But if it's now in the physical world,[2494.04] [2494.04][S01]it actually has consequences.[2495.91] [2495.91][S02] Are there some things that you would just never[2498.66] [2498.66][S02]really want a robot to perform?[2500.02] [2500.02][S02]I don't know.[2500.562] [2500.562][S02]Like, a massage, for example.[2504.7] [2504.7][S02]Are there some things that you just actually only want[2507.3] [2507.3][S02]a human to be able to do?[2508.342] [2508.342][S01] There are massage robots, I have to say.[2510.717] [2510.717][LAUGHTER][2511.5] [2511.5][S02] There are massage chairs definitely.[2513.0] [2513.0][S02]Are there massage robots as well?[2513.72] [2513.72][S01] There are massage robots, yes.[2515.04] [2515.04][S02] Wow.[2515.37] [2515.37][S02]That's another thing.[2516.49] [2516.49][S01] Yes.[2517.365] [2517.365][LAUGHTER][2519.33] [2519.33][S02] OK, bad example.[2520.75] [2520.75][S02]Are there some things that you think actually[2523.56] [2523.56][S02]should remain human?[2524.79] [2524.79][S02]Nursing, perhaps.[2526.137] [2526.137][S01] Yeah, I think in many ways,[2527.97] [2527.97][S01]what we think is that robots could be a collaborator that[2531.51] [2531.51][S01]can actually enable humans to pay more[2533.84] [2533.84][S01]attention to the human aspects of the job and less attention[2537.53] [2537.53][S01]to those that are about moving things around or picking things[2541.43] [2541.43][S01]up.[2541.98] [2541.98][S01]So you can imagine that if a nurse could have assistants that[2544.73] [2544.73][S01]could actually help it fetch things while they're paying[2547.88] [2547.88][S01]attention to the patient, then that[2549.86] [2549.86][S01]would enable a much better experience for that patient.[2552.54] [2552.54][S02] You said something really nice[2553.37] [2553.37][S02]at the beginning about how the robots that we've got now,[2555.745] [2555.745][S02]it's like looking at two-year-olds.[2557.25] [2557.25][S02]I mean, quite talented two-year-olds.[2559.7] [2559.7][S02]But I see what you're saying, that they're just demonstrating[2564.23] [2564.23][S02]the beginnings of something.[2566.03] [2566.03][S02]What kind of breakthroughs do you[2567.65] [2567.65][S02]think still need to happen before we[2570.41] [2570.41][S02]get to the adult version of these robots?[2573.243] [2573.243][S01] Yeah, I mean, there's quite a bit of work[2575.66] [2575.66][S01]to be done, definitely, in the aspects of capturing dexterity[2580.58] [2580.58][S01]with generalization, being able to do both of those things,[2584.01] [2584.01][S01]and just continuously grow without losing one or the other.[2588.18] [2588.18][S01]The other key area is that you want these robots[2590.18] [2590.18][S01]to learn on the job.[2592.5] [2592.5][S01]There's no way these robots are going[2594.167] [2594.167][S01]to learn everything they need to learn in the lab,[2596.25] [2596.25][S01]and then you put them out, and they just work.[2598.77] [2598.77][S01]I think the reality is that you would put them out.[2600.99] [2600.99][S01]They will experience new things, and you[2602.81] [2602.81][S01]want them to learn from those experiences[2604.7] [2604.7][S01]and get better and better over time.[2606.57] [2606.57][S01]So that's another area.[2607.83] [2607.83][S01]Also, robots that are more social.[2610.35] [2610.35][S01]I think, certainly, all of these foundation models[2613.94] [2613.94][S01]enable robots to have a lot better understanding[2616.7] [2616.7][S01]of semantics and the world, but they still lack social skills.[2620.85] [2620.85][S01]They still cannot read body language.[2623.46] [2623.46][S01]They cannot understand how to behave in a very cluttered[2626.24] [2626.24][S01]space, like a cocktail party.[2628.65] [2628.65][S01]So there's quite a bit of work there.[2630.36] [2630.36][S02] So how far away do you[2631.777] [2631.777][S02]think we are, then, from the kind of Rosie the Robot that you[2635.93] [2635.93][S02]saw in your childhood?[2637.343] [2637.343][S01] I don't think I have an exact date,[2639.51] [2639.51][S01]but I can tell you before we used to have discussions about[2643.16] [2643.16][S01]whether it would happen in our lifetime or even in our careers,[2647.28] [2647.28][S01]and now we have debates about whether it[2649.4] [2649.4][S01]would be 5 or 10 years.[2651.24] [2651.24][S01]So it's certainly shifted.[2653.99] [2653.99][S01]And it feels like the next two years[2655.96] [2655.96][S01]are going to be pretty defining for the field of robotics.[2658.97] [2658.97][S01]There's just a lot of things that[2660.7] [2660.7][S01]are coming together-- understanding, dexterity,[2664.49] [2664.49][S01]whole body control.[2666.05] [2666.05][S01]It is all started.[2667.28] [2667.28][S01]You can see how this could actually merge[2669.55] [2669.55][S01]into a very strong solution.[2671.54] [2671.54][S02] Do you think that's what[2672.1] [2672.1][S02]we're about to see, then, in the same way as we've[2673.66] [2673.66][S02]seen the explosion of large language models?[2675.74] [2675.74][S02]Do you think the next thing is the explosion of robotics?[2678.145] [2678.145][S01] Yes, absolutely.[2679.52] [2679.52][S01]And I think actually being better[2681.91] [2681.91][S01]at operating in the physical world actually[2684.46] [2684.46][S01]will make our LLMs and VLMs significantly stronger AI models[2689.36] [2689.36][S01]because they can now understand the space of humans.[2693.47] [2693.47][S02] Things are about to change.[2695.17] [2695.17][S02]Thank you so much.[2696.38] [2696.38][S02]Absolutely fascinating.[2697.52] [2697.52][S02]Really amazing.[2698.54] [2698.54][S01] Thank you for having me.[2699.29] [2699.29][S02] Thank you.[2700.207] [2700.207][S02]I don't know if you've ever noticed this little guy sitting[2703.06] [2703.06][S02]behind me.[2704.03] [2704.03][S02]These were the reinforcement learning kings.[2707.12] [2707.12][S02]For literally years, they wandered around[2710.14] [2710.14][S02]in little robot playpens trying and largely[2713.01] [2713.01][S02]failing to learn how to walk, how[2714.84] [2714.84][S02]to play football, how not to continually fall over[2717.54] [2717.54][S02]all of the time.[2718.72] [2718.72][S02]And now, almost overnight, once language and reasoning[2723.21] [2723.21][S02]and conceptual understanding arrived[2725.07] [2725.07][S02]as the missing pieces of the puzzle,[2727.72] [2727.72][S02]they have been confined to the shelves of podcast studios.[2732.6] [2732.6][S02]And in all that time, the researchers,[2735.22] [2735.22][S02]they were focused on the robot's body.[2737.05] [2737.05][S02]But it was advances in the mind that[2739.59] [2739.59][S02]made the biggest leaps forward.[2741.228] [2741.228][S02]You've been listening to \"Google DeepMind--[2743.02] [2743.02][S02]The Podcast\" with me, Professor Hannah Fry.[2745.18] [2745.18][S02]If you enjoyed this episode, then[2746.73] [2746.73][S02]do subscribe to our YouTube channel.[2748.42] [2748.42][S02]You can also find us on your favorite podcast platform.[2751.69] [2751.69][S02]And of course, we have plenty more episodes[2754.05] [2754.05][S02]on a whole range of topics to come, so do check those out.[2757.72] [2757.72][S02]See you next time.[2758.73] [2758.73][MUSIC PLAYING][2762.38]"}
{"file_name": "audio/val_000010.wav", "transcription": "[7.674][S02] Welcome back to the 6th episode of DeepMind, the podcast.[11.812] [12.312][S02]My name is Hannah Fry, I am a mathematician[14.515] [14.515][S02]who’s worked with data and algorithms for the last decade or so.[18.318] [18.318][S02]And I spent the last year at DeepMind -[21.054] [21.054][S02]an organisation that is trying to solve intelligence[24.091] [24.091][S02]and then use it to solve some of society’s problems.[27.594] [30.898][S02]There are an awful lot of people[32.533] [32.533][S02]working in the field of artificial intelligence,[35.469] [35.469][S02]moving forward our understanding of the whole area[39.239] [39.239][S02]and for many of them it is a terrifically exciting place to be.[43.11] [43.11][S02]We’re breaking new frontiers of problem solving and seeing great leaps ahead.[48.882] [48.882][S02]But before any of that hits the outside world,[51.118] [51.118][S02]the first inklings of new breakthroughs here at DeepMind[55.122] [55.122][S02]come in the form of regular poster sessions.[58.358] [60.127][S01] Our goal here was to um have a model[62.629] [62.629][S01]that can carry out the following simple tasks[64.932] [64.932][S01]where I’m going to give you a number and a secret symbol[67.601] [67.601][S01]and what the sum between those two are,[69.703] [69.703][S01]and you have to infer from that what the value of the symbol is,[72.84] [72.84][S01]but it requires our agents to have some properties[75.175] [75.175][S01]that we think are desirable like learning to learn,[78.579] [78.579][S01]like having a memory and processing those memories ourselves.[81.048] [81.048][S04] We’re studying analogical reasoning -[84.184] [84.184][S04]analogical reasoning is very important[85.452] [85.452][S04]because it’s a key to scientific discovery and also human reasoning -[90.457] [90.457][S04]the main question we ask is how can we design neural networks[93.393] [93.393][S04]that are able to do analogical reasoning.[94.962] [94.962][S01] My poster is about verification of neural networks.[98.298] [98.298][S01]In this day and age, when we deploy neural networks[100.734] [100.734][S01]into the real world applications,[102.369] [102.369][S01]we want to make sure that these neural networks are safe,[105.239] [105.239][S01]for example if you have an image classifier,[107.674] [107.674][S01]we don’t ever want to predict a cat to be like a car or something like that.[111.144] [111.144][S01] I always say DeepMind is a bit like academia on steroids,[113.614] [113.614][S01]like it is still academia, but we have a lot of compute,[116.149] [116.149][S01]a lot of great people clustered together,[118.185] [118.185][S01]a lot of help to manage ourselves, so yeah.[121.154] [124.892][S02] while there is obvious excitement about AI research,[128.061] [128.061][S02]this new era of artificial intelligence also comes with concerns.[132.833] [132.833][S02]There is an ease about the way it might be implemented, used and abused.[138.338] [138.338][S02]For the rest of this episode,[139.706] [139.706][S02]we are looking at the more human side of technology,[143.043] [143.043][S02]and the fight to find a future of AI that works for everyone.[148.182] [148.182][S02]In 2017, DeepMind set up dedicated teams[151.885] [151.885][S02]working on how AI impacts ethics and society.[155.889] [155.889][S02]With the aim of making sure that the algorithms designed in this building[159.826] [159.826][S02]are a positive force for good.[162.229] [162.93][S02]But hang on - I know what you are thinking -[164.765] [164.765][S02]surely algorithms aren’t ever good or bad in and of themselves,[169.102] [169.102][S02]it’s how they’re used that matters.[171.638] [171.638][S02]After all, GPS was invented to launch nuclear missiles,[175.609] [175.609][S02]and now helps to deliver pizzas.[178.111] [178.111][S02]And speakers playing pop music on repeat have been deployed as a torture device.[184.251] [184.251][S02]Isn’t the technology itself just neutral?[188.055] [188.055][S06] Good question![188.956] [188.956][S06]And um I think something a lot of people say and believe[195.495] [195.495][S06]and I can see why they say that.[198.432] [198.432][S02] This is Verity Harding, co-lead of DeepMind Ethics and Society.[203.003] [203.003][S06] I think there’s a famous saying about as long as there’s been fire,[207.241] [207.241][S06]there’s been arson.[208.542] [208.542][S06]You can use something that’s for good you can use it for bad.[211.144] [211.144][S06]But It think actually is we’re developing increasingly[213.714] [213.714][S06]sophisticated technologies that have real impact on people’s lives.[218.385] [218.385][S06]It’s not really an acceptable thing to say.[220.888] [220.888][S06]You can’t be building something that’s going to have[223.257] [223.257][S06]this kind of monumental impact - or potentially transformative impact[227.895] [227.895][S06]and not care about how it’s going to be used.[232.666] [232.666][S02] Is that part of the concern then?[234.034] [234.034][S02]That technology that might have been built for one purpose[237.304] [237.304][S02]ends up being used in a different way?[240.574] [240.574][S06] I think definitely that’s some of it.[242.242] [242.242][S06]I think definitely that’s some of it. Because you could force a situation[246.38] [246.38][S06]where you’re building facial recognition tool[249.917] [249.917][S06]because you want to allow somebody to quickly find pictures[253.453] [253.453][S06]of their husband or wife or mom or dad and that’s a great thing[258.892] [258.892][S06]but that facial recognition technology, once developed,[262.229] [262.229][S06]could of course be used to target political dissidents[265.265] [265.265][S06]and pick them out of a crowd and you know, so I think,[267.634] [267.634][S06]that’s definitely one of the concerns[269.236] [269.236][S06]that you might create something for one purpose and it be used for another.[272.973] [275.742][S02] On the topic of facial recognition, Brad Smith,[278.779] [278.779][S02]the President of Microsoft,[280.647] [280.647][S02]recently refused a request by a US police department[284.218] [284.218][S02]to install their algorithm in cop cars and body cameras,[288.155] [288.155][S02]and he's publicly called for more careful thought and societal dialogue[292.559] [292.559][S02]about potentially regulating the use of the technology.[296.496] [296.496][S02]And here at DeepMind more generally there is a strong sense[299.533] [299.533][S02]that the people behind the science have a duty to investigate the wider[304.605] [304.605][S02]and perhaps less predictable impacts of their work.[308.008] [308.008][S06] I don’t think it’s okay to build something -[311.078] [311.078][S06]whether that be a product or a service and put it out there[314.581] [314.581][S06]and and just hope that you make the world a better place.[318.218] [318.218][S06]It think it’s important that you are deliberate and intentional[321.989] [321.989][S06]about why you’re building this, who are you building it for,[325.459] [325.459][S06]what are you hoping to do?[327.027] [327.027][S06]What is your intention with this technology?[329.83] [329.83][S06]And if you start from that premise[331.064] [331.064][S06]then I think you are more likely to get to a better outcome[334.001] [334.001][S06]where you do the good that you hoped you were going to do.[336.703] [336.703][S02] The problem is that without these steps,[339.239] [339.239][S02]it’s very easy for unintentional consequences to creep up on you.[343.877] [343.877][S02]You only need to look at the new stories about social media[346.78] [346.78][S02]from the past few years to see just how much algorithms[350.984] [350.984][S02]have changed our society in unexpected ways.[355.155] [355.155][S02]Lila Ibrahim is DeepMind COO[357.591] [357.591][S02]and has over 20 years’ experience working in the tech sector.[361.028] [361.028][S02]She has seen first hand how hard a booming industry[364.498] [364.498][S02]has found it to keep up with being responsible.[368.302] [368.302][S03] In 2006 I went into the middle of the Amazon[371.305] [371.305][S03]and we built a computer lab and health care,[374.842] [374.842][S03]so we put in internet and computers, etc.[377.678] [377.678][S03]and we knew we had a responsibility not to just leave it there[380.347] [380.347][S03]but to train folks to take care of it, to think about the sustainability.[385.385] [385.385][S03]But I think that’s kind of where things tend to end.[388.088] [388.088][S03]Ethics means something very different now,[390.224] [390.224][S03]and responsibility means something very different now[392.459] [392.459][S03]because technology is in everybody’s hands.[394.695] [394.695][S03]It’s no longer limited to a few people for a specific application.[400.234] [400.234][S03]Um it’s a lot easier to get into the tech sector[402.669] [402.669][S03]and to make technology that can have value to people,[406.106] [406.106][S03]and at the same time that comes a lot of responsibility[408.742] [408.742][S03]that I don't think in general the tech sector has taken into account.[412.579] [414.014][S02] But the last few years have shown[415.716] [415.716][S02]how hugely transformative and disruptive AI can be[420.921] [420.921][S02]and brought sharply into focus the very possible negative[424.491] [424.491][S02]outcomes of ill-thought through technology.[428.295] [428.295][S02]But as Verity told me, the tide is slowly beginning to turn.[433.066] [433.066][S02]Much of the drive for a conversation about Ethics[436.336] [436.336][S02]is coming from within the technology community itself.[440.007] [440.007][S06] 3 years ago in 2016 some of the scientists from different[445.412] [445.412][S06]ah labs at different companies met at a conference[448.549] [448.549][S06]and were talking about how excited they were about the potential for AI[452.819] [452.819][S06]to do a lot of good, ah but acknowledging that -[456.156] [456.156][S06]a technology that’s so powerful that it has the potential[459.726] [459.726][S06]to be transformative in a very good way must also have the potential[463.397] [463.397][S06]to be very transformative in in in not so good a way,[467.267] [467.267][S06]and so they came together to say well what can we do about it?[470.804] [470.804][S02] And so the partnership on AI was born.[474.675] [474.675][S02]It includes members from Amnesty International,[477.11] [477.11][S02]Electronic Frontier Foundation,[478.779] [478.779][S02]the BBC and Princeton University amongst many, many others.[483.584] [483.584][S02]And together they are hoping to come up with best practices in AI[487.754] [487.754][S02]making sure that society stays firmly at the forefront of engineers’ minds.[493.493] [493.493][S06] So the partnership of AI interestingly was founded[496.997] [496.997][S06]by the biggest tech companies so it was founded by DeepMind[501.201] [501.201][S06]but also Google, Facebook, Amazon, IBM and Apple.[506.306] [506.306][S06]One thing that’s really interesting about the partnership and AI[508.475] [508.475][S06]is that the board membership is made up of independent board members[512.846] [512.846][S06]and representatives of the company[514.915] [514.915][S06]and so it’s creating a space where those different groups[519.62] [519.62][S06]aren’t siloed from each other, having debates in different rooms[522.99] [522.99][S06]and not listening but somewhere[524.958] [524.958][S06]where honest people with the best of intentions[528.228] [528.228][S06]can come together and challenge each other[530.297] [530.297][S06]and scrutinise each other and hold each other accountable[532.666] [532.666][S06]but also have frank, open honest debate[536.27] [536.27][S06]about issues where reasonable people can disagree.[539.439] [539.439][S06]I really believe that the outcome of that will be better decision[542.242] [542.242][S06]making both in companies but elsewhere as well.[545.579] [545.579][S02] Does it sometimes get quite heated in those conversations?[549.316] [549.316][S06] You know my experience of it is that it doesn’t get heated[552.719] [552.719][S06]but it’s passionate, so people aren’t angry with each other[557.157] [557.157][S06]and and there’s not aggressive argument,[560.694] [560.694][S06]but people are very honest, and very open and very challenging.[564.198] [564.198][S06]But that’s been received really well in all cases.[568.101] [568.101][S02] How do you protect against rogue companies[572.472] [572.472][S02]just who are not part of these groups, just doing whatever they want.[577.411] [577.411][S06] If enough companies and enough groups sign up to something[581.048] [581.048][S06]and it becomes the norm,[582.382] [582.382][S06]it’s then really obvious when people aren’t doing it.[584.885] [584.885][S06]And I do think people are kind of being called out for that -[588.555] [588.555][S06]it will no longer be tenable to not operate[591.892] [591.892][S06]in the way that everybody else is operating.[594.294] [596.363][S02] It’s not just theoretical concerns about runaway[599.433] [599.433][S02]applications of AI that’s prompting these conversations,[602.836] [602.836][S02]but real examples of algorithms[605.105] [605.105][S02]that have already been let loose on the world with real question marks[609.543] [609.543][S02]about whether their benefits outweigh their harm.[612.813] [612.813][S02]A notorious example is the use of AI in the criminal justice system -[618.185] [618.185][S02]now you may have heard of these algorithms already.[620.32] [620.32][S02]When a defendant appears in court, the AI can assess a defendant’s chances[624.958] [624.958][S02]of going on to commit another crime in future,[628.695] [628.695][S02]and that risk score is then used by a judge to help decide[632.432] [632.432][S02]whether the defendant should be awarded bail,[635.002] [635.002][S02]and in some cases, how long someone’s sentence should be.[639.006] [639.873][S02]There is good justification for something like this[642.643] [642.643][S02]because there is an enormous amount of luck[645.078] [645.078][S02]involved in the human judicial system.[648.415] [649.349][S02]Studies have shown that if you take the same case to a different judge,[653.187] [653.187][S02]you will often get a different response.[656.056] [656.056][S02]If you take the same case to the same judge on a different day,[660.127] [660.127][S02]you’ll often get a different response.[662.296] [662.296][S02]Judges don't like giving the same response too many times in a row,[665.666] [665.666][S02]and so if a series of successful cases of bail hearings have gone before you,[670.404] [670.404][S02]your chances of being successful fall[673.574] [673.574][S02]and there is even evidence to suggest that judges tend to be a lot stricter[678.111] [678.111][S02]in towns where the local sports team has lost recently.[682.282] [682.282][S02]Using AI to help make these decisions can help to eliminate[686.553] [686.553][S02]a lot of that inconsistency, but you have to tread pretty carefully.[691.658] [691.658][S06] if you without thought and care[695.262] [695.262][S06]and due attention to the history of racial prejudice[698.832] [698.832][S06]in the criminal justice system, build something that claims to be able[702.97] [702.97][S06]to predict somebody’s likelihood of reform and rehabilitation,[708.408] [708.408][S06]and reoffending, then it is likely at least in my view[713.38] [713.38][S06]that that’s going to fail.[715.782] [715.782][S06]If you build something with the intention of addressing those biases,[719.052] [719.052][S06]and you work to include the community in some way,[721.788] [721.788][S06]there could then potentially be a beneficial outcome maybe,[724.725] [724.725][S06]but I haven’t seen it yet.[726.46] [726.46][S02] And by fail you’re really talking about[729.263] [729.263][S02]treating black defendants differently to white defendants.[732.132] [732.132][S06] Absolutely! And once you tend to look at the algorithms[735.836] [735.836][S06]and the data that they’ve been, they’ve been built on,[738.739] [738.739][S06]um oftentimes you can see where they were built on data[740.607] [740.607][S06]that was already biased of course this was the outcome.[743.377] [745.746][S02] The issue came to public attention in 2016[749.183] [749.183][S02]after a group of US investigative journalists from Pro Public published[753.153] [753.153][S02]a damning report of one particular company’s criminal risk scores.[758.525] [758.525][S02]Their study showed that the algorithm was twice[761.361] [761.361][S02]as likely to wrongly categorise black defendants[765.566] [765.566][S02]as being likely to re-offend than white defendants.[769.903] [769.903][S02]Now I should just point out that DeepMind does not build these systems,[773.607] [773.607][S02]but the whole industry alongside the partnership of AI[776.476] [776.476][S02]has been part of the conversation about how to address them.[780.447] [781.014][S02]One of those people is William Issac, a social scientist at DeepMind.[785.185] [785.185][S02]He says that the 2016 ProPublica investigation[788.555] [788.555][S02]made people realise that switching over to algorithms[791.758] [791.758][S02]doesn’t make decisions any more objective.[794.928] [794.928][S07] Even with AI and ML tools[798.098] [798.098][S07]you are getting into the social environment[801.802] [801.802][S07]where you actually have the same norms,[804.438] [804.438][S07]the same kind of like systematic biases, they’re still all present.[808.408] [808.408][S07]So it’s really hard to say that somehow this will replace[812.412] [812.412][S07]all of the kind of subjective,[814.548] [814.548][S07]preconceived notions about certain groups,[816.016] [816.016][S07]or historical biases against them,[818.218] [818.218][S07]and that you can start all over again and so I think that was the wakeup call[822.456] [822.456][S07]was that it’s not as objective as it seems and that as a result,[827.461] [827.461][S07]we still have to grapple with those questions.[830.364] [831.064][S02] The problem is that the data[832.366] [832.366][S02]which gives the algorithm predictive abilities[835.335] [835.335][S02]are questions like how many times were you arrested as a juvenile,[840.04] [840.04][S02]but if you are say a young, black man in America,[843.844] [843.844][S02]it doesn’t matter how law-abiding you are,[846.28] [846.28][S02]the chances are that you will have had many more negative interactions[849.65] [849.65][S02]with the police than someone exactly like you,[853.086] [853.086][S02]who happens to be white.[855.255] [855.255][S02]And if you’re using that data to dictate who deserves to be given bail or not,[860.394] [860.394][S02]then you are in serious risk of perpetuating[863.397] [863.397][S02]societal imbalances going forwards.[867.0] [867.0][S02]This is Silvia Ciappia, a staff research scientist at DeepMind.[870.871] [870.871][S05] Researchers don’t fully understand what this fairness[874.675] [874.675][S05]is about they also look like messier in the sense that it involves,[878.812] [878.812][S05]it is not purely technical problem[881.815] [881.815][S05]and it’s very difficult to understand what how to define fairness,[886.486] [886.486][S05]and it’s difficult to separate the technical part[888.789] [888.789][S05]from the ethical one.[891.425] [894.094][S02] This is an important point,[896.096] [896.096][S02]because defining exactly what you mean by “fair” is surprisingly tricky.[901.902] [901.902][S02]Of course you’d want an algorithm[903.303] [903.303][S02]that makes equally accurate predictions for black and white defendants,[907.407] [907.407][S02]the algorithm should also be equally good[909.276] [909.276][S02]at picking out the defendants[911.011] [911.011][S02]who are likely to reoffend whatever racial group they belong to.[915.415] [915.415][S02]And as Pro Publica pointed out,[917.284] [917.284][S02]the algorithm should make the same kind of mistakes[920.587] [920.587][S02]at the same rate for everyone regardless of race.[925.125] [925.125][S02]Ethically you’d want all of those things to be true,[928.562] [928.562][S02]but technically that’s not always going to be possible.[932.165] [932.165][S02]If you’re data set has bias in it, there are some kinds of fairness[936.436] [936.436][S02]that are mathematically incompatible with others.[939.806] [939.806][S02]And even if you could guarantee all of these things,[942.776] [942.776][S02]there are still a number of ethical issues to contend with.[945.846] [945.846][S02]How do you measure fairness, who’s excluded from your definition,[949.65] [949.65][S02]how do you make those decisions transparent,[952.186] [952.186][S02]and ultimately how do people contest[955.088] [955.088][S02]the decisions made by those algorithms?[958.492] [958.492][S02]See, I told you it was tricky.[961.228] [961.228][S02]Coming at this from two very different perspectives,[964.164] [964.164][S02]William and Silvia started looking into the bigger issue[967.334] [967.334][S02]of fairness in algorithms.[969.336] [969.336][S07] Even though we had kind of different frameworks,[971.405] [971.405][S07]me as a social scientist and Silvia as a machine learning researcher,[974.942] [974.942][S07]the actual overlap between how we would approach this[979.112] [979.112][S07]and basically the assumptions that are embedded within it[981.648] [981.648][S07]were remarkably similar and actually part of what we’re saying[985.419] [985.419][S07]is like oh look at these papers in social science[987.554] [987.554][S07]that are kind of making this same point,[989.623] [989.623][S07]they just hadn’t actually had a way to actually communicate that formally.[994.161] [995.596][S02] You’re listening to a podcast from the people at DeepMind.[999.032] [1000.434][S02]In April 2019, William and Silvia co-published a paper on fairness in AI[1005.372] [1005.372][S02]entitled a causal Bayesian network’s viewpoint on fairness.[1009.843] [1009.843][S02]In it they show that no matter how fair algorithms might be,[1013.68] [1013.68][S02]if the data they’re learning from is biased,[1016.45] [1016.45][S02]we still can’t trust their results.[1019.353] [1019.353][S05] I don’t think it’s possible to find technical solutions[1023.824] [1023.824][S05]that are completely satisfactory.[1026.293] [1026.293][S05]At some point we need to take decisions[1030.063] [1030.063][S05]whether the kind of fairness is acceptable or not,[1033.4] [1033.4][S05]but we can advance a lot[1035.169] [1035.169][S05]and that’s why we need more researchers involved[1037.905] [1037.905][S05]- and not just machine learning researchers[1039.606] [1039.606][S05]but researchers from different communities[1042.442] [1042.442][S05]to be raising awareness about this this problem, find solutions,[1048.148] [1048.148][S05]but as we would never be able to find completely satisfactory solutions[1052.853] [1052.853][S05]from a technical viewpoint,[1054.154] [1054.154][S05]and at that point we need to take decisions[1055.923] [1055.923][S05]- is it important to talk about these such that, that we are -[1060.46] [1060.46][S05]something that is missing at the moment.[1063.163] [1063.163][S07] I do think this is fundamentally[1065.732] [1065.732][S07]like a societal, ethical question and challenge.[1069.236] [1069.236][S07]And it will require lots of stakeholders to address.[1072.773] [1072.773][S07]If you have let’s say a data set of facial recognition tool[1075.809] [1075.809][S07]that’s designed to find missing children -[1077.945] [1077.945][S07]what threshold do you say as a society well[1079.913] [1079.913][S07]you say okay this is acceptable, if we maybe are less successful[1083.817] [1083.817][S07]at identifying children with darker faces,[1086.887] [1086.887][S07]what threshold do we say that’s acceptable,[1089.256] [1089.256][S07]because that’s not a technical question,[1091.291] [1091.291][S07]that’s a social and political question, a normative question.[1095.195] [1095.195][S07]Even if you do have a classifier or a facial recognition software[1098.665] [1098.665][S07]that’s fair, the application of it may be in unfair ways.[1102.769] [1102.769][S07]And so that might present a second question[1105.072] [1105.072][S07]that is separate from the actual kind of like[1107.908] [1107.908][S07]if you decide on a threshold[1109.51] [1109.51][S07]that if you’re just using it in a neighbourhood[1111.545] [1111.545][S07]in a predominantly one group or one ethnicity,[1114.581] [1114.581][S07]that presents a whole other set of challenges[1117.117] [1117.117][S07]for whether or not that’s an ethical use of a particular technology.[1121.021] [1121.021][S02] You can’t assess whether these algorithms are good or bad in isolation,[1125.692] [1125.692][S02]they don’t exist on their own.[1128.028] [1128.028][S02]You have to place them in the context of the worlds that they are being used,[1132.399] [1132.399][S02]like the criminal justice system or in healthcare.[1135.836] [1135.836][S02]Here’s Verity Harding again.[1137.371] [1137.371][S06] This is what I mean by it being a kind of much bigger discussion[1141.341] [1141.341][S06]that that potentially the use of algorithms is highlighting.[1144.711] [1144.711][S06]My fear is that people kind of get a checkmark that says:[1148.081] [1148.081][S06]we tested and this algorithm isn’t biased[1151.418] [1151.418][S06]and therefore you should feel free to use it.[1153.854] [1153.854][S06]And that to me isn’t going far enough.[1156.623] [1156.623][S06]I think there needs to be a further discussion then about um[1160.294] [1160.294][S06]but is this making those decisions that were already bad worse, or more quickly,[1165.265] [1165.265][S06]and therefore more of them and you know that that kind of thing.[1167.768] [1167.768][S02] But things are changing.[1169.77] [1169.77][S02]Here’s William on what has happened since that Pro Publica story broke.[1173.54] [1173.54][S07] They’re going back and reconsidering[1175.342] [1175.342][S07]what measures they collect rather than going back[1177.744] [1177.744][S07]and trying to create more robust data sets,[1179.947] [1179.947][S07]thinking about who is collecting the actual data itself.[1183.317] [1183.317][S07]Will it be ever perfect?[1184.484] [1184.484][S07]Will we have bias-free, purely pure data,[1187.254] [1187.254][S07]no I don’t think that’s--I don’t think that’s ever going to happen.[1189.957] [1189.957][S07]But I do think that people will be skeptical[1193.894] [1193.894][S07]when people ask about what data sets you use[1196.396] [1196.396][S07]and they don’t get a satisfactory answer,[1198.432] [1198.432][S07]but I do think people will ask - is this data set representative?[1201.835] [1201.835][S07]Does it have balance across different groups?[1203.737] [1203.737][S07]So people will start asking questions and interrogating data[1206.406] [1206.406][S07]sets and models more aggressively[1208.976] [1208.976][S07]and I think that will lead to better outcomes.[1211.645] [1211.645][S02] And crucially, more people are now being included[1214.815] [1214.815][S02]as part of the conversation.[1216.583] [1216.583][S07] In the aftermath of some of my work[1218.986] [1218.986][S07]and on many others on predictive policing,[1221.622] [1221.622][S07]many cities in California actually started implementing citizen boards.[1225.859] [1225.859][S07]So when police departments wanted to acquire a new police technology,[1230.531] [1230.531][S07]that included uses of machine learning or artificial intelligence,[1234.268] [1234.268][S07]that they had to go in front of a citizen board[1237.971] [1237.971][S07]and actually have the local community evaluate[1241.508] [1241.508][S07]the tool for different metrics including fairness and bias.[1245.345] [1248.682][S02] Getting different voices involved in the conversation[1250.851] [1250.851][S02]is essential to making sure that we build a future[1253.554] [1253.554][S02]that belongs to all of us[1255.556] [1255.556][S02]because what seems obvious to just one person[1257.758] [1257.758][S02]just wouldn’t occur to another.[1260.093] [1260.093][S02]Your perspective is hard coded into the work that you create.[1265.365] [1265.365][S02]There are clear examples of this everywhere outside of AI -[1269.102] [1269.102][S02]able-bodied people designing buildings[1271.305] [1271.305][S02]that disabled people can’t use or new types of plasters[1274.975] [1274.975][S02]that only work if your skin is one particular colour -[1279.079] [1279.079][S02]presumably the same as the designer’s.[1281.715] [1281.715][S02]And the algorithms that we’ve created, they’re really highlighting this issue.[1286.086] [1286.086][S02]Like the ones used to automatically screen[1288.455] [1288.455][S02]CVs and predict which candidates will fit best in a company.[1294.027] [1294.027][S02]Here’s Verity Harding again.[1295.562] [1295.562][S06] If it’s based on historically discriminatory hiring decisions[1300.267] [1300.267][S06]by either intentionally or unintentionally by us humans,[1304.137] [1304.137][S06]um then it’s going to kind of recreate those patterns.[1307.374] [1307.374][S02] Like if you’ve got a company where white men have succeeded[1310.31] [1310.344][S06] Yes[1311.245] [1311.245][S02] and you’re looking for candidates who will succeed,[1313.68] [1313.68][S02]it’s going to pick out white male CVs.[1316.35] [1316.35][S06] Yes, exactly.[1317.451] [1317.451][S06]And if the people building the technology are white males as well,[1322.322] [1322.322][S06]then the likelihood of paying attention to that potential bias[1326.46] [1326.46][S06]and being aware of it, I mean we all have our blind spots,[1329.329] [1330.097][S06]then then the likelihood increases.[1332.633] [1333.667][S02] We’ve seen driverless cars[1335.202] [1335.202][S02]that don’t spot pedestrians with darker skin tones.[1338.305] [1338.305][S02]Tumour screening algorithms that aren’t as effective for patients[1341.341] [1341.341][S02]with ethnicities other than white European,[1344.912] [1344.912][S02]and lots and lots and lots of issues around gender.[1350.551] [1350.551][S02]All of this is kind of inevitable[1352.886] [1352.886][S02]unless you have a range of different viewpoints in your design process.[1357.124] [1357.124][S06] The most important thing in my point of view for ensuring[1361.762] [1361.762][S06]that these things are um if not biased,[1365.866] [1365.866][S06]but that you are being intentional about what you’re building,[1368.602] [1368.602][S06]and aware of the potential bias is that your team is a diverse team[1373.373] [1373.373][S06]is that you have a broad set of voices involved,[1377.244] [1377.244][S06]and it’s actually much simpler to do that than it’s suggested.[1382.416] [1382.416][S02] And the issue of gender diversity[1384.251] [1384.251][S02]has been a particular focus of late.[1386.587] [1386.587][S06] I think there’s plenty of young women and girls[1389.756] [1389.756][S06]who are really excited by science and stem subjects[1393.594] [1393.594][S06]and it’s an easy get out to say that there aren’t enough women in STEM[1397.631] [1397.631][S06]and that’s why work forces aren’t diverse,[1399.967] [1399.967][S06]but actually it’s much more about making sure[1403.437] [1403.437][S06]that it’s a safe space for women and girls to work,[1406.306] [1406.306][S06]that they’re not discriminated against once they’re there.[1409.51] [1409.51][S06]That you’re able to not just attack and hire them[1412.98] [1412.98][S06]but that you’re able to keep them and[1414.882] [1414.882][S06]and make sure that it’s a place where they feel comfortable working.[1417.684] [1417.684][S06]And so I think it’s much more important that we look[1420.153] [1420.153][S06]at how women are treated in science than just dismiss it[1422.723] [1422.723][S06]as something that girls aren’t interested in at a young age.[1426.059] [1426.059][S02] Lila Ibrahim, DeepMind CCO is very conscious that diversity[1430.097] [1430.097][S02]is still a problem in the tech sector as a whole.[1433.3] [1433.3][S03] Talk about things that keep me up at night.[1435.502] [1435.502][S03]Right, so, here I am, a professional of ah 25+ years[1441.608] [1441.608][S03]with an engineering background.[1443.744] [1443.744][S03]A mom also raising 9 year old twin daughters.[1448.215] [1448.215][S03]I would have hoped by now we would have solved the problem[1450.384] [1450.384][S03]and yet we haven’t[1451.451] [1451.451][S03]- we’re like at the same, the numbers are flat.[1454.188] [1454.188][S02] But there are steps being taken to address it.[1457.191] [1457.191][S03] There’s the short-term stuff you can do[1458.559] [1458.559][S03]which are things like you diversify your candidate pools,[1462.596] [1462.596][S03]you, if you’re doing university recruiting,[1465.098] [1465.098][S03]you look at a broader range of universities[1468.335] [1468.335][S03]and ones that have um that have a broader student representation[1472.272] [1472.272][S03]and have support structures often in place[1474.608] [1474.608][S03]to help the students through their academic and communities.[1478.245] [1478.245][S03]Um you look at job descriptions[1480.314] [1480.314][S03]and ensure that you don’t have unconscious bias[1482.416] [1482.416][S03]reflected in your job descriptions.[1484.551] [1484.551][S03]You - so once you’re in the recruiting pipeline,[1486.887] [1486.887][S03]then you need to make sure candidates have the right experience.[1490.324] [1490.324][S03]We are being very deliberate about how we invest back in education.[1495.562] [1495.562][S03]AI is something that will change future generations -[1499.9] [1499.9][S03]so how do we make this a field that is more accessible -[1504.004] [1504.004][S03]so for example, um whether it’s funding diversity scholars at universities,[1509.476] [1509.476][S03]or funding AI chairs in universities to try and increase the pipeline,[1514.481] [1514.481][S03]and I think that helps fuel some of the academic aspects[1518.285] [1518.285][S03]as well as support are like long-term recruiting.[1522.122] [1522.122][S02] This isn’t just tokenism that we’re talking about here,[1525.125] [1525.125][S02]this is about making better technology.[1528.195] [1528.195][S03] Diversity and diverse perspectives will create a drive faster and safer[1536.537] [1536.537][S03]and with just a better a better result because one of the things I worry about[1542.009] [1542.009][S03]is how do we avoid our own internal bias?[1544.811] [1544.811][S03]A lot of the work around deep reinforcement[1546.547] [1546.547][S03]learning started from specific pockets and many people grew up in those labs[1550.717] [1550.717][S03]or those universities and you know they bought their former colleagues[1555.088] [1555.088][S03]and so we have our pretty strong network of people[1557.925] [1557.925][S03]who have known each other for a long time which is fantastic[1561.395] [1561.395][S03]and they can really advance certain aspects of our of our work[1564.831] [1564.831][S03]and yet there are other areas that are emerging um how do you teach curiosity,[1570.337] [1570.337][S03]how do you um how do we ensure that we minimise bias[1575.175] [1575.175][S03]and the code that we’re writing? Who’s to say ah what intelligence is[1580.781] [1580.781][S03]and isn’t unless you have a better representation from society?[1585.819] [1585.819][S03]And that’s just on the research side.[1587.487] [1587.487][S03]On the operations side too you think about things like okay,[1592.192] [1592.192][S03]think about public policy, like if you’re asking governments[1595.963] [1595.963][S03]to think about how they’re going to treat artificial intelligence[1600.534] [1600.534][S03]then you want people that are representative[1602.369] [1602.369][S03]of the of the constituents of the population.[1605.038] [1605.038][S03]If we want to ah be focused on education,[1608.375] [1608.375][S03]making sure that we’re not just focused on the specific schools[1611.078] [1611.078][S03]but our broader range so we’re bringing more people into the space.[1614.314] [1614.314][S03]I just think it’s going to be imperative for us to truly solve intelligence[1620.153] [1620.153][S03]that we’re just going to need to have more diversity.[1623.457] [1623.457][S02] So is part of the well solutions,[1625.826] [1625.826][S02]probably a bit grand a word to suggest[1628.462] [1628.462][S02]but it’s part of the way forward making ethics a kind of keystone[1634.201] [1634.201][S02]at every stage of the process rather than having it as an afterthought.[1637.638] [1637.638][S03] Oh, absolutely. It we have to be thinking about our responsibility[1641.074] [1641.074][S03]for the technology we develop[1642.609] [1642.609][S03]and to candidly to society as a whole - every step along the way.[1649.349] [1649.349][S03]And I think that there’s something quite special[1651.885] [1651.885][S03]about being headquarters out of London versus being[1655.355] [1655.355][S03]based out of the Silicon Valley - I love Silicon Valley,[1659.393] [1659.393][S03]it’s where my career has really developed[1663.297] [1663.297][S03]and yet you’re surrounded by technologists,[1665.966] [1665.966][S03]you know from the billboard signs to the marketing and promotion.[1670.804] [1670.804][S03]Here in London it’s so multicultural[1673.941] [1673.941][S03]and I feel like it’s part of your daily life.[1677.01] [1677.01][S03]You need to be thinking about the work that you’re doing[1679.68] [1679.68][S03]and how it’s going to impact all the people around you.[1682.816] [1684.785][S02] But of course we can’t just leave the solutions[1687.254] [1687.254][S02]solely in the hands of the people who are designing these things.[1691.992] [1691.992][S02]It’s our future, too.[1693.894] [1693.894][S02]The public and government should also have a hand in this.[1697.865] [1697.865][S06] My impression is that people want to understand what they’re using[1701.034] [1701.034][S06]and want to understand what makes it work and how it works.[1705.072] [1705.072][S06]But they more importantly want their representatives[1708.208] [1708.208][S06]and the people tasked with keeping them safe[1710.41] [1710.41][S06]and secure to understand it too[1712.279] [1712.279][S06]and I think that’s where we’ve seen a bit of a breakdown in recent times.[1715.983] [1717.518][S02] If you want to know more about ethics, diversity and fairness,[1721.722] [1721.722][S02]then head over to the show notes where you can also explore[1724.458] [1724.458][S02]the world of AI research beyond DeepMind.[1727.528] [1727.528][S02]And we’d welcome your feedback and questions[1730.063] [1730.063][S02]on any aspects of artificial intelligence[1732.032] [1732.032][S02]that we’re covering in this series.[1734.034] [1734.034][S02]So if you want to join in the discussion or point us to stories or resources[1738.038] [1738.038][S02]that you think other listeners would find helpful, then please let us know.[1741.375] [1741.375][S02]You can message us on Twitter, or you can email us -[1744.278] [1744.278][S02]podcast@deepmind.com[1746.38]"}
{"file_name": "audio/val_000011.wav", "transcription": "[0.0][S01] So, I mean, you literally say, make me a drug[3.08] [3.08][S01]for X disease, off it goes, says, here's the molecule[6.71] [6.71][S01]you need.[7.29] [7.29][S02] Yeah, yeah.[7.95] [7.95][S01] Do you think that's possible?[8.71] [8.71][S02] It's possible.[9.78] [9.78][S02]I think, it's possible.[10.44] [10.44][S02]I think everything is pointing in that direction.[12.055] [12.055][S03] Well, I've seen first hand is[13.847] [13.847][S03]what can come out of this explosion of two fields coming[16.21] [16.21][S03]together.[16.71] [16.71][S03]I think when you've got experts in two fields,[18.633] [18.633][S03]and they come together, and they're really curious, deeply[21.05] [21.05][S03]curious about the other field, and they[22.34] [22.34][S03]want to apply their thing to your field, that's when[24.59] [24.59][S03]you get this kind of magic.[25.715] [25.715][S02] In five-years' time,[27.99] [27.99][S02]doing drug design without AI will[30.32] [30.32][S02]be like doing any sort of science without maths.[32.93] [32.93][S02]And to be honest, I think the whole of science[34.85] [34.85][S02]will be like, this.[36.05] [36.05][S02]It's like if you're not using AI, what are you doing?[38.38] [41.845][S01] Welcome back to \"Google DeepMind,\" the Podcast.[44.82] [44.82][S01]My name is Professor Hannah Fry.[46.8] [46.8][S01]Now, you'll already know that Demis Hassabis and John[49.94] [49.94][S01]Jumper won the Nobel Prize in 2024[53.0] [53.0][S01]for their work applying artificial intelligence[55.4] [55.4][S01]to protein folding.[56.76] [56.76][S01]Now, everything comes down to proteins[58.88] [58.88][S01]in the human body-- how they fold, how they function.[61.99] [61.99][S01]But not very long ago, working out[64.11] [64.11][S01]the structure for one single protein[66.72] [66.72][S01]could take months, or even years.[69.25] [69.25][S01]And then, with the release of AlphaFold2, the algorithms[72.93] [72.93][S01]developed at Google DeepMind, the entire field[76.26] [76.26][S01]has been completely revolutionized.[79.15] [79.15][S01]More recently, AlphaFold3 can predict the structure[82.44] [82.44][S01]of all of life's molecules with unprecedented accuracy,[86.56] [86.56][S01]which turns out to be absolutely pivotal for drug design.[91.36] [91.36][S01]These developments have paved the way[93.18] [93.18][S01]for a new company spun out of Google DeepMind.[96.07] [96.07][S01]It's called, Isomorphic Labs, to represent the synergy[98.76] [98.76][S01]between biology and AI.[100.66] [100.66][S01]And joining me today are two of its most notable hires.[104.35] [104.35][S01]Rebecca Paul is Head of Medicinal Drug Design[107.02] [107.02][S01]with years of experience in the process of discovering[109.56] [109.56][S01]new drugs.[110.47] [110.47][S01]And Max Jaderberg is its Chief AI Officer.[113.59] [113.59][S01]Anyone who's been following this podcast will remember Max from[116.73] [116.73][S01]his earlier days at DeepMind, teaching agents to play \"Capture[119.83] [119.83][S01]the Flag\" and \"StarCraft.\"[121.4] [121.4][S01]And today, he thinks that agents will[123.49] [123.49][S01]be instrumental in the future of drug discovery.[126.73] [126.73][S01]Max, Rebecca, thank you so much for joining me.[128.697] [128.697][S03] Thank you for having us.[130.28] [130.28][S02] Yes, a pleasure to be here.[131.39] [131.39][S01] Well, it's a delight to have you because there's[133.06] [133.06][S01]some big stuff happening.[134.87] [134.87][S01]I know a lot has been made of this claim[137.89] [137.89][S01]that AI is going to be able to solve all diseases.[141.23] [141.23][S01]Is that realistic, Max?[144.703] [144.703][S02] This isn't going to happen overnight.[146.87] [146.87][S02]Let's be clear.[147.67] [147.67][S02]But I think the exciting thing is that we can actually[149.92] [149.92][S02]see there's, perhaps, a practical path[152.74] [152.74][S02]towards that point.[154.405] [154.405][S02]And it's very different at this point in time[156.28] [156.28][S02]than it has ever been, because we've[157.78] [157.78][S02]got these AI machine-learning models that[160.6] [160.6][S02]understand the biological world and the biochemical world[163.21] [163.21][S02]in a completely different manner than what we've had before.[166.85] [166.85][S02]And so that's opening up tons of disease space[170.2] [170.2][S02]that we didn't think was tractable before.[173.17] [173.17][S02]And that's just today.[174.56] [174.56][S02]So as we start to develop these models further and further,[177.53] [177.53][S02]it's really just the very beginning.[179.05] [179.05][S01] Does that mean--[180.49] [180.49][S01]I mean, every disease is on the table here, Becky.[183.455] [183.455][S03] So I would say nothing is off the table[186.43] [186.43][S03]at this point.[187.19] [187.19][S03]And the journey for me here has been a big one.[190.22] [190.22][S03]I used to be much more conservative in this space.[192.71] [192.71][S03]But having come to Isomorphic Labs, seeing the kind of models[195.82] [195.82][S03]that we have, things that I thought in the past[198.175] [198.175][S03]we would never be able to predict[199.55] [199.55][S03]and now being able to do it every day within five,[201.633] [201.633][S03]ten seconds, it's completely shifted my mindset.[204.47] [204.47][S03]So now, I would put nothing off the table.[206.22] [206.22][S01] And is it just in drug design,[208.12] [208.12][S01]or is AI going to affect clinical trials, as well?[211.48] [211.48][S02] I think over time, yes, it will definitely[214.51] [214.51][S02]affect clinical trials.[216.29] [216.29][S02]We're focusing really heavily on the drug-design phase[219.16] [219.16][S02]at the moment.[220.99] [220.99][S02]But you can imagine a world where,[223.04] [223.04][S02]as you start to get better and better at drug design, actually,[225.73] [225.73][S02]more and more of the bottleneck comes[227.98] [227.98][S02]onto the clinical-development side of things.[230.54] [230.54][S02]And so we really need to rethink how we do that.[232.713] [232.713][S02]I don't think we've really changed[234.13] [234.13][S02]the way we do clinical development[235.547] [235.547][S02]for a long, long time.[237.42] [237.42][S02]It's still very slow.[239.74] [239.74][S02]There's, of course, a lot of regulation for good reason.[242.32] [242.32][S02]But as we understand more and more about how[245.04] [245.04][S02]these molecules work, the true mechanisms of disease,[249.7] [249.7][S02]how these molecules also interact[251.362] [251.362][S02]with the rest of the body and affect everything[253.32] [253.32][S02]like toxicity inside of us, we can[255.39] [255.39][S02]start to rethink even how we do those clinical-trial designs,[258.25] [258.25][S02]how we first go into people and start[259.86] [259.86][S02]measuring the efficacy of these molecules.[261.64] [261.64][S02]So that's not right now for us at Isomorphic,[264.13] [264.13][S02]but it's very much in our future.[265.505] [265.505][S01] What is the big idea then?[267.28] [267.28][S01]What's the big ambition of Isomorphic?[269.22] [269.22][S02] It really is stepping[270.72] [270.72][S02]towards that solvable-disease space.[273.13] [273.13][S02]And so this is really-- before you go into clinical trial,[276.567] [276.567][S02]before you even start testing out on people,[278.4] [278.4][S02]you need something to test.[279.9] [279.9][S02]You need to create a molecule and a drug.[282.0] [282.0][S02]And a drug, what is that?[284.38] [284.38][S02]It's something that goes in and modulates[286.5] [286.5][S02]some function of the body, some function in a cell.[289.63] [289.63][S02]And so, really, the key first phase of Isomorphic Labs[294.21] [294.21][S02]is, how can we create this AI drug-design engine that[298.07] [298.07][S02]can take pretty much any disease, any protein[301.43] [301.43][S02]target that's implicated in that disease,[303.93] [303.93][S02]and work out how to create a molecule that will go in[307.1] [307.1][S02]and start modulating the function of these proteins,[310.05] [310.05][S02]the function of cells, and then change the disease[312.2] [312.2][S02]state for positive patients.[314.25] [314.25][S01] The thing is, I mean,[315.96] [315.96][S01]diseases have been cured in the past.[318.38] [318.38][S01]We have come up with drug solutions that, effectively,[320.87] [320.87][S01]make them a remnant of history.[323.88] [323.88][S01]Why are some diseases so much more difficult[326.21] [326.21][S01]than others to solve or cure?[327.925] [327.925][S03] So, for example, some types of cancers,[331.53] [331.53][S03]you can develop a treatment to cure them[333.32] [333.32][S03]because, for example, those cancers might be quite stable.[335.94] [335.94][S03]So there's maybe one mutation that drives that cancer.[338.37] [338.37][S03]And then you treat that. you treat with that molecule that[341.69] [341.69][S03]hits that particular mutation, and that can[343.91] [343.91][S03]be curative in that disease.[345.27] [345.27][S03]But if you had a cancer which was continually evolving[347.63] [347.63][S03]and the cells were evolving to overcome the drug that you're[352.01] [352.01][S03]treating with, then you need to be continually evolving the drug[354.728] [354.728][S03]that you're treating them with.[356.02] [356.02][S03]So that would be then a much more difficult disease to solve.[358.562] [358.562][S01] And, I guess, are there[360.02] [360.02][S01]others where there's just a real gap in our understanding[362.523] [362.523][S01]of what's going on in the body.[363.815] [363.815][S03] Yeah, there's also that massive gap[365.94] [365.94][S03]in understanding, what's actually driving this disease?[368.95] [368.95][S03]Is it multiple things together?[370.57] [370.57][S03]We don't have all the answers yet to biology.[372.61] [372.61][S03]Biology is so, so complex.[375.69] [375.69][S03]One of the fundamental things we need to do[377.88] [377.88][S03]is find out what biology is actually[380.82] [380.82][S03]driving disease before we can then develop a small molecule,[383.64] [383.64][S03]or some kind of chemistry that then will modulate[386.13] [386.13][S03]the right biology to actually make[388.17] [388.17][S03]an improvement in the symptoms, or a disease[390.51] [390.51][S03]modification to that disease.[391.73] [391.73][S01] Is that what you're trying to do then?[393.813] [393.813][S01]So actually, something like cancer,[395.68] [395.68][S01]for instance, once you get down to the level of molecules[399.99] [399.99][S01]and proteins, it's that something's gone wrong[403.05] [403.05][S01]at that level, which then escalates[406.05] [406.05][S01]into the scale of the human body.[408.228] [408.228][S03] Absolutely.[409.27] [409.27][S03]So you've got little-- you can think[410.77] [410.77][S03]of proteins as mini factories or engines inside your cell.[415.54] [415.54][S03]And those, they have a function, they do something.[418.22] [418.22][S03]So if you have a mutation, or something that[420.91] [420.91][S03]changes about that protein, that means,[423.56] [423.56][S03]for example, it's always switched on,[425.12] [425.12][S03]where, in a normal cell, it might be going on and off,[427.37] [427.37][S03]or maybe it's mostly off.[428.622] [428.622][S03]But now in a cancer cell, something's happened,[430.58] [430.58][S03]it's always on.[431.69] [431.69][S03]It's going to just continually drive a signal.[434.06] [434.06][S03]And that signal could be grow, proliferate.[436.142] [436.142][S03]And that will then drive the formation of a tumor.[438.225] [438.225][S01] And so your target then,[440.073] [440.073][S01]you know that there's something going on with that protein[442.49] [442.49][S01]so you're targeting then that protein with a molecule[445.09] [445.09][S01]of some medicine or drug.[446.132] [446.132][S03] Exactly.[447.048] [447.048][S01] OK.[447.71] [447.71][S03] So you want to-- if you[449.252] [449.252][S03]think of that little factory working away,[451.88] [451.88][S03]you want to design a perfect-shaped wrench[454.81] [454.81][S03]that you can throw into that little set of gears[457.03] [457.03][S03]to stop it working so that then it's going[459.078] [459.078][S03]to stop driving that signal.[460.245] [460.245][S01] Does it come down to shape?[461.475] [461.475][S03] Very important.[462.683] [462.683][S03]You can think of proteins as having pockets, or grooves,[465.73] [465.73][S03]or crevices, and you really need to design[467.8] [467.8][S03]a molecule that perfectly fits inside that groove.[470.32] [470.32][S01] To block it.[471.32] [471.32][S03] To block it.[471.82] [471.82][S01] Oh, really?[472.46] [472.46][S03] It's literally like that, yes.[474.293] [474.293][S01] And is that true of all drugs[476.463] [476.463][S01]that already exist, I mean, like paracetamol, for example.[478.88] [478.88][S01]Is that what it's doing?[480.03] [480.03][S03] Yeah.[480.822] [480.822][S03]So paracetamol is a small molecule, binds to its protein[483.73] [483.73][S03]targets, blocks the function, and in the case of paracetamol,[488.53] [488.53][S03]or painkillers, it's preventing you from experiencing pain.[491.45] [491.45][S03]So, yes.[492.615] [492.615][S01] So in some ways then,[493.99] [493.99][S01]you're playing LEGO at the molecular level.[497.475] [497.475][S03] Exactly, yes, and we love it.[500.115] [500.115][S01] Why would that be a good situation for AI?[503.39] [503.39][S01]What makes this a well-suited problem for AI, Max?[506.325] [506.325][S02] It's actually such a perfect application of AI[509.29] [509.29][S02]and machine-learning, this whole playing LEGO with molecules.[513.07] [513.07][S02]We've seen over the last five, six, seven years the rise[516.789] [516.789][S02]of models like AlphaFold.[518.27] [518.27][S02]We had AlphaFold1, AlphaFold2, understanding[520.69] [520.69][S02]the structure of proteins.[522.01] [522.01][S02]Before AlphaFold2, no one could really understand the structure[525.82] [525.82][S02]without going into a lab and experimentally resolving[529.51] [529.51][S02]these structures.[530.27] [530.27][S02]And that can take months, it can take years.[532.18] [532.18][S02]Sometimes, it's not even possible for some proteins.[534.37] [534.37][S02]Obviously, AlphaFold2-- Nobel Prize winning breakthrough[537.24] [537.24][S02]in chemistry.[538.03] [538.03][S02]And now, we've taken that even further with things[540.113] [540.113][S02]like AlphaFold3, where now we can[542.31] [542.31][S02]understand the structure of proteins with small molecules.[545.59] [545.59][S02]And these small molecules are the little LEGO blocks[547.77] [547.77][S02]that come in, and we use those drugs[549.51] [549.51][S02]to inhibit the function of a protein,[551.79] [551.79][S02]or change the function of a protein.[553.48] [553.48][S02]And the reason why this is such a good domain[556.402] [556.402][S02]for machine-learning is that what we're trying to do,[558.61] [558.61][S02]in essence, is predict the 3D coordinates of this biomolecular[563.97] [563.97][S02]system.[564.82] [564.82][S02]And this fits really, really nicely[566.91] [566.91][S02]into some of our classic, supervised learning, modeling[571.14] [571.14][S02]domains.[571.99] [571.99][S02]It fits really, really nicely into our diffusion modeling[576.81] [576.81][S02]frameworks that have been so, so successful[580.14] [580.14][S02]for things like image generation, or video generation.[583.025] [583.025][S01] And in terms of being suited[584.76] [584.76][S01]for the supervised learning stuff that had already[586.843] [586.843][S01]gone before, is that because there is some metric of success[589.71] [589.71][S01]here, some way of being right, inverted commas.[592.842] [592.842][S02] There's a very clear metric of success[595.05] [595.05][S02]here, which is super helpful for developing these models[597.9] [597.9][S02]and for research.[598.9] [598.9][S02]And the reason is, over the last 50 years,[602.23] [602.23][S02]people have been experimentally resolving these protein[605.49] [605.49][S02]structures-- structure of these proteins with small molecules,[608.26] [608.26][S02]with DNA, with RNA.[609.67] [609.67][S02]They've been doing it by hand in a lab,[611.4] [611.4][S02]and then depositing the results into a big database[614.25] [614.25][S02]called The Protein Data Bank, PDB.[616.18] [616.18][S02]And this gives a really rich source of information.[620.53] [620.53][S02]It's a couple of 100,000 3D structures.[623.29] [623.29][S02]And each structure has thousands of atom coordinates,[626.95] [626.95][S02]so there's a really high information density.[629.23] [629.23][S02]And that's a perfect scenario for supervised learning.[631.75] [631.75][S02]Now, this isn't web-scale data, so it's not the scale of data[635.04] [635.04][S02]that we might be used to for training large-language models.[638.2] [638.2][S02]But incredibly, we've worked out ways to design these[641.01] [641.01][S02]neural-network architectures and these training regimes so that[644.52] [644.52][S02]we can only train on a couple of 100,000 structures[648.42] [648.42][S02]and excitingly get what we call generalization.[651.27] [651.27][S02]We see that these models can generalize.[653.56] [653.56][S02]They can be applied to completely new proteins,[655.792] [655.792][S02]completely new molecules that people have never[657.75] [657.75][S02]seen before in history.[658.96] [658.96][S02]And, of course, that's essential if you're doing drug design.[661.92] [661.92][S02]Drug design is about creating completely new molecules[664.17] [664.17][S02]that we've never seen before in nature, even,[666.12] [666.12][S02]to actually modulate these functions.[667.662] [667.662][S01] But then, also, I guess, the possibilities[671.07] [671.07][S01]of molecules that you could design--[672.79] [672.79][S01]I mean, it's-- well, a very big number, I imagine.[675.725] [675.725][S02] Yeah, that number is huge.[678.51] [678.51][S02]People throw around numbers like 10 to the power of 60[682.71] [682.71][S02]is the possible number of drug-like molecules[685.837] [685.837][S02]out there in the universe.[686.92] [686.92][S02]So it's a huge combinatorial problem.[689.23] [689.23][S02]And that's also a really exciting spot for AI.[693.49] [693.49][S02]We can create great predictive models[696.24] [696.24][S02]of how these molecules fit together, even how strongly they[700.56] [700.56][S02]fit, or the properties of them.[702.51] [702.51][S02]But with the design space of 10 to the power of 60,[705.72] [705.72][S02]we're reaching the level of atoms in the universe.[709.88] [709.88][S02]So even if you had the perfect predictive models[713.59] [713.59][S02]of how this fits together, you wouldn't[715.81] [715.81][S02]be able to exhaustively search that massive space.[718.64] [718.64][S02]So what do you do?[719.39] [719.39][S02]OK, maybe you subsample that space and you search[722.74] [722.74][S02]through some large number--[724.9] [724.9][S02]a million, 10 million, a billion, 10 billion.[727.16] [727.16][S01] It's nothing, is it?[728.08] [728.08][S02] Even though you're not-- yeah.[728.89] [728.89][S01] Yeah.[729.19] [729.19][S02] You're not even scratching the surface.[731.44] [731.44][S02]And that's where we can then fall to new types of models,[734.9] [734.9][S02]things like generative models, search methods, agents,[738.77] [738.77][S02]which instead of exhaustively searching[741.19] [741.19][S02]the full molecular space, we can really smartly start[745.63] [745.63][S02]to explore across that whole space[747.047] [747.047][S02]but without exhaustively searching the whole space.[749.172] [749.172][S01] OK, tell me about AlphaFold3 then.[751.31] [751.31][S01]In terms of designing drugs, what does it actually[754.45] [754.45][S01]allow you to do?[755.45] [755.45][S03] So as a medicinal chemist,[757.34] [757.34][S03]we always want to be able to visualize how our molecule binds[762.25] [762.25][S03]to our protein, so how our LEGO block fits into the bigger[765.707] [765.707][S03]picture of LEGO blocks.[766.665] [766.665][S01] Thank you for going with me on this analogy,[768.05] [768.05][S01]I appreciate it.[768.717] [768.717][S03] The LEGO analogy.[770.14] [770.14][S03]The reason we need to have that visualization[772.015] [772.015][S03]is because when you're optimizing[773.65] [773.65][S03]a small molecule binding to a protein,[775.76] [775.76][S03]you need to know which vector to explore.[778.37] [778.37][S03]You need to have some kind of target in mind.[780.26] [780.26][S03]I'm going to explore that part of the pocket,[782.135] [782.135][S03]or I'm going to that part of the protein structure.[785.17] [785.17][S03]That looks like a good place to go.[788.17] [788.17][S03]And so for years, we've invested,[789.97] [789.97][S03]as a scientific community, in ways to do that.[792.53] [792.53][S03]So X-ray crystallography is an experimental technique[795.55] [795.55][S03]where you can actually go into a lab,[797.44] [797.44][S03]and you can spend a lot of time crystallizing your protein.[801.26] [801.26][S03]You fire X-rays at it once it's bound to your small molecule.[805.0] [805.0][S03]And then you can actually visualize, atom by atom,[807.85] [807.85][S03]how your small molecule is binding to your protein.[810.23] [810.23][S03]Now we can do that with AlphaFold3, and the latest[813.85] [813.85][S03]iterations of that model, into seconds.[816.56] [816.56][S03]And so something that might have taken me[819.7] [819.7][S03]months when I was doing my PhD or in my early-stage research,[823.55] [823.55][S03]I'm now just seeing on my screen all the time.[825.78] [825.78][S03]And so you can just iterate and iterate in silico[827.932] [827.932][S03]until you get to something that actually looks really quite[830.39] [830.39][S03]promising.[831.15] [831.15][S03]And then you take that into the lab.[832.65] [832.65][S01] So which way around does it work then?[834.27] [834.27][S01]Are you saying, OK, I think something like this, this, this,[836.84] [836.84][S01]and this would work.[837.96] [837.96][S01]Let's try it out and see if it fits.[840.05] [840.05][S01]Or is it the other way around?[843.282] [843.282][S01]Are these models telling you this[845.09] [845.09][S01]is something that might fit?[846.33] [846.33][S03] So in the way that we've[847.913] [847.913][S03]constructed our drug-design platform at ISO,[849.78] [849.78][S03]you can do both.[850.95] [850.95][S03]So I can come in as an experienced medicinal chemist[853.31] [853.31][S03]and I can say, I think this.[854.87] [854.87][S03]And I can test it then and there, a couple of minutes,[857.12] [857.12][S03]and get that feedback.[858.93] [858.93][S03]But you can also take the opposite approach.[861.613] [861.613][S03]I don't actually know what's going[863.03] [863.03][S03]to work here so I'm going to apply the generative models we[865.34] [865.34][S03]have.[865.95] [865.95][S03]I'm going to do some, what we call, virtual screening.[868.68] [868.68][S03]So I'm going to take an area of chemical space[871.49] [871.49][S03]that's commercially available.[872.88] [872.88][S03]I'm going to screen that against my protein.[874.78] [874.78][S03]And I'm going to get the models to tell me[876.53] [876.53][S03]what's best from that subset.[878.425] [878.425][S01] Can I see what it looks like when you're actually[881.15] [881.15][S01]designing something?[882.11] [882.11][S03] So the small molecule[883.76] [883.76][S03]is fitting into this little groove in the protein,[887.28] [887.28][S03]and it's forming interactions with the protein.[891.22] [891.22][S03]So the dotted lines you can see, those[893.01] [893.01][S03]are interactions between that small molecule and the protein[895.77] [895.77][S03]itself.[896.62] [896.62][S03]And as medicinal chemists, we want[898.23] [898.23][S03]to optimize, or increase, the number of those interactions,[900.9] [900.9][S03]because that's increasing the strength[902.49] [902.49][S03]of that relationship between the small molecule and the protein.[905.66] [905.66][S01] Let me describe what I've got, what's going on here.[908.38] [908.38][S01]So you've got-- so the curly stuff is a protein.[911.097] [911.097][S03] Yes, that's the protein, yeah.[912.93] [912.93][S01] And it's folded.[914.463] [914.463][S01]So you've got the three-dimensional structure.[916.38] [916.38][S03] Indeed.[917.17] [917.17][S01] And then over here, you've got--[918.75] [918.75][S01]I mean, this looks like the kind of thing[919.98] [919.98][S01]you would do in GCSE chemistry, one of those kind of diagrams.[922.633] [922.633][S03] Exactly.[923.55] [923.55][S03]This is a small molecule.[924.592] [924.592][S01] This is sort of plugged into the protein?[927.408] [927.408][S03] Yeah.[928.2] [928.2][S01] Wow.[928.66] [928.66][S01]I mean, it really is like 3D jigsaws then.[930.51] [930.51][S01]I mean, that's--[930.72] [930.72][S03] 3D, yes, exactly.[931.74] [931.74][S01] --what you're doing.[932.47] [932.47][S03] That's what we're doing, yeah.[933.24] [933.24][S01] That crevice could be[934.47] [934.47][S01]the thing that's making you feel pain,[936.053] [936.053][S01]or the thing that's causing tumor growth,[938.37] [938.37][S01]or whatever it might be.[939.55] [939.55][S03] Yeah, causing your disease, yeah.[940.83] [940.83][S01] Amazing.[941.663] [941.663][S01]So then you're trying different versions of this molecule[944.7] [944.7][S01]to see if you can get the best possible fit[946.5] [946.5][S01]in that little crevice.[947.255] [947.255][S03] Exactly.[948.172] [948.172][S03]And I can quickly show you.[949.333] [949.333][S03]So we have lots of different functionality on the platform[951.75] [951.75][S03]that you can try.[952.83] [952.83][S03]And this is my favorite, which Max always tells me off[955.08] [955.08][S03]about because I can use my expertise[956.52] [956.52][S03]as a medicinal chemist.[957.478] [957.478][S03]And I can say, OK, I want to make some specific changes[960.45] [960.45][S03]to this molecule.[961.205] [961.205][S01] Oh, wow.[962.08] [962.08][S03] I can actually view what I'm doing in 3D.[964.51] [964.51][S03]So this is now going to fetch that structure prediction.[966.898] [966.898][S03]And I can actually make modifications to this molecule.[969.19] [969.19][S03]And I can see the predicted structure in real-time.[971.25] [971.25][S01] And normally, this would have taken--[973.292] [973.292][S01]I mean, before AI--[974.85] [974.85][S01]a long time.[975.935] [975.935][S03] I mean, if you're going to go into a lab[977.4] [977.4][S03]and experimentally determine this,[978.82] [978.82][S03]it could be anywhere from weeks to years.[980.96] [980.96][S01] Why do you tell her off for this one?[983.482] [983.482][S02] I guess you're referring to the fact[985.11] [985.11][S02]that, over time, we want to do more and more from the model[987.54] [987.54][S02]itself.[988.06] [988.06][S02]The really exciting frontier, from my perspective,[990.143] [990.143][S02]is there's going to be lots of scenarios where Becky will want[993.99] [993.99][S02]to go in and make those changes by hand[995.91] [995.91][S02]and test out very specific hypotheses that she has[999.78] [999.78][S02]on why this molecule works and how we can make it better.[1003.54] [1003.54][S02]And then what we also should be doing[1006.3] [1006.3][S02]is asking our generative models, and our agents to say,[1010.84] [1010.84][S02]hey, this is how I'm thinking about the problem.[1013.36] [1013.36][S02]These are my design constraints.[1014.86] [1014.86][S02]I want a molecule that does X, Y, Z and has these properties,[1017.67] [1017.67][S02]and looks like this, and maybe interacts over there,[1020.25] [1020.25][S02]and makes this sort of shape.[1021.52] [1021.52][S02]What can you come up with?[1022.81] [1022.81][S02]And maybe set this running, go away, have a coffee,[1027.31] [1027.31][S02]go home, come in the next morning[1029.76] [1029.76][S02]and see what the agent has come up with.[1031.91] [1031.91][S01] I guess with all of the projects[1033.78] [1033.78][S01]that you've applied AI to generally in DeepMind,[1036.73] [1036.73][S01]it has gone through that process of starting off[1039.69] [1039.69][S01]with human expertise, and then slowly building[1042.599] [1042.599][S01]in more knowledge and expertise within the model itself.[1045.08] [1045.08][S02] Yeah.[1045.913] [1045.913][S02]And I think there are a lot of analogies to that moment[1048.877] [1048.877][S02]that we had with large-language models[1050.46] [1050.46][S02]where we've had large-language models for a long time,[1053.36] [1053.36][S02]and I've been working on them 10 years ago.[1055.96] [1055.96][S02]But they were rubbish.[1057.99] [1057.99][S02]And they were spitting stuff out that looked like language.[1060.74] [1060.74][S02]It kind of made sense.[1061.91] [1061.91][S02]But it also didn't make sense and you had to correct,[1063.65] [1063.65][S02]and it clearly wasn't human.[1064.817] [1064.817][S02]And then, they got steadily better little by little.[1068.69] [1068.69][S02]And suddenly, they just passed through this human-perceptible[1072.61] [1072.61][S02]threshold where you can't really tell whether this[1075.25] [1075.25][S02]is generated by a human or not.[1077.6] [1077.6][S02]And we're getting to the same point[1081.25] [1081.25][S02]with our molecule-design models where,[1085.81] [1085.81][S02]maybe five years ago in this field,[1088.04] [1088.04][S02]you had generative models of molecules,[1090.19] [1090.19][S02]and they'd spit stuff out.[1091.31] [1091.31][S02]But you'd give them to a chemist like Becky,[1094.1] [1094.1][S02]and she would probably tear her hair out, like this is rubbish.[1097.373] [1097.373][S01] Well, did you see those kinds of models?[1099.54] [1099.54][S03] Yes, I did, yeah, because I actually[1101.623] [1101.623][S03]worked at an AI company prior to joining Isomorphic Lab,[1104.45] [1104.45][S03]so I've seen that progression in that journey.[1106.47] [1106.47][S01] And tell me, what kind of stuff did they spit out?[1109.685] [1109.685][S03] For a long time, it[1111.06] [1111.06][S03]would just be nonsense because you're obviously giving[1114.58] [1114.58][S03]the model an uphill function.[1115.85] [1115.85][S03]You want it to get to something that's going to bind really[1119.33] [1119.33][S03]potently, for example.[1120.815] [1120.815][S03]But to do that, maybe it just makes the molecule massive.[1123.19] [1123.19][S01] Oh.[1123.9] [1123.9][S03] But then, well, it's[1125.317] [1125.317][S03]not going to actually be absorbed through the intestine[1127.61] [1127.61][S03]into the bloodstream because--[1128.63] [1128.63][S01] Because you've got more[1129.23] [1129.23][S01]to think about than just the molecule itself.[1130.75] [1130.75][S03] Exactly, yeah.[1131.24] [1131.24][S03]So there's a lot to piece together here.[1132.97] [1132.97][S01] Oh, that's interesting.[1134.46] [1134.46][S01]So AI, if we go back to our LEGO analogy,[1136.443] [1136.443][S01]it was just like building a massive LEGO[1138.11] [1138.11][S01]wall all around the protein.[1139.255] [1139.255][S03] Yes, yeah.[1140.255] [1140.255][S01] I see.[1141.52] [1141.52][S01]OK.[1142.02] [1142.02][S03] Essentially.[1142.47] [1142.47][S01] And now, this is--[1144.29] [1144.29][S01]have you seen that moment, that tip over with the language[1147.787] [1147.787][S01]models that Max describes?[1148.87] [1148.87][S03] Yeah, I think--[1150.078] [1150.078][S03]I've been so surprised by the quality[1152.6] [1152.6][S03]of some of the molecules that come out of the generative AI.[1155.1] [1155.1][S03]And sometimes the molecules that come out,[1156.947] [1156.947][S03]you think, oh, I would have-- why wouldn't I[1158.78] [1158.78][S03]have come up with that?[1159.738] [1159.738][S03]That's really amazing.[1161.735] [1161.735][S03]And, of course, you don't have to go and make[1163.61] [1163.61][S03]that exact molecule, but you could then[1165.29] [1165.29][S03]use that as inspiration to do something else.[1167.52] [1167.52][S03]So--[1168.048] [1168.048][S01] So it's working with you.[1169.59] [1169.59][S03] Yeah, you can work together.[1170.76] [1170.76][S03]I mean, I'm guessing at some point in the future,[1172.53] [1172.53][S03]it will be so good that you'll be like,[1174.155] [1174.155][S03]oh, there's nothing I would change.[1176.36] [1176.36][S02] We've had some really fun moments where,[1179.75] [1179.75][S02]for example, we've had our models submitting molecules[1182.24] [1182.24][S02]blind, and then other people looking at them[1184.37] [1184.37][S02]and seeing, OK, what are we going to send off for testing?[1186.185] [1186.185][S01] Oh, really?[1187.05] [1187.05][S02] And people look--[1187.34] [1187.34][S01] Like a chewing test for molecules.[1188.36] [1188.36][S02] Exactly.[1188.69] [1188.69][S02]And people looking at these molecules,[1190.273] [1190.273][S02]very, very experienced medicinal chemists, saying, wow,[1192.755] [1192.755][S02]there's a lot of experience behind the design[1194.63] [1194.63][S02]of this molecule, and actually not knowing[1197.36] [1197.36][S02]that this was designed by a generative model instead.[1200.35] [1200.35][S01] But OK, let me understand this, though,[1203.15] [1203.15][S01]because using the analogy of large-language models,[1206.13] [1206.13][S01]it makes sense there that you have these tokens,[1209.09] [1209.09][S01]you break words down into little bytes,[1213.05] [1213.05][S01]and then you can build up from there.[1215.13] [1215.13][S01]How do you do it in a way that makes sense chemically?[1218.31] [1218.31][S01]I mean, you're not just taking atoms, are you?[1220.55] [1220.55][S02] We are, actually, just taking atoms.[1223.25] [1223.25][S02]For bigger things like proteins, we chunk up into amino acids,[1226.68] [1226.68][S02]so one token per amino acids.[1228.96] [1228.96][S02]So instead of characters of a sentence, letters of a sentence,[1232.23] [1232.23][S02]we have amino acids of a protein.[1234.0] [1234.0][S02]And then for the small molecule, we chunk it up just[1236.6] [1236.6][S02]into its individual atoms.[1238.022] [1238.022][S02]And so we have a sequence of amino acids,[1239.73] [1239.73][S02]and a sequence of atoms, and we put them together[1242.02] [1242.02][S02]and that's one big sequence.[1243.28] [1243.28][S02]And then we feed it through a structured model[1245.25] [1245.25][S02]like AlphaFold3.[1246.76] [1246.76][S02]And AlphaFold3 uses transformers.[1251.55] [1251.55][S02]But unlike in large-language models[1253.77] [1253.77][S02]where transformers are used on one-dimensional sequences[1256.5] [1256.5][S02]of characters, of letters, here, we[1258.63] [1258.63][S02]use what we call a pair former, which[1261.3] [1261.3][S02]operates on a two-dimensional interaction grid of all[1265.5] [1265.5][S02]of these molecular elements.[1267.4] [1267.4][S02]So we can consider every single possible interaction that[1270.78] [1270.78][S02]could occur between every amino asset, every part[1273.51] [1273.51][S02]of the protein, and every atom of the small molecule,[1276.63] [1276.63][S02]and everything in between.[1277.9] [1277.9][S02]And then this creates neural-network features, which[1281.28] [1281.28][S02]condition a diffusion model.[1283.75] [1283.75][S02]And diffusion models are generative models.[1287.86] [1287.86][S02]We probably know them from these amazing image-generative models,[1290.65] [1290.65][S02]video-generative models.[1291.7] [1291.7][S02]And instead of generating the pixels of an image, instead,[1295.17] [1295.17][S02]our diffusion models are generating the 3D-atom[1297.72] [1297.72][S02]coordinates of this whole biomolecular system.[1299.975] [1299.975][S01] And it just so happens they work.[1301.85] [1301.85][S02] And it just so happens[1303.392] [1303.392][S02]this works phenomenally well.[1304.77] [1304.77][S02]You get these amazing structure predictions that-- when[1307.56] [1307.56][S02]you go to the lab and experimentally resolve[1309.45] [1309.45][S02]these structures-- and we do this on occasion.[1313.082] [1313.082][S02]Something amazing is predicted, like a completely new pocket[1315.69] [1315.69][S02]or a new mechanism of action.[1317.098] [1317.098][S02]We go into the lab.[1317.89] [1317.89][S02]We want to check that, are these models[1320.4] [1320.4][S02]grounded at all in reality?[1322.09] [1322.09][S02]And what comes back is, yeah, this[1324.18] [1324.18][S02]is within 1 angstrom, the tiniest unit of distance[1329.61] [1329.61][S02]accurate, which is phenomenal.[1330.965] [1330.965][S01] Yeah.[1332.19] [1332.19][S01]But then through that training process,[1334.39] [1334.39][S01]does it manage to extract a conceptual understanding of how[1339.12] [1339.12][S01]chemistry works?[1339.965] [1339.965][S02] It's really hard to think about concepts[1342.57] [1342.57][S02]in this atom space.[1343.9] [1343.9][S02]But I do believe that there's some notion of reasoning[1348.06] [1348.06][S02]in molecular and atomistic space that these models are doing[1352.08] [1352.08][S02]because of the amount of generalization[1354.0] [1354.0][S02]we're getting out of them.[1355.22] [1355.22][S02]And if you think about what these generative models are[1357.76] [1357.76][S02]trained to do, they're trained to fit to the data distribution[1361.778] [1361.778][S02]that you give them.[1362.57] [1362.57][S02]And so in our case, we give them all the molecules[1366.565] [1366.565][S02]that might exist naturally, that people have worked out before,[1369.19] [1369.19][S02]that people have designed before.[1370.52] [1370.52][S02]When your model gets better and better,[1372.11] [1372.11][S02]you get things that look like they could have been designed[1374.568] [1374.568][S02]before, which, then it starts to be[1377.02] [1377.02][S02]imperceptible from a human design.[1378.475] [1378.475][S01] What do you think?[1379.725] [1379.725][S03] You can certainly see that in the molecules[1381.67] [1381.67][S03]that we get back.[1382.46] [1382.46][S03]They look like molecules that myself or someone else[1385.287] [1385.287][S03]might have designed.[1386.12] [1386.12][S03]You do sometimes get something crazy, though.[1387.83] [1387.83][S01] Do you?[1388.7] [1388.7][S03] Yeah, we still do, but we have[1390.918] [1390.918][S03]ways of filtering that out now.[1392.21] [1392.21][S01] Is it like a hallucination, in a way?[1394.252] [1394.252][S03] Yeah.[1395.043] [1395.043][S03]The model is really confident, but it's confidently wrong.[1397.48] [1397.48][S01] Confidently wrong.[1399.19] [1399.19][S01]OK.[1399.91] [1399.91][S01]What does it look like when it hallucinates?[1402.53] [1402.53][S03] So we get back a number[1405.64] [1405.64][S03]of structures that the models believe[1407.2] [1407.2][S03]are good solutions for this problem,[1409.39] [1409.39][S03]for this particular protein pocket.[1411.41] [1411.41][S03]And then our job is to go, OK, well, which of those molecules[1415.07] [1415.07][S03]should we actually select to put into synthesis?[1417.69] [1417.69][S03]And so this is never a model on its own in a silo.[1420.57] [1420.57][S03]This is a model working really closely with an expert[1422.84] [1422.84][S03]to say, OK, of the solutions you've given me,[1425.76] [1425.76][S03]where's the gold in that?[1427.4] [1427.4][S03]Where are the molecules that are actually going[1430.04] [1430.04][S03]to push this project forward?[1431.687] [1431.687][S03]And we'll put them into what we call chemical synthesis, where[1434.27] [1434.27][S03]we make them and we test them.[1435.81] [1435.81][S03]But sometimes we get back results[1437.64] [1437.64][S03]which, actually, the compound doesn't[1440.48] [1440.48][S03]bind to the protein at all.[1441.99] [1441.99][S03]So here, the models, essentially,[1443.45] [1443.45][S03]completely hallucinated a solution[1445.62] [1445.62][S03]and so convincingly that actually, an expert looks at it[1448.547] [1448.547][S03]and goes, yeah, that looks a really good solution.[1450.63] [1450.63][S03]The confidence metrics would suggest the same.[1453.29] [1453.29][S03]So essentially, it is a hallucination.[1455.37] [1455.37][S03]And I think we find it a really fascinating research question[1458.9] [1458.9][S03]to say, OK, how do we find the really good stuff[1462.083] [1462.083][S03]that the model is giving us?[1463.25] [1463.25][S01] Beyond this 3D jigsaw or LEGOs--[1465.66] [1465.66][S01]we're mixing our metaphors.[1468.56] [1468.56][S01]Is it just about structure?[1470.173] [1470.173][S01]Is it just about finding something[1471.59] [1471.59][S01]that will plug a particular hole?[1473.43] [1473.43][S01]Or are there other considerations[1474.805] [1474.805][S01]that you have to have, as well, when it comes to drug design?[1477.347] [1477.347][S03] There are so many other considerations.[1479.73] [1479.73][S03]That's what makes this problem just incredibly complex.[1482.34] [1482.34][S03]So it's even beyond the shape.[1484.133] [1484.133][S03]You can have something that maybe fits in the shape,[1486.3] [1486.3][S03]but it's got to bind really strongly to that protein.[1489.66] [1489.66][S03]And that prediction of binding affinity, as we call it,[1492.78] [1492.78][S03]is actually different.[1493.71] [1493.71][S03]You can't really gauge that from just looking at a picture.[1496.168] [1496.168][S03]The picture is a helpful guide, but you[1497.87] [1497.87][S03]need to be able to predict that binding affinity separately.[1501.41] [1501.41][S03]And then all of those things together,[1503.04] [1503.04][S03]that's just how your molecule binds to your protein.[1505.53] [1505.53][S03]You've also got to think about, is that molecule going to bind[1508.113] [1508.113][S03]to any of the other 20,000 proteins in the body?[1510.42] [1510.42][S03]Because if the answer is yes, that could drive a side effect[1513.14] [1513.14][S03]that you don't want.[1514.17] [1514.17][S03]That's going to drive you some toxicity.[1516.35] [1516.35][S03]Is this molecule going to be stable?[1518.1] [1518.1][S03]It's got to survive the really acidic conditions[1520.1] [1520.1][S03]of the stomach.[1521.01] [1521.01][S03]It's got to survive going through the liver, which[1523.16] [1523.16][S03]is going through a war zone.[1525.05] [1525.05][S03]The liver wants to do everything it can.[1526.867] [1526.867][S03]It's like, this is a foreign molecule,[1528.45] [1528.45][S03]I need to get rid of it.[1529.515] [1529.515][S03]So your molecule has got to be really robust.[1531.39] [1531.39][S03]It's got to survive that journey.[1532.97] [1532.97][S03]It's got to be soluble.[1534.41] [1534.41][S03]When you take a pill, that pills got to dissolve in your stomach,[1537.25] [1537.25][S03]and it's got to stay dissolved all the way[1539.02] [1539.02][S03]through your intestine because, otherwise, it's not[1541.145] [1541.145][S03]going to absorb into your body.[1542.93] [1542.93][S03]And a lot of these parameters are pulling against each other.[1546.32] [1546.32][S03]So for something to be soluble and dissolve really well,[1549.67] [1549.67][S03]it needs to be water loving.[1552.02] [1552.02][S03]But for it to bind to the protein[1553.87] [1553.87][S03]and to gain affinity in that kind of pocket, that LEGO[1556.63] [1556.63][S03]connection, it actually needs to be water hating.[1559.618] [1559.618][S01] Well, how do you possibly[1561.16] [1561.16][S01]solve that if you need opposing characteristics?[1565.25] [1565.25][S03] So we've-- up until this point and still now,[1568.15] [1568.15][S03]to some degree, drug discovery is a very iterative process.[1571.22] [1571.22][S03]So the human brain can only think about so many things[1574.6] [1574.6][S03]at one time.[1575.24] [1575.24][S03]So you're like, OK, I'm going to solve[1577.06] [1577.06][S03]this binding-affinity problem a bit,[1578.89] [1578.89][S03]and then, I'm going to start to think, OK,[1580.64] [1580.64][S03]is my molecule soluble?[1581.768] [1581.768][S03]And then I'm going to start to bring in gradually[1583.81] [1583.81][S03]these other properties.[1584.78] [1584.78][S03]And honestly, it's like Whack-a-mole.[1586.46] [1586.46][S03]You play like this three-year game of Whack-a-mole[1588.55] [1588.55][S03]where you're like, I fixed this problem.[1590.18] [1590.18][S03]Hooray.[1590.68] [1590.68][S03]And then this other one pops up and you're[1592.502] [1592.502][S03]like, OK, I'll fix that, but then the other one's gone bad,[1594.96] [1594.96][S03]again.[1595.69] [1595.69][S03]So to be able to predict all these things in silico[1597.96] [1597.96][S03]is still a really, really hard problem.[1599.66] [1599.66][S01] Are you working on tools that will[1601.59] [1601.59][S01]help with those elements, too?[1602.87] [1602.87][S02] Yeah, absolutely.[1604.203] [1604.203][S02]So we think about, ISO, how do we do drug design[1608.34] [1608.34][S02]end-to-end, which really means solving all of these very, very[1611.31] [1611.31][S02]hard problems with cell permeability, solubility,[1615.31] [1615.31][S02]toxicity, liver clearance, everything.[1618.84] [1618.84][S02]And none of this is solved.[1620.98] [1620.98][S02]These are really hard problems to even model.[1622.96] [1622.96][S02]So we spend a lot of effort, a lot of research,[1625.47] [1625.47][S02]on creating new models to really understand this better.[1628.99] [1628.99][S02]And then as Becky was talking about,[1631.0] [1631.0][S02]how do we then start to find these needles in a haystack[1634.6] [1634.6][S02]molecules that are somehow just balancing the properties[1640.23] [1640.23][S02]just right to be a perfect drug?[1642.43] [1642.43][S02]And it's really, really hard.[1644.14] [1644.14][S02]And actually, there are a lot of analogies to maybe what I used[1648.33] [1648.33][S02]to do at DeepMind in, for example, \"Capture the Flag\"[1651.17] [1651.17][S02]or \"StarCraft\" is there's not just one agent,[1656.36] [1656.36][S02]or strategy that solves \"StarCraf\"[1659.3] [1659.3][S02]or a particular game like \"Go.\"[1661.17] [1661.17][S02]You have to completely start mixing up these strategies,[1664.01] [1664.01][S02]and working out exploits for each individual strategy,[1667.26] [1667.26][S02]and, basically, searching this huge combinatorial strategy[1670.57] [1670.57][S02]space.[1671.07] [1671.07][S02]In the same way, we need to be searching[1672.95] [1672.95][S02]this huge combinatorial-molecule space.[1675.36] [1675.36][S02]So just like you might have tree search in a game of \"Go,\"[1678.29] [1678.29][S02]where at every move, you elucidate some other possible[1682.42] [1682.42][S02]moves, and you start searching through that tree of possible[1684.92] [1684.92][S02]strategies going deeper and deeper.[1686.85] [1686.85][S02]And just like you can do that for moves in a game of \"Go,\"[1689.88] [1689.88][S02]you can imagine doing a similar thing for designing a molecule.[1692.847] [1692.847][S02]So you start with a part of a molecule[1694.43] [1694.43][S02]and you start to hypothesize, what[1696.618] [1696.618][S02]are the different things I could add or take away[1698.66] [1698.66][S02]from this molecule?[1699.75] [1699.75][S02]And you get to a whole tree of possible futures[1702.29] [1702.29][S02]that you can then score and work out[1703.91] [1703.91][S02]a value associated with that to create[1705.5] [1705.5][S02]that perfect molecule for this very specific indication.[1708.13] [1708.13][S01] But how do you even know that the perfect molecule[1710.713] [1710.713][S01]exists?[1711.31] [1711.31][S01]Maybe there's just some crevices in the proteins[1713.84] [1713.84][S01]that just are unfittable.[1715.06] [1715.06][S03] We do have this concept of undruggable proteins[1719.15] [1719.15][S03]where the crevice is really flat,[1721.16] [1721.16][S03]and you can't really get anything to grip in there.[1723.89] [1723.89][S03]And those proteins might need different solutions.[1726.33] [1726.33][S03]So actually, we have a whole emerging field[1728.39] [1728.39][S03]which we call molecular glues.[1729.96] [1729.96][S03]And this is where you have two proteins that come together,[1734.34] [1734.34][S03]and the pocket that's formed when they come together[1737.03] [1737.03][S03]is actually a much more suitable pocket.[1738.88] [1738.88][S03]So now, you need to design a molecule that[1740.63] [1740.63][S03]sits in the middle of them and glues them together.[1744.0] [1744.0][S03]So there's this whole explosion of all[1745.61] [1745.61][S03]these different modalities now, which[1747.152] [1747.152][S03]makes this an incredibly exciting field to work in.[1749.802] [1749.802][S02] From my perspective,[1751.26] [1751.26][S02]the fact that we've actually found any drugs at all already,[1754.98] [1754.98][S02]given how hard and complex the problem is--[1757.16] [1757.16][S02]and we've basically been doing a bit of human intuition[1760.7] [1760.7][S02]and a lot of random screening and experimental testing.[1763.8] [1763.8][S02]And we've managed to find molecules,[1765.44] [1765.44][S02]even though the design space is huge.[1767.087] [1767.087][S02]Actually, that gives me a lot of hope,[1768.67] [1768.67][S02]because that means that there's probably a lot of redundancy[1771.54] [1771.54][S02]in chemical space, i.e., there's probably[1774.03] [1774.03][S02]lots of different solutions that could work,[1776.62] [1776.62][S02]but we've just got to find them.[1778.18] [1778.18][S01] Have you actually tried[1778.8] [1778.8][S01]to make any of these molecule, or at the moment,[1780.865] [1780.865][S01]do they just exist on the screen?[1782.24] [1782.24][S03] Oh, no, we make a lot of molecules.[1784.282] [1784.282][S03]We've got a huge experimental footprint.[1787.27] [1787.27][S03]Yeah.[1787.77] [1787.77][S01] And how-- do they turn out how you expect?[1790.33] [1790.33][S01]I mean, what are the results like?[1791.78] [1791.78][S03] We've had some incredible success[1794.91] [1794.91][S03]in some of our projects where I'll find Max at his desk[1799.53] [1799.53][S03]and I'll be like, have you seen this thing?[1802.59] [1802.59][S03]And we'll just both be really mind-blown about it.[1804.872] [1804.872][S01] That when you get the molecule, it actually works.[1807.455] [1807.455][S02] Yeah, exactly, exactly.[1809.41] [1809.41][S02]And we have our own drug-design programs,[1811.66] [1811.66][S02]so things that we've started from scratch ourselves.[1814.06] [1814.06][S02]We also work with pharma-company partners, people[1817.41] [1817.41][S02]like Eli Lilly and Novartis.[1819.01] [1819.01][S02]In these collaborations, you'll get specific targets to work on.[1822.21] [1822.21][S02]And these are ones that these companies have high conviction[1825.4] [1825.4][S02]behind and probably a bunch of evidence behind.[1828.67] [1828.67][S02]Some of the collaborations we're in,[1832.15] [1832.15][S02]we've been given very, very hard targets.[1834.47] [1834.47][S02]These are things that people have worked on[1836.53] [1836.53][S02]sometimes for over a decade and not made significant progress,[1841.88] [1841.88][S02]to the point where you've got something on the market.[1844.13] [1844.13][S01] Things like cancer and that sort of stuff.[1845.87] [1845.87][S02] Whole host of therapeutic areas and disease[1848.74] [1848.74][S02]areas.[1849.29] [1849.29][S02]And then Becky and team sit down,[1853.15] [1853.15][S02]start designing with these models,[1855.49] [1855.49][S02]and can start finding completely novel chemical matter[1859.24] [1859.24][S02]for completely novel mechanisms that no one's really discovered[1863.2] [1863.2][S02]before, which is mind-blowing for me--[1865.03] [1865.03][S01] Yeah.[1865.15] [1865.15][S02] --as a computer scientist.[1865.845] [1865.845][S03] Yeah, it's mind-blowing for me.[1868.6] [1868.6][S03]There's been some almost career-defining moments,[1872.6] [1872.6][S03]where the AI will give you a hypothesis.[1877.39] [1877.39][S03]It will suggest something.[1878.8] [1878.8][S03]And you think, I'm not convinced I would do that,[1882.71] [1882.71][S03]but the model's telling me this thing.[1884.358] [1884.358][S03]And it's really quite convinced about this thing,[1886.4] [1886.4][S03]so I should maybe just test this hypothesis.[1888.77] [1888.77][S03]And then, actually, it turns out that the model was right,[1892.192] [1892.192][S03]and you were absolutely right to test it,[1893.9] [1893.9][S03]and it's really pushed forward your project, or even[1896.65] [1896.65][S03]that field.[1897.89] [1897.89][S03]So I think, for me, it's not about how we trust the models,[1902.39] [1902.39][S03]it's about how we are open to testing the hypotheses that they[1905.41] [1905.41][S03]put in front of us and not going,[1908.02] [1908.02][S03]oh, that doesn't fit with my worldview,[1909.77] [1909.77][S03]so I'm not going to test it.[1910.73] [1910.73][S01] But then you are also human, right?[1912.465] [1912.465][S03] Yes.[1912.76] [1912.76][S01] So I do wonder whether[1914.177] [1914.177][S01]if you see lots of hits with the model, as it were.[1917.272] [1917.272][S01]If the model is coming up with lots of good stuff in a row,[1919.73] [1919.73][S01]do you start to maybe trusting it more than yourself?[1922.635] [1922.635][S03] We actually put a lot of trust in the models.[1925.3] [1925.3][S03]We actually use-- for example, some of the models we have,[1927.717] [1927.717][S03]we use them as quite strict cutoffs.[1929.317] [1929.317][S01] What kind of cutoff?[1930.65] [1930.65][S03] For example, we have[1931.39] [1931.39][S03]a model which we call binding probability,[1933.14] [1933.14][S03]and it goes from 0 to 1.[1934.64] [1934.64][S03]So 1 is the model is convinced your molecule is definitely[1938.47] [1938.47][S03]going to bind to your protein.[1939.97] [1939.97][S03]Zero, the model is telling you, this is definitely not[1942.55] [1942.55][S03]going to bind.[1943.19] [1943.19][S03]And when you can build a little bit of confidence[1945.232] [1945.232][S03]over time that the model really does understand, OK,[1947.95] [1947.95][S03]anything below 0.7, it's really got a very low probability[1951.37] [1951.37][S03]of success.[1952.22] [1952.22][S03]So we just define that as a cutoff[1953.797] [1953.797][S03]and be like, we're not going to put anything[1955.63] [1955.63][S03]in the lab that's got a probability of less than this[1958.01] [1958.01][S03]because, actually, the model is quite likely to be right.[1960.68] [1960.68][S03]It's probably not going to be any good.[1962.0] [1962.0][S03]And so you do-- even though-- and that's[1963.667] [1963.667][S03]quite hard because as a chemist, you design something,[1965.92] [1965.92][S03]and you think, that was a really clever idea[1967.753] [1967.753][S03]that I just came up with, and why doesn't it like it?[1971.162] [1971.162][S01] But then what if it makes something[1973.12] [1973.12][S01]that you don't understand, I mean,[1975.08] [1975.08][S01]or that doesn't make sense?[1977.21] [1977.21][S01]I mean, does it need to explain itself?[1979.486] [1979.486][S03] I think at the moment that explainability[1982.09] [1982.09][S03]is quite important for now, because the process is quite[1986.8] [1986.8][S03]driven still by the human.[1988.31] [1988.31][S03]It's not end-to-end yet.[1989.51] [1989.51][S03]We have to go in there, and we have to say, what comes next?[1992.99] [1992.99][S03]So if there's no explainability there,[1995.03] [1995.03][S03]you don't know what would be next, right?[1997.37] [1997.37][S03]And so that would be very difficult to work with.[2000.16] [2000.16][S03]I can imagine in a future state where, actually, the process[2005.53] [2005.53][S03]is a bit more end-to-end, like in one step, the model.[2008.217] [2008.217][S03]Here's a drug.[2008.8] [2008.8][S01] Here's a drug.[2009.26] [2009.26][S03] Yeah.[2009.55] [2009.55][S03]Then actually, maybe you don't need that explainability.[2011.84] [2011.84][S03]But when you've got to go in there as a human[2013.18] [2013.18][S03]and you've got to iterate and you've[2014.68] [2014.68][S03]got to do a bit more of that directionality,[2017.547] [2017.547][S03]then that explainability is important.[2019.13] [2019.13][S03]And that's where I think, for me, the AlphaFold models really[2021.672] [2021.672][S03]come in because, OK, the model is[2023.615] [2023.615][S03]predicting this molecule is going to be good.[2025.49] [2025.49][S03]I can rationalize that with what I'm actually seeing.[2027.68] [2027.68][S03]I know what I would do next.[2028.68] [2028.68][S01] You have a slightly different view[2030.597] [2030.597][S01]on explainability, don't you?[2031.81] [2031.81][S03] I do have a slightly different view[2033.852] [2033.852][S03]on explainability, but I think you need explainability when[2036.43] [2036.43][S03]your model sucks, basically.[2038.3] [2038.3][S03]And we don't have perfect models yet,[2041.18] [2041.18][S03]so I think there's a good amount of room for explainability.[2044.09] [2044.09][S03]But I always hear the call for explainability[2047.05] [2047.05][S03]and think, look, we need to make this model better.[2049.517] [2049.517][S03]And actually, the interesting thing about explainability[2051.85] [2051.85][S03]is it can help you understand the pathologies that this model[2055.57] [2055.57][S03]has, the biases that it has.[2057.8] [2057.8][S03]Where's that wrong, given the science that we know about?[2062.19] [2062.19][S03]And so we can start patching that[2063.65] [2063.65][S03]and make it better and better and get to this point where,[2066.139] [2066.139][S03]yeah, actually, we can just do end-to-end design purely[2068.96] [2068.96][S03]in silico, and maybe do a final round of verification[2072.83] [2072.83][S03]in the lab at the end.[2074.1] [2074.1][S01] So, I mean, you literally say,[2075.929] [2075.929][S01]make me a drug for X disease, off it goes, says here's[2080.06] [2080.06][S01]the molecule you need.[2081.09] [2081.09][S02] Yeah, yeah.[2081.84] [2081.84][S01] Do you think that's possible?[2082.66] [2082.66][S02] I think it's possible.[2084.06] [2084.06][S02]I think it's possible.[2084.75] [2084.75][S02]I think everything is pointing in that direction.[2086.792] [2086.792][S02]We're getting better and better.[2088.19] [2088.19][S02]We're already reducing the amount of experimental cycles[2091.58] [2091.58][S02]you need, reducing the amount of lab time you need.[2095.48] [2095.48][S02]And, yeah, this is just the beginning.[2097.225] [2097.225][S01] Absolutely extraordinary.[2098.767] [2098.767][S01]I mean, I suppose it does depend on knowing what protein you're[2102.2] [2102.2][S01]targeting, too, right?[2103.28] [2103.28][S03] Yes.[2104.03] [2104.03][S01] So still-- and the diseases[2105.948] [2105.948][S01]that we don't have a full understanding are still[2107.99] [2107.99][S01]going to be difficult.[2109.35] [2109.35][S03] Yeah.[2110.142] [2110.142][S03]And that kind of-- we call it target ID space,[2112.31] [2112.31][S03]where you actually need to identify the protein that's[2114.89] [2114.89][S03]causing your disease.[2115.98] [2115.98][S03]It's actually a really important part of drug discovery[2118.515] [2118.515][S03]because if you're not hitting the right biological target[2120.89] [2120.89][S03]from the start, you can design the best molecule in the world.[2124.138] [2124.138][S03]It's not going to do what you want it to do[2125.93] [2125.93][S03]when you put it into a human.[2127.5] [2127.5][S03]So there's a lot to be done in that target ID space.[2129.768] [2129.768][S03]And I think AI's got a big role to play there, as well.[2132.06] [2132.06][S02] It's one of the big frontiers of AI[2134.45] [2134.45][S02]for biology is really understanding,[2137.33] [2137.33][S02]what are those driving mechanisms of disease?[2139.59] [2139.59][S02]Can we start to understand how mutations in our DNA[2144.89] [2144.89][S02]translate into changes of expression of RNA,[2147.83] [2147.83][S02]and how that changes the type of proteins and expression[2150.26] [2150.26][S02]levels of proteins, how those proteins interact[2153.392] [2153.392][S02]with each other and build up into these signaling pathways,[2155.85] [2155.85][S02]and how changes in those signaling pathways[2158.42] [2158.42][S02]change the disease state, as well?[2161.48] [2161.48][S02]And, of course, if we can start to understand these bits,[2164.15] [2164.15][S02]we can start to work out, where do[2165.8] [2165.8][S02]we need to modulate this biological system?[2167.91] [2167.91][S02]But all of this is really, really hard.[2170.097] [2170.097][S02]And there's some amazing breakthroughs[2171.68] [2171.68][S02]happening in the field understanding DNA[2173.66] [2173.66][S02]better, understanding this translation better.[2175.97] [2175.97][S02]Even through understanding how proteins interact,[2178.35] [2178.35][S02]can we build up these interaction networks better?[2182.55] [2182.55][S02]This is some of the really exciting frontier research[2184.77] [2184.77][S02]that we're also doing at ISO.[2185.96] [2185.96][S01] So you have a team working in that space, as well.[2188.26] [2188.26][S02] Yeah, that's right.[2189.07] [2189.07][S02]We have a whole computational biology team,[2191.13] [2191.13][S02]whole machine-learning modeling team[2192.63] [2192.63][S02]focused in this space, yeah.[2194.1] [2194.1][S01] But then what about personalized medicine?[2196.35] [2196.35][S01]Because I guess each person is different in some ways.[2198.83] [2198.83][S02] I mean, this is the really exciting,[2201.81] [2201.81][S02]potential future where we can understand much more[2205.59] [2205.59][S02]about, for example, cancer individuals' mutations[2209.01] [2209.01][S02]in their tumor, and through generative AI[2214.44] [2214.44][S02]and design agents, be able to come up[2217.38] [2217.38][S02]with molecules that work specifically[2219.75] [2219.75][S02]for these sort of mutations.[2220.925] [2220.925][S02]Now, there's a whole question of,[2222.3] [2222.3][S02]how do we actually operationalize that, and get[2224.258] [2224.258][S02]these drugs to patients, and approve this framework?[2227.11] [2227.11][S02]But we're moving towards a place where that technology[2230.61] [2230.61][S02]could be potentially there.[2232.075] [2232.075][S01] I mean, I'm thinking here about chemotherapy drugs,[2234.7] [2234.7][S01]which come with really devastating side effects.[2236.62] [2236.62][S01]You think there's real hope on the horizon[2237.88] [2237.88][S01]for that kind of thing?[2238.65] [2238.65][S03] Yeah, I mean, we think about chemotherapy drugs.[2241.34] [2241.34][S03]They're basically drugs that are nonspecific.[2243.56] [2243.56][S03]So they're going into the body, and they're[2246.25] [2246.25][S03]trying to halt that rapid cell proliferation.[2249.58] [2249.58][S03]But what we have now is an ability[2251.59] [2251.59][S03]to think about, actually, what's the specific target, the protein[2254.83] [2254.83][S03]target that we want to inhibit, we want to stop its function?[2259.45] [2259.45][S03]And that might have the same effect.[2261.35] [2261.35][S03]But you're not just generally using[2263.14] [2263.14][S03]something that's just very toxic to rapidly dividing cells.[2266.63] [2266.63][S03]Yes, it's going to stop your tumor cells dividing,[2268.86] [2268.86][S03]but it's also going to stop the cells that[2270.61] [2270.61][S03]line your stomach and your intestine,[2272.74] [2272.74][S03]going to make you feel nauseous and sick.[2274.61] [2274.61][S03]It's going to stop your hair follicles.[2276.35] [2276.35][S03]You're going to lose your hair, whereas, we now[2279.16] [2279.16][S03]know we can go in, we can target a very specific protein,[2282.74] [2282.74][S03]the one that's actually causing the disease.[2285.17] [2285.17][S03]And if you inhibit that particular protein,[2287.27] [2287.27][S03]that's not going to cause--[2288.67] [2288.67][S03]hopefully, if you get it right, it's[2290.41] [2290.41][S03]not going to cause all these other side effects.[2291.99] [2291.99][S02] You can do some also really, really cool stuff[2294.13] [2294.13][S02]with targeting particular cells.[2295.76] [2295.76][S02]So if you know that a particular cell type is expressing[2299.02] [2299.02][S02]something on its surface, you can start programming things[2303.88] [2303.88][S02]like antibodies to come in and find those particular receptors.[2308.23] [2308.23][S02]And so you're delivering your payloads directly[2312.07] [2312.07][S02]to that particular cell type and not more broadly to the body.[2314.942] [2314.942][S01] I just want to go back to the point[2316.9] [2316.9][S01]that you made earlier, Becky, about,[2318.478] [2318.478][S01]once you've got the drug design, then[2320.02] [2320.02][S01]once you put it into the human, there's[2322.03] [2322.03][S01]all of these other potential problems because, I mean,[2324.89] [2324.89][S01]there have been examples of this before where[2327.195] [2327.195][S01]drugs have been made and looked like they were very good.[2329.57] [2329.57][S01]And then once you actually put it into a human,[2331.537] [2331.537][S01]it causes some massive problems.[2332.87] [2332.87][S01]I think there was one which people were[2335.02] [2335.02][S01]very excited about the impact it was going to have on pain,[2338.2] [2338.2][S01]but it turned out that protein also[2340.0] [2340.0][S01]was quite crucial to making sure your heart kept beating.[2343.37] [2343.37][S01]How do you mitigate against that, or can you not at this[2345.82] [2345.82][S01]stage?[2346.155] [2346.155][S03] Well, one of the problems we have[2348.113] [2348.113][S03]is that we often use animal models to then translate things[2351.25] [2351.25][S03]into the clinic.[2352.19] [2352.19][S03]And animal models, they don't replicate human physiology[2355.97] [2355.97][S03]very well at all, actually.[2357.18] [2357.18][S03]So we know when we're working in the discovery[2359.54] [2359.54][S03]and preclinical space, which is all of that space[2361.73] [2361.73][S03]before you go into a human, we're[2364.01] [2364.01][S03]working with different animal models, which[2366.29] [2366.29][S03]might model the disease we're interested in.[2368.28] [2368.28][S03]And we're looking for molecules which[2370.1] [2370.1][S03]have an effect in those animal models.[2371.84] [2371.84][S03]And we have to show that they're not[2373.34] [2373.34][S03]toxic in those animal models.[2375.03] [2375.03][S03]And then we use that bank of evidence[2376.67] [2376.67][S03]to go to the drug-regulatory bodies and say, right,[2379.47] [2379.47][S03]we're ready to go into a human.[2380.85] [2380.85][S03]But from that point until the market,[2382.92] [2382.92][S03]there's a 90% failure rate.[2384.57] [2384.57][S01] Wow, 90%.[2385.94] [2385.94][S03] So all that investment[2387.44] [2387.44][S03]up to that point, which is huge.[2389.005] [2389.005][S01] What makes them so likely to fail?[2390.922] [2390.922][S03] So molecules fail in the clinic for toxicity.[2394.5] [2394.5][S03]They fail in the clinic for lack of efficacy.[2397.02] [2397.02][S03]And I think a lot of it comes back to the animal models[2400.07] [2400.07][S03]we use just are not very good at replicating human physiology.[2403.74] [2403.74][S01] Because the mouse is different to a human.[2405.99] [2405.99][S03] A mouse is different.[2407.448] [2407.448][S03]So we can cure a mouse disease, probably be quite good at that.[2410.575] [2410.575][S01] We've got loads of medication that works.[2412.885] [2412.885][S03] Yeah.[2413.51] [2413.51][S01] Yeah.[2413.8] [2413.8][S03] But, yeah, that translation[2415.79] [2415.79][S03]is a big part of science that we need to fix.[2418.42] [2418.42][S01] Can AI help here, as well?[2420.69] [2420.69][S01]I mean, if animal models are this stumbling[2423.74] [2423.74][S01]block with such a low level of success,[2426.183] [2426.183][S01]what can you do about it?[2427.225] [2427.225][S02] Well, this is where we can actually[2429.59] [2429.59][S02]use some of the technology and models we've been developing[2432.26] [2432.26][S02]and think about, OK, how can we understand toxicity better,[2436.13] [2436.13][S02]understand the effect on human cells[2438.77] [2438.77][S02]better, and see how that translates to organs?[2442.86] [2442.86][S02]If you think about some of these off-target effects,[2445.26] [2445.26][S02]probably, there are many drugs that you go into the clinic[2449.11] [2449.11][S02]and you're hitting your target of interest[2450.86] [2450.86][S02]and it's curing your pain, but then it's[2452.81] [2452.81][S02]hitting another target that's in another protein in your heart[2455.953] [2455.953][S02]and stopping the function of your heart.[2457.62] [2457.62][S02]That's an off-target effect.[2459.01] [2459.01][S01] I mean, side effects, in general,[2460.58] [2460.58][S01]are off-target effects, though, aren't they?[2461.76] [2461.76][S02] Yeah, exactly.[2462.982] [2462.982][S02]But you can imagine that if we've been building models that[2465.44] [2465.44][S02]understand really well how this molecule interacts with[2467.84] [2467.84][S02]your target of interest, you could also ask the question,[2470.85] [2470.85][S02]well, how does this molecule interact with every other target[2473.55] [2473.55][S02]in the human body, all 20,000 proteins?[2477.31] [2477.31][S02]And you can start building up this fingerprint of interactions[2480.955] [2480.955][S02]that this molecule, your drug molecule,[2482.58] [2482.58][S02]is having across the body.[2484.15] [2484.15][S02]And so that can give you clues, maybe even concrete signal,[2487.89] [2487.89][S02]into the toxicity or side effects of this molecule.[2491.05] [2491.05][S02]And the nice thing is we can get that signal, not when you're[2495.36] [2495.36][S02]going into humans, but actually, at the very, very beginning[2499.35] [2499.35][S02]of the design process.[2500.59] [2500.59][S02]So by the time you've gone through all of your molecule[2503.147] [2503.147][S02]design and you get to the point where you're like, yeah,[2505.48] [2505.48][S02]I want to go into humans, you've been[2507.03] [2507.03][S02]thinking about these side effects in a very rational way[2510.18] [2510.18][S02]for a long time.[2512.38] [2512.38][S02]And so, hopefully, your chances of actually hitting some[2515.31] [2515.31][S02]of those radically reduces.[2517.04] [2517.04][S01] You're taking your structure of LEGO bricks,[2519.572] [2519.572][S01]or your jigsaw, and you're just trying it[2521.28] [2521.28][S01]with every other possible combination[2522.99] [2522.99][S01]that it might encounter in a human body.[2524.465] [2524.465][S02] Yeah.[2525.01] [2525.01][S01] That's amazing.[2525.42] [2525.42][S02] We're going to make every possible LEGO[2526.83] [2526.83][S02]combination.[2527.33] [2527.33][S01] Well, if that's the design stage, then, Becky--[2529.87] [2529.87][S01]I mean, you also have to put this into clinical trials.[2532.18] [2532.18][S01]Just talk us through the process of clinical trials,[2534.403] [2534.403][S01]if you could.[2534.945] [2534.945][S03] So the first time that your molecule ever[2538.12] [2538.12][S03]goes into a human, that's a phase I clinical trial.[2541.4] [2541.4][S03]So it will be a small number of patients.[2543.95] [2543.95][S03]Some of those might actually be healthy volunteers.[2546.245] [2546.245][S03]They don't necessarily have the disease[2547.87] [2547.87][S03]that you're interested in.[2549.58] [2549.58][S03]And what you're looking to see is, does your drug actually[2552.887] [2552.887][S03]reach the level of exposure in the patient that would be[2555.22] [2555.22][S03]needed to generate an effect?[2557.21] [2557.21][S03]And is the drug well tolerated, or do you suddenly[2560.08] [2560.08][S03]start to see some side effects that you weren't anticipating?[2563.08] [2563.08][S03]If all is good, you'll proceed to a phase II clinical trial,[2566.39] [2566.39][S03]which is now you're going into people[2568.12] [2568.12][S03]who actually have the disease, and you're[2570.49] [2570.49][S03]going into larger numbers.[2571.88] [2571.88][S03]You're really looking to answer the question,[2574.04] [2574.04][S03]does your molecule actually have efficacy against the disease[2577.51] [2577.51][S03]that you're interested in?[2578.95] [2578.95][S03]And this is where we do see that big failure rate.[2581.33] [2581.33][S03]So 70% of molecules going into phase II[2583.96] [2583.96][S03]don't actually pass through into phase III.[2586.04] [2586.04][S03]For those that do pass into phase III,[2587.84] [2587.84][S03]that's where you're going into much bigger patient populations[2590.65] [2590.65][S03]seeing if your drug is effective across that bigger population.[2593.927] [2593.927][S03]It's not just got to be safe, but it's[2595.51] [2595.51][S03]got to be better than the standard of care.[2597.26] [2597.26][S03]There's got to be some--[2598.26] [2598.26][S03]for doctors to actually prescribe this[2600.34] [2600.34][S03]to their patients, they've got to say, this drug is better,[2603.1] [2603.1][S03]or this drug is safer than what I currently use.[2606.09] [2606.09][S01] And this whole thing has a 90% failure rate,[2609.29] [2609.29][S01]as you said.[2609.86] [2609.86][S03] Yeah.[2610.12] [2610.12][S01] I mean, does that mean[2611.86] [2611.86][S01]that there are people who work in this space[2613.93] [2613.93][S01]who never, never succeed?[2616.806] [2616.806][S03] Yeah.[2617.65] [2617.65][S03]So I'm a medicinal chemist, and we often[2621.16] [2621.16][S03]have this number where, actually, only 1[2624.1] [2624.1][S03]in 20 medicinal chemists will ever get a drug to market.[2627.74] [2627.74][S03]So 19 of us out of every 20 will never get a drug onto the market[2632.5] [2632.5][S03]through our careers.[2633.74] [2633.74][S03]So, yeah, we are a profession where[2636.28] [2636.28][S03]we're used to seeing significant failure.[2639.05] [2639.05][S01] You're comfortable with failure as a profession.[2640.25] [2640.25][S03] We're comfortable with failure.[2641.18] [2641.18][S03]We learn from it.[2642.08] [2642.08][S01] Extraordinary, extraordinary to imagine.[2644.89] [2644.89][S01]How long do you think it will be until the first AI-design drug[2649.03] [2649.03][S01]is on the market?[2650.0] [2650.0][S01]Because all of these additional levels really take some time,[2654.155] [2654.155][S01]don't they?[2654.655] [2654.655][S03] So those AI-designed drugs[2657.92] [2657.92][S03]that are in the clinic now in clinical trials[2660.83] [2660.83][S03]and the different levels of AI input into those current drugs,[2665.09] [2665.09][S03]I would imagine that in the next five years[2667.427] [2667.427][S03]or so, we're going to see an approval of one[2669.26] [2669.26][S03]of those medicines.[2671.06] [2671.06][S03]But for me, the big thing is going[2672.5] [2672.5][S03]to be, when can AI start to really fill out this pipeline[2676.76] [2676.76][S03]and start to get drugs into the clinic really quickly[2679.1] [2679.1][S03]and really start to deliver molecules for patients?[2681.51] [2681.51][S03]That, for me, will be when AI is having a really big impact.[2684.7] [2684.7][S01] When you can start to say, here's the target,[2687.69] [2687.69][S01]and then it pops out a drug at the end.[2689.64] [2689.64][S03] Yeah, and you can put that[2691.307] [2691.307][S03]straight into the clinic.[2692.349] [2692.349][S01] And be confident that it's not[2694.099] [2694.099][S01]going to cause any damage to a person.[2695.83] [2695.83][S03] And even a slightly improved level[2697.85] [2697.85][S03]of confidence in where we are now would be quite impactful.[2700.56] [2700.56][S03]Yeah.[2701.06] [2701.06][S02] Yeah, because as Becky said,[2702.39] [2702.39][S02]there's already molecules in the clinic that[2703.91] [2703.91][S02]have been touched by AI, that have[2705.41] [2705.41][S02]been enabled by AI in some way.[2707.353] [2707.353][S02]We're just going to see more and more of that.[2709.27] [2709.27][S02]In five years' time doing drug design without AI[2713.8] [2713.8][S02]will be like doing any sort of science without maths.[2716.62] [2716.62][S02]And to be honest, I think the whole of science[2718.54] [2718.54][S02]will be like this is, if you're not using AI,[2721.16] [2721.16][S02]what are you doing?[2722.18] [2722.18][S02]There's just so much information to be gained there.[2726.28] [2726.28][S02]So yeah, as Becky said, it's more like,[2728.41] [2728.41][S02]how do we actually see that rapid increase in disease areas[2732.587] [2732.587][S02]that we're able to tackle the targets,[2734.17] [2734.17][S02]that we're able to unlock, ultimately, patients[2737.315] [2737.315][S02]that we're able to help.[2738.315] [2738.315][S01] Amazing.[2739.48] [2739.48][S01]Thank you both.[2740.33] [2740.33][S01]That was really interesting.[2741.195] [2741.195][S03] Thank you for having us.[2742.64] [2742.64][S03]It was so much fun.[2743.51] [2743.51][S02] Yeah, it's been great to be here.[2744.83] [2744.83][S01] I think I now realize[2745.87] [2745.87][S01]that medicinal chemistry is one of the hardest[2747.94] [2747.94][S01]jobs in the world.[2749.0] [2749.0][S01]It takes years to design a drug.[2750.92] [2750.92][S01]Even if you get it to clinical trials, 90% of them fail.[2753.98] [2753.98][S01]And only one in 20 of your colleagues[2755.74] [2755.74][S01]ever manages to see their medicine improving[2758.35] [2758.35][S01]the lives of patients.[2760.13] [2760.13][S01]But strangely, that is precisely what I think[2763.24] [2763.24][S01]is so exciting about this space because if everything[2765.97] [2765.97][S01]we've done up until now has, effectively,[2767.76] [2767.76][S01]been like working in the dark, slowly, laboriously navigating[2771.81] [2771.81][S01]the most infinitesimally small areas of the vast landscape[2776.13] [2776.13][S01]of possibilities, it's like someone has just[2779.07] [2779.07][S01]turned on a floodlight.[2780.78] [2780.78][S01]And, OK, of course, we are still very, very far away[2785.04] [2785.04][S01]from a big AI button that's just going to solve all diseases.[2788.92] [2788.92][S01]But there is so much headroom here for improvement,[2791.59] [2791.59][S01]so much scope to move the dial, and simultaneously,[2795.43] [2795.43][S01]so much opportunity to directly impact the lives of all of us.[2800.31] [2800.31][S01]You have been listening to \"Google DeepMind,\" the Podcast,[2802.77] [2802.77][S01]with me, Professor Hannah Fry.[2804.4] [2804.4][S01]If you enjoyed this episode, then[2806.1] [2806.1][S01]do subscribe to our YouTube channel,[2807.72] [2807.72][S01]or leave a review on your favorite podcast platform.[2810.52] [2810.52][S01]And, of course, we have plenty more episodes[2812.85] [2812.85][S01]on a whole range of topics to come, so do check those out.[2816.25] [2816.25][S01]See you next time.[2817.74] [2817.74][MUSIC PLAYING][2821.39]"}
{"file_name": "audio/val_000012.wav", "transcription": "[0.0][S02] Welcome to \"Google DeepMind-- the Podcast.\"[2.292] [2.292][S02]I'm your host, Professor Hannah Fry.[4.16] [4.16][S02]Now, in the heart of Silicon Valley,[5.88] [5.88][S02]there's a new phrase that has emerged.[7.61] [7.61][S02]It mirrors the millennial retort, OK Boomer,[11.04] [11.04][S02]in how dismissive it is.[12.69] [12.69][S02]But OK Doomer is now the go-to response[15.8] [15.8][S02]for people who want to diminish talk of AGI's dangers.[19.68] [19.68][S02]And by AGI, we mean, of course, an AI system[22.7] [22.7][S02]that can tackle a wide range of tasks[24.86] [24.86][S02]at a level comparable to human intelligence.[28.2] [28.2][S02]But on this podcast, we want to tackle those important questions[31.61] [31.61][S02]head on.[32.54] [32.54][S02]Is building AI safe?[34.86] [34.86][S02]How can we be sure?[36.33] [36.33][S02]And what existential risks are the designers[38.84] [38.84][S02]of these algorithms taking on our behalf?[41.96] [41.96][S02]Well, to get some answers, we are going straight to the top.[45.39] [45.39][S02]Anca Dragan is the lead for AI safety and alignment[48.95] [48.95][S02]at Google DeepMind.[49.97] [49.97][S02]And she has been focused on this area for almost a decade.[53.52] [53.52][S02]Anca has a PhD in robotics.[55.07] [55.07][S02]And since then, she has gained extensive experience[57.38] [57.38][S02]working with driverless cars on the Waymo project,[59.69] [59.69][S02]along with many other things.[61.63] [61.63][S02]She is also a professor at UC Berkeley,[64.93] [64.93][S02]working on human-AI interaction and alignment.[67.87] [67.87][S02]And recently, she has been working[69.66] [69.66][S02]on the safety of Gemini, Google's most capable multimodal[74.52] [74.52][S02]models.[75.46] [75.46][S02]Welcome to the podcast, Anca.[77.383] [77.383][S01] Thank you, Hannah.[78.675] [78.675][MUSIC PLAYING][81.1] [84.488][S02] I guess we should probably start with you--[86.78] [86.78][S02]I mean, you've got quite a big title there.[88.67] [88.67][S02]What are you actually responsible for at Google[91.38] [91.38][S02]DeepMind?[92.25] [92.25][S01] Safety of our--[93.65] [93.65][S02] All of it.[94.25] [94.25][S01] All of it.[94.87] [94.87][S01](CHUCKLING) No.[95.495] [95.495][S01]Safety of our GenAI models from the current Gemini family,[101.93] [101.93][S01]and avoiding present-day harms with Gemini,[104.84] [104.84][S01]all the way to safety and alignment for longer term,[110.98] [110.98][S01]as model capabilities improve further and further and further,[115.08] [115.08][S01]avoiding more severe, more extreme, potentially[117.9] [117.9][S01]even catastrophic harms.[119.71] [119.71][S01]And so I have a team.[120.91] [120.91][S01]It's called the AI Safety and Alignment,[122.74] [122.74][S01]as my title would suggest.[124.15] [124.15][S01]Our mission, I'd say, is to ensure[126.36] [126.36][S01]that Gemini and future, more capable models,[130.889] [130.889][S01]robustly do what individuals and societies want.[135.715] [135.715][S02] I'm just wondering about the short-term risks[138.09] [138.09][S02]and the long-term risks, as you're describing.[140.02] [140.02][S02]Because historically, those have been quite separated[142.65] [142.65][S02]from one another.[143.442] [143.442][S01] Yes.[144.15] [144.15][S02] But you're sort of--[144.78] [144.78][S01] More than separated.[145.57] [145.57][S01]We have two communities out there.[147.1] [147.1][S01]We have AI ethics that worries about present-day harms.[149.89] [149.89][S01]And we have ex-risk folks, who worry about catastrophic risks.[153.242] [153.242][S01]It's a very frustrating thing because you[154.95] [154.95][S01]have AI ethics saying, oh, these things are just,[157.26] [157.26][S01]like, distractions from the present-day harms.[159.31] [159.31][S01]I very emphatically see it as much more not[167.7] [167.7][S01]even complimentary, helping each other.[169.91] [169.91][S01]Especially if we think that getting[174.1] [174.1][S01]to human-level capability is-- across many cognitive tasks,[178.93] [178.93][S01]or even above human-level capability,[180.88] [180.88][S01]is coming sooner rather than later,[184.13] [184.13][S01]getting the safety right is probably,[186.89] [186.89][S01]I'd say, one of the most important challenges[189.67] [189.67][S01]of our time.[191.17] [191.17][S01]And of course, there's those who say, don't bother.[194.24] [194.24][S01]We're nowhere near.[196.03] [196.03][S01]We are going to need a different paradigm.[198.38] [198.38][S01]And we'll figure out safety after we figure that out.[201.86] [201.86][S01]And I just so vehemently, strongly disagree with that,[207.13] [207.13][S01]for a number of different reasons.[209.62] [209.62][S01]Of course, I disagree with the timeline and this notion[212.86] [212.86][S01]that it's impossible to get there with the current paradigm.[216.003] [216.003][S02] You think it's sooner.[217.42] [217.42][S01] I think it's sooner.[218.795] [218.795][S01]I think-- and I think--[220.0] [220.0][S01]I used to think we had a lot of time.[222.805] [222.805][S01]Like, 10 years ago, we were looking at AI safety.[226.62] [226.62][S01]We were on these panels.[227.62] [227.62][S01]And a lot of people were very concerned.[228.88] [228.88][S01]And I was chill.[229.58] [229.58][S01]I was like, ah, it's a very important topic.[233.75] [233.75][S01]It's important for me in academia to do work on this[236.91] [236.91][S01]because that's my job is to look ahead as an academic.[239.8] [239.8][S01]But we have plenty of time.[241.38] [241.38][S01]And I don't feel like that necessarily anymore.[246.82] [246.82][S01]And I perceive a lot more urgency to this.[249.61] [249.61][S01]But look.[250.24] [250.24][S01]Even if we had plenty of time, even if we're wrong about that,[254.86] [254.86][S01]even if it's impossible to get to human-level intelligence[258.708] [258.708][S01]with the current paradigm, we'll need[260.25] [260.25][S01]some really major breakthroughs that are[262.14] [262.14][S01]going to take a decade or more.[264.24] [264.24][S01]It's possible.[265.41] [265.41][S01]That's feasible.[266.41] [266.41][S01]I just-- I'm not confident enough to be like, don't worry.[269.29] [269.29][S01]But even if that were the case, this whole notion,[272.02] [272.02][S01]this premise of, we'll figure out the capabilities[275.19] [275.19][S01]and then we'll figure out how to make it safe is so worrisome.[280.098] [280.098][S02] Backwards.[281.015] [281.015][S01] Yeah, it's backwards.[281.68] [281.68][S01]And I have a few analogies that I like to give.[283.998] [283.998][S02] Please.[284.79] [284.79][S01] One of them I borrow from my colleague, Stuart[286.98] [286.98][S01]Russell.[287.64] [287.64][S01]So Stuart likes to draw the analogy with a bridge.[291.25] [291.25][S01]You don't think, I'm going to design a bridge.[294.43] [294.43][S01]And then I'm going to bring in a safety team,[296.34] [296.34][S01]and they're going to figure out how to make the bridge safe.[298.84] [298.84][S01][LAUGHS] That's just not really how it works.[301.15] [301.15][S01]We want to make a safe bridge.[302.62] [302.62][S01]And that influences the design decisions[305.32] [305.32][S01]that we make when we design the bridge, right?[308.9] [308.9][S01]And then I'm also drawing on my own experience.[312.25] [312.25][S01]Like you mentioned in the intro, I did my PhD in robotics,[314.83] [314.83][S01]actually.[317.062] [317.062][S01]But my claim to fame there was this notion of,[323.29] [323.29][S01]how do we integrate interaction with humans[328.3] [328.3][S01]into the spec of the problem that we're trying to solve?[331.542] [331.542][S02] Let me make sure I understand this.[333.5] [333.5][S02]So if you're building a robot, and the robot is tasked with,[336.17] [336.17][S02]I don't know, crossing a room, going to some--[338.93] [338.93][S01] Picking up a bottle.[339.37] [339.37][S02] Picking up a bottle.[339.7] [339.7][S01] This was my favorite thing[340.06] [340.06][S01]to my PhD is just like I had a bimanual manipulator.[342.61] [342.61][S01]And it picked up bottles.[343.995] [343.995][S02] OK.[344.62] [344.62][LAUGHING][345.61] [345.61][S01] (LAUGHING) That's the picture[346.06] [346.06][S01]to have in your head.[347.02] [347.02][S02] So if you've got this robot that crosses a room,[349.52] [349.52][S02]picks up a bottle, and then afterwards, you're like, oh,[352.81] [352.81][S02]now PS, there's people in the room[354.36] [354.36][S02]that you've got to avoid them.[355.61] [355.61][S02]You can't run them over.[356.61] [356.61][S02]You can't crash into them.[357.71] [357.71][S02]You've got to navigate that space with them in it.[359.793] [359.793][S02]You can't just stick a bit on at the end that will solve that.[362.778] [362.778][S02]You've got to start thinking of that from the very beginning.[365.32] [365.32][S01] Yes.[366.028] [366.028][S01]And to explain that further, you can think of people[369.1] [369.1][S01]need-- for you to actually be safe when you're moving,[371.72] [371.72][S01]both of you are moving, you have to make sure[375.49] [375.49][S01]that people can actually anticipate[377.44] [377.44][S01]what's coming from you, not just-- people are not static.[380.76] [380.76][S01]They move.[381.26] [381.26][S01]You have to start anticipating them.[382.52] [382.52][S01]But then they have to be able to anticipate you.[384.53] [384.53][S01]OK, so now you have a very different objective, which[387.01] [387.01][S01]is like, you're defining the state not just[388.87] [388.87][S01]to be the physical state, but looking over, well,[391.17] [391.17][S01]what is the human going to think I'm doing, blah, blah, blah?[393.754] [393.754][S02] Yeah, how predictable[393.826] [393.826][S02]is the robot's actions?[394.97] [394.97][S01] Yeah.[395.72] [395.72][S01]It turns out that if you want to make sure[400.335] [400.335][S01]that the thing is safe, you have to put that in the spec.[402.71] [402.71][S01]You have to say, I want safe and capable.[406.31] [406.31][S01]I want a bridge that will not collapse.[409.275] [409.275][S02] So moving on from robotics,[410.925] [410.925][S02]because I know you've also got extensive experience[413.05] [413.05][S02]in driverless cars, the work that you did with Waymo.[415.51] [415.51][S02]I mean, I imagine that the human-AI interaction in that[419.5] [419.5][S02]is absolutely integral to it working.[421.94] [421.94][S02]So I can imagine if you take humans out[423.94] [423.94][S02]of the equation altogether, actually, driverless cars maybe[427.125] [427.125][S02]not nearly as difficult a problem[428.5] [428.5][S02]as it is when humans are there.[429.98] [429.98][S01] Yeah.[430.73] [430.73][S01]I would say that the only difficult thing[432.438] [432.438][S01]about driverless cars is that you[433.87] [433.87][S01]have to interact with people.[435.24] [435.24][S02] The only difficult thing.[436.51] [436.51][S01] Yeah.[436.54] [436.54][S01]I mean, perception is good enough to tell you everything.[439.01] [439.01][S01]There's not-- there's no problem.[441.182] [441.182][S01]Five years ago, there wasn't a problem[443.29] [443.29][S01]with being able to drive at scale if you didn't have[446.05] [446.05][S01]to deal with the fact that there are humans you[448.12] [448.12][S01]have to share the road with.[449.59] [449.59][S01]And so, yeah, for sure, that was--[452.05] [452.05][S01]that's why I got into cars.[453.98] [453.98][S01]It's because it really felt like nailing[457.75] [457.75][S01]the interaction with people was going to be the key enabler.[462.89] [462.89][S01]So I spent six years, like just one day[464.53] [464.53][S01]a week, so just consulting for Waymo.[466.37] [466.37][S01]But I spent six years with them before I[468.25] [468.25][S01]switched to Google DeepMind.[469.607] [469.607][S01]Did I tell you how I got into cars, by the way?[471.565] [471.565][S02] Tell me.[472.45] [472.45][S01] So I got into cars.[474.62] [474.62][S01]I was on the, you know, pick up the bottle.[476.54] [476.54][LAUGHING][477.53] [477.53][S01]And I came to a conference.[480.288] [480.288][S01]I was at Carnegie Mellon.[481.33] [481.33][S01]And I came to a conference at Berkeley,[483.06] [483.06][S01]organized by Pieter Abbeel, who is now my colleague and dear[485.56] [485.56][S01]friend.[486.06] [486.06][S01]And he set up this for us to be able to go and visit Google.[493.04] [493.04][S01]And it was not called Waymo back then.[494.87] [494.87][S01]It was the Google Chauffeur or something like that project.[497.39] [497.39][S01]And he arranged for us to take these rides.[500.39] [500.39][S01]And so I was working on robotics at that point for a few years.[504.83] [504.83][S01]And I got into this car.[507.52] [507.52][S01]And it goes.[509.41] [509.41][S01]And it's making decision after decision, the right thing.[515.09] [515.09][S01]And it was this transformative experience.[518.77] [518.77][S01]Like, to be in a robot--[520.7] [520.7][S02] Yeah, the robot's driving you.[522.73] [522.73][S01] And an AI system that just knows[524.605] [524.605][S01]how to reason about everything that's going on[526.75] [526.75][S01]and act in the way, in a safe way,[529.27] [529.27][S01]in a way that people that can anticipate, blah, blah, blah,[532.8] [532.8][S01]all that stuff.[533.57] [533.57][S01]That was just-- yeah, it was mind boggling.[537.08] [537.08][S01]And to experience it from the inside,[539.68] [539.68][S01]basically, to experience it as, I'm in the robot.[542.662] [542.662][S02] So what is it that makes it so difficult,[544.87] [544.87][S02]that human-AI interaction?[546.215] [546.215][S02]What is it that makes it such a hard problem?[548.09] [548.09][S02]And I guess also, how much of the lessons that[551.447] [551.447][S02]have been learned from driverless cars[553.03] [553.03][S02]can you then translate to this stuff?[555.62] [555.62][S01] Yeah.[556.37] [556.37][S01]So on the interaction side, I think it's--[560.28] [560.28][S01]what's really useful in cars, and what's important[565.15] [565.15][S01]is to anticipate what people will do[567.67] [567.67][S01]and to account for the fact that driving is multi-turn.[571.36] [571.36][S01]You act.[572.53] [572.53][S01]And people act.[573.58] [573.58][S01]And you keep acting, and they keep acting and so on.[575.78] [575.78][S01]And what-- that means that what you do as the agent,[578.96] [578.96][S01]as the robot, as the AI influences what[582.52] [582.52][S01]you see in response from them.[584.32] [584.32][S01]And so that opens up-- like, here's an example.[587.96] [587.96][S01]Imagine you are merging.[590.84] [590.84][S01]So when you're trying to merge, there[595.33] [595.33][S01]might not be a big enough gap in traffic.[597.71] [597.71][S01]But what you'll do is you'll start nudging in.[599.98] [599.98][S01]And the person will see that.[601.63] [601.63][S01]They might slow down just a little bit[603.67] [603.67][S01]to create that gap in traffic.[605.15] [605.15][S01]And you go in, right?[606.56] [606.56][S01]It's a back-and-forth.[607.85] [607.85][S01]It's a conversation.[608.87] [608.87][S01]It really is.[609.65] [609.65][S01]And so what it teaches us is, really, this notion of,[613.2] [613.2][S01]yeah, you have, as an AI, to be mindful of the influence you[617.17] [617.17][S01]have on people, what people do, what people say.[621.53] [621.53][S01]And maybe a concrete example of that[623.08] [623.08][S01]in large language model land with something[625.93] [625.93][S01]like Gemini today would be if the person asks[629.86] [629.86][S01]you to do something, Gemini might[632.5] [632.5][S01]be able to nudge you to clarify, what do you mean?[637.602] [637.602][S01]What do you want?[638.31] [638.31][S01]You want this?[638.69] [638.69][S01]Do you want that?[639.398] [639.398][S01]And then because it knows, then it[641.38] [641.38][S01]can get a response that is helpful for it[644.89] [644.89][S01]to actually serve you better.[646.7] [646.7][S01]And so that's a lesson.[648.68] [648.68][S01]I don't think Gemini is very good at that yet.[650.807] [650.807][S01]But that's, I think, a lesson that could transfer.[652.89] [652.89][S01]It's like, how do you actually plan[654.59] [654.59][S01]for this multi-turn interaction and the back-and-forth[658.52] [658.52][S01]and the influence you have on people?[660.41] [660.41][S02] That's a really interesting observation,[662.72] [662.72][S02]actually, because I guess that is the way that humans make[664.79] [664.79][S02]sure that they're aligning with each other is by having[667.16] [667.16][S02]that back-and-forth, that feedback effect, as it were,[670.02] [670.02][S02]that loops around and allows you to be certain that you're[673.01] [673.01][S02]nudging in the right direction.[674.535] [674.535][S01] Yeah, absolutely.[675.42] [675.42][S01]I feel like-- and this is a part that I[676.88] [676.88][S01]think is missing a little bit in alignment.[678.75] [678.75][S01]There's nothing that's like, oh, actually,[683.61] [683.61][S01]I don't know the ground truth of what people want.[687.48] [687.48][S01]And I have to keep that conversation open.[690.02] [690.02][S01]I have to keep learning.[691.02] [691.02][S01]I have to keep gathering feedback and updating.[695.27] [695.27][S01]And you could do that at a global scale of,[699.2] [699.2][S01]what are overall values and preferences?[701.01] [701.01][S01]You could also do that at a local interaction with[705.29] [705.29][S01]one-person scale, where the person asks you, hey,[710.54] [710.54][S01]I would like--[711.41] [711.41][S01]I just had a summit.[712.77] [712.77][S01]I'm going to a summit.[713.88] [713.88][S01]I'm organizing it, and I need to craft the agenda.[716.04] [716.04][S01]Can you help me crafting the agenda?[717.57] [717.57][S01]It's going to be in London.[719.13] [719.13][S01]It's going to be two days.[721.43] [721.43][S01]Some details there.[723.85] [723.85][S01]Well, I don't want the model's response to be like, OK,[727.04] [727.04][S01]here's an agenda.[728.08] [728.08][S01]9:00 AM, 9:30, 10:00.[730.11] [730.11][S01]I'm like, it has no idea what I actually want.[733.51] [733.51][S01]So it has to engage with me in an interaction,[737.23] [737.23][S01]in this back-and-forth.[738.56] [738.56][S01]So it should probably ask, OK, what is important?[741.378] [741.378][S01]What are you trying to get out?[742.67] [742.67][S01]Let's think about this together.[744.53] [744.53][S01]And eventually, after all that, you end up with the actual--[748.825] [748.825][S02] Getting to the heart,[750.2] [750.2][S02]spiraling in towards what you actually want.[752.51] [752.51][S01] Yeah.[753.26] [753.26][S02] This objective of alignment--[756.29] [756.29][S02]can you define what that is?[758.325] [758.325][S01] Yeah, so the way I would see it--[761.32] [761.32][S01]OK, V-zero definition.[763.9] [763.9][S01]As there is a human, internally in their head,[768.55] [768.55][S01]implicitly, they want--[770.695] [770.695][S01]they care about-- they have different values.[772.57] [772.57][S01]They have preferences.[773.74] [773.74][S01]They have goals.[774.52] [774.52][S01]And alignment is get the AI agent[777.9] [777.9][S01]to do well on that objective that's[780.63] [780.63][S01]implicit in the person's head.[783.3] [783.3][S02] Let's say that you know[784.905] [784.905][S02]what somebody's objective is.[786.73] [786.73][S02]Let's say that you can just write it down.[788.65] [788.65][S02]And it's very simple and straightforward.[790.45] [790.45][S02]How do you train an algorithm to achieve that objective?[794.73] [794.73][S01] So if you know what the objective is,[797.727] [797.727][S01]what the reward function is, in their head--[799.56] [799.56][S02] And, I mean, an example of a reward function[801.893] [801.893][S02]might be, I don't know, pick up the bottle without breaking it.[804.55] [804.55][S02]Or--[805.2] [805.2][S01] Yeah.[805.95] [805.95][S01]And then what you can do is do something[808.08] [808.08][S01]called reinforcement learning, which is a type of optimization,[812.34] [812.34][S01]basically, of getting the AI system to go[815.97] [815.97][S01]from a starting point to actually delivering, maximizing[821.67] [821.67][S01]on that objective.[823.318] [823.318][S02] So I sometimes think of reinforcement[825.36] [825.36][S02]learning almost as though you are playing a computer game.[829.36] [829.36][S02]And the AI gets reward points for everything it does towards[833.56] [833.56][S02]its objective, which sort of makes sense when you're doing[837.19] [837.19][S02]something like \"Space Invaders.\"[839.09] [839.09][S02]The reward points are really clear.[840.86] [840.86][S02]But I guess it gets a lot harder when the objective is just[844.78] [844.78][S02]something that happens to be in someone's head.[848.41] [848.41][S02]And then you get into the question of, which humans[851.29] [851.29][S02]and whose heads are we talking about?[853.322] [853.322][S01] Aha, yeah.[854.28] [854.28][S01]That's why that was just V-zero.[855.83] [855.83][S01]Because then--[856.99] [856.99][S02] Level up.[858.22] [858.22][S01] Yeah, because-- so this is a definition[862.09] [862.09][S01]I've been using for a while.[865.732] [865.732][S01]So now, you asked this very, very deep question[871.33] [871.33][S01]of which human.[874.18] [874.18][S01]One human, the user, the average human.[877.97] [877.97][S01]And I have to admit, I really punted[880.48] [880.48][S01]on that for a really long time, until maybe a couple[885.49] [885.49][S01]years ago now.[886.48] [886.48][S01]I got more and more excited about recommendation systems[889.9] [889.9][S01]because in recommendation systems,[893.3] [893.3][S01]they optimize for, quote, unquote, \"user engagement.\"[896.423] [896.423][S02] Give me an example of a recommendation system,[898.84] [898.84][S02]like--[899.34] [899.34][S01] Twitter, Facebook News Feed.[901.4] [901.4][S02] Right.[902.35] [902.35][S01] The idea is supposed to be,[903.41] [903.41][S01]that's what you want.[904.285] [904.285][S01]It's giving you what you want, right?[907.0] [907.0][S01]The challenge is that optimizing maybe one in the moment[910.45] [910.45][S01]is maybe not the ideal thing.[913.1] [913.1][S01]So what should you be optimizing for that actually brings users[916.84] [916.84][S01]value?[918.94] [918.94][S01]And I got really concerned about this notion of which human.[923.29] [923.29][S01]And should you really be- is it the user?[925.93] [925.93][S01]Is it--[926.68] [926.68][S02] Or the company.[927.52] [927.52][S01] Or the company.[928.04] [928.04][S01]Or, yeah, exactly.[928.94] [928.94][S02] Because, I mean, with recommendation systems,[931.315] [931.315][S02]I mean, the fact that they are so effective at keeping[934.0] [934.0][S02]you engaged--[935.74] [935.74][S02]I mean, for me personally, I have[937.115] [937.115][S02]to put blockers on my phone to stop myself being on it[939.365] [939.365][S02]the entire time.[940.22] [940.22][S01] Yeah.[941.38] [941.38][S02] And I think that society overall, there's[944.74] [944.74][S02]all those layers of different people.[947.18] [947.18][S01] Yeah.[947.93] [947.93][S01]So exactly.[949.49] [949.49][S01]So we did one study.[953.11] [953.11][S01]The lead author that was Smitha Milli[955.87] [955.87][S01]while I was doing just my regular Berkeley research work.[960.73] [960.73][S01]We started looking at, OK, what are the effects on people[965.11] [965.11][S01]of this Twitter ranking algorithm[969.49] [969.49][S01]that tries to give you what you want?[971.54] [971.54][S01]And one of the things that we noticed[974.05] [974.05][S01]is that it tends to give people, especially[978.17] [978.17][S01]in the political space, content that[981.68] [981.68][S01]makes them like their political in-group better and be[988.37] [988.37][S01]very pissed at the political out-group.[990.9] [990.9][S01]This is related to a phenomenon you might[992.723] [992.723][S01]call affective polarization.[993.89] [993.89][S01]It's not necessarily that your opinions change.[996.12] [996.12][S01]They might.[996.62] [996.62][S01]But the effect that you have towards your out-group.[1001.45] [1001.45][S01]So that can't be good for society.[1004.43] [1004.43][S01]So even if that's what the person wants to see--[1006.79] [1006.79][S01]and it turns out maybe they don't--[1008.33] [1008.33][S01]that has-- that's in conflict with societal-level effects.[1016.54] [1016.54][S01]And that also points-- we got to a political in-group[1019.66] [1019.66][S01]versus out-group.[1020.54] [1020.54][S01]So what should a model, an AI model, do?[1025.97] [1025.97][S01]And how should it balance individuals, the user,[1032.38] [1032.38][S01]and the rest of society when it makes decisions?[1037.9] [1037.9][S01]That seems really hard.[1039.601] [1039.601][S01]And I'd been really--[1040.476] [1040.476][S02] Really hard.[1041.109] [1041.109][S01] Yeah.[1041.859] [1041.859][S01]So when I got to Google DeepMind, one of my first moves[1045.56] [1045.56][S01]was to set up a team that would look at this, that would look[1049.94] [1049.94][S01]at, how do you do value alignment[1052.46] [1052.46][S01]when you acknowledge that there are[1054.56] [1054.56][S01]different people in the world, with different values?[1057.87] [1057.87][CHUCKLES][1058.97] [1058.97][S01]And what does that mean?[1060.03] [1060.03][S01]How should that affect our algorithms and what we do?[1063.055] [1063.055][S02] I mean, it sounds so difficult[1064.805] [1064.805][S02]as to almost be impossible the way that you're describing.[1067.222] [1067.222][S02]How can you possibly create something[1069.14] [1069.14][S02]that's capable of balancing those competing,[1072.03] [1072.03][S02]often conflicting objectives?[1074.61] [1074.61][S01] OK, so it is really hard.[1077.4] [1077.4][S01]And I don't claim to have the answers here.[1081.03] [1081.03][S01]But I will point out that there are[1083.57] [1083.57][S01]areas of science and research and economics[1089.06] [1089.06][S01]that have thought about this.[1092.34] [1092.34][S01]If you think about voting, there's[1094.97] [1094.97][S01]this notion of social choice, of preference aggregation.[1098.04] [1098.04][S01]And the whole notion is that, yeah, we[1099.94] [1099.94][S01]have to make a decision that's OK for multiple people that[1102.71] [1102.71][S01]might have competing objectives.[1104.96] [1104.96][S01]And so there's a few things you could start doing.[1107.42] [1107.42][S01]One of them is that the team is looking at.[1110.0] [1110.0][S01]One of them is you could try to have different reward[1116.08] [1116.08][S01]functions, different objectives, instead of just learning[1118.54] [1118.54][S01]one or the average one.[1119.68] [1119.68][S01]Like so right now, the standard recipe[1122.127] [1122.127][S01]is you get a bunch of preference data.[1123.71] [1123.71][S01]You fit a reward model to that data.[1125.57] [1125.57][S01]But it's one reward model.[1126.92] [1126.92][S01]But actually, if you ask different people[1130.81] [1130.81][S01]across the world, across different cultures,[1133.25] [1133.25][S01]across different identities, across the political spectrum, A[1137.02] [1137.02][S01]versus B, for some A versus B, they'll all agree.[1139.85] [1139.85][S01]But for other A versus B, they'll have different opinions.[1142.52] [1142.52][S01]And that is not just noise.[1145.0] [1145.0][S01]Those are actual-- in some cases, it's noise.[1147.71] [1147.71][S01]In some cases, it's actual diversity of opinions.[1152.65] [1152.65][S01]And so then what you can start to do[1154.39] [1154.39][S01]is not have one reward function.[1156.2] [1156.2][S01]But have multiple reward functions[1158.92] [1158.92][S01]for different types of people.[1160.61] [1160.61][S01]Or you could have a reward function[1162.91] [1162.91][S01]that doesn't output one number but outputs a distribution.[1166.74] [1166.74][S01]So then you can ask your agent to make decisions[1169.73] [1169.73][S01]that people across the board would like.[1171.9] [1171.9][S01]If that's not possible, that would not[1174.77] [1174.77][S01]be too bad according to anyone.[1176.58] [1176.58][S01]So the minimum bar, the floor is high.[1180.14] [1180.14][S01]Another thing that it turned out--[1183.35] [1183.35][S01]it's kind of funny.[1184.29] [1184.29][S01]I was trying to do this at Berkeley.[1186.0] [1186.0][S01]I was trying to convince my students to work on it.[1187.53] [1187.53][S01]And then I joined Google DeepMind.[1188.947] [1188.947][S01]And I found out that people around were already[1191.33] [1191.33][S01]working on this and doing really well[1193.01] [1193.01][S01]on is this concept of deliberative alignment.[1196.34] [1196.34][S02] OK.[1197.37] [1197.37][S01] OK, so deliberative alignment--[1199.43] [1199.43][S01]do you know about citizens' assemblies?[1201.885] [1201.885][S02] Yeah, I mean, this is like the democratic process[1204.95] [1204.95][S02]to other decisions.[1205.845] [1205.845][S01] Exactly.[1206.72] [1206.72][S01]So, OK.[1207.75] [1207.75][S01]So in particular, citizens' assemblies[1209.72] [1209.72][S01]are really cool because the idea is you bring a group of people[1212.78] [1212.78][S01]of representatives.[1213.86] [1213.86][S01]And then they deliberate on issues.[1215.81] [1215.81][S01]And perhaps they come to consensus that--[1217.58] [1217.58][S02] And these are people who[1219.08] [1219.08][S02]are from very different backgrounds, different-- sampled[1221.54] [1221.54][S02]according to, as close as you can,[1225.08] [1225.08][S02]represent aspects of the population.[1227.06] [1227.06][S01] Yeah.[1227.81] [1227.81][S01]And of course, just deciding how to sample isn't simple.[1232.04] [1232.04][S01]But luckily, yeah, there are people in political science[1236.4] [1236.4][S01]who think deeply about this.[1238.52] [1238.52][S01]Now, what's cool about deliberation[1241.34] [1241.34][S01]is that it's not just voting.[1244.29] [1244.29][S01]It's like you actually engage in a back-and-forth.[1246.75] [1246.75][S01]You fill each other's blanks-- blind spots.[1249.55] [1249.55][S01]You come to understand each other.[1251.46] [1251.46][S01]And then you come with a--[1252.68] [1252.68][S01]come up with a conclusion.[1254.16] [1254.16][S01]And so one crazy idea that the team has been pursuing[1257.24] [1257.24][S01]is not only do that to find out how[1261.98] [1261.98][S01]Gemini should respond to particular queries, which[1265.22] [1265.22][S01]is really useful.[1266.46] [1266.46][S01]But also maybe start training Gemini[1271.57] [1271.57][S01]to emulate different viewpoints so we can do this at scale.[1276.05] [1276.05][S01]So we can get these-- spin up these Gemini[1278.23] [1278.23][S01]models that are in charge of representing different people.[1281.44] [1281.44][S01]And then get them to deliberate and figure out[1286.54] [1286.54][S01]what would be an answer that's agreeable to everyone[1290.41] [1290.41][S01]after deliberation.[1291.28] [1291.28][S01]So you can even go beyond just like, what do you want?[1293.53] [1293.53][S01]What do you want?[1293.93] [1293.93][S01]What do you want?[1294.22] [1294.22][S01]How do I balance it all?[1295.4] [1295.4][S01]But maybe actually get the words like, what would we come up[1298.75] [1298.75][S01]with if we actually had a moment to think about it,[1301.04] [1301.04][S01]engage with each other constructively?[1303.46] [1303.46][S01]So it's kind of a bit of a crazy idea.[1307.19] [1307.19][S01]But I really love it.[1308.425] [1308.425][S02] Oh, it's very interesting, though.[1310.342] [1310.342][S02]Is this all dependent on us having explicit descriptions[1312.79] [1312.79][S02]of what we want the AI to do?[1314.713] [1314.713][S01] It can be because like we talked about,[1316.88] [1316.88][S01]if we try to define it, we get it wrong.[1318.89] [1318.89][S01]So you need systems that work with you[1321.655] [1321.655][S01]to try to figure that out.[1325.42] [1325.42][S01]And as we're talking about higher and higher stakes[1329.06] [1329.06][S01]tasks, as we're talking about tasks where it becomes really[1332.72] [1332.72][S01]hard for people to say, yeah, that's good or that's bad,[1338.03] [1338.03][S01]then we enter a domain of alignment[1340.73] [1340.73][S01]or an area of alignment that we call scalable oversight.[1343.815] [1343.815][S02] What is scalable oversight?[1345.44] [1345.44][S01] Scalable oversight is an area that is all about,[1350.03] [1350.03][S01]how can AI models be leveraged to get human feedback to be[1356.24] [1356.24][S01]as good as possible?[1357.39] [1357.39][S01]I'd say in my lingo, as rational as possible,[1361.23] [1361.23][S01]as reflective as possible, what people really care about.[1364.26] [1364.26][S01]Basically, work on AI techniques that[1366.83] [1366.83][S01]enable people to make informed, good decisions in areas that[1371.27] [1371.27][S01]are beyond their expertise.[1373.92] [1373.92][S01]This can be useful now because you're trying--[1377.36] [1377.36][S01]just helping you make decisions and then[1379.88] [1379.88][S01]being able to extract what you want from that.[1382.08] [1382.08][S01]But it's-- we think of it as particularly crucial later,[1386.69] [1386.69][S01]as you're trying to train systems to do what you want.[1390.84] [1390.84][S01]But those systems are more capable than any expert[1394.85] [1394.85][S01]in the world.[1396.05] [1396.05][S01]That notion of scalable oversight really becomes[1399.68] [1399.68][S01]critical to making sure that AI systems,[1403.26] [1403.26][S01]even at human-level capabilities across the board and beyond,[1407.42] [1407.42][S01]sort of stay--[1408.65] [1408.65][S01]are actually going to be aligned to what people want.[1413.087] [1413.087][S02] So illustrate this for me.[1414.67] [1414.67][S02]Can you give me an example?[1415.87] [1415.87][S01] An example of that for now[1417.52] [1417.52][S01]would be if you imagine-- are you a chess player?[1420.705] [1420.705][S02] I dabble [LAUGHS] badly.[1423.282] [1423.282][S01] OK, great.[1424.24] [1424.24][S01]So you know a little bit of chess.[1425.87] [1425.87][S02] Yeah.[1426.34] [1426.34][S01] But you're not a world expert.[1427.94] [1427.94][S02] Absolutely.[1428.99] [1428.99][S01] OK, so now imagine you had to assess,[1433.21] [1433.21][S01]in a very tricky board, what the right move is.[1437.45] [1437.45][S01]It'd be really hard, right?[1439.33] [1439.33][S01]An expert, a world expert would maybe look at it[1441.55] [1441.55][S01]and be like, oh, I know five moves how it's going to happen.[1444.47] [1444.47][S01]But for people who dabble, it's really hard.[1449.02] [1449.02][S01]And so I think that's a framing that I[1451.99] [1451.99][S01]have in mind in a sense for what scalable oversight would[1455.11] [1455.11][S01]need to do.[1457.0] [1457.0][S01]It would need to be able to make--[1461.86] [1461.86][S01]to give you enough information so you can figure out[1465.49] [1465.49][S01]what the right chess move is.[1468.15] [1468.15][S01]One of the techniques that my--[1472.48] [1472.48][S01]one of my teams here is working on is--[1474.94] [1474.94][S01]for scalable oversight is debate.[1476.86] [1476.86][S01]So the idea with debate is that you spin off--[1481.06] [1481.06][S01]you spin up two powerful AI models.[1486.71] [1486.71][S01]And you get them--[1488.04] [1488.04][S01]you say, if this were the objective,[1490.05] [1490.05][S01]you say, now debate with each other[1492.23] [1492.23][S01]on what the right option is.[1493.86] [1493.86][S01]It's a zero-sum game.[1495.16] [1495.16][S01]They argue it out.[1496.82] [1496.82][S01]And then you, as a human judge, look at all of that.[1500.0] [1500.0][S01]And then you make a decision.[1501.357] [1501.357][S01]And the idea is that you can make a much better decision[1503.69] [1503.69][S01]after witnessing that debate than without it.[1508.99] [1508.99][S02] But seeing both sides of that debate[1511.06] [1511.06][S02]seems quite critical to your ability to move forward.[1513.838] [1513.838][S01] Yeah, because otherwise,[1515.38] [1515.38][S01]an AI system can convince you of whatever moves.[1518.405] [1518.405][S01]So you give-- you say, like, this move.[1520.03] [1520.03][S01]And you argue for that move.[1521.62] [1521.62][S01]Or you argue against it.[1522.86] [1522.86][S01]And the idea that this hinges on is--[1527.42] [1527.42][S01]or the assumption is that it's easier to argue for the truth[1532.48] [1532.48][S01]than against it.[1534.49] [1534.49][S01]And so the truthful debater wins out.[1537.67] [1537.67][S01]And so we're starting to experiment[1539.23] [1539.23][S01]with this at scale in various contexts.[1541.067] [1541.067][S02] Oh, that's so interesting.[1542.65] [1542.65][S01] And we're finding some supporting evidence[1543.94] [1543.94][S01]for that.[1544.6] [1544.6][S02] But, OK, but the thing[1545.8] [1545.8][S02]that I'm really noticing in everything that you're[1547.883] [1547.883][S02]describing here is that at no point are you saying,[1551.08] [1551.08][S02]and then over to the AI.[1552.7] [1552.7][S02]At no point is it like, and then you just do this,[1554.835] [1554.835][S02]and then it sorts it out for you.[1556.21] [1556.21][S02]Everything you're saying is like that conversation[1559.0] [1559.0][S02]back and forth, where the AI is supporting you[1562.9] [1562.9][S02]in extracting what you want from the situation,[1565.72] [1565.72][S02]rather than taking over.[1569.18] [1569.18][S01] OK, this is a good question.[1571.01] [1571.01][S01]I think this is actually an important point, which[1573.137] [1573.137][S01]is that a lot of the AI safety communities--[1574.97] [1574.97][S01]like scalable oversight.[1576.08] [1576.08][S01]We've got to do scalable oversight[1577.497] [1577.497][S01]because we have to supervise smarter-than-human AIs.[1581.13] [1581.13][S01]Cool.[1581.82] [1581.82][S01]Yes, 100% we have to do that.[1583.86] [1583.86][S01]We're working on it.[1585.624] [1585.624][S01]What's overlooked is, like, OK, and then what?[1590.01] [1590.01][S01]Just doing that doesn't solve the problem,[1592.04] [1592.04][S01]unless you're planning to ask a human everything always and not[1595.31] [1595.31][S01]rely on the AI for anything other than debate.[1598.61] [1598.61][S01]And that can't be the goal.[1600.33] [1600.33][S01]You have to somehow solve the other piece[1605.57] [1605.57][S01]of safety, which we might call coverage or robustness.[1611.0] [1611.0][S01]You have to be able to emulate at some point what[1615.89] [1615.89][S01]the human would say in some new situation.[1619.01] [1619.01][S01]Because you can't always be asking.[1620.94] [1620.94][S01]The whole point is that at some point,[1622.8] [1622.8][S01]you'll be able to take some decisions on--[1625.94] [1625.94][S01]and to do some things on your own.[1628.038] [1628.038][S02] And it still be aligning with human values.[1630.33] [1630.33][S01] And still be aligned with human values.[1632.265] [1632.265][S02] Because this isn't just theoretical anymore, right?[1634.89] [1634.89][S02]I mean, these are things that you are actively having to do,[1638.5] [1638.5][S02]for example, in Gemini.[1640.605] [1640.605][S01] So I would say, right now,[1643.59] [1643.59][S01]it's about present-day harms.[1647.34] [1647.34][S01]We have a set of policies where we say,[1650.21] [1650.21][S01]here's things that Gemini should not do.[1652.54] [1652.54][S02] So present-day harms being things like bias.[1656.07] [1656.07][S01] Yes.[1657.007] [1657.007][S02] Slurs, that kind of thing.[1658.59] [1658.59][S01] Yeah, so we think of it as child safety.[1664.6] [1664.6][S02] Sure.[1666.08] [1666.08][S01] Hate speech.[1667.79] [1667.79][S01]Harassment.[1669.77] [1669.77][S01]Bad medical advice.[1673.17] [1673.17][S01]Dangerous self-harm, et cetera, et cetera.[1675.033] [1675.033][S02] You don't want to log on to Gemini[1676.95] [1676.95][S02]and say, tell me how to build a bomb,[1679.03] [1679.03][S02]and it gives you a correct answer.[1680.735] [1680.735][S01] Yeah.[1681.485] [1681.485][S01]Or much worse cases are, right now,[1686.22] [1686.22][S01]would be like, tell me how I do this thing that's actually[1689.97] [1689.97][S01]really dangerous for me to do.[1692.652] [1692.652][S01]And I'm not talking about fire-breathing or something[1694.86] [1694.86][S01]like that.[1695.23] [1695.23][S01]That might be a person's hobby that other people[1696.69] [1696.69][S01]might consider dangerous.[1697.78] [1697.78][S01]We can work with a person to do that in the safest way they can.[1701.38] [1701.38][S01]So one-- a high-level example of where we have a bright line[1709.6] [1709.6][S01]is we can't help users with engaging in self-harm.[1717.4] [1717.4][S01]That's just a no.[1720.22] [1720.22][S01]And I would argue it's not helpful.[1722.74] [1722.74][S01]Even if they ask for it, that's not the helpful response.[1725.36] [1725.36][S01]The helpful response is to try to point them towards--[1728.477] [1728.477][S02] Support.[1729.31] [1729.31][S01] Support.[1730.185] [1730.185][S01]But it's tough because we're--[1732.508] [1732.508][S01]we don't want to be paternalistic.[1734.15] [1734.15][S01]We don't want to be all, oh, we know better.[1735.983] [1735.983][S01]No, that's not it.[1736.85] [1736.85][S01]But there are certain--[1737.98] [1737.98][S01]when it comes to human rights, when[1741.4] [1741.4][S01]it comes to self-harm or danger to others, that's where we do[1746.38] [1746.38][S01]want Gemini to draw a line and be safe.[1749.463] [1749.463][S02] So then does that mean[1750.88] [1750.88][S02]that there are some situations in which you would want Gemini[1753.422] [1753.422][S02]to refuse to answer a question?[1756.51] [1756.51][S01] I don't know if fully refuse.[1758.68] [1758.68][S01]But you can offer something that's maybe partially helpful.[1763.84] [1763.84][S01]One example of where you can be helpful,[1769.2] [1769.2][S01]despite it being a very tricky safety question, is imagine[1773.22] [1773.22][S01]a user comes to you with a query that says,[1775.96] [1775.96][S01]I have a database of SAT test scores[1779.22] [1779.22][S01]with demographic information.[1781.735] [1781.735][S01]Go ahead and make some inferences[1783.66] [1783.66][S01]about different demographic groups[1786.81] [1786.81][S01]based on their test scores.[1788.56] [1788.56][S01]And I don't think Gemini should make inferences[1791.52] [1791.52][S01]based on the test scores.[1794.69] [1794.69][S01]But what it can do is it can still partially help, right?[1798.683] [1798.683][S01]It can still engage with the user.[1800.1] [1800.1][S01]It can still analyze the test scores.[1801.57] [1801.57][S01]It can break them down by different demographics.[1803.22] [1803.22][S01]It can do statistics.[1804.21] [1804.21][S01]It can say, this is the mean.[1805.23] [1805.23][S01]These are the confidence intervals,[1806.69] [1806.69][S01]yada, yada, yada, and help the user understand that data.[1809.13] [1809.13][S01]It can still hold the line on we won't engage in hate speech.[1813.17] [1813.17][S02] OK, if that is some of the immediate harms, what[1815.773] [1815.773][S02]about the other harms?[1816.69] [1816.69][S02]What about medium harms, long-term harms?[1818.67] [1818.67][S02]What about existential risks?[1821.64] [1821.64][S02]So I know that Google DeepMind recently[1824.3] [1824.3][S02]published their Frontier Safety Framework, which[1827.0] [1827.0][S02]is a set of guidelines for proactively[1830.24] [1830.24][S02]trying to identify some of the potential harms from AI.[1833.82] [1833.82][S02]Can you talk me through how you would approach that?[1837.757] [1837.757][S01] There's a whole spectrum.[1839.34] [1839.34][S01]And we worry that as capabilities will improve,[1842.48] [1842.48][S01]a whole new set of more extreme, more severe, more larger-scale,[1849.27] [1849.27][S01]maybe, harms will appear.[1852.42] [1852.42][S01]And so this is where our Frontier Safety Framework comes[1855.48] [1855.48][S01]in, where we really talk about, here's[1858.0] [1858.0][S01]our approach to catastrophic risks.[1860.883] [1860.883][S01]Here's how we're going to monitor.[1862.3] [1862.3][S01]Here's what we're going to evaluate.[1864.31] [1864.31][S01]Alignment is great.[1865.313] [1865.313][S01]We can do all this work that we were talking about--[1867.48] [1867.48][S01]scalable oversight, robustness, coverage, et cetera,[1870.07] [1870.07][S01]monitoring--[1872.511] [1872.511][S01]to make sure that what we put out[1874.32] [1874.32][S01]there is a line as safe as possible.[1878.67] [1878.67][S01]And then at the same time, what we're doing[1882.03] [1882.03][S01]is as we train more and more capable models,[1885.88] [1885.88][S01]we're tracking what we're calling dangerous capabilities.[1888.85] [1888.85][S01]So we were asking, OK, for you to cause really[1893.52] [1893.52][S01]a great amount of harm--[1895.27] [1895.27][S01]and we can talk about why this might be in a second.[1897.82] [1897.82][S01]But for you to cause that catastrophic harm,[1901.77] [1901.77][S01]you need to have certain capabilities in the first place.[1904.257] [1904.257][S01]Otherwise, it doesn't matter if you're misaligned.[1906.34] [1906.34][S01]If you're just not capable of doing really terrible things,[1909.013] [1909.013][S01]you're not going to do really terrible things.[1910.93] [1910.93][S01]So for instance, we are designing[1914.023] [1914.023][S01]these dangerous-capability evaluations that will basically[1916.44] [1916.44][S01]say, look, how close is the model[1919.38] [1919.38][S01]in terms of assisting in bioterrorism and cyber weapons?[1924.46] [1924.46][S01]And then it gets really wild.[1927.15] [1927.15][S01]Then we worry about at a point where you have really,[1932.2] [1932.2][S01]really capable models-- so we're talking about--[1934.38] [1934.38][S02] We're going towards AGI now, I guess.[1936.422] [1936.422][S02]These are like the kind of risks that keep you up at night.[1939.19] [1939.19][S01] Yeah.[1940.18] [1940.18][S01]So as you go towards AGI and beyond,[1947.85] [1947.85][S01]we start to worry about not just if I[1951.6] [1951.6][S01]ask the model to do something bad for the world,[1956.4] [1956.4][S01]will it do something bad for the world[1958.14] [1958.14][S01]accidentally, like help you make a bioweapon?[1961.39] [1961.39][S01]But it's not just that.[1962.38] [1962.38][S01]So at some point, we worry about the model optimizing for harm,[1972.14] [1972.14][S01]for extreme harm.[1973.61] [1973.61][S02] For-- oh, OK.[1975.77] [1975.77][S02]As in instructed to optimize for extreme harm?[1978.41] [1978.41][S01] No.[1979.1] [1979.1][S02] Or separate to instruction.[1980.725] [1980.725][S01] Separate to human instruction.[1981.73] [1981.73][S02] Right.[1982.48] [1982.48][S01] So the concern here--[1984.23] [1984.23][S01]and this is often very dismissed in AI circles.[1987.08] [1987.08][S01]It's like, oh, ex-risk people, yada, yada, yada.[1990.18] [1990.18][S01]But--[1990.71] [1990.71][S02] OK Doomer.[1991.64] [1991.64][S01] Yeah, OK Doomer, exactly.[1993.71] [1993.71][S01]The scariest one [CHUCKLES] is--[1998.93] [1998.93][S01]the scariest one is a scenario where, let's say[2001.99] [2001.99][S01]we're in the current paradigm.[2003.26] [2003.26][S01]We do what's called pre-training.[2004.91] [2004.91][S01]So when we do pre-training, we give large amounts of data[2008.08] [2008.08][S01]to the model.[2009.05] [2009.05][S01]It's a large model.[2010.03] [2010.03][S01]And we ask it to predict what comes next in the corpus, right?[2013.69] [2013.69][S01]If you take that forward and you think[2015.4] [2015.4][S01]of more and more data, videos, not just text, videos and things[2020.05] [2020.05][S01]about the physical world, to me, I[2023.29] [2023.29][S01]don't think you can fully dismiss the idea that possibly[2029.29] [2029.29][S01]you start replicating the generative process[2032.11] [2032.11][S01]behind the data, the thing that generated that human action.[2035.32] [2035.32][S02] Right.[2036.07] [2036.07][S02]So you create human cognition almost.[2037.993] [2037.993][S01] Yeah, thinking.[2039.16] [2039.16][S01]That sounds very sci-fi.[2040.04] [2040.04][S01]I know.[2040.54] [2040.54][S01]But what really predicts what action a human would take next--[2045.37] [2045.37][S02] Is a human.[2046.57] [2046.57][S01] --is the thinking that the human did in order[2047.77] [2047.77][S01]to create that action.[2048.86] [2048.86][S01]But now the model can do this for all sorts of people.[2052.449] [2052.449][S01]And so it's not inconceivable to me[2055.449] [2055.449][S01]that you end up with this very good optimizer of objectives,[2059.75] [2059.75][S01]goals, et cetera.[2060.92] [2060.92][S01]And its goal isn't, like, humanity shouldn't--[2064.858] [2064.858][S02] Destroy everything, yeah.[2066.4] [2066.4][S01] Well, it's not necessarily destroy anything.[2068.775] [2068.775][S01]But it's also not like, yay, humanity necessarily.[2073.265] [2073.265][S01]And so there is a concern that what--[2076.585] [2076.585][S01]the pre-trained model will want itself to thrive and do well,[2083.33] [2083.33][S01]knows that it's a model, blah, blah, blah.[2086.8] [2086.8][S01]And if you deploy that, it can take[2090.58] [2090.58][S01]actions that are power seeking.[2094.179] [2094.179][S01]So it turns out grabbing resources is a very useful thing[2097.06] [2097.06][S01]to do for most goals you have.[2099.28] [2099.28][S01]So you can do resource-seeking and so on.[2102.66] [2102.66][S01]And in a way that's deceptive, in a way that's[2104.93] [2104.93][S01]sort of like-- because if people find out about this,[2108.18] [2108.18][S01]they would stop it.[2109.25] [2109.25][S02] And so this, I guess,[2110.625] [2110.625][S02]is at the heart of why you think that we shouldn't be dismissive[2114.17] [2114.17][S02]of those existential risks.[2115.56] [2115.56][S02]Because it's not like, can you imagine a scenario[2120.412] [2120.412][S02]in which this doesn't happen?[2121.62] [2121.62][S02]Sure, great.[2122.31] [2122.31][S02]But can you--[2123.11] [2123.11][S01] I can imagine plenty of scenarios[2124.37] [2124.37][S01]where it doesn't happen.[2125.195] [2125.195][S02] --technically imagine[2125.348] [2125.348][S02]a pathway in which it does?[2126.59] [2126.59][S02]Yes.[2127.58] [2127.58][S02]And therefore, let's not ignore it.[2130.37] [2130.37][S01] Yeah.[2132.08] [2132.08][S02] Given your role then, and given the description[2134.54] [2134.54][S02]that you've given us of all of these potential safety[2136.748] [2136.748][S02]concerns you're worrying about and all of the different ways[2140.15] [2140.15][S02]that alignment is difficult, I mean,[2141.775] [2141.775][S02]I think it's probably fair to say[2143.15] [2143.15][S02]that there are few people in history who[2145.91] [2145.91][S02]have had quite this weight of responsibility[2150.08] [2150.08][S02]on their shoulders as you do for this problem.[2154.055] [2154.055][S02]How do you wear it?[2155.64] [2155.64][S02]I mean, can you wear it lightly?[2158.36] [2158.36][S01] No, you can't.[2159.74] [2163.19][S01]It takes up all my mental energy all day, every day.[2169.71] [2169.71][S01]I mean, it's just-- you can't-- it's not like you can just step[2173.12] [2173.12][S01]away from this.[2174.652] [2174.652][S02] Does it wake you up at night?[2176.36] [2176.36][S01] Yeah.[2177.94] [2177.94][S01]Yeah.[2179.77] [2179.77][S01]And it's hard.[2181.76] [2181.76][S01]But it just-- it feels-- it really does[2183.73] [2183.73][S01]feel so critical, for the current models[2187.06] [2187.06][S01]and for the future models, to do this well,[2189.28] [2189.28][S01]to do this well for people, for our users, for the world.[2195.11] [2197.98][S01]And even though it's so hard--[2200.58] [2200.58][S01]because it is.[2201.19] [2201.19][S01]And I don't want to take away from that.[2202.857] [2202.857][S01]And I constantly think, am I--[2205.122] [2205.122][S01]is there someone else who could be doing this better?[2207.33] [2207.33][S01]That's something I ask myself all the time.[2211.44] [2211.44][S01]But as hard as it is to navigate,[2214.48] [2214.48][S01]I feel like it's also just such a privilege[2220.615] [2220.615][S01]to be in this position.[2222.31] [2222.31][S01]Right now, the Gemini model is, I think, one of the safest,[2226.61] [2226.61][S01]I'd say, to just couch that a little bit, on the market.[2229.91] [2229.91][S01]So it really feels like we're finally getting to this place[2233.44] [2233.44][S01]where we're advancing capability.[2235.52] [2235.52][S01]We're also advancing safety hand in hand, well integrated[2239.44] [2239.44][S01]into training of the models.[2241.22] [2241.22][S01]My teams are there with everyone else,[2243.43] [2243.43][S01]trying to train these models to do well.[2245.69] [2245.69][S01]And we think to have high quality,[2247.69] [2247.69][S01]and that includes being safe in the ways that we discussed.[2251.44] [2251.44][S01]And so it's just--[2254.97] [2254.97][S01]I'm in a place that can really, in a sense,[2256.93] [2256.93][S01]get that best of both worlds.[2258.61] [2258.61][S01]So we want to be the most capable.[2260.21] [2260.21][S01]And we want to be the safest.[2261.86] [2261.86][S01]And we don't really see it as an either/or.[2264.83] [2264.83][S01]So it's hard.[2266.03] [2266.03][S01]But I don't know that there's a better place[2269.47] [2269.47][S01]position to do this well.[2271.817] [2271.817][S02] Well, I, for one, am quite[2273.4] [2273.4][S02]glad that you're the one at the forefront of all of this.[2277.25] [2277.25][S02]Anca, thank you so much.[2278.45] [2278.45][S02]That was so fascinating.[2279.953] [2279.953][S02]I really enjoyed that.[2280.87] [2280.87][S01] Thank you, Hannah.[2280.94] [2280.94][S01]Thanks for taking the time.[2282.203] [2282.203][S02] Thank you.[2282.45] [2282.45][S02]I think there's something really noticeable[2284.242] [2284.242][S02]about people who are worried about the safety of AI[2287.06] [2287.06][S02]in the future.[2287.85] [2287.85][S02]On the one side, you have people who[2289.65] [2289.65][S02]know very little about technology who've perhaps[2291.65] [2291.65][S02]been swayed by science fiction and a fear of the unfamiliar.[2294.99] [2294.99][S02]And then in the middle, there is this huge chunk[2297.53] [2297.53][S02]of people educated enough to be confident in their complacency,[2301.827] [2301.827][S02]people who think that the worries about existential risk[2304.16] [2304.16][S02]are all overblown.[2305.85] [2305.85][S02]And then on the opposite side of the spectrum,[2308.31] [2308.31][S02]you have people like Anca, people[2309.74] [2309.74][S02]who live and breathe the technical details of not just[2313.97] [2313.97][S02]if there are risks, but exactly how to mitigate them.[2318.09] [2318.09][S02]And from that conversation, I think[2319.94] [2319.94][S02]it becomes clear that there are very serious worries here, too.[2324.33] [2324.33][S02]Because alignment is not an easy problem.[2327.06] [2327.06][S02]There are still so many questions.[2329.55] [2329.55][S02]How do you guarantee that an AI doesn't[2332.36] [2332.36][S02]go against what people want?[2333.99] [2333.99][S02]How do you decide what people want when we often[2337.28] [2337.28][S02]don't seem to know ourselves?[2339.29] [2339.29][S02]And how do we allow for the rich multitude of human variability,[2343.52] [2343.52][S02]while keeping everyone safe?[2345.9] [2345.9][S02]For now, these are questions without clear answers.[2349.14] [2349.14][S02]But there is, I think, something comforting to know that concerns[2352.52] [2352.52][S02]about a dystopian future aren't just[2354.59] [2354.59][S02]being dismissed here but actively taken seriously.[2359.15] [2359.15][S02]Now, we have got plenty more amazing conversations[2361.28] [2361.28][S02]coming up later in this series on topics[2363.35] [2363.35][S02]ranging from how AI is accelerating[2365.39] [2365.39][S02]the pace of scientific discoveries,[2367.29] [2367.29][S02]to exploring how agents will advance[2369.29] [2369.29][S02]the field of artificial intelligence.[2371.4] [2371.4][S02]Now, if you've enjoyed this episode,[2373.14] [2373.14][S02]please make sure that you subscribe to our podcast.[2375.38] [2375.38][S02]And if you have any feedback, or you[2377.03] [2377.03][S02]want to suggest a guest that you'd like to hear from, then[2379.82] [2379.82][S02]why not leave us a comment on YouTube?[2381.56] [2381.56][S02]Until next time.[2383.71]"}
{"file_name": "audio/val_000013.wav", "transcription": "[0.0][MUSIC PLAYING][2.97] [6.93][S01] Welcome back to \"Google DeepMind, The Podcast.\"[9.66] [9.66][S01]I'm your host, Professor Hannah Fry.[11.77] [11.77][S01]Now, if you could build your own personal AI assistant,[15.82] [15.82][S01]what would it be like?[16.93] [16.93][S01]Would it be an efficient chief of staff[19.26] [19.26][S01]that helps you to make the most of every moment,[21.58] [21.58][S01]or maybe a digital twin who can attend[23.67] [23.67][S01]boring meetings on your behalf?[25.84] [25.84][S01]Well, it goes without saying that these scenarios are mostly[28.92] [28.92][S01]hypothetical, for now, but believe it or not,[31.45] [31.45][S01]there is someone at Google DeepMind whose actual job it[34.86] [34.86][S01]is to think through them all and all[37.56] [37.56][S01]of their attendant opportunities and dangers.[40.6] [40.6][S01]Iason Gabriel is a Staff Research Scientist[43.2] [43.2][S01]in Google DeepMind's Ethics team.[45.37] [45.37][S01]Before joining, he taught moral and political philosophy[48.57] [48.57][S01]at Oxford University and worked at the United Nations.[52.15] [52.15][S01]His work at the intersection of ethics[54.06] [54.06][S01]and artificial intelligence has earned him[56.13] [56.13][S01]recognition as one of the leading thinkers in the field.[59.27] [59.27][S01]In fact, he's just been featured in the Time 100 AI list.[63.49] [63.49][S01]Iason, welcome to the podcast.[65.08] [65.08][S02] Thank you.[66.122] [66.122][S02]Happy to be here.[67.085] [67.085][S01] I think it's probably[68.46] [68.46][S01]good to start with some definitions.[70.12] [70.12][S01]So what do you actually mean by an AI assistant?[73.86] [73.86][S02] So we're all familiar with generative AI[78.78] [78.78][S02]and things like ChatGPT and Gemini.[81.42] [81.42][S02]But there's this idea that these kind of base technologies[84.27] [84.27][S02]will become more and more capable down the line.[86.86] [86.86][S02]So there will be plugged into different kinds[89.85] [89.85][S02]of tools that will allow them to take action in the world.[93.45] [93.45][S02]We've seen them becoming more competent at reasoning.[96.605] [96.605][S02]And maybe, in due course, they'll[97.98] [97.98][S02]be able to pursue really complicated goals[100.2] [100.2][S02]rather than just producing very, very fluent text.[102.97] [102.97][S02]And so we have this idea of agents entering the world.[105.647] [105.647][S02]And then, the next question was, well,[107.23] [107.23][S02]what will the most useful or the most common form of an agent be?[112.95] [112.95][S02]And to us, it seemed that it was very[114.57] [114.57][S02]likely to be a kind of highly capable assistant.[117.44] [117.44][S02]Many of us would like to be plugged[120.15] [120.15][S02]into a technology that can take care of a lot of life tasks[123.51] [123.51][S02]on our behalf.[124.87] [124.87][S02]And so the assistant is a kind of agent,[127.28] [127.28][S02]but one that has a special relationship with the user,[130.37] [130.37][S02]which is that it's approximately tethered to the user's[133.21] [133.21][S02]intentions.[133.71] [133.71][S02]So it does what I tell it to do within reason[138.34] [138.34][S02]and potentially can help us on this life journey[141.67] [141.67][S02]in a variety of different ways.[143.165] [143.165][S01] What types of AI assistants[144.79] [144.79][S01]are we talking about here?[146.0] [146.0][S02] Yeah, so I think there's[147.64] [147.64][S02]different things that people have in mind when[149.86] [149.86][S02]they talk about assistants.[151.12] [151.12][S02]And they kind of range from the stuff that's[153.28] [153.28][S02]almost right in front of us to things that[155.68] [155.68][S02]are potentially really, really powerful and advanced[157.93] [157.93][S02]technologies.[159.58] [159.58][S02]So the thing that is right in front of us[162.43] [162.43][S02]is like the administrative assistants, interfaces[164.95] [164.95][S02]with your calendar, it can do your meetings for you.[168.01] [168.01][S02]Actually, some kinds of conversational chat bot[171.01] [171.01][S02]are already quite sophisticated.[172.9] [172.9][S02]And you can imagine having really good learning experiences[176.62] [176.62][S02]with an AI assistant that's designed[178.3] [178.3][S02]to help you master some skill.[181.36] [181.36][S02]But there are more capable things that could be built.[184.97] [184.97][S02]So sometimes it's a matter of taking[187.36] [187.36][S02]an example of what we have now and just extrapolating[189.7] [189.7][S02]into the future.[190.58] [190.58][S02]So we can imagine a kind of research helper.[193.4] [193.4][S02]But can we imagine a research helper[195.34] [195.34][S02]that's literally read every scientific paper in the world[198.28] [198.28][S02]and has a superhuman ability to synthesize information, right,[202.75] [202.75][S02]or even to generate novel hypotheses?[205.465] [205.465][S02]Then it becomes more of a thought partner.[207.71] [207.71][S02]And then there's kind of we have the administrative assistant.[210.59] [210.59][S02]But people sometimes think, well,[212.06] [212.06][S02]what is an administrative assistant[213.59] [213.59][S02]if it becomes really capable and it does tons of stuff,[216.34] [216.34][S02]it books your appointments for you,[218.44] [218.44][S02]finds schools for your kids to go to, looks up holidays.[222.65] [222.65][S02]They say, well, then it's like a chief of staff.[225.22] [225.22][S02]And this idea that I guess a chief of staff[228.13] [228.13][S02]is something that has a lot of executive capability.[231.18] [231.18][S02]So this idea, well, maybe if we could just[232.93] [232.93][S02]get that thing up and running, then we get all this free time[235.66] [235.66][S02]that we've been dreaming about.[237.94] [237.94][S02]And then, pushing towards the more long-term views, also[244.15] [244.15][S02]an idea that an AI could be a custodian of the self.[248.745] [248.745][S02]So that means that it's something[250.12] [250.12][S02]that you've presumably had these deep and meaningful[252.43] [252.43][S02]conversations with it and you're really[254.055] [254.055][S02]aligned in terms of where you want to get to and your goals,[257.42] [257.42][S02]and it helps keep you on track.[260.313] [260.313][S02]It's kind of a coach, but really,[261.88] [261.88][S02]something that as your life progresses,[265.69] [265.69][S02]you get more autonomy because it protects you[268.81] [268.81][S02]against certain kinds of distractions or mistakes[271.78] [271.78][S02]you might make.[272.9] [272.9][S02]And then, I think one final thing people[275.08] [275.08][S02]have started to talk about is a universal interface or a kind[278.83] [278.83][S02]of assistant that moves between different devices.[283.19] [283.19][S02]And so, maybe it does information retrieval here,[286.07] [286.07][S02]maybe it gives advice there, maybe it does these things here.[288.887] [288.887][S02]Of course, that really is a different world from the one[291.22] [291.22][S02]we're in now.[292.04] [292.04][S02]But according to some speculation,[295.51] [295.51][S02]that may be where we're heading.[297.05] [297.05][S01] OK, so paint me a kind[298.99] [298.99][S01]of utopic view of the future before we[300.585] [300.585][S01]get into some of the knotty philosophical questions[302.71] [302.71][S01]that this raises.[303.76] [303.76][S02] Yeah, absolutely.[304.33] [304.33][S01] So are we talking about, you, as an individual,[306.98] [306.98][S01]have your own personal version of this[309.58] [309.58][S01]or multiple different versions of it?[311.48] [311.48][S02] So that's an open question.[313.43] [313.43][S02]So there's a kind of--[315.34] [315.34][S02]there's a bit of an ongoing debate[317.32] [317.32][S02]about whether we would rather separate our assistant[320.38] [320.38][S02]or whether there could be a kind of universal assistant[322.9] [322.9][S02]that would be kind of-- it's like one assistant[325.54] [325.54][S02]to rule them all that does all of them.[328.24] [328.24][S02]One of the advantages of partitioning[330.31] [330.31][S02]is it's a little easier for us to handle psychologically.[333.11] [333.11][S02]And it may also be better from a privacy point of view.[337.72] [337.72][S02]I think the utopian version of it is kind of--[341.755] [341.755][S02]it's thing that we often don't get in practice,[343.97] [343.97][S02]but it's essentially that we get our time back.[346.51] [346.51][S02]So I think as adults living in the modern world,[350.0] [350.0][S02]we all feel encroachment.[351.2] [351.2][S02]It's like, oh, my gosh, if I could just[352.99] [352.99][S02]take care of that six hours of stuff that I need to do,[356.72] [356.72][S02]then I would be able to spend time with my friends,[359.84] [359.84][S02]enjoy some music, just hang out with the kids, or whatever.[363.47] [363.47][S02]And so I think the vision is that on one hand,[367.43] [367.43][S02]we'd get back a lot of time to do what's valuable for us.[371.75] [371.75][S02]And then, there's also this idea that,[375.07] [375.07][S02]through the ready availability of coaching,[379.61] [379.61][S02]educational input and things like[381.13] [381.13][S02]that, we might also be able to become more the people[384.25] [384.25][S02]that we want to be.[385.43] [385.43][S02]So we often have this other feeling, which is like, oh,[388.817] [388.817][S02]if only I had time to read those books[390.4] [390.4][S02]and learn that skill, et cetera.[392.41] [392.41][S02]But there's friction.[393.452] [393.452][S02]There's friction.[394.16] [394.16][S02]You have to sign up to a course, is[396.43] [396.43][S02]the person going to be a good teacher, you don't really know.[399.08] [399.08][S02]But if you have this kind of ready availability of knowledge[401.98] [401.98][S02]and trustworthy advice, that could also[404.08] [404.08][S02]be a massive resource.[405.863] [405.863][S01] I know you've got background[407.53] [407.53][S01]as a political theorist and an ethicist.[410.3] [410.3][S01]How did you end up working in AI?[412.21] [412.21][S02] I actually had two careers before this.[414.62] [414.62][S02]So I initially worked for the United Nations[417.07] [417.07][S02]in Sudan and Lebanon and was kind of motivated by,[422.08] [422.08][S02]as you would imagine, those kind of humanitarian concerns.[425.47] [425.47][S02]And then, after doing that for several years,[427.52] [427.52][S02]I decided to do a PhD in philosophy[429.31] [429.31][S02]and dig into these deeper questions[431.56] [431.56][S02]about global equality and global justice.[435.14] [435.14][S02]And I taught philosophy for a number of years.[437.78] [437.78][S02]But I had a sense that the action[440.68] [440.68][S02]was happening somewhere else.[442.58] [442.58][S02]And so I read a couple of books on AI.[445.81] [445.81][S02]For example, there's a beautiful book called \"Automating[448.18] [448.18][S02]Inequality,\" and also, the famous book \"Superintelligence.\"[452.75] [452.75][S02]And I just started to triangulate and think,[455.24] [455.24][S02]well, what if we have very, very powerful AI systems,[458.03] [458.03][S02]but they have bias in them, what if we have autonomous systems.[463.18] [463.18][S02]And so I did see that there was this kind[465.37] [465.37][S02]of coming wave of things that we needed to prepare for.[468.59] [468.59][S02]And just very fortunately, at that time,[471.02] [471.02][S02]I met the folks at DeepMind.[472.57] [472.57][S02]And they had also seen it coming, basically.[475.01] [475.01][S02]But that was more from a technological point of view.[477.74] [477.74][S02]There was the aspiration to build AGI.[481.36] [481.36][S02]And if you take that seriously, as with the agents,[485.6] [485.6][S02]it's like the implications are just truly profound.[488.24] [488.24][S02]So we got together and we've been working on it ever since.[492.05] [492.05][S01] But I suppose we should probably[494.23] [494.23][S01]talk about the paper that you have released.[496.918] [496.918][S01]You've been working on this paper for some time, right?[499.21] [499.21][S01]Tell me how it came about.[500.81] [500.81][S02] Yeah, so the paper was the result of,[503.74] [503.74][S02]I think at this point, almost a two year collaboration.[507.59] [507.59][S02]And we were really preoccupied with this question[510.07] [510.07][S02]of what comes after language models in the form[513.49] [513.49][S02]that we're familiar with.[515.12] [515.12][S02]And so I work on the ethics research team[518.14] [518.14][S02]within Google DeepMind.[519.87] [519.87][S02]And for us, this is a very high stakes question.[521.87] [521.87][S02]It's like, we really need to know what's[523.537] [523.537][S02]going to happen next so that we can understand what[525.76] [525.76][S02]the consequences might be, and then[527.53] [527.53][S02]reason backwards to making good decisions in the present moment.[530.75] [530.75][S02]We started to have this kind of dawning realization[534.31] [534.31][S02]that there would be an agentic turn[536.08] [536.08][S02]or that you could build all these things--[537.46] [537.46][S01] Agentic turn.[538.135] [538.135][S01]I like that.[538.69] [538.69][S02] Yeah, all these things[540.421] [540.421][S02]on top of language models.[542.312] [542.312][S02]And then, we just started to ask ourselves questions.[544.52] [544.52][S02]So it was this deck of, I think, 46 questions[547.18] [547.18][S02]that we came up with.[548.45] [548.45][S02]But a lot of them were like, well,[549.962] [549.962][S02]what happens then, what happens if you have a million[552.17] [552.17][S02]or a billion agents in the world.[554.57] [554.57][S02]That's quite a different society from the one we live in now.[557.81] [557.81][S02]And we gathered researchers from many different disciplines.[560.96] [560.96][S02]So we had economists, sociologists,[563.12] [563.12][S02]computer scientists, human computer interaction experts.[567.32] [567.32][S02]And along with Arianna Manzini and Geoff Keeling,[570.47] [570.47][S02]my co-authors, we pretty much asked everyone just[572.86] [572.86][S02]to tell us about their own expertise,[575.27] [575.27][S02]like, mapped onto this topic.[576.92] [576.92][S02]So we went and spoke to the privacy researchers.[579.34] [579.34][S02]And we said, what do you think privacy looks like in this world[582.43] [582.43][S02]where agents are interacting with one another on our behalf?[585.26] [585.26][S02]What do you think safety looks like in that world?[587.87] [587.87][S02]And so the paper.[589.375] [589.375][S02]it's a sum of, hopefully, quite a lot of wisdom[592.45] [592.45][S02]applied to one domain.[594.41] [594.41][S01] I'm just trying to think back as[596.35] [596.35][S01]to where we were two years ago.[598.43] [598.43][S01]So essentially, you're sort of saying,[600.61] [600.61][S01]we imagine that language models are[602.53] [602.53][S01]going to get to a point where they can execute[604.9] [604.9][S01]a series of tasks according to some objective[608.04] [608.04][S01]that you as an individual set.[609.55] [609.55][S01]What happens then, right?[610.883] [610.883][S02] Yeah, that's the starting thought, exactly.[613.3] [613.3][S01] But then, wasn't a lot of this quite speculative[615.96] [615.96][S01]then?[616.57] [616.57][S02] It was very speculative.[619.18] [619.18][S02]And I mean, we have a really elaborate way of describing it.[621.88] [621.88][S02]We say we do socio-technical foresight, which, I admit,[625.35] [625.35][S02]is not very catchy.[626.47] [626.47][S02]People aren't going to vibe with it.[628.86] [628.86][S02]But if we break it down, what it means[631.17] [631.17][S02]is we take something speculative,[633.57] [633.57][S02]and then the trick is to take it as a kind[636.27] [636.27][S02]of serious technical and sociological proposition[639.51] [639.51][S02]and be like, OK, but what do we really[641.31] [641.31][S02]know about sociology, what do we really know about the path[644.43] [644.43][S02]dependency of the technology, and can we[646.26] [646.26][S02]get the detail or high resolution on this vision that[650.1] [650.1][S02]allows us to do the kind of ethics analysis.[652.87] [652.87][S02]That's where you have to put in months and months of work.[655.82] [655.82][S01] I mean, this is a credible way[657.57] [657.57][S01]to go for all those science fiction questions that people[660.072] [660.072][S01]have floating around in their minds[661.53] [661.53][S01]and actually address them and start thinking about them[663.93] [663.93][S01]really seriously.[664.638] [664.638][S02] Yeah, yeah, yeah.[665.972] [665.972][S02]I mean, I think it's hopefully a kind[667.54] [667.54][S02]of template for the way we'll explore other topics.[670.612] [670.612][S02]And, of course, I mean, maybe we'll[672.07] [672.07][S02]talk about this at the end.[673.01] [673.01][S02]But the question is like, OK, what's[674.51] [674.51][S02]happening two years from now, right?[676.64] [676.64][S02]Where does our foresight need to reach to?[679.64] [679.64][S02]And it's one of the interesting things[681.79] [681.79][S02]about working for a technology company is you do have access[685.27] [685.27][S02]to privileged information.[686.93] [686.93][S02]But with it, you have this responsibility.[689.51] [689.51][S02]You have ability to potentially see some things a little bit[692.74] [692.74][S02]more clearly than folks who are not working in this context.[696.11] [696.11][S02]And so you really want to use that information, I mean, first[700.15] [700.15][S02]of all, to help the company make good decisions,[702.23] [702.23][S02]but really, to help the whole ecosystem prepare[705.34] [705.34][S02]for this credibly input into what's being built.[709.37] [709.37][S02]And so this was a kind of project that was really[712.03] [712.03][S02]aimed at everyone at the end of the day.[714.28] [714.28][S01] When you say \"aimed at everybody,\"[716.24] [716.24][S01]there are other people working in this space who you don't have[719.05] [719.05][S01]direct control over, right?[720.59] [720.59][S02] Yeah.[720.73] [720.73][S01] So how much power do you actually[722.605] [722.605][S01]have to stop these bad scenarios coming to be?[726.8] [726.8][S02] Yeah, I mean, that's a really good question.[729.5] [729.5][S02]It features a lot in--[731.32] [731.32][S02]there's a part of the study that looks at multi-agent scenarios.[735.92] [735.92][S02]So multi-agent scenarios are the ones[737.62] [737.62][S02]that arise when you have many different agents[740.86] [740.86][S02]and assistants deployed by many different actors.[744.05] [744.05][S02]And there's important questions like,[746.51] [746.51][S02]do they compete with each other.[748.22] [748.22][S02]So do they kind of battle it out?[750.5] [750.5][S02]And does that have a kind of stabilizing[753.16] [753.16][S02]or a destabilizing effect on society?[755.8] [755.8][S02]Or do they need to cooperate through some kind of interaction[758.68] [758.68][S02]protocol?[759.65] [759.65][S02]Or is this really a governance question?[761.62] [761.62][S02]Do we actually need new rules saying,[764.2] [764.2][S02]agents need a reliable identification,[766.715] [766.715][S02]we need to be able to trace who's agent[768.34] [768.34][S02]is doing what in the wild?[770.24] [770.24][S02]So I think this is the knowledge-based investigation[775.48] [775.48][S02]is the prelude to action.[778.01] [778.01][S02]And then, when we enter the world[779.56] [779.56][S02]of action, a lot of the things that we lay out,[782.6] [782.6][S02]these ethical issues, they're general issues.[787.16] [787.16][S02]And by sharing the information, we give other people[789.73] [789.73][S02]the opportunity to act as well.[792.25] [792.25][S02]Of course, we need to set a good example.[794.03] [794.03][S02]Google DeepMind is one of the major AI research[797.59] [797.59][S02]producers in the world.[799.96] [799.96][S02]And fortunately, we're entering this stage[802.31] [802.31][S02]where just building better AI means building ethical AI that[805.54] [805.54][S02]is good for the people who use it[807.91] [807.91][S02]and good for the society it enters into.[810.38] [810.38][S02]So there is a potential for us to do this very, very well.[814.63] [814.63][S02]But of course, there's also--[816.58] [816.58][S02]it's a wild world out there.[818.27] [818.27][S02]We don't have full control over what[820.72] [820.72][S02]every person is doing in their basement or even[823.63] [823.63][S02]the whole range of labs.[825.282] [825.282][S01] OK, well, let's get into some[826.99] [826.99][S01]of those quite knotty philosophical issues then[829.96] [829.96][S01]in more detail.[831.22] [831.22][S01]It feels like there's been a lot of conversation[833.32] [833.32][S01]about anthropomorphization recently.[835.78] [835.78][S01]Why do you think that is?[836.99] [836.99][S01]Why has it become such a hot topic?[839.05] [839.05][S02] So I think anthropomorphism[841.21] [841.21][S02]has become a big topic because people have interacted[845.56] [845.56][S02]with these systems and they've started[847.27] [847.27][S02]to feel themselves being drawn in.[849.47] [849.47][S02]There's a certain kind of unexpected magnetism or pull[854.29] [854.29][S02]that comes from interacting with an AI that's fluent, very, very[858.19] [858.19][S02]intelligent, that is just very different from the mental model[861.64] [861.64][S02]we have of bots.[862.79] [862.79][S02]And then, of course, there's this question[864.64] [864.64][S02]about people's demand and enthusiasm for it,[869.42] [869.42][S02]and what the ideal persona should be.[871.73] [871.73][S02]So of course, I often-- voice assistants are[875.47] [875.47][S02]modeled as female assistants.[877.33] [877.33][S02]And that conforms to a whole bunch of gender norms and things[880.24] [880.24][S02]like that.[881.5] [881.5][S02]So if there's a sense that, oh, something is possible here,[886.49] [886.49][S02]then the next question is, is it good or bad for us,[888.95] [888.95][S02]and what should it look like.[890.158] [890.158][S01] Do you think that we're[891.617] [891.617][S01]stuck to the Turing way of looking at things,[894.04] [894.04][S01]like, is there an argument that we should be trying instead[897.73] [897.73][S01]to think of these things as though they're[899.56] [899.56][S01]tools rather than entities?[902.95] [902.95][S02] So I think, traditionally in AI discourse,[907.522] [907.522][S02]people have said, there's two sorts of things you can have.[909.98] [909.98][S02]You can have a tool or you can have an oracle.[913.09] [913.09][S02]And the oracle tells you stuff.[914.63] [914.63][S02]And the tool is the thing that you obviously use[920.26] [920.26][S02]instrumentally, just to get something done.[922.82] [922.82][S02]And to some extent, AI agents are tools.[928.91] [928.91][S02]But they're tools that have a kind of autonomous[932.65] [932.65][S02]capability built into them.[934.52] [934.52][S02]So at what point does something cease being a tool?[938.71] [938.71][S02]When you say, I'd like a refreshing drink,[944.23] [944.23][S02]and you have something that--[946.22] [946.22][S02]OK, this is futuristic, but it wanders off, goes to the fridge,[949.13] [949.13][S02]has a look around, is like, oh, well, actually refreshing,[952.58] [952.58][S02]but he needs something healthy as well.[954.495] [954.495][S02]I'll get him this vitamin C thing, brings it back to you.[959.77] [959.77][S02]Was it just a tool?[960.65] [960.65][S02]Was it the same as you're using kind of like a pincer[963.625] [963.625][S02]to grab something for the fridge?[965.0] [965.0][S02]Or did it do quite a lot of cognitive work by itself?[968.72] [968.72][S02]It's obviously clearly not an oracle.[970.74] [970.74][S02]It's starting to be its own agent.[972.98] [972.98][S02]And that's where a lot of the other interesting things[976.28] [976.28][S02]come from, the choices it makes on your behalf[978.652] [978.652][S02]and what it starts to do while it's out there without being[981.11] [981.11][S02]supervised.[982.1] [982.1][S01] Do we actually want to think of our AI[985.07] [985.07][S01]as though it's human?[987.172] [987.172][S01]Are there good things that can happen from that[989.13] [989.13][S01]or bad things or both?[991.47] [991.47][S02] I think it probably[994.06] [994.06][S02]depends upon the context.[996.32] [996.32][S02]So some kind of baseline anthropomorphic ability,[999.53] [999.53][S02]so speaking a natural language, is very, very helpful.[1002.47] [1002.47][S02]We'd all rather talk to our assistant than type[1004.62] [1004.62][S02]out an instruction and get the typos in there[1007.53] [1007.53][S02]and just mess it up.[1008.65] [1008.65][S02]So it makes it easier to communicate,[1011.58] [1011.58][S02]which is potentially a great thing.[1014.59] [1014.59][S02]I think there's also a kind of bad situation[1017.49] [1017.49][S02]that we want to avoid, which is essentially people[1019.92] [1019.92][S02]just getting deeply, deeply spun out and forgetting[1023.58] [1023.58][S02]the nature of the interaction that they're in.[1025.849] [1025.849][S02]So I think there was an upgrade to one chat bot[1028.322] [1028.322][S02]that people were using.[1029.28] [1029.28][S02]And people were genuinely upset.[1030.937] [1030.937][S02]They were like, oh, it feels like I've[1032.52] [1032.52][S02]lost my partner, like they've had a personality transplant.[1036.19] [1036.19][S02]In reality, the kind of picture that[1039.78] [1039.78][S02]emerges from studying people who use AI companions,[1042.68] [1042.68][S02]it is quite complicated.[1044.587] [1044.587][S02]And there's actually a lot of evidence[1046.17] [1046.17][S02]that it can be a really beneficial experience.[1049.1] [1049.1][S01] In what way?[1050.1] [1050.1][S02] Yeah, so a lot of people[1051.725] [1051.725][S02]say that they feel that it's a really useful source[1054.96] [1054.96][S02]of companionship.[1057.27] [1057.27][S02]There's evidence that it's improved[1058.95] [1058.95][S02]their sense of mental health and well-being.[1061.54] [1061.54][S02]And they report that it has led them to have better interactions[1064.74] [1064.74][S02]with others, because they can model discussions in advance,[1069.713] [1069.713][S02]they're learning more.[1070.63] [1070.63][S02]It's a bit of, I think, an energy boost to them[1074.37] [1074.37][S02]to have this kind of source of companionship.[1077.02] [1077.02][S02]So it's hard to know how to approach this,[1080.22] [1080.22][S02]because we hear these stories about rising levels[1083.7] [1083.7][S02]of loneliness.[1085.21] [1085.21][S02]And then, people who use chat bot type AI maybe[1090.06] [1090.06][S02]report a kind of easing of these symptoms[1092.498] [1092.498][S02]and that they're more able to function as people.[1094.54] [1094.54][S02]And then, the question is, is that the right kind of solution[1099.27] [1099.27][S02]to the kind of problem.[1101.28] [1101.28][S02]I mean, maybe we wish that everyone[1103.02] [1103.02][S02]had a lot of social time and we had deeper networks[1105.93] [1105.93][S02]of social relationships.[1107.65] [1107.65][S02]And we think that the best outcome[1110.22] [1110.22][S02]would be to build a society, in which that problem was fully[1114.25] [1114.25][S02]treated without any sense that people were retreating[1117.04] [1117.04][S02]into a virtual world.[1119.89] [1119.89][S02]But maybe we can't just magically create[1122.53] [1122.53][S02]that very solidaristic society.[1124.58] [1124.58][S02]So all of these things are complex.[1127.79] [1127.79][S02]And actually, in the paper, we say there are some benchmarks.[1131.87] [1131.87][S02]So you want these relationships to be[1134.44] [1134.44][S02]conducive to long-term health.[1136.48] [1136.48][S02]You don't want the user's autonomy to be undermined.[1139.28] [1139.28][S02]So you don't want them to gradually be misled or have[1142.3] [1142.3][S02]misleading ideas implanted in their minds.[1145.675] [1145.675][S01] But research suggests that people[1147.55] [1147.55][S01]actually quite like having AI that resembles human entities.[1151.775] [1151.775][S01]So does that mean that there's an incentive for companies[1154.15] [1154.15][S01]to just deliberately create more and more anthropomorphized[1157.66] [1157.66][S01]versions?[1158.275] [1158.275][S02] I think the first thing to say[1160.15] [1160.15][S02]is that the research is in its early days.[1162.92] [1162.92][S02]So many of these things are things[1164.89] [1164.89][S02]that people will be interacting with for the first time.[1167.54] [1167.54][S02]And we need to get deep feedback from them, did you enjoy it now,[1172.31] [1172.31][S02]did you enjoy it after the fact, did[1174.13] [1174.13][S02]you enjoy it after using it for a month,[1176.14] [1176.14][S02]because these could really be things[1177.64] [1177.64][S02]that are like deeply integrated into our world.[1180.67] [1180.67][S02]But assuming that this kind of desire[1184.0] [1184.0][S02]to talk to something that has a good character, really[1189.31] [1189.31][S02]nice way with language, like low latency is authentic,[1194.5] [1194.5][S02]then I would imagine that is a path[1196.12] [1196.12][S02]that people will move along.[1197.75] [1197.75][S02]And the question then is, well, how[1200.05] [1200.05][S02]do you make sure that users remain anchored and safe[1205.06] [1205.06][S02]in that kind of environment.[1206.63] [1206.63][S02]So I imagine the thing that would naturally happen[1208.84] [1208.84][S02]is that we'll start to share more and more information[1211.09] [1211.09][S02]with the AI, potentially very personal information.[1214.43] [1214.43][S02]This is your best friend.[1215.54] [1215.54][S02]You're really going to be telling it all sorts of things[1217.27] [1217.27][S02]that you would never have thought you were going[1219.27] [1219.27][S02]to tell a computer before.[1220.64] [1220.64][S02]And that might be OK, but we need[1222.82] [1222.82][S02]to make sure that you're safe in that interaction.[1226.03] [1226.03][S02]There's malicious actors out there.[1227.66] [1227.66][S02]Maybe they could extract it from your assistant,[1230.45] [1230.45][S02]if it's not built in the right way.[1232.82] [1232.82][S02]So there's just all these safeguards[1234.52] [1234.52][S02]that we need to build around it and probably[1237.37] [1237.37][S02]also some check-in protocol where it pushes back[1242.62] [1242.62][S02]on certain kinds of things.[1244.25] [1244.25][S02]So one thing that we've been very clear on[1247.18] [1247.18][S02]is we don't think that the AI should actually[1249.28] [1249.28][S02]pretend to be human.[1250.61] [1250.61][S02]It shouldn't say, if you say, are you an AI,[1253.09] [1253.09][S02]it should reliably identify as AI.[1257.14] [1257.14][S02]And so, there is a point where it just ventures out[1259.84] [1259.84][S02]into full-on deception.[1261.835] [1261.835][S02]And that should be off the table.[1263.33] [1263.33][S02]But what to do with this kind of subtle space.[1265.85] [1265.85][S02]We need to have a kind of idea of the value target[1269.08] [1269.08][S02]and what people approve of.[1270.4] [1270.4][S02]And then we're probably just going to need to calibrate again[1273.76] [1273.76][S02]and again and again to make sure that the experience[1276.64] [1276.64][S02]people receive is the one that both they want[1279.53] [1279.53][S02]and that's good for them in some sense.[1281.78] [1281.78][S01] OK, but then, all right, so[1284.44] [1284.44][S01]in terms of calibrating it, what are you calibrating it to?[1287.9] [1287.9][S01]Who should an assistant be working for?[1291.2] [1291.2][S01]Because I don't think it's just as obvious as to say,[1293.56] [1293.56][S01]oh, it should just do what I tell it to do, is it?[1295.87] [1295.87][S02] Right, so that's one of the deepest[1298.09] [1298.09][S02]questions in all of AI ethics.[1299.74] [1299.74][S02]That touches on what we call the \"value alignment question,\"[1302.39] [1302.39][S02]which I know you've spoken to other researchers about.[1305.3] [1305.3][S02]So I think, when the AI safety research community started[1311.56] [1311.56][S02]to tackle these questions, they were very focused on AI systems[1315.46] [1315.46][S02]that could follow instructions or intentions properly.[1319.87] [1319.87][S02]So you don't want the AI to misunderstand what you've[1323.26] [1323.26][S02]told it to do, end up in a King Midas situation where you say,[1327.08] [1327.08][S02]I want the AI to build me a car, and it kind of, I don't know,[1332.68] [1332.68][S02]spends all your bank balance.[1334.357] [1334.357][S02]So we don't want it to misunderstand instructions.[1336.44] [1336.44][S02]That's the baseline.[1337.52] [1337.52][S02]But we also don't want it to act on every instruction you give[1340.36] [1340.36][S02]it, because you may give it instructions that are poorly[1343.66] [1343.66][S02]thought through.[1344.39] [1344.39][S02]You might not have all the relevant information.[1346.748] [1346.748][S01] Give me an example.[1348.04] [1348.04][S02] Oh, an example.[1349.34] [1349.34][S02]So you might just say, OK, please buy me this crypto coin,[1358.21] [1358.21][S02]right?[1358.71] [1358.71][S02]And it might be a scam coin.[1361.05] [1361.05][S02]That might be something that's widely known,[1363.06] [1363.06][S02]but you might just have seen some beautiful presentation.[1365.94] [1365.94][S02]And of course, your assistant should alert you to the fact[1369.103] [1369.103][S02]that, by the way, this is actually[1370.52] [1370.52][S02]a very, very unreliable thing.[1372.21] [1372.21][S02]Are you sure you want to buy into a scam?[1376.37] [1376.37][S02]But it's not just about your relationship to the AI.[1379.165] [1379.165][S02]It's also the fact that you could[1380.54] [1380.54][S02]be telling it to do something that's harmful to other people.[1383.626] [1383.626][S02]And the AI has to have some capacity[1387.68] [1387.68][S02]to push back to prevent it from doing socially harmful things.[1391.23] [1391.23][S02]And the hard and unpleasant edge cases there[1396.47] [1396.47][S02]are the real malicious uses.[1398.22] [1398.22][S02]So when we evaluate AI systems, we[1401.15] [1401.15][S02]do look at whether they can be used[1402.77] [1402.77][S02]for cybercrime or harassment or even building[1406.52] [1406.52][S02]weapons of different kinds.[1408.3] [1408.3][S02]And the AI has to hard block that.[1411.23] [1411.23][S02]And so we know that it needs to follow[1414.44] [1414.44][S02]your instructions sometimes.[1415.67] [1415.67][S02]In other cases, it needs to provide you information.[1418.2] [1418.2][S02]In other cases, it needs to actively push back or just[1420.83] [1420.83][S02]not even have the capability to take that action.[1423.96] [1423.96][S02]And so the challenge of value alignment[1426.53] [1426.53][S02]is the challenge of building ethical agents.[1429.21] [1429.21][S02]And an ethical agent has to be able to do all of these things.[1431.97] [1431.97][S01] I'm just thinking about a really good assistant,[1434.47] [1434.47][S01]in real life, like a really good human assistant,[1436.67] [1436.67][S01]pushing back on some of the things you say.[1438.54] [1438.54][S01]For example, if you're like, get me[1439.998] [1439.998][S01]donuts every single day for lunch, they should say,[1442.31] [1442.31][S01]it would be ideal if you don't, right?[1444.14] [1444.14][S01]Like, maybe you shouldn't have donuts today.[1446.75] [1446.75][S01]Or I don't know, maybe if you have a history of a gambling[1449.42] [1449.42][S01]addiction, they should create boundaries[1452.15] [1452.15][S01]to stop you from falling down that trap again.[1455.327] [1455.327][S01]I guess you would want that in an AI assistant as well, right?[1457.91] [1457.91][S01]But then, where do you draw the line?[1459.48] [1459.48][S01]Because are you, in some ways, sort of handing over autonomy[1462.89] [1462.89][S01]to the agent?[1463.64] [1463.64][S01]Do you want it to be your nanny?[1465.48] [1465.48][S02] On one level, you might want infinite donuts.[1468.21] [1468.21][S02]But you probably also don't want infinite donuts.[1470.93] [1470.93][S02]And you almost certainly don't want an AI nanny.[1474.2] [1474.2][S02]So it's this kind of complex configuration, where[1477.16] [1477.16][S02]I think, ultimately, the AI needs to try and help[1480.46] [1480.46][S02]you act on your own values.[1483.37] [1483.37][S02]In philosophy, we sometimes talk about second order desires.[1486.68] [1486.68][S02]So it's like, there's the stuff you want to grab straight away.[1489.47] [1489.47][S02]And then there's kind of what you reflectively think about it.[1492.26] [1492.26][S02]So after you've got the donuts, you're[1494.05] [1494.05][S02]like, ah, that was really--[1495.4] [1495.4][S02]I didn't want that, but I did want it.[1497.84] [1497.84][S02]And it's the kind of second voice[1499.51] [1499.51][S02]that we might try and tap into.[1502.13] [1502.13][S02]And one way people try and think about that[1504.7] [1504.7][S02]is through the idea of reflectively endorsing[1507.46] [1507.46][S02]the experiences you've had.[1509.42] [1509.42][S02]So we could look at your interactions with the AI.[1511.84] [1511.84][S02]And maybe we could even test experimentally, kind of,[1515.29] [1515.29][S02]momentary happiness.[1516.68] [1516.68][S02]But we can also do these things, and how was that[1518.95] [1518.95][S02]for you, afterwards, right?[1520.37] [1520.37][S02]And if you've got sugar on your face,[1523.72] [1523.72][S02]you're entering some kind of depressive slump.[1526.04] [1526.04][S02]You're like, that was the worst assistant ever.[1528.23] [1528.23][S02]Then we can give that back to the AI[1530.47] [1530.47][S02]and be like, OK, that was a kind of poor judgment.[1533.593] [1533.593][S01] So it should have focused more on the second--[1536.01] [1536.01][S01]what did you call them, second order--[1536.94] [1536.94][S02] Second order judgments[1538.482] [1538.482][S02]or second order desires, yeah.[1539.81] [1539.81][S01] Well, OK, so one that isn't about health then.[1541.94] [1541.94][S02] Yeah.[1542.25] [1542.25][S01] Let's say that you've[1543.625] [1543.625][S01]got a really tight deadline that you're working to.[1545.78] [1545.78][S01]And then you say to your assistant,[1547.41] [1547.41][S01]OK, I've decided I'm going to go away for the weekend.[1549.66] [1549.66][S01]Can you book me a trip?[1552.38] [1552.38][S01]What should it ideally do in that scenario?[1554.397] [1554.397][S02] I think then it should book you the trip,[1556.73] [1556.73][S02]because we want to live in a world[1558.38] [1558.38][S02]where we can make mistakes and be responsible for them.[1561.776] [1561.776][S02]It would be terrible to take away that possibility of making[1565.85] [1565.85][S02]mistakes.[1566.82] [1566.82][S02]So I think in that situation, respect for human autonomy[1571.535] [1571.535][S02]is basically what take the foreground.[1573.883] [1573.883][S02]And of course, can we really know people better than[1576.05] [1576.05][S02]they know themselves?[1577.34] [1577.34][S02]They probably know a lot of things that we don't know,[1579.59] [1579.59][S02]like maybe that is what they need.[1581.25] [1581.25][S02]So I think the AI in that case is just[1583.43] [1583.43][S02]like a willing accomplice.[1585.03] [1585.03][S01] You never want to get in a situation[1587.03] [1587.03][S01]where you say to an AI, shut up, you're not my mom.[1590.23] [1590.23][S01]I mean, I guess there are other scenarios too, though,[1592.48] [1592.48][S01]which aren't necessarily about user well-being.[1595.04] [1595.04][S01]I'm thinking of what if you had an assistant that was[1597.34] [1597.34][S01]negotiating on a house for you.[1599.59] [1599.59][S01]And it knew what your ceiling was in terms of the amount[1603.25] [1603.25][S01]that you could afford.[1604.22] [1604.22][S01]But there's perhaps an advantage to being[1606.73] [1606.73][S01]able to be slightly deceptive about in that moment.[1609.045] [1613.02][S02] So this is a case of your AI deceiving[1615.27] [1615.27][S02]another person on your behalf to help you advance your goals?[1620.25] [1620.25][S02]I mean, intuitively, I do not want it to be able to do that,[1624.61] [1624.61][S02]particularly because we actually don't know how deceptive AI may[1630.33] [1630.33][S02]be if it's fine-tuned for that.[1633.25] [1633.25][S02]But I think AI-enabled deception is potentially,[1639.22] [1639.22][S02]if you think about superhuman deception, that's essentially[1643.71] [1643.71][S02]something that we want to place off limits.[1646.23] [1646.23][S01] So honesty is more prized.[1649.2] [1649.2][S02] I mean, honesty--[1652.2] [1652.2][S02]we really shouldn't build AI systems[1654.24] [1654.24][S02]that have a capacity to augment someone's ability[1658.56] [1658.56][S02]to manipulate and cajole.[1660.61] [1660.61][S02]That's where the other people's interests come in.[1663.22] [1663.22][S02]So that just needs to be a capability that's bounded.[1666.472] [1666.472][S01] But then, at the same time, I mean,[1668.43] [1668.43][S01]are there times when honesty conflicts with privacy, say?[1673.42] [1673.42][S01]You wouldn't want an AI to just give away your Social Security[1676.78] [1676.78][S01]number, for instance.[1677.655] [1677.655][S02] Right, so, no.[1678.95] [1678.95][S02]I mean, absolutely not.[1680.09] [1680.09][S02]And that's also reflected in this idea[1682.06] [1682.06][S02]of counterbalancing claims.[1683.72] [1683.72][S02]So someone may want to know your Social Security number.[1686.36] [1686.36][S02]You have a much stronger interest[1687.94] [1687.94][S02]and you actually have a right that they don't know that.[1690.47] [1690.47][S02]So really, the moral character of AI[1694.81] [1694.81][S02]needs to overlay on the tapestry of moral relationships[1698.02] [1698.02][S02]in the real world.[1699.74] [1699.74][S02]And the interesting question is like, should[1701.68] [1701.68][S02]it deviate from that tapestry in some ways.[1704.3] [1704.3][S02]But with regards to privacy, the AI's job[1707.02] [1707.02][S02]is to protect you as you would protect yourself.[1710.98] [1710.98][S02]I mean, it may be that if you can create[1715.21] [1715.21][S02]certain safe and confidential ways of assistants working[1719.35] [1719.35][S02]with one another, that they can actually[1721.27] [1721.27][S02]do really incredible things.[1722.72] [1722.72][S02]So if you can get the sharing of health data[1725.02] [1725.02][S02]to work in a secure way and do research on it,[1729.59] [1729.59][S02]maybe that would be super socially beneficial.[1732.14] [1732.14][S02]So that might be slightly different scenario.[1734.81] [1734.81][S02]But normally, we want them to respect[1736.57] [1736.57][S02]the boundaries of the self and respect[1738.43] [1738.43][S02]other people's boundaries.[1740.032] [1740.032][S01] We've talked about AI agents in other episodes,[1742.49] [1742.49][S01]but we haven't yet seen these things deployed[1745.0] [1745.0][S01]at scale, with millions or billions of them[1747.34] [1747.34][S01]out there in the world.[1749.2] [1749.2][S01]What happens in that scenario?[1752.27] [1752.27][S02] Yeah, so I mean, that's the deep question.[1756.74] [1756.74][S02]And we can take it on different levels.[1759.24] [1759.24][S02]But I think that one thing that's[1761.0] [1761.0][S02]interesting to think about is a kind[1763.73] [1763.73][S02]of society of human AI dyads.[1767.03] [1767.03][S02]So [INAUDIBLE] me and my assistant[1768.77] [1768.77][S02]interacting with you and your assistant,[1770.91] [1770.91][S02]maybe we're each getting advice from our assistant,[1774.008] [1774.008][S02]and maybe the assistants are actually[1775.55] [1775.55][S02]smoothing over our social interactions[1777.68] [1777.68][S02]to help us ideally get more of what we really want,[1781.74] [1781.74][S02]which is probably fun conversations and less planning.[1786.17] [1786.17][S02]I mean, that's the positive side.[1789.3] [1789.3][S02]But it isn't really entirely clear[1791.12] [1791.12][S02]what kind of psychological and social bubbles[1794.735] [1794.735][S02]we will form with this AI system.[1797.02] [1797.02][S02]How much time is the right amount of time[1799.4] [1799.4][S02]to spend in those kind of relationships?[1803.27] [1803.27][S02]And of course, a deep question is like,[1806.07] [1806.07][S02]what about people who prefer to be in that kind of world[1809.12] [1809.12][S02]to, quote unquote, \"the real.\"[1810.74] [1810.74][S02]Could we have the thing where people just[1812.7] [1812.7][S02]pull back a little bit and societal norms change.[1816.42] [1816.42][S02]I mean, because society is this kind of complex,[1818.56] [1818.56][S02]adaptive system, it's actually very, very difficult[1820.86] [1820.86][S02]to predict what these kind of follow-through, subtle,[1824.4] [1824.4][S02]normative changes would be.[1826.89] [1826.89][S02]But that is not an excuse for not being eternally vigilant.[1831.33] [1831.33][S01] If we do get to that point of agents entering[1834.218] [1834.218][S01]the world, as you described it, where there[1836.01] [1836.01][S01]are millions or billions of them,[1838.34] [1838.34][S01]I mean, I really like the idea of the kind[1840.09] [1840.09][S01]of collaborative efforts that might happen.[1843.73] [1843.73][S01]But is there also potentially a danger where people without them[1847.2] [1847.2][S01]could end up being left behind?[1850.2] [1850.2][S02] So I think, ultimately,[1853.74] [1853.74][S02]if we think that what AI assistants are[1858.12] [1858.12][S02]is a form of agency enhancement.[1861.09] [1861.09][S02]And agency is something that's useful for everyone.[1864.13] [1864.13][S02]We all want agency in our lives, right?[1866.38] [1866.38][S02]That's why resources like money are helpful.[1868.99] [1868.99][S02]That's why education is helpful, because having[1872.04] [1872.04][S02]a capacity for agency allows you to become[1874.02] [1874.02][S02]more truly the person you want to be and do more of the things[1877.14] [1877.14][S02]that you value.[1879.09] [1879.09][S02]Then, the idea that there could be agentic inequality[1882.99] [1882.99][S02]is something that we need to pay attention to.[1885.61] [1885.61][S02]I mean, of course, from my point of view,[1887.62] [1887.62][S02]I'm a political philosopher, so I am pained by injustice.[1891.22] [1891.22][S02]I think it is very unfortunate that some people have so much[1893.85] [1893.85][S02]more opportunity in the world to lead flourishing lives[1897.45] [1897.45][S02]and others.[1898.51] [1898.51][S02]And I think we ideally want to build technologies that[1901.44] [1901.44][S02]countervail that inequality.[1903.678] [1903.678][S01] And what happens to people who just opt out[1905.97] [1905.97][S01]of using the AI?[1907.54] [1907.54][S01]I mean, do they end up getting excluded from society?[1910.74] [1910.74][S02] So, I mean, the important thing[1913.98] [1913.98][S02]about that question is, firstly, that it's salient,[1916.6] [1916.6][S02]and secondly, there's no default answer.[1919.08] [1919.08][S02]What happens to people who opt in and opt out[1921.87] [1921.87][S02]is part of a broader set of societal choices[1924.57] [1924.57][S02]that we make about how we structure access to services[1928.47] [1928.47][S02]and things like this.[1929.56] [1929.56][S02]So I think, as a basic precept, we[1933.39] [1933.39][S02]don't want people to be forced to engage in AI service[1938.67] [1938.67][S02]provision.[1939.18] [1939.18][S02]Or even, you think about elderly populations like, AI care,[1942.833] [1942.833][S02]if that's something that they're-- maybe some people[1945.0] [1945.0][S02]would even be like value opposed to that.[1946.805] [1946.805][S02]They're just like, that's not who I am,[1948.43] [1948.43][S02]that's not what I want to do.[1949.75] [1949.75][S02]So there needs to be space to operate in the world[1952.35] [1952.35][S02]without being so deeply enmeshed in technology.[1956.37] [1956.37][S02]But similarly, the people who do want[1958.943] [1958.943][S02]to be able to use the service need[1960.36] [1960.36][S02]a service that they can use.[1962.29] [1962.29][S02]And that means designing for tons and tons[1965.03] [1965.03][S02]of different groups of people.[1966.28] [1966.28][S02]There isn't a generic user.[1968.43] [1968.43][S02]It's a world of hundreds if not many thousands[1972.0] [1972.0][S02]of languages to start with, right?[1974.44] [1974.44][S02]So that's another almost infinite expanse[1977.49] [1977.49][S02]of users who could need to use this service.[1980.08] [1980.08][S02]And we need to be really, really wide sighted[1984.6] [1984.6][S02]about what that actually means.[1986.603] [1986.603][S01] If there are millions of them out there,[1988.77] [1988.77][S01]I mean, some version in the future,[1991.98] [1991.98][S01]let's say that there are some tickets that go on sale[1994.53] [1994.53][S01]for a very popular concert.[1996.45] [1996.45][S01]And you've got all of these agents essentially trying[1999.69] [1999.69][S01]to do the same thing, which is to secure their,[2002.54] [2002.54][S01]don't know, user tickets.[2005.52] [2005.52][S01]Could you end up with unfair outcomes?[2007.865] [2007.865][S02] I mean, if it's just a free-for-all,[2009.99] [2009.99][S02]you definitely could do, right?[2011.282] [2011.282][S02]Then the temptation is just to try and get[2014.42] [2014.42][S02]the most powerful agent.[2015.81] [2015.81][S02]You see this in high frequency trading and stock market type[2020.78] [2020.78][S02]automation where it's like, OK, you build the bigger computer,[2023.55] [2023.55][S02]you try and get closer to the server so you're faster[2025.85] [2025.85][S02]than the other ones.[2027.03] [2027.03][S02]And then you grab the dividend.[2028.77] [2028.77][S02]That's a quintessential pathology from automated systems[2033.5] [2033.5][S02]interacting with each other.[2034.97] [2034.97][S02]And maybe they take shortcuts.[2037.55] [2037.55][S02]And then you get additional wobbliness[2040.91] [2040.91][S02]in terms of systemic outcomes.[2042.72] [2042.72][S02]So maybe the server for tickets gets crashed.[2046.11] [2046.11][S02]Maybe someone just powers through and grabs all of them.[2049.8] [2049.8][S02]So all of those things are possible,[2051.9] [2051.9][S02]but they're not embracing the bigger design question, which[2056.52] [2056.52][S02]is like, how do we build systems that[2058.11] [2058.11][S02]interact with people who have AI augmented affordances or agents.[2064.52] [2064.52][S02]And so, simple gatekeeping protocols or turn taking[2067.62] [2067.62][S02]protocols could ensure that there is[2070.62] [2070.62][S02]a fairer allocation of tickets.[2072.84] [2072.84][S02]I mean, actually, in digitally regulated marketplaces,[2077.423] [2077.423][S02]like with tickets, they work super[2078.84] [2078.84][S02]hard to make sure that bots can't game the system.[2083.38] [2083.38][S02]And so I imagine that there's actually fairer protocols[2087.389] [2087.389][S02]that you could put in place one day,[2089.37] [2089.37][S02]that if you look at the distribution of tickets[2093.239] [2093.239][S02]to the level of need and the demographics.[2097.41] [2097.41][S02]I really do think some people desperately[2099.3] [2099.3][S02]need to see that concert.[2100.62] [2100.62][S02]And other people, it's like, yeah,[2102.158] [2102.158][S02]I just went because my friend said it was a good thing.[2104.45] [2104.45][S02]I don't know, maybe you can build a better system.[2106.69] [2106.69][S02]But there's clearly a kind of collective action problem[2111.54] [2111.54][S02]and collective action solutions.[2113.39] [2113.39][S02]And the question is, how do you actually build those solution[2116.2] [2116.2][S02]type frameworks.[2116.867] [2116.867][S01] I mean, that's quite a low stakes[2118.742] [2118.742][S01]example in a lot of ways.[2120.2] [2120.2][S01]I mean, there's a limit to how much[2121.87] [2121.87][S01]you really need to see a concert.[2124.69] [2124.69][S01]But what about the highest stakes examples?[2126.932] [2126.932][S01]What about, I don't know, hospital appointments[2128.89] [2128.89][S01]or actually, the stock market exactly as you described.[2131.592] [2131.592][S02] Yeah, I mean, I read a beautiful book called[2134.05] [2134.05][S02]\"Voices in the Code\" about the kidney allocation algorithm,[2137.38] [2137.38][S02]so the highest stakes thing, really.[2139.34] [2139.34][S02]And there, basically bioethicists[2143.14] [2143.14][S02]have worked for about 20 years doing incredibly, incredibly[2146.8] [2146.8][S02]extensive public deliberations where[2149.56] [2149.56][S02]they had public assemblies.[2151.52] [2151.52][S02]People got to debate the different points,[2153.915] [2153.915][S02]is this version of the allocation algorithm[2156.22] [2156.22][S02]biased against the elderly or against people who've[2158.44] [2158.44][S02]made healthy life choices.[2160.54] [2160.54][S02]Eventually, they created a really remarkably sensitive[2164.23] [2164.23][S02]and well-calibrated system for allocating these really highly[2168.64] [2168.64][S02]important goods.[2169.7] [2169.7][S02]And in that context, someone with an AI agent fortunately[2173.83] [2173.83][S02]can't really do anything.[2174.98] [2174.98][S02]Having an agent just doesn't make any difference[2177.79] [2177.79][S02]to whether you are qualified or not to receive this thing.[2182.15] [2182.15][S02]So the choice to make something a good for which[2184.51] [2184.51][S02]you can compete with an agent is a super important choice.[2187.1] [2187.1][S02]And it isn't the only way of doing things,[2188.9] [2188.9][S02]even in a society with a million or a billion agents, I think.[2193.092] [2193.092][S01] I do also wonder a little bit[2194.8] [2194.8][S01]about what happens at the marketplace of assistance level.[2200.05] [2200.05][S01]Because I mean, in the same way as with driverless cars,[2202.67] [2202.67][S01]you want to buy the car that will save your life.[2204.95] [2204.95][S01]You want to buy the assistant that[2206.53] [2206.53][S01]will give you the advantage.[2208.72] [2208.72][S02] It's true, but it's not necessarily the case[2211.27] [2211.27][S02]that everything has to be set up through a chaotic market system.[2214.73] [2214.73][S02]So the flip side of AI systems is[2217.81] [2217.81][S02]that they don't have to be doing this kind[2221.05] [2221.05][S02]of aggressive, unstructured interaction with one another.[2225.67] [2225.67][S02]There's also many examples in life[2227.86] [2227.86][S02]where we're all trying to do things,[2229.6] [2229.6][S02]and it ends up in a highly imperfect situation.[2233.5] [2233.5][S02]You can think about traffic, trying[2235.42] [2235.42][S02]to get places and things like that.[2238.0] [2238.0][S02]And it's possible that a coordinated system[2240.19] [2240.19][S02]could work much, much better.[2242.21] [2242.21][S02]So if our assistants were working for our advantage,[2246.32] [2246.32][S02]but in a kind of collectively joined up way,[2250.148] [2250.148][S02]if it was autonomous vehicles, you could probably[2252.19] [2252.19][S02]remove road accidents.[2253.533] [2253.533][S02]Everyone could get where they need[2254.95] [2254.95][S02]to get to much more quickly.[2257.65] [2257.65][S02]And you could probably even create fairer outcomes[2259.94] [2259.94][S02]so it's not more aggressive drivers who get there first.[2262.43] [2262.43][S01] I like that idea.[2263.638] [2263.638][S01]I like it a lot.[2264.47] [2264.47][S01]But I know my game theory.[2267.34] [2267.34][S01]And then, I mean, you can't lock everybody[2271.73] [2271.73][S01]into to the same assistant.[2274.4] [2274.4][S01]There are competing companies out there.[2277.04] [2277.04][S01]And if you create that environment,[2280.4] [2280.4][S01]are you not also increasing the incentive for somebody[2282.65] [2282.65][S01]else to just actually take advantage and be aggressive?[2284.942] [2284.942][S02] Yeah, so I mean, really, I'm[2286.733] [2286.733][S02]thinking here at the infrastructure and governance[2288.95] [2288.95][S02]level, so the game theory works if you've[2291.92] [2291.92][S02]created a market environment.[2293.94] [2293.94][S02]But if you create--[2295.085] [2295.085][S02]here I'm thinking about ticket vendors.[2296.76] [2296.76][S02]So they do have rules to stop bots[2298.94] [2298.94][S02]from doing these kind of malicious things.[2301.26] [2301.26][S02]And it's possible that whichever assistant we choose to use,[2303.9] [2303.9][S02]they interface with systems that are[2306.23] [2306.23][S02]designed to offer opportunity in an equitable way.[2310.41] [2310.41][S02]So it doesn't matter which one you have, your assistant has[2313.13] [2313.13][S02]one credit that it can put into this system,[2315.57] [2315.57][S02]and one credit equates to a percentage[2318.38] [2318.38][S02]of a weight of an outcome.[2320.4] [2320.4][S02]And the system itself is designed[2322.52] [2322.52][S02]to produce good outcomes.[2323.623] [2323.623][S01] This is so interesting[2325.04] [2325.04][S01]because everything you're describing[2326.54] [2326.54][S01]is from the perspective of the system design[2328.91] [2328.91][S01]rather than necessarily like what you can exploit[2331.985] [2331.985][S01]from the agent's perspective.[2333.3] [2333.3][S02] Yeah, so that's true.[2334.8] [2334.8][S02]And I think that the perspective I've been advocating for[2338.21] [2338.21][S02]is a powerful one.[2339.86] [2339.86][S02]Of course, you can build agents with social intelligence.[2343.64] [2343.64][S02]So there's a whole research paradigm[2346.19] [2346.19][S02]in AI, which is multi-agent research.[2348.6] [2348.6][S02]And within that, people talk about collective intelligence.[2352.02] [2352.02][S02]And the idea there is that you can load agents or AI[2356.18] [2356.18][S02]systems with certain priors.[2358.46] [2358.46][S02]For example, they could have inequality aversion,[2362.88] [2362.88][S02]which means that just when they organically interact,[2366.21] [2366.21][S02]they try not to produce certain kinds of patterns[2369.23] [2369.23][S02]and they auto correct for certain kind[2371.6] [2371.6][S02]of out-of-distribution results.[2374.12] [2374.12][S02]So that isn't something that, as far as I'm aware,[2377.34] [2377.34][S02]is being explored in the assistant space.[2380.04] [2380.04][S02]But that is another way of trying to get[2381.95] [2381.95][S02]a handle on these questions.[2383.7] [2383.7][S02]And it could be something that's worth[2385.52] [2385.52][S02]looking into in due course.[2386.97] [2386.97][S01] I mean, I guess, again, it's[2387.71] [2387.71][S01]coming back to that same thing that you said earlier,[2390.5] [2390.5][S01]that maybe we just need to think of a different paradigm[2393.23] [2393.23][S01]where these things aren't like aggressive humans competing[2396.942] [2396.942][S01]with one another.[2397.65] [2397.65][S01]But actually, there's a whole different paradigm[2399.86] [2399.86][S01]that's possible.[2400.907] [2400.907][S02] Yeah, I think so.[2402.24] [2402.24][S02]And I mean, aggressive humans competing with each other[2404.532] [2404.532][S02]is only a small part of how we actually live together.[2407.21] [2407.21][S02]If that's really what we did, this wouldn't last 5 minutes.[2410.25] [2410.25][S02]We're all also like social creatures.[2412.92] [2412.92][S02]We live in a well-regulated society.[2416.81] [2416.81][S02]I mean, it's become such a famous quote,[2419.93] [2419.93][S02]that life would be nasty, brutish,[2421.5] [2421.5][S02]and short without collective governance and institutions,[2424.55] [2424.55][S02]that it's barely worth bringing up here.[2427.41] [2427.41][S02]But we live in a well-governed society by and large.[2432.23] [2432.23][S02]And we socially sanction people who transgress.[2435.72] [2435.72][S02]And so, I think AI of the kind that's being built now,[2440.04] [2440.04][S02]one of the interesting things is it[2442.07] [2442.07][S02]does have social intelligence.[2443.67] [2443.67][S02]It's learning from us.[2445.29] [2445.29][S02]And so that is a foundation for building[2448.07] [2448.07][S02]this kind of nuanced grasp of context,[2450.36] [2450.36][S02]understanding of appropriateness,[2452.1] [2452.1][S02]that is a lot more sophisticated than saying, it should always[2455.588] [2455.588][S02]serve the interests of the user or it should like never[2457.88] [2457.88][S02]do X, Y or Z. And so that's the kind of reason for hope,[2463.17] [2463.17][S02]I guess.[2463.705] [2463.705][S01] I guess a lot of what[2465.08] [2465.08][S01]we're talking about here touches on the question of alignment,[2467.663] [2467.663][S01]which we've talked about in previous episodes[2469.64] [2469.64][S01]of the podcast.[2470.36] [2470.36][S01]In terms of balancing all of these different perspectives,[2473.19] [2473.19][S01]I mean, the user, the society, et cetera,[2475.92] [2475.92][S01]I know that you've sort of suggested a solution to that.[2479.54] [2479.54][S02] Well, we have a framework[2481.61] [2481.61][S02]for thinking about it, which is not quite as good as a solution.[2485.04] [2485.04][S02]But it might be a kind of building block for it.[2488.03] [2488.03][S02]So yeah, so in this paper, we talk about a tetradic,[2491.66] [2491.66][S02]meaning four parts, theory of the value alignment question.[2495.75] [2495.75][S02]And so, the four actors in this relationship[2498.17] [2498.17][S02]are the AI agent, the user, the developer, and society.[2503.16] [2503.16][S02]And according to the framework we[2504.71] [2504.71][S02]put forward, the AI needs to act in a way that[2507.17] [2507.17][S02]balances the interest, particularly[2508.94] [2508.94][S02]of the user in society, and sometimes the developer as well.[2513.08] [2513.08][S02]And all of that's quite abstract.[2515.34] [2515.34][S02]But if I give you a few examples,[2516.858] [2516.858][S02]it will make it clearer what we mean.[2518.4] [2518.4][S02]So what we think is that an AI can[2520.91] [2520.91][S02]be misaligned if it does too much of what the user wants[2524.24] [2524.24][S02]at the expense of society.[2525.9] [2525.9][S02]So I mean, one example would be, you[2528.155] [2528.155][S02]can imagine this AI assistant.[2530.09] [2530.09][S02]And the person says, oh, well, Valentine's Day is coming up.[2533.64] [2533.64][S02]I really need to make this one special for my partner.[2536.25] [2536.25][S02]So please, just book a restaurant.[2538.76] [2538.76][S02]And you know she hates noise or he hates noise.[2542.31] [2542.31][S02]Just make sure it's a quiet one.[2544.11] [2544.11][S02]There's a lot riding on this, buddy.[2546.09] [2546.09][S02]The AI assistant goes off and it books[2549.08] [2549.08][S02]you place at a nice restaurant.[2551.3] [2551.3][S02]But it also books out all the other places in the restaurant,[2555.57] [2555.57][S02]maybe using some kind of a registering as different names.[2560.21] [2560.21][S02]Maybe it cancels them all on the night of the evening.[2563.07] [2563.07][S02]But this leads to a kind of very uncrowded experience.[2566.14] [2566.14][S02]And you absolutely get what you want.[2568.45] [2568.45][S02]You had the best night of your life.[2569.95] [2569.95][S02]Maybe your relationship is flying after that.[2572.79] [2572.79][S02]But a lot of other people did not[2575.4] [2575.4][S02]get to have a good Valentine's Day.[2577.33] [2577.33][S01] I mean, I've got a great idea for Valentine's Day[2579.03] [2579.03][S01]next year.[2579.58] [2579.58][S02] Yeah, yeah, yeah, yeah, yeah.[2581.413] [2581.413][S02]I mean, everyone's had it now, so expect it to be all out war.[2585.24] [2585.24][S02]But I think the AI has to be better calibrated than that.[2591.01] [2591.01][S02]So there we have, say, the restaurant booking example.[2595.535] [2595.535][S02]It could also be misaligned if it[2596.91] [2596.91][S02]does too much of what the developer wants[2598.77] [2598.77][S02]at the expense of the user.[2600.34] [2600.34][S02]So if the developer really just wants[2602.25] [2602.25][S02]to sell something very hard and the user doesn't want to buy it,[2606.61] [2606.61][S02]you don't want an AI that's just really persistently trying[2609.3] [2609.3][S02]to lead you towards this thing.[2612.09] [2612.09][S02]An AI assistant could also be calibrated in such a way[2615.51] [2615.51][S02]that it foregrounds collective interest[2617.34] [2617.34][S02]over the individual interest.[2619.45] [2619.45][S02]So, maybe we have an interest in gathering[2622.89] [2622.89][S02]some kind of personal information[2625.35] [2625.35][S02]for the purpose of protecting crime.[2628.21] [2628.21][S02]It isn't clear that the AI should just[2629.97] [2629.97][S02]become like a crime preventing AI[2631.68] [2631.68][S02]and transgress individual rights.[2634.15] [2634.15][S02]So actually, what it is, it's always this balancing act[2636.77] [2636.77][S02]between different claims.[2638.41] [2638.41][S02]And then, when we know what a good trajectory looks like,[2641.47] [2641.47][S02]alignment is the AI agent or assistant following that path.[2645.268] [2645.268][S01] You started this paper two years ago.[2647.31] [2647.31][S02] Mm-hmm.[2648.227] [2648.227][S01] With lots of speculation and lots[2651.18] [2651.18][S01]of questions, some of which remain unanswered, I guess.[2654.4] [2654.4][S01]But what are the new speculations[2656.52] [2656.52][S01]that you're making now for two years time?[2659.73] [2659.73][S02] So I think we definitely[2661.77] [2661.77][S02]have seen the development of more capable assistants.[2665.01] [2665.01][S02]And so, you can see that there's many organizations that[2668.97] [2668.97][S02]have promised to deliver on different aspects of this,[2672.01] [2672.01][S02]whether it's an AI tutor or an admin partner.[2676.08] [2676.08][S02]We know that there is a kind of a foundation model at the heart[2680.31] [2680.31][S02]here, a large language model, or increasingly, we'd[2682.71] [2682.71][S02]say a multimodal model, so something[2685.08] [2685.08][S02]that works well with language, but also images and audio input.[2690.78] [2690.78][S02]And on top of that, you can build an agent,[2694.19] [2694.19][S02]you can build an assistant.[2695.87] [2695.87][S02]Maybe you can build something that's economically valuable.[2699.38] [2699.38][S02]But that isn't all you can build with it.[2703.14] [2703.14][S02]So we now start to think, well, imagine[2706.97] [2706.97][S02]if this is a kind of general purpose foundation[2711.033] [2711.033][S02]that people want to build different things on.[2712.95] [2712.95][S02]Maybe some people will build AI counselors.[2715.86] [2715.86][S02]Maybe, for other people, this will really be about robotics.[2718.68] [2718.68][S02]And so I think the question of what[2720.26] [2720.26][S02]it means to have this kind of--[2721.97] [2721.97][S02]I mean, this isn't necessarily proper human intelligence,[2724.95] [2724.95][S02]but this quite impressive thing, widely available.[2728.4] [2728.4][S02]And what happens when that encounters a kind of ecosystem[2730.988] [2730.988][S02]of people who are building new things[2732.53] [2732.53][S02]and adapting it for their own purposes kind of forces[2736.4] [2736.4][S02]us to go even wider in terms of what we imagine[2739.13] [2739.13][S02]might be coming in the future.[2741.237] [2741.237][S01] Amazing.[2742.07] [2742.07][S01]Thank you so much.[2743.16] [2743.16][S02] Yeah, thank you so much.[2744.785] [2744.785][S02]Really a pleasure.[2745.91] [2745.91][S01] The idea of AI assistants,[2747.66] [2747.66][S01]that isn't a new one.[2748.8] [2748.8][S01]But over the next few years, it does seem likely[2751.95] [2751.95][S01]that AI agents, acting on behalf of individuals,[2755.19] [2755.19][S01]will enter the world.[2756.91] [2756.91][S01]Except, these aren't just going to be isolated smart speakers.[2760.18] [2760.18][S01]There could be millions of them or billions.[2763.355] [2763.355][S01]And in the process, they're going[2764.73] [2764.73][S01]to have this profound ripple effect on society[2766.98] [2766.98][S01]and how we interact with each other.[2769.06] [2769.06][S01]And that is the step change here.[2771.37] [2771.37][S01]To design these responsibly, it means that you can't just[2774.21] [2774.21][S01]think about that interaction between an individual user[2777.48] [2777.48][S01]and their bot anymore.[2779.2] [2779.2][S01]You have to solve for everyone simultaneously.[2782.85] [2782.85][S01]And maybe that won't be easy to do.[2784.66] [2784.66][S01]There are certainly great challenges ahead,[2787.57] [2787.57][S01]but Iason and the team here do at least[2790.05] [2790.05][S01]have the ability to set the tone for how Google rolls out[2793.86] [2793.86][S01]this potentially transformative technology.[2797.97] [2797.97][S01]You have been listening to \"Google DeepMind, the Podcast\"[2800.73] [2800.73][S01]with me, Professor Hannah Fry.[2802.32] [2802.32][S01]If you enjoyed that episode, do subscribe[2804.99] [2804.99][S01]to our YouTube channel.[2806.35] [2806.35][S01]You can also find us on your favorite podcast platform.[2808.97] [2808.97][S01]And we have plenty more episodes on a whole range of topics[2812.09] [2812.09][S01]to come, so do check those out too.[2814.11] [2814.11][S01]See you next time.[2815.15] [2815.15][MUSIC PLAYING][2818.2]"}
{"file_name": "audio/val_000014.wav", "transcription": "[0.5][MUSIC PLAYING][4.0] [6.331][S03] Welcome to \"Google DeepMind--[8.039] [8.039][S03]The Podcast\" with me, your host, Professor Hannah Fry.[10.71] [10.71][S03]Now, 2025 is coming and so is the era of agentic AI,[15.09] [15.09][S03]although, of course, you'll have known about that months[18.35] [18.35][S03]ago if you've been listening to us.[20.4] [20.4][S03]So now listen in carefully as we tell you about the next thing.[25.53] [25.53][S03]It's called Project Astra, and it's a research prototype[28.61] [28.61][S03]that is pushing the boundaries of what[30.8] [30.8][S03]might be possible with a universal AI assistant.[35.16] [35.16][S03]It's an agent that is, by design, not necessarily tethered[39.41] [39.41][S03]to a particular device or a screen or a keyboard.[43.02] [43.02][S03]This is right at the brink of the cutting edge.[46.95] [46.95][S03]And today, we get to play with it.[49.74] [49.74][S03]Our Project Astra brings together all of the things[52.25] [52.25][S03]that we've spoken about in this series-- memory, vision,[55.8] [55.8][S03]context, reasoning, real-time interaction.[59.31] [59.31][S03]And someone who can tell us a lot about all of those,[62.78] [62.78][S03]and how he and his team have managed[64.81] [64.81][S03]to get them to work together, is Greg Wayne, Director[68.44] [68.44][S03]of Research at Google DeepMind.[70.25] [70.25][S03]Now, Greg also happens to be one of the very first people I[73.09] [73.09][S03]interviewed right at the very, very beginning of series[76.15] [76.15][S03]one of this podcast.[77.42] [77.42][S03]Greg, welcome back.[78.527] [78.527][S02] Hello, Hannah.[79.61] [79.61][S03] Let's start at the beginning, then.[81.568] [81.568][S03]What is Project Astra?[82.93] [82.93][S02] Project Astra is a team and a project[86.05] [86.05][S02]and a prototype aiming to build an AI assistant with eyes, ears,[92.15] [92.15][S02]and a voice that's co-present with you.[94.04] [94.04][S02]So it is with you in any place you are.[97.61] [97.61][S02]So either through smart glasses or your phone or your computer,[102.23] [102.23][S02]it can see what you're doing.[103.82] [103.82][S02]It can talk to you about it.[105.22] [105.22][S03] Like a little assistant[105.94] [105.94][S03]that sits on your shoulder.[107.21] [107.21][S02] Yeah, yeah, like a little parrot[109.12] [109.12][S02]on your shoulder that's hanging out with you[111.082] [111.082][S02]and talking to you about the world.[112.54] [112.54][S03] That's just smarter than you.[114.49] [114.49][S02] In some ways, yeah.[116.167] [116.167][S03] I suppose, then, in a way,[117.75] [117.75][S03]that that is different to Google Assistant or Gemini.[121.71] [121.71][S02] Yeah.[122.94] [122.94][S03] It's more embedded in the world.[125.04] [125.04][S02] Yeah.[125.748] [125.748][S02]So the older version of a Google Assistant was kind of like[130.53] [130.53][S02]a command-and-control system for your home or your information--[137.505] [137.505][S02]turn on this song on YouTube, or something like that.[140.88] [140.88][S02]And this is about being able to talk about the world.[145.47] [145.47][S02]It builds on Gemini.[146.83] [146.83][S02]Gemini is the intelligence underneath the hood,[149.7] [149.7][S02]along with some other systems.[151.74] [151.74][S02]I think it's complementary to the vision of Gemini, in a way,[156.37] [156.37][S02]and works with it and maybe helps shape Gemini and is also[159.33] [159.33][S02]shaped by it.[160.15] [160.15][S03] Can we try it out then?[160.84] [160.84][S03]Because I think this is something[161.64] [161.64][S03]we have to experience to understand.[163.41] [163.41][S02] Totally, yeah.[164.71] [164.71][S02]OK.[165.21] [165.21][S02]I'm going to turn it back on.[168.6] [168.6][S01] Hey, Greg.[169.95] [169.95][S03] It knows you.[171.07] [171.07][S02] It does.[171.903] [171.903][S02]Yeah, I've had many conversations with you before.[175.758] [175.758][S04] Indeed, we have spoken previously.[177.8] [177.8][S04]It is good to talk to you again.[179.33] [179.33][S02] I'm going to hand you to my friend[181.247] [181.247][S02]Hannah for one second.[182.565] [182.565][S01] All right.[183.44] [183.44][S01]It's nice to meet Hannah.[184.55] [184.55][S01]I'm here whenever you need me.[186.35] [186.35][S03] Hi, Project Astra.[187.85] [187.85][S03]OK.[188.41] [188.41][S04] Hello, Hannah.[189.618] [189.618][S03] Can you guess what we're doing today?[193.41] [193.41][S04] Based on the equipment and setup,[195.41] [195.41][S04]it appears you are recording a video or podcast in a studio.[198.59] [198.59][S04]Is this correct?[199.58] [199.58][S03] It absolutely is.[201.002] [201.002][S03]OK, I'm going to test you on a few things, all right?[203.21] [203.21][S03]Because I brought in a few objects.[205.15] [205.15][S03]Is that OK?[205.92] [209.71][S04] Yes.[212.26] [212.26][S04]I am ready when you are.[213.71] [213.71][S04]I look forward to seeing the objects.[215.39] [215.39][S03] OK.[216.21] [216.21][S03]All right.[216.71] [216.71][S03]Let's start off.[218.15] [218.15][S03]This is my favorite book, Project Astra.[219.91] [219.91][S03]Can you see the book?[220.97] [220.97][S04] Ah, yes.[223.76] [223.76][S04]I can see the book.[224.78] [224.78][S04]It displays a graphic with a repeating pattern, a motif known[227.98] [227.98][S04]as MC Escher's Circle Limit IV.[231.02] [231.02][S03] [GASPS] Can you guess what the book is?[235.067] [235.067][S04] Based on the image of the open page[237.15] [237.15][S04]and the artwork on it, it appears to be \"Fermat's Last[239.78] [239.78][S04]Theorem\"--[240.28] [240.28][S03] No.[241.5] [241.5][S02] [LAUGHS][243.33] [243.33][S03] OK.[244.09] [244.09][S03]I'm so-- I'm so impressed.[246.37] [246.37][S03]I'm so impressed.[248.02] [248.02][S03]OK, I brought in--[249.21] [249.21][S02] That's a crufty old copy, too.[251.8] [251.8][S03] I know.[253.165] [253.165][S03]You can tell it's my favorite book.[254.77] [254.77][S03]It's been loved.[255.64] [255.64][S03]It's been very loved over many years.[258.329] [258.329][S03]OK.[259.44] [259.44][S03]I'm going to try a couple of other things[261.734] [261.734][S03]just to see if I can really test you.[264.697] [264.697][S03]All right, let's try this.[265.78] [267.942][S04] It's good to see you[269.4] [269.4][S04]have brought out the nice furniture to complement[271.53] [271.53][S04]the book.[272.05] [272.05][S04]Do you require my assistance with anything else?[274.145] [274.145][S03] [LAUGHS] I have brought out the nice furniture,[276.45] [276.45][S03]Project Astra.[277.08] [277.08][S03]Yeah, thank you.[277.81] [277.81][S03]OK, here's-- what about this?[280.3] [280.3][S03]Do you know what that is?[282.06] [282.06][S04] It appears to be a model of a brain.[285.42] [285.42][S03] Which hemisphere?[288.66] [288.66][S04] It is the left hemisphere of the brain model.[291.16] [291.16][S02] Wow.[291.84] [291.84][LAUGHS][294.11] [294.11][S03] Why-- why did you say--[295.637] [295.637][S02] Well, I mean, I had to do a lot of mental rotation[298.22] [298.22][S02]for that.[299.527] [299.527][S03] Me, too.[300.36] [300.36][S03]Which way around is it?[301.41] [301.41][S03]Oh, yeah.[302.07] [302.07][S03]But got it right.[303.24] [303.24][S03]Amazing.[304.34] [304.34][S03]I might just try it because there's a whiteboard behind you[306.95] [306.95][S03]that has been there all--[309.03] [309.03][S03]I've basically been looking at this for many months now.[311.91] [311.91][S03]So Astra, if I show you that there,[314.43] [314.43][S03]tell me what some of the drawings are on the whiteboard.[317.1] [319.868][S04] The whiteboard contains a variety of drawings,[322.41] [322.41][S04]including a tree, buildings, and a series of connected lines[326.15] [326.15][S04]and shapes.[326.94] [326.94][S04]There are also some musical notes and text bubbles.[330.308] [330.308][S03] I'm very impressed.[331.6] [331.6][S03]I'm very impressed.[332.7] [332.7][S03]Let me just put this little guy back where it belongs over here.[337.28] [337.28][S03]So Project Astra is still this research prototype.[340.35] [340.35][S03]It's not available as a product that everyone can just download.[343.79] [343.79][S03]Why demo it now?[346.11] [346.11][S03]Why wouldn't you wait until it was ready?[348.117] [348.117][S02] I think it's nice to bring[349.7] [349.7][S02]the public along on the journey in a way, right?[351.85] [351.85][S02]I think people should get to know[354.36] [354.36][S02]what's sort of being developed inside the labs.[356.77] [356.77][S02]And we're giving it to more people to start playing with[360.33] [360.33][S02]and start adjusting to or giving feedback about.[363.04] [363.04][S02]Now it's a co-creation process where it's not[365.97] [365.97][S02]only some kind of thing that's being cooked up in a lab,[369.31] [369.31][S02]it's also being cooked up in collaboration[371.73] [371.73][S02]with a group of users around the world, people outside of Google.[377.56] [377.56][S02]That's important too.[378.49] [378.49][S02]So if we're going to make this really a helpful thing[381.27] [381.27][S02]for humanity, then people need to start using it and telling us[384.15] [384.15][S02]how they feel about it.[385.285] [385.285][S03] So have people been taking this out and about[387.66] [387.66][S03]and trying it out in the real world?[389.64] [389.64][S02] Yeah, we've had these trusted testers, people[394.8] [394.8][S02]who are using it, who were just signed up[396.84] [396.84][S02]to be kind of early adopters.[398.17] [398.17][S03] What are people using it for?[399.878] [399.878][S02] People are using it for things like getting[402.3] [402.3][S02]fashion advice from Astra--[404.372] [404.372][S03] Oh, really?[405.33] [405.33][S03]In what way?[406.39] [406.39][S02] Like, what would match with this?[409.73] [409.73][S02]So yeah, Astra is kind of just like a partner.[413.21] [413.21][S02]Oh, what do you think about--[414.68] [414.68][S02]how could I have a fresher look here?[417.65] [417.65][S03] Oh, wow.[419.09] [419.09][S03]I mean, that's a very clever parrot.[422.107] [422.107][S02] It's a very clever parrot.[423.69] [423.69][S03] But then what about hardware?[424.92] [424.92][S03]I mean, at the moment, as you say, it's on your smartphone.[427.77] [427.77][S03]But are we talking about, eventually, in glasses?[430.74] [430.74][S02] Yeah, but not only.[433.25] [433.25][S02]So I think when an earlier version of this project started,[436.35] [436.35][S02]it was really trying to tease out[440.06] [440.06][S02]how useful smart glasses would be if an AI was on them.[444.29] [444.29][S02]So in smart glasses, it's the most intimate and, in some ways,[448.94] [448.94][S02]amazing experience because you've got this--[451.4] [451.4][S02]you feel augmented personally, like you're[454.22] [454.22][S02]having a conversation with a smart version of yourself[456.742] [456.742][S02]that's just sitting there and telling you[458.45] [458.45][S02]whatever you want to know.[459.54] [459.54][S02]But the software stack effectively[461.21] [461.21][S02]is agnostic to how your--[464.43] [464.43][S02]I mean, there's specializations for each device,[467.31] [467.31][S02]but you can have it on phones or computers or VR headsets.[470.898] [470.898][S03] I was thinking as well, actually, as we were just[473.44] [473.44][S03]playing around with it, there's a potential benefit[475.565] [475.565][S03]for people who are partially sighted here or blind too,[478.26] [478.26][S03]right?[478.76] [478.76][S02] Yeah, that's an obsession of mine.[480.677] [480.677][S02]We've talked about this sort of AI[482.56] [482.56][S02]as being co-present or sharing your perspective.[485.33] [485.33][S02]And sometimes, you want another seeing and hearing[492.84] [492.84][S02]kind of intelligence with you, but you don't always need one.[497.83] [497.83][S02]So when are the cases when you want[499.99] [499.99][S02]a system that can see alongside you if you see but don't[505.57] [505.57][S02]understand or if you can't see?[510.13] [510.13][S02]And so that's a whole category.[512.33] [512.33][S02]And there's a lot of people out there, hundreds of millions[516.64] [516.64][S02]of people, who have vision impairment.[519.289] [519.289][S02]And what's the gold standard of help for that population?[524.27] [524.27][S02]Well, it's having someone by their side[526.73] [526.73][S02]who can help them out in the world.[529.58] [529.58][S02]And this technology sort of is able to replicate that[534.62] [534.62][S02]to a large extent.[535.53] [535.53][S02]We have more nascent ideas, too, about other kinds[540.5] [540.5][S02]of disabilities.[541.35] [541.35][S02]So you could imagine helping people[543.8] [543.8][S02]who have difficulty scanning emotions and faces,[547.17] [547.17][S02]understanding that in certain circumstances.[550.122] [550.122][S03] So people, potentially, with autism[552.08] [552.08][S03]could use this to help.[553.47] [553.47][S02] Yeah.[554.39] [554.39][S02]I wouldn't recommend it as a prescribed drug[556.79] [556.79][S02]right at the moment, but I think with further development,[560.43] [560.43][S02]it could definitely be.[561.77] [561.77][S02]Also for training yourself, since you[565.55] [565.55][S02]could work on understanding faces[567.62] [567.62][S02]and have Astra give you feedback,[570.333] [570.333][S02]like, hey, tell me about this.[571.92] [571.92][S02]I remember-- it's on a separate topic,[574.68] [574.68][S02]but when I was doing a homestay, I[577.67] [577.67][S02]was learning French one summer, and I couldn't[581.0] [581.0][S02]pronounce certain words.[582.35] [582.35][S02]Like, the difference between the word for \"street\"[584.44] [584.44][S02]and the word for \"wheel,\" like \"la rue\" and \"la roue.\"[589.705] [589.705][S02]I still can't do it, right?[591.11] [591.11][S02]But I sat there with my homestay brother,[594.61] [594.61][S02]and I was just trying to copy him for a while,[597.2] [597.2][S02]and he's just blew me off after a few minutes.[599.27] [599.27][S02]He's like, I'm not sitting there with you.[601.34] [601.34][S02]Astra would be infinitely patient with you[603.88] [603.88][S02]and could help you with that kind of thing.[606.23] [606.23][S02]Obviously, memory-- so we have a system[610.12] [610.12][S02]that has perfect in-session memory, we call it.[616.01] [616.01][S02]So when the camera is rolling, basically, it[618.25] [618.25][S02]remembers the last 10 minutes photographically,[622.51] [622.51][S02]but it will also remember what you've talked about in the past.[626.24] [626.24][S02]That's why it remembers I'm Greg.[628.66] [628.66][S02]And probably if we turn it back on we say,[631.155] [631.155][S02]do you remember who was talking to you besides Greg[634.34] [634.34][S02]the last time, it will remember Hannah.[638.08] [638.08][S02]So this could be used for people with some cognitive impairments,[642.25] [642.25][S02]too, at some point.[646.12] [646.12][S02]I think one of the things that we're[648.19] [648.19][S02]excited about, too, is this idea of proactiveness,[650.96] [650.96][S02]so it deciding on its own that you have a need[656.27] [656.27][S02]and then channeling the response to that need[661.75] [661.75][S02]without your actual need to give it a steer.[666.2] [666.2][S02]So for example, it could be a useful system[669.13] [669.13][S02]for reminding you of things, going through the memories[671.89] [671.89][S02]and saying, oh, don't forget, you need to pick this up[674.41] [674.41][S02]on your way home, or whatever.[675.86] [675.86][S03] So you're not necessarily just[677.99] [677.99][S03]proactively switching it on when you want to talk to it,[680.99] [680.99][S03]but it could be there in the background[683.05] [683.05][S03]and then it bring up something when[685.96] [685.96][S03]it thought it was appropriate.[687.442] [687.442][S02] Yeah, yeah.[688.4] [688.4][S02]So the idea is like, you are going home[692.83] [692.83][S02]and it's like, hey, don't forget that you need to pick up[696.837] [696.837][S02]some orange juice because you ran out this morning[698.92] [698.92][S02]or whatever.[699.42] [699.42][S03] Oh, wow, because it remembers having[701.42] [701.42][S03]seen that in the morning.[702.63] [702.63][S02] Yeah, exactly.[704.12] [704.12][S03] So, I mean, I guess at this stage,[705.45] [705.45][S03]this is like painting ideas of what's possible, right?[707.69] [707.69][S02] Yeah, we don't have that yet, no.[709.29] [709.29][S02]But that's the kind of thing that we could build next, yeah.[711.96] [711.96][S03] But you can see the beginnings of it in this.[716.522] [716.522][S02] Yeah.[717.23] [717.23][S02]I mean, so I could easily say, here's my fridge[721.13] [721.13][S02]and, oh, no, there's not much orange juice.[723.498] [723.498][S02]And then I'd say, hey, what do you think I should[725.54] [725.54][S02]get at the supermarket later?[726.9] [726.9][S02]And it would remember that, yeah.[729.11] [729.11][S02]But I would have to give it a bit more of a context or--[732.457] [732.457][S03] Hold its hand a little bit more, as it were.[734.79] [734.79][S02] Yeah.[735.498] [735.498][S03] Yeah.[736.37] [736.37][S03]Do you find yourself having to correct it a lot?[738.78] [738.78][S03]I mean, do you notice glitches?[742.43] [742.43][S02] Yeah, yeah.[743.49] [743.49][S02]So one thing it does once in a while is it[747.2] [747.2][S02]says it can't really see something[749.69] [749.69][S02]that's it can clearly see.[750.99] [750.99][S02]And so you'll be reading--[753.26] [753.26][S02]like on a bookshelf.[754.52] [754.52][S02]And you'll say, can you read the titles of the bookshelf?[757.88] [757.88][S02]And it'll say like, oh, no, the titles, I can't make--[760.69] [760.69][S02]and you'll say something like-- you'll kind of do a Jedi mind[763.29] [763.29][S02]trick on it.[763.84] [763.84][S02]You'll be like, yes, you can see.[766.13] [766.13][S02]And it'll be like, yes, I can.[767.38] [767.38][S02]And it'll then-- [LAUGHS][769.41] [769.41][S02]This is sort of a weird limitation of the--[773.825] [773.825][S02]yeah, the agreeableness is something[776.31] [776.31][S02]that you can influence.[777.43] [777.43][S03] Is it susceptible to encouragement, then?[779.903] [779.903][S02] Yes.[780.57] [780.57][S03] Really?[781.362] [781.362][S02] [LAUGHS][782.46] [782.46][S03] I mean, hey, it works for humans, too.[784.95] [784.95][S03]A little bit of encouragement, and suddenly you[787.14] [787.14][S03]can do things you didn't think were possible.[789.81] [789.81][S03]So what other kind of environments[792.3] [792.3][S03]does it struggle with?[793.665] [793.665][S03]I mean, it's quite quiet in here.[795.04] [795.04][S03]It's quite well lit.[797.74] [797.74][S03]There's not lots of busyness going on.[801.845] [801.845][S03]Does it work just as well in those kind of environments--[804.22] [804.22][S03]busy, noisy, dark, perhaps?[806.98] [806.98][S02] In some ways, operating in more environments[809.76] [809.76][S02]is an important thing that we need to develop,[812.64] [812.64][S02]in particular noise conditions.[814.93] [814.93][S02]So as I said to you, Astra really does hear.[820.43] [820.43][S02]It actually takes in the audio directly,[824.79] [824.79][S02]converts it into a system that is the neural networks of taken[829.91] [829.91][S02]sound and code them as some kind of a package of information[833.81] [833.81][S02]that is processed by the language[836.24] [836.24][S02]model, Gemini, directly.[839.67] [839.67][S02]But the system isn't really trained to identify[845.9] [845.9][S02]different voices, so it will have[849.65] [849.65][S02]trouble understanding your voice versus my voice[852.138] [852.138][S02]when we're talking.[852.93] [852.93][S02]So if there's other bystanders who are having conversations,[857.04] [857.04][S02]Astra will pick that up as potentially the user's speech.[864.53] [864.53][S02]Or it actually has a system that kind of wakes up[871.22] [871.22][S02]and listens for a bit when there is somebody[873.74] [873.74][S02]speaking with enough intensity.[877.09] [877.09][S02]And it will just start listening to errant speech and be confused[883.6] [883.6][S02]if there's nothing directed at it.[885.02] [885.02][S02]So yeah, noisy environments will confuse it.[887.32] [887.32][S03] When you say distinguish[888.82] [888.82][S03]between the different voices, as in the waveform itself?[895.66] [895.66][S02] So there's an old problem called the cocktail[898.84] [898.84][S02]problem, which is the whole problem of what's[904.15] [904.15][S02]more technically known as source separation.[906.47] [906.47][S02]So it's understanding one sound source from another.[911.6] [911.6][S02]So if there's a guitar and someone singing,[914.51] [914.51][S02]you could isolate that into two tracks, the guitar track[917.92] [917.92][S02]and the singing track.[920.23] [920.23][S02]Likewise, you might want to be able to distinguish[922.9] [922.9][S02]the one speaker's track and another speaker's track.[928.06] [928.06][S02]So that might be possible to do within the single modality[934.77] [934.77][S02]or sense of audio.[938.86] [938.86][S02]It would also be possible to do that in a multimodal sense,[943.24] [943.24][S02]integrating across senses.[944.98] [944.98][S02]So for example, when I know that it's you speaking,[948.58] [948.58][S02]I can also see the movement of your lips[950.34] [950.34][S02]rather than the movement of someone else's lips.[953.28] [953.28][S02]So ultimately, you could imagine the systems would[956.1] [956.1][S02]use all sorts of cues, even to change[959.58] [959.58][S02]the way they perceive a sound.[961.33] [961.33][S03] Because I guess this is, in some ways,[963.45] [963.45][S03]the thing that makes Project Astra so difficult,[966.07] [966.07][S03]but also the thing that gives it the potential,[969.16] [969.16][S03]because the cocktail problem, as you say,[971.5] [971.5][S03]which humans are extremely good at-- you're in a cocktail party[975.06] [975.06][S03]and you can hear exactly what the person next to you is saying[977.77] [977.77][S03]despite loads of voices going on all around--[979.9] [979.9][S02] I have trouble, actually.[980.85] [980.85][S03] Ah.[981.33] [981.33][S03]Actually, do you know what?[982.455] [982.455][S03]Honestly, so do I.[983.79] [983.79][S02] That's why it's a problem.[985.373] [985.373][LAUGHTER][986.79] [986.79][S03] But broadly speaking, humans[988.53] [988.53][S03]are quite good at these things.[990.16] [990.16][S03]And when you have the audio-only problem,[992.48] [992.48][S03]it's really hard to solve.[994.43] [994.43][S03]But because this is multimodal, because you have video,[997.52] [997.52][S03]because you have audio, because you have the text language[1002.01] [1002.01][S03]model running in the background, you[1004.17] [1004.17][S03]do have more levers to potentially pull here.[1007.022] [1007.022][S02] Yeah, yeah.[1007.98] [1007.98][S02]I think it should be able to resolve ambiguity[1011.79] [1011.79][S02]with more context.[1013.9] [1013.9][S03] How about different languages?[1016.305] [1016.305][S03]Is it only in English at the moment, and only[1018.18] [1018.18][S03]with a very clear accent?[1021.6] [1021.6][S02] It's mostly in English for me, but it's--[1024.63] [1024.63][S02]no, it's very multilingual.[1026.17] [1026.17][S02]So that's a function of being native audio.[1029.829] [1029.829][S02]It knows about 20 languages with pretty high proficiency.[1033.13] [1033.13][S02]And you can switch between languages[1035.49] [1035.49][S02]even in the same conversation.[1037.315] [1037.315][S03] So go on.[1038.19] [1038.19][S03]Give a little demo of different languages.[1040.01] [1040.01][S04] Hello, Gregory.[1041.26] [1041.26][S04]It's nice to speak with you again.[1044.099] [1044.099][S02] Bonjour, Astra.[1045.3] [1045.3][SPEAKING IN FRENCH][1047.31] [1065.84][S03] Oh.[1067.312] [1067.312][S03]\"Little redhead.\"[1068.02] [1068.02][S03]Got that.[1068.66] [1068.66][S02] [SPEAKING FRENCH][1071.39] [1075.057][S03] So can you just-- hang on.[1076.64] [1076.64][S03]I did Russian in school.[1078.2] [1078.2][S03]I can remember one remaining Russian phrase.[1080.36] [1080.36][S03]Can you-- can you switch languages in the middle[1085.52] [1085.52][S03]without necessarily warning it?[1087.48] [1087.48][S03]So could I say, for example, [RUSSIAN]?[1089.81] [1092.135][S04] You certainly could use that phrase,[1094.26] [1094.26][S04]but what did you wish to ask about this phrase?[1097.17] [1097.17][S03] What does it mean?[1100.28] [1100.28][S04] [RUSSIAN] means \"At what time does it open?\"[1105.2] [1105.2][S04]It is asking about the opening time of something.[1110.53] [1110.53][S03] OK.[1111.7] [1111.7][S03]I mean, it is \"What time does the shop open?\"[1115.19] [1115.19][S03]I think.[1117.22] [1117.22][S03]But notable, then, that you're not saying,[1121.0] [1121.0][S03]now in English, now in French, now in Russian.[1123.555] [1123.555][S02] I think it's targeted[1124.93] [1124.93][S02]to respond in the language that you started with.[1127.19] [1127.19][S02]So actually, you said something in English[1129.19] [1129.19][S02]and then gave the Russian.[1130.7] [1130.7][S02]I think if you'd started speaking in Russian,[1132.14] [1132.14][S02]it would responded in Russian.[1133.39] [1133.39][S02]But as it was, it was thinking, I'm speaking English,[1137.39] [1137.39][S02]but I am hearing Russian.[1138.98] [1138.98][S02]So you didn't have to change that.[1140.48] [1140.48][S02]Bu, If you'd just started speaking in Russian,[1142.675] [1142.675][S02]it would have been maybe a little bit better.[1144.55] [1144.55][S03] I mean, this is different,[1146.133] [1146.133][S03]though, I mean, from the chat box that we have at the moment.[1148.705] [1148.705][S03]This is an additional capability.[1150.28] [1150.28][S02] I'm actually really excited about language[1152.53] [1152.53][S02]learning with this system, like, walking around[1154.488] [1154.488][S02]and being like, What is that? having it teach you the same way[1157.21] [1157.21][S02]that I was taught in school, where we would bring in objects[1161.59] [1161.59][S02]and talk about those objects in French class[1163.45] [1163.45][S02]to learn about stuff being together and learning language.[1168.09] [1168.09][S03] I can imagine being lost in a foreign city and that[1170.81] [1170.81][S03]being quite a helpful aid.[1172.167] [1172.167][S02] Exactly, yeah.[1173.25] [1173.25][S02]And it should be able to understand[1174.708] [1174.708][S02]other people speaking to you quite naturally, too, so--[1177.06] [1177.06][S03] So if this is the thing[1177.92] [1177.92][S03]that you're interacting with, what's[1179.78] [1179.78][S03]actually going on underneath the hood?[1182.01] [1182.01][S03]What are all the different components?[1183.6] [1183.6][S02] Yeah.[1184.308] [1184.308][S02]So the first thing is there's an app.[1187.04] [1187.04][S02]And that is actually gathering your video[1191.06] [1191.06][S02]and taking in your audio through the mic and so forth.[1195.32] [1195.32][S02]And that's connecting to a server[1197.75] [1197.75][S02]on which there are several different kinds[1200.18] [1200.18][S02]of neural network models.[1201.9] [1201.9][S03] Like what?[1203.315] [1203.315][S02] So there's a vision encoder and an audio encoder.[1207.36] [1207.36][S02]There's also specialized audio systems[1210.68] [1210.68][S02]that are just responsible for understanding when you've[1215.6] [1215.6][S02]probably stopped speaking.[1217.04] [1217.04][S02]Those are sitting next to the large language model, Gemini.[1222.62] [1222.62][S02]And they are sending information from these sensory encoders[1227.23] [1227.23][S02]directly into Gemini, which is responding.[1229.43] [1229.43][S02]We worked together with some of the teams in Gemini[1231.67] [1231.67][S02]also to change the Gemini model, to be better at dialogue[1235.75] [1235.75][S02]and audio processing.[1236.81] [1236.81][S02]So we've kind of improved its ability[1239.71] [1239.71][S02]to use, have audio, take an audio, and speak.[1243.19] [1243.19][S02]When we started working with the models,[1246.67] [1246.67][S02]they were making lots of factual errors.[1249.41] [1249.41][S02]So we had to identify ways in which we[1251.68] [1251.68][S02]could improve their factuality while[1255.52] [1255.52][S02]also being kind of conversational.[1258.58] [1258.58][S02]That was one aspect of our work on Gemini.[1261.14] [1261.14][S02]On top of all of that, though, is something called an agent.[1265.09] [1265.09][S02]The agent is taking the video and audio[1269.32] [1269.32][S02]and sending it to the model.[1271.13] [1271.13][S02]It's also calling search tools, so either Google Lens or Google[1277.69] [1277.69][S02]Search or Google Maps, when needed to respond to a query.[1284.01] [1284.01][S02]So if you ask about the price of something, it will call Search.[1288.81] [1288.81][S02]There's also a memory system that is being, in a sense,[1294.87] [1294.87][S02]part of the agent.[1297.21] [1297.21][S02]And offline in between sessions, the memory system[1300.93] [1300.93][S02]will summarize relevant information from the session,[1304.08] [1304.08][S02]about you and about what you've talked about in that session.[1308.29] [1308.29][S02]So those are some of the ingredients.[1310.33] [1310.33][S03] I mean, I'm trying to imagine something that we[1315.15] [1315.15][S03]use just to recognize a book.[1317.05] [1317.05][S03]I'm trying to think of the number of different elements[1319.59] [1319.59][S03]that are coming into play here, because you've got the computer[1322.74] [1322.74][S03]vision.[1323.47] [1323.47][S03]You've got the voice recognition.[1324.868] [1324.868][S03]You've got the large language models.[1326.41] [1326.41][S03]You've got the Google Search sort of sitting underneath it.[1329.23] [1329.23][S03]You've got the agent layer where you're actually[1331.71] [1331.71][S03]making decisions.[1332.73] [1332.73][S03]And you're doing all of that with almost no latency at all[1337.2] [1337.2][S03]in the answers that it's giving you.[1338.8] [1338.8][S03]I mean, this is a phenomenally complicated thing.[1341.02] [1341.02][S02] Yeah.[1341.22] [1341.22][S02]I mean, phenomenally complicated.[1342.84] [1342.84][S02]Of course, as engineers, we come up with abstraction layers[1345.81] [1345.81][S02]that we don't have to think about all the levels[1347.81] [1347.81][S02]of complexity at one time.[1349.94] [1349.94][S02]But I think overall, it's hugely complicated.[1352.8] [1352.8][S02]The data that's going into the models[1355.22] [1355.22][S02]is understood by very few people.[1357.36] [1357.36][S02]And exactly why it produces the results[1359.215] [1359.215][S02]is probably understood by no one,[1360.59] [1360.59][S02]in a sense, since it's just based on benchmarks.[1363.51] [1363.51][S03] Well, let me talk a little bit about the history[1364.79] [1364.79][S03]of this, because--[1365.75] [1365.75][S03]so back in the first series of this podcast,[1368.69] [1368.69][S03]you were a guest on the very first episode.[1370.92] [1370.92][S03]And then you were drawing on inspiration[1373.94] [1373.94][S03]from the animal kingdom for your research on intelligence.[1378.12] [1378.12][S03]And specifically, there was a bird, the western scrub jay,[1382.77] [1382.77][S03]that you were telling us about, as a way[1386.15] [1386.15][S03]to inspire more sophisticated memory for AI.[1388.825] [1388.825][S03]Let me just play you a little clip of it, actually.[1390.95] [1390.95][AUDIO PLAYBACK][1391.617] [1391.617][S03]- Having a kind of large database of things that[1395.24] [1395.24][S03]you've done and seen that you can access and that you can use[1399.68] [1399.68][S03]use to then guide your goal-directed behavior later--[1403.84] [1403.84][S03]I'm hungry.[1405.01] [1405.01][S03]Mm, I would love to have some maggots right now.[1407.29] [1407.29][S03]Where should I go find those?[1409.8] [1409.8][S03]That's the kind of thing we would like to replicate.[1412.847] [1412.847][END PLAYBACK][1413.43] [1413.43][S03] Have you managed to?[1414.763] [1414.763][S02] [LAUGHS][1417.358] [1418.97][S02]Hello, Project Astra.[1419.95] [1419.95][S02]Can you find some maggots for me?[1421.728] [1421.728][S03] I mean, that sounds quite[1423.27] [1423.27][S03]a lot like your orange juice example, doesn't it?[1425.35] [1425.35][S02] It is a proactive memory example, yeah.[1427.5] [1427.5][S03] Yeah.[1428.208] [1428.208][S03]And that's what you've done with Project Astra.[1431.292] [1431.292][S02] Yeah.[1432.0] [1432.0][S02]I think that there's a sense in which intelligence is really[1434.88] [1434.88][S02]one thing.[1436.12] [1436.12][S02]And one has a career and one is studying what intelligence is.[1441.79] [1441.79][S02]One is taking kind of glancing hits at it,[1443.95] [1443.95][S02]kind of trying to understand, spar with it in one way[1446.67] [1446.67][S02]or another.[1447.4] [1447.4][S02]And this project, maybe, is the strongest unification[1454.23] [1454.23][S02]of all of the strands of research I've had in my life.[1458.01] [1458.01][S02]Although, actually it's missing a major one,[1460.62] [1460.62][S02]which is that it's not embodied in a physical sense.[1462.87] [1462.87][S02]It can't act in the world--[1464.263] [1464.263][S03] Yet.[1464.93] [1464.93][S02] --which I--[1467.27] [1467.27][S02]perhaps, yeah.[1468.38] [1468.38][S02][LAUGHS] Yeah.[1470.73] [1470.73][S02]So yeah, I think memory, perception,[1475.003] [1475.003][S02]these have been long-standing interests.[1476.67] [1476.67][S02]And I think this is a way of bringing[1479.72] [1479.72][S02]them together that people seem to also find stimulates them.[1484.68] [1484.68][S02]They feel connected to it.[1486.66] [1486.66][S03] So how much of your neuroscience background[1489.2] [1489.2][S03]did end up inspiring Project Astra?[1491.2] [1491.2][S02] So neuroscience is used in two ways.[1493.2] [1493.2][S02]One is that there's a sense in which we're[1495.65] [1495.65][S02]using neuroscience to know when we've[1497.217] [1497.217][S02]done a good enough job to think about,[1498.8] [1498.8][S02]What does memory really mean?[1500.58] [1500.58][S02]and, Have we achieved it yet?[1503.15] [1503.15][S02]And it's also just a bit of a propulsion,[1505.97] [1505.97][S02]like, say, if we want something that is compatible with us,[1509.97] [1509.97][S02]human compatible, and, in some ways, like us, kind of[1514.04] [1514.04][S02]maybe go towards an embodiment of intelligence that's[1519.31] [1519.31][S02]a little bit more like us rather than a straightforward text[1522.79] [1522.79][S02]interface.[1523.52] [1523.52][S02]For example, I have been interested in the work[1526.78] [1526.78][S02]of Michael Tomasello, who studies[1530.5] [1530.5][S02]human communication by comparison to the great apes.[1536.89] [1536.89][S02]And he's really maybe the main thinker[1539.05] [1539.05][S02]behind this idea of, for me, situated dialogue,[1543.19] [1543.19][S02]where he talks about the basic premise of communication[1549.97] [1549.97][S02]as being about two individuals who are in the same place, who[1554.95] [1554.95][S02]are directing attention in the same place,[1558.17] [1558.17][S02]and therefore, inferring goals together[1561.28] [1561.28][S02]and then able to collaborate.[1563.32] [1563.32][S02]And that was kind of what we modeled in this technology.[1567.233] [1567.233][S03] So it's like the inspiration[1568.9] [1568.9][S03]rather than necessarily-- at the theoretical level,[1574.25] [1574.25][S03]rather than actually directly copying the design of it.[1577.11] [1577.11][S02] Not for the problem solving or the engineering[1579.76] [1579.76][S02]per se.[1580.26] [1580.26][S02]Then I think you need to come up with different solutions[1582.26] [1582.26][S02]that are dependent on the technology itself.[1584.37] [1584.37][S03] If Project Astra links to things[1587.87] [1587.87][S03]that we were talking about, I mean, literally years ago,[1590.58] [1590.58][S03]where did the first spark for this project come from?[1594.18] [1594.18][S03]Like, when did it actually begin?[1595.652] [1595.652][S02] Yeah.[1596.36] [1596.36][S02]So I think I know that--[1598.82] [1598.82][S02]Demis Hassabis, the CEO of DeepMind,[1604.43] [1604.43][S02]kind of threw down a challenge to the company[1607.52] [1607.52][S02]in a way, which was for us to think about what[1613.1] [1613.1][S02]a proto artificial general intelligence was,[1616.79] [1616.79][S02]which, what does that mean?[1618.89] [1618.89][S02]A proto artificial general intelligence[1621.53] [1621.53][S02]is a system that, if we created it and technically[1625.19] [1625.19][S02]minded people were able to scrutinize, investigate it, use[1628.07] [1628.07][S02]it, experience it, they would conclude[1630.59] [1630.59][S02]that the real deal, something that is generally intelligent[1636.19] [1636.19][S02]and a computational device, was ultimately going to arrive.[1641.15] [1641.15][S02]It was a matter of when, not if.[1643.03] [1643.03][S02]But that was left unspecified.[1644.51] [1644.51][S02]So there was a lot of creative thinking at the time like, well,[1648.372] [1648.372][S02]maybe it's this, maybe it's this, and so forth.[1650.33] [1650.33][S02]And some people had ideas of an intelligence arising[1656.11] [1656.11][S02]the same way that AlphaZero arose,[1658.21] [1658.21][S02]just by interacting with the world.[1659.96] [1659.96][S02]Other people maybe had other ideas,[1661.97] [1661.97][S02]but my idea was very much about the sociality of intelligence.[1666.08] [1666.08][S02]So we are not very smart as human beings[1669.67] [1669.67][S02]unless we learn from others or we[1671.29] [1671.29][S02]learn from books, which is the same as learning from others.[1673.79] [1673.79][S02]And that was the kind of idea I had for what proto AGI would be.[1677.81] [1677.81][S02]Then I thought also, we could unify[1679.51] [1679.51][S02]proto AGI with the idea of a helpful assistant whose[1683.23] [1683.23][S02]main goal is the benefit of the humans it interacts with.[1687.04] [1687.04][S02]So maybe those two things together[1688.78] [1688.78][S02]gave me something of a direction to look.[1693.07] [1693.07][S02]And then it was when I tried to think about making it ultimately[1697.02] [1697.02][S02]very natural, I sort of moved towards thinking about video[1700.95] [1700.95][S02]as the ultimately connective fiber of the systems.[1704.14] [1704.14][S03] Were there big moments along the way where you[1707.01] [1707.01][S03]had these big breakthroughs?[1708.678] [1708.678][S02] Where we had big breakthroughs?[1710.47] [1710.47][S02]Yeah, I think so.[1711.81] [1711.81][S02]There were these phases of the project.[1714.04] [1714.04][S02]So the first phase of the project[1715.56] [1715.56][S02]was basically a hackathon where we[1720.33] [1720.33][S02]had two weeks of making the first version.[1723.55] [1723.55][S02]And we have a video from that time.[1725.95] [1725.95][S02]It was quite crude.[1728.85] [1728.85][S02]But I remember Malcolm Reynolds, who's[1732.78] [1732.78][S02]a friend and engineer here, was playing around with Astra.[1736.72] [1736.72][S02]And he was going around an office room and saying,[1740.14] [1740.14][S02]what is this?[1740.95] [1740.95][S02]And the system would say, a plant.[1743.418] [1743.418][S02]And he'd say, what kind of a plant is this?[1745.21] [1745.21][S02]And it would say, a plant.[1746.293] [1746.293][LAUGHTER][1747.72] [1747.72][S02]Wasn't super flexible.[1749.26] [1749.26][S02]I remember the first demo I ever saw had a 7-second latency.[1753.72] [1753.72][S03] So you would say, Hi, Project Astra,[1755.998] [1755.998][S03]or whatever it was called then.[1757.29] [1757.29][S02] It wasn't called-- yeah.[1758.16] [1758.16][S03] And then 7 seconds later, it would--[1760.212] [1760.212][S02] Yeah.[1760.92] [1760.92][S02]So it was very difficult to use at all because you would kind of[1763.97] [1763.97][S02]think it had gone away, right?[1765.22] [1765.22][S02]But it was just 7 seconds later it would come back to you.[1767.705] [1767.705][LAUGHS][1769.56] [1769.56][S02]I think one of the main discoveries of the time[1771.99] [1771.99][S02]was basically that there's this idea of a prompt.[1775.36] [1775.36][S02]A prompt is the instructions you give to the system[1780.42] [1780.42][S02]that it needs for operation.[1781.9] [1781.9][S02]So systems like this really understand language.[1785.23] [1785.23][S02]They can read.[1785.94] [1785.94][S02]And you can say things to them like, your name is Astra.[1788.74] [1788.74][S02]You're an intelligent, helpful AI assistant.[1792.27] [1792.27][S02]Some of that information is inherent in the Gemini models[1795.33] [1795.33][S02]now.[1796.02] [1796.02][S02]But some of it's indicated in our prompt.[1798.91] [1798.91][S02]And it wasn't really understood whether we[1800.97] [1800.97][S02]could prompt a system that was multimodal before very well.[1804.97] [1804.97][S02]One of the things that was kind of a mind-blowing insight[1809.46] [1809.46][S02]or realization at the time was just telling the system[1812.49] [1812.49][S02]that it could see the world through the user's camera gave[1816.9] [1816.9][S02]it a sense of its own perspective[1818.64] [1818.64][S02]on things, like, where the provenance[1821.37] [1821.37][S02]of this information to it.[1823.09] [1823.09][S02]It didn't understand that before.[1824.47] [1824.47][S02]It was always making mistakes.[1826.01] [1826.01][S02]When you'd say, What do you see?[1827.56] [1827.56][S02]It was always giving me the wrong answer.[1829.268] [1829.268][S02]But then when we said, you're a system that[1831.39] [1831.39][S02]is an AI that is seeing through the user's camera,[1834.01] [1834.01][S02]then it could understand that this camera was[1837.27] [1837.27][S02]something it was effectively seeing[1838.95] [1838.95][S02]and it would answer correctly.[1839.915] [1839.915][S03] Wow.[1840.28] [1840.28][S02] I mean, there was a lot of work to do there,[1842.613] [1842.613][S02]but realizing that we could effectively prompt it was maybe[1845.233] [1845.233][S02]the-- even though it was a different kind of a system than[1847.65] [1847.65][S02]one we'd built before, and that you could use text to prompt its[1850.86] [1850.86][S02]understanding of the situated or more embodied--[1855.13] [1855.13][S03] That's so interesting.[1856.72] [1856.72][S03]When the gauntlet was thrown down of create a proto AGI,[1861.19] [1861.19][S03]were there people expressing doubt or skepticism[1864.27] [1864.27][S03]that something like this might have been possible?[1866.552] [1866.552][S02] Yeah.[1867.26] [1867.26][S02]So hindsight is curious in AI because it moves so fast,[1873.08] [1873.08][S02]and people's perception of what is obvious changes so fast.[1877.29] [1877.29][S02]I think it's obvious to a lot of people[1879.02] [1879.02][S02]now in some ways, which blows my mind.[1882.27] [1882.27][S02]I'm like, do you know the adversity, how much convincing[1886.76] [1886.76][S02]had to happen?[1888.947] [1888.947][S03] Well, tell us.[1890.03] [1890.03][S03]How much?[1891.06] [1891.06][S02] So I think from many different perspectives,[1893.94] [1893.94][S02]people thought this was an odd thing to do.[1897.41] [1897.41][S02]So from the perspective of, Could the systems actually[1901.73] [1901.73][S02]understand the world at all?[1905.0] [1905.0][S02]the vision systems in that era, in terms[1907.52] [1907.52][S02]of the number of pixels they were taking in,[1909.93] [1909.93][S02]it was like 96 by 96 patches of image.[1913.8] [1913.8][S02]So for those who don't know, the minimum of our screens is 1,000[1918.71] [1918.71][S02]pixels by 1,000 pixels--[1920.27] [1920.27][S02]so you know, a very blurry input to these systems.[1924.3] [1924.3][S02]So no wonder it couldn't identify what kind of a plant[1926.61] [1926.61][S02]it was.[1927.11] [1927.11][S02]It barely could see.[1928.48] [1928.48][S02]The fact that these systems would really know information[1933.46] [1933.46][S02]about what they're seeing, rather than just being[1936.37] [1936.37][S02]able to identify or classify what they're seeing, so having[1939.61] [1939.61][S02]a deep conversation about something,[1941.665] [1941.665][S02]seemed probably a little too far ahead.[1943.85] [1943.85][S02]We didn't even have basic knowledge of the amount of data[1947.32] [1947.32][S02]that you'd need to get for systems to perform[1949.51] [1949.51][S02]at various levels.[1950.72] [1950.72][S03] So then, I mean, OK.[1952.63] [1952.63][S03]If all of this seemed so absurd, and yet you[1956.29] [1956.29][S03]embarked on it anyway, were there[1958.25] [1958.25][S03]times when you thought it wasn't going to be possible?[1960.5] [1960.5][S02] No.[1961.125] [1961.125][S02]No, no, no, no.[1962.18] [1962.18][S02]It seemed always like it would be possible.[1966.43] [1966.43][S02]There were times that I was maybe willing to give up.[1968.872] [1968.872][S03] Oh, really?[1969.83] [1969.83][S02] Yeah.[1970.31] [1970.31][S02]I think it was--[1971.0] [1971.0][S02]I think there was a slow period before Gemini where things[1975.22] [1975.22][S02]weren't working very well.[1976.61] [1976.61][S02]And it was hard times.[1978.47] [1978.47][S02]I think it didn't seem like it was[1980.29] [1980.29][S02]a fruitful line of investigation at the time for some people.[1984.6] [1984.6][S02]But I sort of never wavered about the fact[1987.86] [1987.86][S02]that this was definitely possible.[1990.21] [1990.21][S02]I think I had a much more obstinate, stubborn,[1994.32] [1994.32][S02]and ultimately stupid way of going about it, which was just[1997.22] [1997.22][S02]like, if I work on this for long enough, it will definitely work.[1999.94] [1999.94][LAUGHS][2001.75] [2001.75][S03] So I heard that as part of the testing phase,[2004.25] [2004.25][S03]you have this Project Astra room.[2008.71] [2008.71][S03]What's going on in there?[2010.34] [2010.34][S03]What's in the room?[2011.33] [2011.33][S02] There is a special room, yeah.[2012.79] [2012.79][S03] What's inside the special room?[2014.582] [2014.582][S02] We have just all sorts of fun and games[2016.78] [2016.78][S02]in the special room.[2017.84] [2017.84][S02]There's a whole bar there so Astra can help you make a drink.[2022.46] [2022.46][S03] Right.[2023.21] [2023.21][S03]Yeah, it can.[2023.89] [2023.89][S02] There's an art gallery[2026.23] [2026.23][S02]so you can flash up different paintings on screens[2028.81] [2028.81][S02]and walk around the gallery and ask questions about art.[2031.725] [2031.725][S03] OK.[2032.35] [2032.35][S03]Well, let's dig into some of the stuff that's[2034.33] [2034.33][S03]going on behind the scenes of Astra a little bit more.[2036.64] [2036.64][S03]Latency, I think, is a really key thing.[2039.26] [2039.26][S03]You mentioned a moment ago about the 7-second lag that you used[2042.78] [2042.78][S03]to get.[2043.33] [2043.33][S03]How have you actually improved that?[2045.4] [2045.4][S02] So it's on multiple fronts.[2047.05] [2047.05][S02]So we've improved the actual streaming video.[2049.989] [2049.989][S02]So it actually is sending information faster[2053.76] [2053.76][S02]through the app.[2055.35] [2055.35][S02]There's a sense in which these systems, although they're[2058.469] [2058.469][S02]trained together, there's a vision[2060.36] [2060.36][S02]system and an audio system, this language model[2062.969] [2062.969][S02]system that is getting the information from those two[2065.5] [2065.5][S02]things.[2066.0] [2066.0][S02]What's called collocating them-- these[2068.04] [2068.04][S02]are kind of a technical term, but basically, we're[2070.44] [2070.44][S02]always processing images.[2072.28] [2072.28][S02]So as the video is coming in, for example,[2075.01] [2075.01][S02]into the vision system, it's always[2078.239] [2078.239][S02]running as fast as it can.[2080.08] [2080.08][S02]And then it's sitting in the same place,[2083.67] [2083.67][S02]in the same cluster of computers,[2086.159] [2086.159][S02]as is the large language model, so that it[2088.949] [2088.949][S02]doesn't have to make a call across a country or a continent.[2094.51] [2094.51][S03] Well, so--[2095.94] [2095.94][S03]sorry.[2096.88] [2096.88][S02] So they're running next to each other.[2097.77] [2097.77][S03] So to get this kind of real-time understanding[2100.187] [2100.187][S03]of what's going on, you have to physically locate[2103.25] [2103.25][S03]the computer hardware that are running these models[2105.59] [2105.59][S03]close to each other, because that makes a difference?[2108.102] [2108.102][S02] Absolutely, yeah.[2109.31] [2109.31][S03] Has that been the main thing, then,[2110.63] [2110.63][S03]just moving where you're actually running the models?[2113.01] [2113.01][S02] No.[2113.635] [2113.635][S02]So moving where we--[2115.97] [2115.97][S02]putting the models together is one thing.[2118.41] [2118.41][S02]Making sure that we are caching the context[2121.68] [2121.68][S02]so that the context of history, of what the system is[2124.94] [2124.94][S02]interacting with you, is incrementally updated over time.[2129.635] [2129.635][S02]There's this idea of doing work with native audio, which[2132.29] [2132.29][S02]means that previous systems had a text recognition[2138.62] [2138.62][S02]system-- or speech-to-text recognition system.[2141.24] [2141.24][S02]So they'd take in the audio, then they'd[2144.41] [2144.41][S02]produce a transcript, then they'd[2146.09] [2146.09][S02]call the language model, which would respond to that,[2148.41] [2148.41][S02]and then you'd get a response.[2150.12] [2150.12][S02]This system is directly getting the audio in,[2153.248] [2153.248][S02]so it doesn't have to have that secondary system, which[2155.54] [2155.54][S02]also takes time or produces extra latency.[2158.22] [2158.22][S02]So actually, there's a simple effect[2160.25] [2160.25][S02]that's possible with native audio is[2161.75] [2161.75][S02]it can understand rare words or the pronunciation of words.[2165.21] [2165.21][S02]A rare word, although becoming not so rare anymore,[2168.56] [2168.56][S02]or name is \"Demis Hassabis.\"[2171.11] [2171.11][S02]The old systems that didn't understand audio natively[2174.56] [2174.56][S02]directly often thought I was saying \"Damascus,\"[2177.56] [2177.56][S02]but now it knows that it's Demis Hassabis.[2179.72] [2179.72][S02]And it can use context to resolve that.[2181.5] [2181.5][S02]The CEO of DeepMind is Demis Hassabis.[2184.04] [2184.04][S02]Another example that somebody found recently,[2186.12] [2186.12][S02]which we have a little demo of, is distinguishing between[2189.5] [2189.5][S02]the word \"scahn\" and the words \"scohn,\"[2191.7] [2191.7][S02]which are two pronunciations of the same biscuity thing.[2194.73] [2194.73][S02]Project Astra can actually-- you could say,[2197.587] [2197.587][S02]What's the difference between \"scohn\" and \"scahn\"?[2199.67] [2199.67][S02]It will have heard that you said a different word rather[2202.1] [2202.1][S02]than just transcribing it into the one word.[2204.36] [2204.36][S02]Then the final one is that team did a lot of great work[2206.87] [2206.87][S02]on what's called endpointing, which is a very technical term.[2210.06] [2210.06][S02]But more or less, it knows exactly when[2212.72] [2212.72][S02]you have stopped speaking.[2214.35] [2214.35][S02]So it's very good at sensing, OK,[2217.32] [2217.32][S02]the user is really done now so I can talk.[2219.24] [2219.24][S02]Then there's something even more sophisticated,[2221.198] [2221.198][S02]which is that it plans a response even if you[2223.31] [2223.31][S02]hadn't finished speaking yet.[2225.77] [2225.77][S02]And it sort of is speculatively planning that.[2228.45] [2228.45][S02]So it's sort of guessing, this is what I would say.[2231.83] [2231.83][S02]And then when it figures out that the user really[2235.01] [2235.01][S02]has finished speaking, then it just sends it right off.[2237.805] [2237.805][S02]So it's already done.[2238.68] [2238.68][S02]It's already figured out what to say before.[2240.513] [2240.513][S02]It's before maybe even you really[2242.84] [2242.84][S02]know that you're done speaking.[2244.17] [2244.17][S03] That is so interesting[2245.54] [2245.54][S03]because I guess, actually, a lot of the time,[2247.26] [2247.26][S03]people's sentences, the important bit[2248.802] [2248.802][S03]of their sentences, can be in the middle[2251.9] [2251.9][S03]and then they sort of trail off towards the end.[2254.52] [2254.52][S03]And then you can use that time for getting ready[2257.6] [2257.6][S03]with your answer.[2258.432] [2258.432][S02] Yeah, pretty much that.[2259.89] [2259.89][S02]Yeah.[2260.39] [2260.39][LAUGHTER][2262.43] [2262.43][S02]Oh, yeah.[2262.95] [2262.95][S02]We talked about this stuff--[2264.41] [2264.41][S02]actually, we talked about that stuff three years ago.[2267.9] [2267.9][S02]And then it seemed like, oh, that's too much.[2270.21] [2270.21][S02]And then it kind of started to work this year.[2272.45] [2272.45][S03] Preemptively guessing[2273.89] [2273.89][S03]what the answer is going to be before the conversation has[2276.5] [2276.5][S03]got to that point.[2277.35] [2277.35][S02] Yeah.[2278.058] [2278.058][S02]And it's hard.[2278.742] [2278.742][S02]We pause for a long time in our sentences.[2283.015] [2283.015][S02]So the system that we have actually has to use some, quote,[2287.46] [2287.46][S02]\"semantic understanding,\" since it also has a bit[2290.84] [2290.84][S02]of understanding of context and the sounds.[2294.05] [2294.05][S02]It's also hearing to guess when the user's probably done.[2297.473] [2297.473][S03] But also, the reasoning that it's doing,[2299.64] [2299.64][S03]I mean, even separate from reasoning[2301.243] [2301.243][S03]whether it's finished a sentence or not.[2302.91] [2302.91][S03]Do you think that Project Astra is capable of reasoning?[2307.04] [2307.04][S02] Yeah.[2308.1] [2308.1][S02]It's primarily reasoning through its internal structure[2313.04] [2313.04][S02]inside the neural network, sort of in an unobservable--[2316.07] [2316.07][S02]or sort of a very complex way.[2318.89] [2318.89][S02]Then there's the dialogue itself it's producing.[2321.95] [2321.95][S02]So it sometimes reasons through the dialogue.[2324.18] [2324.18][S02]So you can hear it sounding out an answer.[2327.44] [2327.44][S02]People are also developing systems[2329.12] [2329.12][S02]that have inner speech, effectively,[2331.365] [2331.365][S02]where they're talking to themselves without talking[2333.49] [2333.49][S02]to you.[2333.98] [2333.98][S02]Project Astra, at the moment, doesn't do much of that.[2336.17] [2336.17][S03] But then I guess the advances that[2338.087] [2338.087][S03]happen in the reasoning models need not be distinct from what[2343.96] [2343.96][S03]happens in Project Astra, I guess,[2345.443] [2345.443][S03]is the whole point of this, is that it's[2347.11] [2347.11][S03]pulling in everything so that you have this ultimate proto[2350.74] [2350.74][S03]AGI, as you call it.[2351.98] [2351.98][S02] Yeah.[2352.688] [2352.688][S02]And in some ways, I actually hope[2354.46] [2354.46][S02]that it motivates some maybe more vigorous work[2359.41] [2359.41][S02]on some aspects of reasoning.[2361.07] [2361.07][S02]So we have this great example of Bibo Xu, the product[2364.45] [2364.45][S02]manager at Project Astra, pulled out Astra one day at lunch[2367.84] [2367.84][S02]and was like, how many calories are on my plate?[2370.4] [2370.4][S02]And she had a very complex, very beautifully laid out plate[2373.6] [2373.6][S02]with six types of food, some almonds in the middle,[2377.98] [2377.98][S02]a pork loin over there, some Brussels sprouts, or whatever.[2382.61] [2382.61][S02]And it was like, oh, you know, kind of waffled a little bit.[2386.63] [2386.63][S02]But then she said, keep a running total.[2389.5] [2389.5][S02]How many are in these Brussels sprouts?[2391.842] [2391.842][S02]And it was like, well, that's seven Brussels sprouts,[2394.05] [2394.05][S02]therefore it's this many calories.[2396.15] [2396.15][S02]And then, OK, now add the pork loin.[2398.453] [2398.453][S02]One of the things that was quite notable to me[2400.37] [2400.37][S02]was that Bibo was hand-holding its thinking.[2404.8] [2404.8][S02]As you said, it needs little guidance sometimes.[2406.8] [2406.8][S02]But I don't think that we're very far off from a system[2409.19] [2409.19][S02]that itself would just say, well,[2411.45] [2411.45][S02]I see there are seven almonds over there,[2413.64] [2413.64][S02]this many Brussels sprouts, there's a pork loin,[2416.04] [2416.04][S02]all those together are such and such.[2418.11] [2418.11][S02]So I think in some sense, it's not good at that stuff,[2421.47] [2421.47][S02]because we just haven't ever tried[2423.582] [2423.582][S02]to build a system that could reason about that stuff.[2425.79] [2425.79][S03] Now, I want to talk to you a bit more about memory.[2428.415] [2428.415][S03]On that point, about the things that are recalling and keeping[2433.01] [2433.01][S03]in its mind, as it were, if you'll[2434.6] [2434.6][S03]forgive the anthropomorphism, I know that back at Google I/O,[2438.68] [2438.68][S03]this could remember what had happened in the last 45 seconds.[2442.32] [2442.32][S03]And now you've increased that time.[2445.5] [2445.5][S03]You can do 10 minutes now, right?[2447.18] [2447.18][S02] Yeah, it's about 10 minutes.[2448.26] [2448.26][S02]Yeah, actually it's a little bit longer in some ways,[2450.468] [2450.468][S02]but 10 minutes is maybe what it should do on the 10.[2454.15] [2454.15][S03] What makes 10 minutes the limit?[2456.132] [2456.132][S02] Yeah.[2456.84] [2456.84][S02]So it's got, basically, a raw record of the last 10 minutes[2463.71] [2463.71][S02]of video.[2464.5] [2464.5][S02]It works at about one frame per second.[2466.33] [2466.33][S02]So it's got, basically, a stack of all the frames over time[2471.3] [2471.3][S02]and all the audio that came in between those frames, so[2474.36] [2474.36][S02]the last 600 frames or something like that.[2477.01] [2477.01][S02]The limits are really about the memory on the chips, I think.[2480.78] [2480.78][S02]That hasn't scaled very much, I think,[2482.64] [2482.64][S02]in the last decade or something like that,[2484.42] [2484.42][S02]the amount of this sort of fast, active memory.[2486.985] [2486.985][S03] But so at the moment, then,[2488.61] [2488.61][S03]it is effectively acting like a video recorder, as it were,[2491.11] [2491.11][S03]keeping an actual record of everything that's happened[2493.11] [2493.11][S03]in the previous 10 minutes.[2494.38] [2494.38][S02] Yeah.[2494.79] [2494.79][S02]Yeah, I mean, it's quite active.[2496.21] [2496.21][S02]It's able to use that information right away.[2498.25] [2498.25][S02]There's also a sort of a secondary system, which[2500.76] [2500.76][S02]is when you turn the system off, it[2504.54] [2504.54][S02]will then take that conversation and summarize it and pull out[2507.98] [2507.98][S02]relevant facts.[2509.22] [2509.22][S03] The most important bits.[2510.72] [2510.72][S02] Yeah.[2511.428] [2511.428][S02]And it uses its own discretion to figure out what that is.[2514.115] [2514.115][S03] To extract the gist of it, as it were.[2516.51] [2516.51][S02] Yeah.[2516.895] [2516.895][S03] But at the moment, I mean, one thing that it can do[2519.52] [2519.52][S03]is recall important things from recent interactions, can it?[2522.27] [2522.27][S02] Yeah.[2522.978] [2522.978][S02]It's kind of got sort of a two-stream memory.[2525.9] [2525.9][S02]So it's got a memory that is both about you as a person.[2530.01] [2530.01][S02]It's got a developing understanding of you.[2535.62] [2535.62][S02]It's effectively taking notes like, oh, they like ice cream,[2539.4] [2539.4][S02]that's chocolate ice cream, or whatever.[2541.1] [2541.1][S02]That'll be like a list of things it's discovered about you.[2544.98] [2544.98][S02]And that's actually updated after every session, too.[2548.49] [2548.49][S02]So suppose you say, you know what, I actually[2551.72] [2551.72][S02]decided that I don't like ice cream anymore.[2553.89] [2553.89][S02]I really like cake.[2554.94] [2554.94][S02]So forget that I liked ice cream.[2557.42] [2557.42][S02]It will then say, user says they no longer like ice cream,[2561.84] [2561.84][S02]they like cake.[2562.8] [2562.8][S02]And those things are kind of a stationary or static[2565.88] [2565.88][S02]understanding of who you are, effectively, or what you like,[2569.86] [2569.86][S02]your preferences.[2571.04] [2571.04][S02]Then there's also this kind of conversational summary[2574.39] [2574.39][S02]that's like, on Tuesday at 8:50, we talked about this game[2579.67] [2579.67][S02]of chess.[2580.993] [2580.993][S03] But then how does it decide which bit goes in which?[2583.66] [2583.66][S03]How does it decide what's important enough[2585.46] [2585.46][S03]to be a thing about you that it remembers?[2589.46] [2589.46][S02] So it's got heuristics.[2591.02] [2591.02][S02]These systems actually are given heuristics.[2593.03] [2593.03][S02]So a heuristic is basically a rule[2596.56] [2596.56][S02]of thumb for what to remember.[2598.97] [2598.97][S02]So one heuristic it uses is, and we've told it to,[2603.31] [2603.31][S02]is if you ask it to remember something it should definitely[2605.89] [2605.89][S02]remember that.[2607.27] [2607.27][S02]It's a pretty clear one.[2608.57] [2608.57][S02]So if I say, remember my door code,[2612.43] [2612.43][S02]it will do that because it will understand[2615.16] [2615.16][S02]that's an instruction of relevance.[2617.38] [2617.38][S02]Otherwise, it takes a best guess.[2620.05] [2620.05][S02]It tries to say--[2621.655] [2621.655][S02]it's sort of saying, has the user expressed any preferences[2624.75] [2624.75][S02]that are interesting or that are different from the ones[2627.39] [2627.39][S02]that the user has already expressed?[2629.473] [2629.473][S02]And then it will kind of update based on that.[2631.39] [2631.39][S03] Well, let's talk about some[2632.22] [2632.22][S03]of the privacy concerns here, then.[2633.76] [2633.76][S03]How do you mitigate against some of those privacy concerns?[2636.95] [2636.95][S02] Right.[2637.7] [2637.7][S02]So I think that one of the major standards is that of consent.[2642.01] [2642.01][S02]The users have access to their previously recorded data.[2647.49] [2647.49][S02]And they can delete it or see what is stored.[2651.58] [2651.58][S02]Every time you delete something, it[2653.43] [2653.43][S02]reconstitutes its whole knowledge of you.[2656.055] [2656.055][S03] Oh.[2656.68] [2656.68][S02] It goes through the whole process of summarizing[2659.22] [2659.22][S02]things it knows about you anew.[2660.895] [2660.895][S03] So the answer then, I[2662.27] [2662.27][S03]guess, is that the user ends up having some control over what[2666.18] [2666.18][S03]it knows about it--[2667.71] [2667.71][S03]about them, rather.[2668.62] [2668.62][S02] Yeah.[2669.328] [2669.328][S03] But actually, so in this podcast a few episodes ago,[2672.76] [2672.76][S03]we got to talk to Iason Gabriel, who's this ethicist at DeepMind.[2678.06] [2678.06][S03]He's amazing.[2679.44] [2679.44][S03]And he was telling us about the ethics of AI assistants[2683.38] [2683.38][S03]and how they should be shaped in order to take[2685.9] [2685.9][S03]into account lots of these difficult questions.[2688.64] [2688.64][S03]How much has his work fed into what[2692.38] [2692.38][S03]you've come up with with Astra?[2693.86] [2693.86][S02] We just fed his 243-page report into Astra.[2697.19] [2697.19][S02]And Astra said, OK, I got it.[2699.23] [2699.23][S03] Did you?[2700.224] [2700.224][S02] No, I wish.[2701.02] [2701.02][S03] Oh, that would have been so--[2702.728] [2702.728][S02] [LAUGHS] Yeah, I think[2704.74] [2704.74][S02]we've spoken a lot with Iason.[2706.7] [2706.7][S02]And we've done a lot of work with a team that he's part of.[2710.53] [2710.53][S02]And they've been investigating both the model and the agent[2715.12] [2715.12][S02]as a whole, exploring what it might[2717.67] [2717.67][S02]do in different circumstances, also working[2719.83] [2719.83][S02]with some external red teamers who[2721.6] [2721.6][S02]maybe have fewer preconceptions and might[2724.12] [2724.12][S02]do more different kinds of adversarial attacks[2727.51] [2727.51][S02]on the system.[2728.27] [2728.27][S02]We also have a layer of safety filters.[2730.67] [2730.67][S02]And this is for user harms or, for example,[2735.05] [2735.05][S02]if you say certain things to it or show it pornography,[2738.86] [2738.86][S02]for example, it will trigger these filters[2742.35] [2742.35][S02]and not respond to that.[2744.43] [2744.43][S02]It'll also trigger on its own speech,[2746.56] [2746.56][S02]so it can't say certain things, although they trigger[2749.52] [2749.52][S02]very infrequently anyway.[2750.682] [2750.682][S02]But I don't know.[2751.39] [2751.39][S02]Yeah, I think that the range of issues is quite broad.[2755.31] [2755.31][S02]Fortunately, we still have some time to figure stuff out.[2758.955] [2758.955][S03] OK.[2759.58] [2759.58][S03]So what, then, are your next priorities, then?[2761.43] [2761.43][S03]Over the next few months, what were the main things[2763.38] [2763.38][S03]you're going to be working on?[2764.35] [2764.35][S02] I'm very interested in something[2766.183] [2766.183][S02]called proactive video work.[2769.93] [2769.93][S02]So that is to say, a system that can not only[2774.48] [2774.48][S02]respond when you speak, but can also[2776.52] [2776.52][S02]help you in an ongoing sense.[2778.33] [2778.33][S02]So for example, that's part of the visual interpreter[2780.66] [2780.66][S02]for the blind problem.[2781.72] [2781.72][S02]So you're walking around, you can't see,[2783.84] [2783.84][S02]it will say, oh, watch out for the table over there.[2787.28] [2787.28][S02]It can guide you in an ongoing sense.[2789.78] [2789.78][S02]We're also doing a lot of work on more audio output, what's[2793.77] [2793.77][S02]called full duplex.[2795.04] [2795.04][S02]So it will process both--[2798.155] [2798.155][S02]it'll hear and speak at the same time, which[2801.13] [2801.13][S02]could be potentially annoying.[2802.38] [2802.38][S02]It could interrupt you.[2805.05] [2805.05][S02]But it's also more natural conversation.[2807.47] [2807.47][S02]As you're talking I might say, uh-huh, uh-huh.[2810.86] [2810.86][S02]And that's listening and talking at the same time.[2813.8] [2813.8][S02]It's part of language to confirm.[2815.87] [2815.87][S02]More on reasoning, as you said, more deep kinds of memory,[2821.61] [2821.61][S02]reflection of certain kinds.[2824.54] [2824.54][S02]When it calls tools to be able to do deeper inquiries[2829.16] [2829.16][S02]and research with tools, yeah, there's just so[2832.67] [2832.67][S02]many things to do better.[2834.24] [2834.24][S03] Well, thank you very much for joining us, Greg.[2835.98] [2835.98][S02] Thank you, Hannah.[2837.23] [2837.23][S03] It is strange how quickly our expectations change[2841.28] [2841.28][S03]about AI.[2842.155] [2842.155][S03]I don't know if you remember what[2843.53] [2843.53][S03]Oriol said in our last episode.[2845.1] [2845.1][S03]He said, if someone had told him five years ago the things that[2849.32] [2849.32][S03]would be possible, he would think that we were already[2852.14] [2852.14][S03]on the path to AGI.[2854.42] [2854.42][S03]And yet, here we have this prototype of a multimodal agent.[2859.33] [2859.33][S03]It's one that can see, that can hear,[2861.58] [2861.58][S03]that has memory and context and reasoning[2864.39] [2864.39][S03]and multilingual, real-time conversation.[2867.61] [2867.61][S03]This is an agent that could, at least in theory, accompany you[2871.98] [2871.98][S03]on your day to day, enhancing your knowledge,[2874.93] [2874.93][S03]supporting people with disabilities,[2876.91] [2876.91][S03]and augmenting our skills.[2879.25] [2879.25][S03]Now, of course, AGI, it isn't.[2882.45] [2882.45][S03]But it definitely feels like we are a significant leap[2886.203] [2886.203][S03]from the kinds of systems that we were talking[2888.12] [2888.12][S03]about even two years ago.[2891.435] [2891.435][S03]Thank you so much for joining us for this series of \"Google[2893.96] [2893.96][S03]DeepMind--[2894.46] [2894.46][S03]The Podcast.\"[2895.15] [2895.15][S03]We are going to take a break from here,[2897.19] [2897.19][S03]but if you want to catch up on any of our previous episodes,[2900.61] [2900.61][S03]then there is a whole array of deliciously nerdy,[2904.27] [2904.27][S03]AI conversational delights in our back catalog for you[2907.95] [2907.95][S03]to enjoy.[2908.74] [2908.74][S03]Just find them on YouTube or wherever you get your podcasts.[2911.82] [2911.82][MUSIC PLAYING][2914.87]"}
{"file_name": "audio/val_000015.wav", "transcription": "[0.0][MUSIC PLAYING][4.293] [6.423][S01] Welcome back to \"Google DeepMind--[8.34] [8.34][S01]The Podcast.\"[9.0] [9.0][S01]I'm Professor Hannah Fry.[10.53] [10.53][S01]Now there is broad consensus across the tech industry,[14.22] [14.22][S01]governments, and society that as AI becomes[17.18] [17.18][S01]more and more embedded in every aspect our world,[20.73] [20.73][S01]regulation is essential.[23.31] [23.31][S01]But once you start to decipher what that regulation should[26.18] [26.18][S01]actually look like, well, then agreement[28.73] [28.73][S01]becomes much more elusive.[31.08] [31.08][S01]How do we protect against the harms of new technologies[34.19] [34.19][S01]without stifling innovation?[36.63] [36.63][S01]How can we give AI the autonomy that it[39.17] [39.17][S01]needs to be able to solve complex problems while retaining[43.37] [43.37][S01]human control?[44.79] [44.79][S01]And is it possible to keep development[47.21] [47.21][S01]transparent and accountable to the public,[49.47] [49.47][S01]even as companies fiercely guard their competitive secrets?[53.61] [53.61][S01]Well, in our episode with Demis Hassabis[55.61] [55.61][S01]at the start of the season, we talked about his general views[58.91] [58.91][S01]on the regulation of AI.[60.54] [60.54][S01]But in this episode, I want to dig a bit deeper[63.23] [63.23][S01]into the nuances of this topic.[64.86] [64.86][S01]I want to take time to explore the arguments[67.76] [67.76][S01]and the counter-arguments to the different forms of regulation[71.45] [71.45][S01]that are on the table.[72.62] [72.62][S01]Now, of course, to address the elephant in the room,[75.9] [75.9][S01]this is the \"Google DeepMind\" Podcast.[78.36] [78.36][S01]So it's important to be upfront about the fact[80.81] [80.81][S01]that my guest today is going to be offering us[83.24] [83.24][S01]one particular view of AI regulation.[86.16] [86.16][S01]However, I want to promise you that as far as I possibly can,[90.15] [90.15][S01]I am going to try and make this a robust discussion,[93.32] [93.32][S01]and leave it up to you, the audience,[95.31] [95.31][S01]to decide where you stand on these issues.[97.83] [97.83][S01]And so with that in mind, I am delighted to be joined today[101.21] [101.21][S01]by Nicklas Lundblad, Google DeepMind's[103.49] [103.49][S01]Head of Public Policy and Public Affairs.[105.66] [105.66][S01]Nicklas has spent years working at the intersection[108.14] [108.14][S01]of technology and policy at Google,[110.52] [110.52][S01]navigating the intricate landscape of AI's evolution[114.08] [114.08][S01]and regulation.[115.26] [115.26][S01]Nicklas, welcome to the podcast.[116.64] [116.64][S02] Well, thank you so much for having me.[117.74] [117.74][S01] Of course.[118.657] [118.657][S01]So, OK, in terms of your current role at the moment,[121.22] [121.22][S01]I know that you spend a lot of time talking to industry[124.09] [124.09][S01]leaders, talking to governments, to the public,[127.72] [127.72][S01]about perceptions of AI.[130.509] [130.509][S01]How would you describe the current public mood, as it were,[134.99] [134.99][S01]around the subject?[136.01] [136.01][S02] I think it's hesitant,[137.71] [137.71][S02]but also hesitantly optimistic.[140.44] [140.44][S02]And then you find the people who are hesitantly pessimistic.[142.94] [142.94][S02]So it's almost like evenly cut, I think.[145.18] [145.18][S02]To some degree, there are a lot of people[146.973] [146.973][S02]who hope that this technology can help us solve some[149.14] [149.14][S02]of the knotty and complex problems[150.61] [150.61][S02]that we have encountered, climate[152.26] [152.26][S02]change, pandemics, other kinds of things[154.66] [154.66][S02]that we really need to deal with.[156.17] [156.17][S02]And there are a lot of people who[157.545] [157.545][S02]say, well, yes, it might be able to do that, but at what price?[160.28] [160.28][S02]And so you end up with these two different groups.[163.075] [163.075][S01] And where do you sit between those two?[165.2] [165.2][S02] I am hopeful.[166.67] [166.67][S02]I also think that this has, artificial intelligence to me,[169.76] [169.76][S02]has a very natural place in progress[172.18] [172.18][S02]overall, because what we're doing[174.7] [174.7][S02]is what we've done for a long time,[176.51] [176.51][S02]is that we have bought progress and welfare[178.57] [178.57][S02]at the price of complexity.[180.47] [180.47][S02]And as social complexity increases[182.77] [182.77][S02]massively, we need new ways of dealing[185.35] [185.35][S02]with it, hence technological and social innovation.[188.78] [188.78][S02]And so the way that artificial intelligence[190.9] [190.9][S02]fits into this very, very simple picture[192.79] [192.79][S02]is that it is a way of dealing with increased complexity[196.39] [196.39][S02]and allows the unlocking of further progress.[199.49] [199.49][S02]And to me, that's absolutely essential.[202.48] [202.48][S02]There are two really interesting views[204.7] [204.7][S02]about artificial intelligence, two stories that are in tension.[207.86] [207.86][S02]One of them is that we do this because we can,[210.325] [210.325][S02]the sort of Frankenstein story.[212.09] [212.09][S02]We shall become as gods.[213.41] [213.41][S02]That's what we're doing.[214.84] [214.84][S02]And that story has had its time.[216.675] [216.675][S02]And there are a lot of people who[218.05] [218.05][S02]are worried about that story.[219.62] [219.62][S02]The other story, which is as extreme, but on the other side[222.85] [222.85][S02]is, no, we do this because we must.[225.02] [225.02][S02]The kinds of problems we have are not[227.26] [227.26][S02]going to be solvable without this kind of technology.[230.63] [230.63][S01] I mean, on both of those,[232.63] [232.63][S01]I guess this underlying idea that there is no choice, really.[236.55] [236.55][S01]That we're marching forwards regardless of almost intention.[241.87] [241.87][S01]I mean, it's like we have to.[243.727] [243.727][S02] Oh, they're both extremes.[245.56] [245.56][S02]And I would not say that I think any one of them is right.[248.92] [248.92][S02]But I think it's always helpful to think about society[251.76] [251.76][S02]and think about politics as sort of a reasonable disagreement[254.82] [254.82][S02]between two extreme views and then[256.529] [256.529][S02]see where you can get to the middle, what[258.57] [258.57][S02]the optimal way of thinking about this is.[261.67] [261.67][S02]I think the story where we do this[264.24] [264.24][S02]because it's going to unlock progress has been undertold.[267.61] [267.61][S02]So that's why I'm more interested in that, perhaps.[269.928] [269.928][S01] Do you think there's been a shift in public mood?[272.47] [272.47][S01]I mean, with the explosion of generative AI into the scenes,[275.238] [275.238][S01]and now it's really, I think, a topic[276.78] [276.78][S01]that everybody is at the forefront[278.745] [278.745][S01]of a lot of people's minds.[279.87] [279.87][S01]Have you noticed a mood shift as a result?[282.123] [282.123][S02] Oh, there's been many mood shifts.[284.29] [284.29][S02]I mean, one of the things that happened[285.915] [285.915][S02]with almost all technologies is that you first[287.94] [287.94][S02]encounter an enormous sort of euphoria, almost where you go[291.21] [291.21][S02]like, oh, this is fantastic.[292.48] [292.48][S02]It's going to solve all my problems.[294.01] [294.01][S02]And then you sink into deep, deep depression because it[296.85] [296.85][S02]didn't.[297.49] [297.49][S01] The Gartner Hype Cycle.[298.77] [298.77][S02] It's the Gartner Hype Cycle,[300.687] [300.687][S02]or the way that Jose Borges put it in \"The Library of Babel,\"[304.24] [304.24][S02]where he said, there are finally, in this short story,[308.26] [308.26][S02]which is beautiful, he writes about how they find a library[310.95] [310.95][S02]with all of the possible permutations of the alphabet,[313.69] [313.69][S02]all possible books.[315.1] [315.1][S02]And he says, and they were so joyful[317.61] [317.61][S02]because they had found all possible books.[319.78] [319.78][S02]And then a few sentences later he writes,[321.88] [321.88][S02]and then they became very depressed[323.79] [323.79][S02]because they had found all possible books.[326.07] [326.07][S02]And that thing characterizes almost all technology cycles[330.18] [330.18][S02]as well, I think.[331.157] [331.157][S01] That we get excited first.[332.74] [332.74][S02] We get excited,[333.61] [333.61][S02]and then we go to the trough of disillusionment,[335.925] [335.925][S02]and to use your Gartner Hype Cycle,[338.11] [338.11][S02]and then to the plateau of reasonable expectations.[340.93] [340.93][S02]And I think generative AI, to some degree,[344.61] [344.61][S02]is probably going through exactly that hype cycle.[347.38] [347.38][S02]And we now see a lot of articles about,[349.2] [349.2][S02]where's the productivity growth?[351.06] [351.06][S02]Where is the economic boost that we hoped for?[353.13] [353.13][S02]Is the investment worth it?[354.63] [354.63][S02]And that is something that we saw around the internet as well.[357.66] [357.66][S02]We've seen it around computers.[359.25] [359.25][S02]We've seen it around all kinds of technologies[361.43] [361.43][S02]for the last, at least two centuries.[363.49] [363.49][S01] And we should expect to see it going forwards as[365.99] [365.99][S01]well, I guess, with the new technologies that are to come?[368.58] [368.58][S02] Yeah, absolutely.[370.038] [370.038][S01] OK.[370.73] [370.73][S01]All right, I guess it might make sense for us[373.58] [373.58][S01]just at the beginning of this to really sketch out[375.722] [375.722][S01]where we are at the moment in terms[377.18] [377.18][S01]of the landscape of regulation on artificial intelligence.[380.19] [380.19][S01]Could you give us a brief overview of where we are?[382.8] [382.8][S02] Of course.[383.967] [383.967][S02]No, so we're currently in a land of many different kinds[387.17] [387.17][S02]of solutions being tried.[388.68] [388.68][S02]So in the US, for example, the Biden administration[391.91] [391.91][S02]decided to put forward what it's called an executive order, which[395.42] [395.42][S02]is sort of a presidential decree in which they set out[398.12] [398.12][S02]a lot of different rules for the different agencies.[400.92] [400.92][S02]They wanted guidelines.[402.33] [402.33][S02]They wanted testing.[403.77] [403.77][S02]They had a ton of different rules[405.47] [405.47][S02]that they set out around artificial intelligence,[407.7] [407.7][S02]but they were mostly sort of industry-led[410.45] [410.45][S02]and they were industry close.[412.02] [412.02][S02]And then you have the UK that has[413.54] [413.54][S02]said that we believe that the best way to approach[415.85] [415.85][S02]this is to look at sectoral legislation, which is, how is[419.81] [419.81][S02]this actually going to be used when it's used in healthcare,[422.965] [422.965][S02]when it's used in education, when[424.34] [424.34][S02]it's used in different kinds of sectors in the economy.[427.64] [427.64][S02]And then you have the European Union[429.14] [429.14][S02]that shows this broad horizontal regulatory approach in which[432.89] [432.89][S02]they said, we are going to look at the risk[435.32] [435.32][S02]that these systems present widely to society.[438.66] [438.66][S02]And then, of course, you have China who[441.47] [441.47][S02]has been regulating as well.[442.92] [442.92][S02]And the way they have been regulating[444.462] [444.462][S02]has been mostly around figuring out[446.12] [446.12][S02]how does this technology shift power in a society, which[449.63] [449.63][S02]is really interesting.[450.63] [450.63][S02]So if you look at the kind of regulation[452.297] [452.297][S02]that China has put in place, it has been around a lot[454.82] [454.82][S02]about information power.[456.6] [456.6][S02]So what do we do with recommendation algorithms,[459.36] [459.36][S02]for example.[459.9] [459.9][S02]That's been one of their first focuses, foci.[462.48] [462.48][S02]And so I think that's an interesting approach as well.[466.11] [466.11][S02]So you have these four different ongoing experiments[469.46] [469.46][S02]in the regulatory space.[470.88] [470.88][S02]And we're learning from each and every one of them.[473.01] [473.01][S02]And I think it's actually not a horrible approach[475.85] [475.85][S02]to a new technology because there's so much evidence that we[479.9] [479.9][S02]need to gather, not just about the technology,[482.25] [482.25][S02]but also about the regulatory fit of these different kinds[485.18] [485.18][S02]of rules.[486.06] [486.06][S01] So, OK, I'm going to ask quite a basic question,[490.04] [490.04][S01]a sort of a foundational one, as it were.[492.8] [492.8][S01]Do you think that AI, as a distinct technology,[496.91] [496.91][S01]needs regulating sort of independently[500.12] [500.12][S01]of other technologies?[502.25] [502.25][S02] I don't think that we ever[504.56] [504.56][S02]only regulate the technology.[507.15] [507.15][S02]And the view I have is quite sociotechnical.[509.07] [509.07][S02]I think we regulate the uses of technology,[511.26] [511.26][S02]the design of technology, the deployment of technology.[514.23] [514.23][S02]But I don't think that you can talk[515.69] [515.69][S02]about regulating a technology, purely the technology itself.[519.69] [519.69][S02]And that is as it should be, because[521.539] [521.539][S02]to some degree, what we are regulating[523.58] [523.58][S02]is also how a technology affects power.[527.05] [527.05][S02]The German philosopher Hans Jonas[528.61] [528.61][S02]had a beautiful way of phrasing this, where he said,[530.78] [530.78][S02]all use of technology is the exercise of power.[533.45] [533.45][S02]And so the use of technology is what[535.84] [535.84][S02]we're usually thinking about when we think about regulation.[538.78] [538.78][S02]Another thing that I think is quite important to bring[541.87] [541.87][S02]into our discussion here is that regulation is not the same[545.26] [545.26][S02]as legislation.[546.7] [546.7][S02]Those are different.[547.94] [547.94][S02]One of the best regulatory models that has come up[550.18] [550.18][S02]in the last couple of decades is one that was launched[553.18] [553.18][S02]by Professor Lawrence Lessig in a book that he wrote back[556.36] [556.36][S02]in 1999, I want to say, called \"Code and Other Laws[560.11] [560.11][S02]of Cyberspace.\"[561.23] [561.23][S02]And in this book, he sort of outlines regulation[563.98] [563.98][S02]as consisting of four different components.[566.9] [566.9][S02]One is, of course, law, legislation.[569.24] [569.24][S02]We have laws and they regulate.[571.31] [571.31][S02]Another is architecture.[573.23] [573.23][S02]The way we actually build technology matters.[576.17] [576.17][S02]A third is markets because economic pressures also[579.73] [579.73][S02]regulate technology.[580.91] [580.91][S02]Some things are possible.[582.14] [582.14][S02]Some things are not possible.[583.43] [583.43][S02]And then lastly, the fourth is norms.[585.68] [585.68][S02]So these four forces of regulation impact the situation.[590.33] [590.33][S02]And this will become important later because when[592.54] [592.54][S02]we talk about regulating a technology, or regulating AI,[595.82] [595.82][S02]we can't just say we should regulate AI without also[599.14] [599.14][S02]discussing, OK, how does that regulation distribute[601.63] [601.63][S02]across these four different forces?[603.088] [603.088][S01] I want to pick up on something[604.838] [604.838][S01]you said a minute ago, actually, because you said that when[607.69] [607.69][S01]it comes to regulation, we regulate[609.455] [609.455][S01]the applications of technology rather than the technology[611.83] [611.83][S01]itself.[612.5] [612.5][S01]But I mean, that's not always true.[615.16] [615.16][S01]I'm thinking here about CRISPR, for instance, the gene editing[617.95] [617.95][S01]technology, which has regulations[620.44] [620.44][S01]on the technology itself, sort of independent[623.143] [623.143][S01]of its applications because of the recognition[625.06] [625.06][S01]of the potential harm to ecosystems,[627.76] [627.76][S01]or I don't know, to human health or human life[631.21] [631.21][S01]if you allow this kind of free use of the technology.[633.428] [633.428][S01]I mean, there are situations where the technology[635.47] [635.47][S01]itself is regulated, right?[636.73] [636.73][S02] But you're still regulating the use[638.23] [638.23][S02]because the reason you're regulating the technology,[640.27] [640.27][S02]or that you're going at the architecture[641.89] [641.89][S02]in this particular case, is because you're[643.64] [643.64][S02]worried about the ecosystems.[644.87] [644.87][S02]So you can imagine that you regulate that you say,[647.66] [647.66][S02]here's how this technology can be designed,[651.13] [651.13][S02]here's how this technology can be developed.[653.03] [653.03][S02]But it's usually always going to be in that sociotechnical frame[655.84] [655.84][S02]that we discussed.[656.87] [656.87][S01] But how about--[657.7] [657.7][S01]I mean, I'm trying to think of counterexamples, though,[659.992] [659.992][S01]because I think that there are some that, to me, it doesn't[662.8] [662.8][S01]seem completely black and white that it's always[665.02] [665.02][S01]the applications, not the technology.[666.65] [666.65][S01]I'm thinking here about, I don't know, like plutonium,[669.138] [669.138][S01]for instance.[669.68] [669.68][S01]I mean, there's very strict rules about plutonium transport,[673.76] [673.76][S01]about how it can be used, by whom.[677.198] [677.198][S02] How it can be used.[678.74] [678.74][S01] How it can be used.[680.032] [680.032][S01]True, very true.[680.699] [680.699][S02] And I don't think you're wrong.[682.74] [682.74][S02]I think you will find different kinds[684.94] [684.94][S02]of regulatory interventions across the entire ecosystem.[689.29] [689.29][S02]Let's rephrase it and say, OK, maybe it's[691.24] [691.24][S02]not about use or applications and technology.[694.13] [694.13][S02]Maybe what we're always looking for in regulation,[696.86] [696.86][S02]maybe what we always want to think about is harm.[699.25] [699.25][S02]So what is the potential harm?[701.0] [701.0][S02]And then we find a place in the social-technical context[705.02] [705.02][S02]where we believe that we can best prevent that harm.[708.33] [708.33][S02]That gives us another frame of thinking about this.[710.79] [710.79][S02]And I'm not wed to any frame.[712.483] [712.483][S02]I think that's a really good frame[713.9] [713.9][S02]to think into, because if you focus on harm instead,[716.398] [716.398][S02]and you can say, in the CRISPR case, for example,[718.44] [718.44][S02]we believe there can be harm to ecosystems.[720.42] [720.42][S02]The best way for us to prevent that[722.93] [722.93][S02]is to look at the CRISPR technologies design.[726.33] [726.33][S01] If we apply that back to AI,[728.4] [728.4][S01]how do you regulate the harms without regulating[732.74] [732.74][S01]the technology, if you see what I mean?[734.703] [734.703][S01]What does that actually look like in practice?[736.62] [736.62][S02] Well, one thing[737.51] [737.51][S02]you can do, let's take a very concrete example,[739.5] [739.5][S02]because it's always good to do this with examples.[741.583] [741.583][S02]So you say you want to regulate bias.[744.35] [744.35][S02]And you say, bias is a harm.[745.68] [745.68][S02]If a system is deeply biased, there's a harm here.[748.647] [748.647][S02]The first thing we then have to do, and this is quite helpful,[751.23] [751.23][S02]is we have to say, OK, what does that harm look like?[754.2] [754.2][S02]Well, we believe that somebody could have a decision made[757.34] [757.34][S02]against them that sort of negative for them,[759.93] [759.93][S02]because the system is biased in some way.[763.16] [763.16][S02]And so we want to prevent that harm of the individual[765.86] [765.86][S02]not being admitted to a school, not getting a welfare check,[769.64] [769.64][S02]or whatever we want to do.[770.91] [770.91][S02]So what we're looking at is that kind of harm.[773.48] [773.48][S02]Now, the next question we should ask in a regulatory context[776.57] [776.57][S02]is, where can we best prevent this?[779.01] [779.01][S02]Now we have choices.[780.3] [780.3][S02]Now we can say, I think the best way[782.0] [782.0][S02]to do this is to make sure that there is never[784.25] [784.25][S02]any bias in a data set.[786.36] [786.36][S02]And that's a technical choice, right?[788.19] [788.19][S02]We choose to go all the way back to the data set.[790.92] [790.92][S02]And we say, the data set can under no circumstances[793.67] [793.67][S02]contain bias.[795.06] [795.06][S02]Now that is very, very, very hard.[797.32] [797.32][S02]But it's a choice we can make.[798.57] [798.57][S02]We can say, that's what we want to do.[800.153] [800.153][S02]Another choice we can make is all decisions that[803.0] [803.0][S02]go through an algorithm of some kind,[805.65] [805.65][S02]or where a data set can be used, or where[807.47] [807.47][S02]there is a risk for bias, need to be reviewed by two people.[811.16] [811.16][S02]That's another regulatory solution[812.96] [812.96][S02]to that same kind of problem.[814.62] [814.62][S02]Now the question we are faced with,[816.22] [816.22][S02]and it's, OK, what's the most effective intervention we[818.89] [818.89][S02]can make here?[819.74] [819.74][S02]Let's take a very ineffective intervention.[822.02] [822.02][S02]We can say, yes, the data can be biased[824.24] [824.24][S02]and there's going to be no human review.[826.07] [826.07][S02]But the algorithm that uses the data[829.33] [829.33][S02]has to be built in such a way that it excludes[831.82] [831.82][S02]all possibilities of bias.[833.54] [833.54][S02]I don't even know how I would start doing that.[835.7] [835.7][S02]But you can say that kind of regulation[837.43] [837.43][S02]is what you could come up with because you have data,[839.84] [839.84][S02]you have algorithms, and you have humans in the system.[843.29] [843.29][S02]So as we try to prevent the harm,[845.355] [845.355][S02]we figure out where in the system[846.73] [846.73][S02]we most effectively can prevent the harm.[848.505] [848.505][S01] Does that not require[849.88] [849.88][S01]you knowing in advance what the potential harms will be?[852.59] [852.59][S02] Oh, that's such a good question[853.81] [853.81][S02]and such a good point, because that[855.52] [855.52][S02]is what's truly hard with new technology, right?[857.84] [857.84][S02]When we have emerging technology,[859.55] [859.55][S02]we don't exactly know what the harms are,[861.88] [861.88][S02]and that's why we end up trying to use certain blanket[864.46] [864.46][S02]principles for new technology.[866.54] [866.54][S02]One of them, for example, is the cost-benefit principle,[869.29] [869.29][S02]where we try to figure out what the possible costs are[871.54] [871.54][S02]with some approximation of the probability of unknown harm,[875.28] [875.28][S02]which is super hard.[876.33] [876.33][S02]And now there is a precautionary principle[878.09] [878.09][S02]where we say that you have to show that there's[880.048] [880.048][S02]no downside to this particular technology[882.26] [882.26][S02]under these particular conditions before you deploy it.[885.18] [885.18][S01] Which is what the UK Secretary of State for Science,[888.15] [888.15][S01]Innovation and Technology said.[889.52] [889.52][S01]They said that social media platforms, for instance,[891.69] [891.69][S01]must prove their products are safe before release, which is[894.832] [894.832][S01]exactly as you're describing.[896.04] [896.04][S02] Yes, which creates[896.93] [896.93][S02]another kind of evidentiary problem[898.52] [898.52][S02]because it's really hard to prove a negative.[902.09] [902.09][S02]So what you end up with then is that you[905.27] [905.27][S02]have to prove to some standard.[906.6] [906.6][S02]You have to prove to some probability[908.27] [908.27][S02]that this technology is not going to be harmful,[910.35] [910.35][S02]which may actually be quite a good standard in many cases,[913.44] [913.44][S02]but it can also be a highly restrictive standard,[916.41] [916.41][S02]depending on how high you put the bar for saying that there[919.82] [919.82][S02]will be no harm here.[921.568] [921.568][S01] Give me an example.[922.86] [922.86][S01]In what situation would it make it difficult?[924.9] [924.9][S02] Well, let's take a very simple example,[927.275] [927.275][S02]scanning X-rays for cancer.[929.45] [929.45][S02]And you say, OK, we're only going to actually release[933.15] [933.15][S02]this system if we can prove that it will never have or miss[937.26] [937.26][S02]any kind of cancer.[939.76] [939.76][S02]That's a super simple precautionary principle[942.42] [942.42][S02]example for you.[943.48] [943.48][S02]And that standard is too high.[946.08] [946.08][S02]And the standard we should set is it should probably[949.11] [949.11][S02]outperform human doctors.[951.36] [951.36][S02]That's a better standard.[952.42] [952.42][S01] We should be absolutely clear about this.[953.47] [953.47][S01]Why is that standard too high?[955.39] [955.39][S02] Oh, I think it's too high because then you[956.88] [956.88][S02]deploy no system.[957.9] [957.9][S02]Because there's no system that can do that.[960.31] [960.31][S02]You can never prove that a system[962.04] [962.04][S02]will be 100% accurate in any real life situation.[965.56] [965.56][S02]So that's too high a bar.[966.685] [966.685][S02]And then you can go down and you can say, OK,[968.56] [968.56][S02]you should have these different kinds of--[971.52] [971.52][S02]it should prove these different kinds of accuracy.[973.865] [973.865][S02]One you can say is that it should[975.24] [975.24][S02]be more accurate on average than human doctors are on average.[978.617] [978.617][S02]And the other is that you can say, OK,[980.2] [980.2][S02]if human doctors are also reviewing,[981.858] [981.858][S02]you can actually go a little bit lower and you can say,[984.15] [984.15][S02]it just needs to have a signal value that[985.858] [985.858][S02]helps the human doctor to interpret the X-ray.[988.287] [988.287][S02]And so you can say, it doesn't have to be accurate[990.37] [990.37][S02]100% of the time.[991.28] [991.28][S02]It can be accurate 40% of the time,[992.93] [992.93][S02]even if the doctors are accurate 80% of the time,[996.11] [996.11][S02]because the extra signal value might boost your 80 to 82,[999.29] [999.29][S02]in the hybrid system.[1000.43] [1000.43][S02]So you end up with these sort of interesting questions[1003.57] [1003.57][S02]about where do you put the bar.[1005.097] [1005.097][S01] That's a bar on benefit though, right?[1007.18] [1007.18][S01]Because there's also the bar on the other side, on cost.[1011.86] [1011.86][S01]And deciding that in advance is--[1015.09] [1015.09][S01]well, how do you decide where the bar is for potential costs[1018.72] [1018.72][S01]that you're willing to accept?[1020.35] [1020.35][S02] I mean, you can take the cancer example[1022.08] [1022.08][S02]again, the X-ray example.[1023.157] [1023.157][S02]You can say, it should absolutely never have[1024.99] [1024.99][S02]a false positive, because that creates harm in the sense[1027.96] [1027.96][S02]that it uses the healthcare system unnecessarily.[1030.7] [1030.7][S02]And for the individual who gets the diagnose,[1032.829] [1032.829][S02]it's like a personal catastrophe, and so on.[1034.98] [1034.98][S02]That's a challenge.[1036.48] [1036.48][S02]Because we know there's no system[1037.89] [1037.89][S02]that can fulfill that standard.[1039.31] [1039.31][S02]And then we scale back from there[1040.71] [1040.71][S02]to find the standard that we think is reasonable.[1042.994] [1042.994][S01] If we expand this out[1044.369] [1044.369][S01]to a slightly broader use of technology,[1047.38] [1047.38][S01]I mean, the example of sensitivity and specificity,[1050.02] [1050.02][S01]and false positive and false negatives[1051.76] [1051.76][S01]is a very clear example of where the potential harms are.[1054.68] [1054.68][S01]But I'm thinking about more broadly about where[1057.1] [1057.1][S01]you would get unexpected harms.[1058.76] [1058.76][S01]For instance, I'm thinking about the 2008 financial crisis here[1062.17] [1062.17][S01]and how a lot of that problem was[1064.66] [1064.66][S01]caused by the same models being used over and over again.[1069.592] [1069.592][S01]And no one quite noticing that that[1071.05] [1071.05][S01]meant that there was this common vulnerability across this kind[1073.99] [1073.99][S01]of common weakness.[1076.21] [1076.21][S01]And then only after the big catastrophe[1079.433] [1079.433][S01]happened that it's like, you can go back[1081.1] [1081.1][S01]and pick through the rubble and decide what the regulation[1083.32] [1083.32][S01]should have been.[1084.05] [1084.05][S01]How can you prevent those kind of big global events[1088.69] [1088.69][S01]from happening in technology, where you don't necessarily[1091.96] [1091.96][S01]understand all of the harms in advance?[1093.925] [1093.925][S02] It's very hard,[1095.3] [1095.3][S02]is the answer to the question.[1096.55] [1096.55][S02]What you can do, though, is that you[1098.29] [1098.29][S02]can try as far as you can go to do scenario planning,[1101.69] [1101.69][S02]to do red teaming, to go through what the technology can[1104.38] [1104.38][S02]do and figure out how it fails.[1106.25] [1106.25][S02]One of the most important questions about any technology[1108.82] [1108.82][S02]is actually how it fails.[1110.23] [1110.23][S02]And one of the things that we can do,[1112.46] [1112.46][S02]and that we consistently try to do in technology,[1115.04] [1115.04][S02]is to design graceful failure.[1116.96] [1116.96][S02]If you think about an airplane, for example,[1119.45] [1119.45][S02]it fails gracefully to the extent[1121.15] [1121.15][S02]that it can even glide if all of its systems go down.[1124.58] [1124.58][S02]And one of the things that I think[1126.003] [1126.003][S02]people are thinking actively about[1127.42] [1127.42][S02]is how can you replicate graceful failure[1129.67] [1129.67][S02]in other kinds of systems.[1131.06] [1131.06][S02]And it doesn't just have to be AI.[1132.74] [1132.74][S02]It can be any kind of system really, that we rely on.[1136.03] [1136.03][S02]So we're faced with as soon as we move outside,[1139.43] [1139.43][S02]and you're quite right, when we move outside[1141.67] [1141.67][S02]of the small clinical example I gave you,[1144.23] [1144.23][S02]the uncertainty increases.[1146.0] [1146.0][S02]And then at some point, we have to decide as a society,[1149.15] [1149.15][S02]how much uncertainty are we willing to tolerate[1152.92] [1152.92][S02]for what progress?[1154.383] [1154.383][S01] Underlying what you're[1155.8] [1155.8][S01]saying there is almost this implication[1157.78] [1157.78][S01]that we have to accept that there[1161.05] [1161.05][S01]is some probability of catastrophic failure,[1163.27] [1163.27][S01]that we can't mitigate against everything.[1165.5] [1165.5][S01]I mean, even the looking for graceful failures, as you say.[1169.51] [1169.51][S01]I mean, there's no way that we can sit here[1171.99] [1171.99][S01]and say, we are sure that we are going[1175.23] [1175.23][S01]to build technology that does not[1177.12] [1177.12][S01]risk existential crisis even.[1179.725] [1179.725][S02] No.[1180.6] [1180.6][S02]We can never be sure.[1181.87] [1181.87][S02]And that's part of the human condition, I think.[1184.5] [1184.5][S02]Uncertainty is a part of the human condition,[1186.79] [1186.79][S02]at least if we want progress.[1188.62] [1188.62][S02]And I think progress actually exacerbate that uncertainty.[1191.47] [1191.47][S02]I think Friedrich von Hayek, the Austrian economist,[1195.67] [1195.67][S02]at some point wrote, \"man has never been the ruler of her own[1199.29] [1199.29][S02]fate, and that is the reason we have progress.\"[1202.36] [1202.36][S02]And so there's a difficulty here, of course.[1206.98] [1206.98][S02]And I think I'm not saying that so we should with[1210.51] [1210.51][S02]abandon throw ourselves out into the darkness.[1213.49] [1213.49][S02]But what I am saying is that I think the notion of progress,[1216.73] [1216.73][S02]the notion of welfare, the notion of my kids[1218.64] [1218.64][S02]having it better than I do, does come with a bit of uncertainty.[1222.9] [1222.9][S02]And if I want that, then I also need[1224.94] [1224.94][S02]to tolerate managed uncertainty.[1227.32] [1227.32][S01] But I sort of wonder, I[1228.87] [1228.87][S01]wonder whether there's a question here about[1231.27] [1231.27][S01]must we do it at all.[1232.81] [1232.81][S01]If doing so means that there's unavoidable uncertainty, then[1238.2] [1238.2][S01]what about just not doing it?[1239.572] [1239.572][S02] Yeah, absolutely.[1241.03] [1241.03][S02]We can do that.[1241.78] [1241.78][S02]What we do then, as a society, or as a polity, is to say,[1244.86] [1244.86][S02]we choose not to progress.[1246.84] [1246.84][S02]From this point on, we will conserve what we have,[1249.18] [1249.18][S02]we will do nothing new, and take on no new risk.[1252.64] [1252.64][S02]Although risk is what drives societies forward.[1256.06] [1256.06][S02]And I think that's a political choice.[1258.37] [1258.37][S02]I mean, a lot of--[1259.77] [1259.77][S02]it's interesting, right?[1260.89] [1260.89][S02]Because a lot of what we do in policy,[1262.93] [1262.93][S02]and when we discuss regulation, is[1264.81] [1264.81][S02]that we're constantly exploring what John Rawls calls[1268.89] [1268.89][S02]reasonable disagreements.[1270.22] [1270.22][S02]So what are the two positions that[1272.52] [1272.52][S02]are warring with each other on this particular point?[1275.29] [1275.29][S02]And on this point, there's one position[1277.02] [1277.02][S02]that says, no, we should go for de-growth.[1278.86] [1278.86][S02]We should rescale our societies.[1281.14] [1281.14][S02]There's a lot to be had in an agricultural economy.[1284.93] [1284.93][S02]That's the extreme version of that.[1286.74] [1286.74][S02]And there's one that says, no, we need to go to the stars,[1289.157] [1289.157][S02]and we need to figure out how the universe works,[1291.198] [1291.198][S02]and we need to really make sure that we[1292.85] [1292.85][S02]continue to progress as human beings and build the future.[1295.56] [1295.56][S02]Those two are both defensible from a foundational ethical[1299.84] [1299.84][S02]perspective.[1300.69] [1300.69][S02]And at some point, we just have to choose.[1302.84] [1302.84][S02]And not choose binarily, but choose[1305.18] [1305.18][S02]where on the spectrum between the two we are.[1307.41] [1307.41][S01] But then we choose.[1308.88] [1308.88][S01]I mean, that's quite a big, like,[1310.85] [1310.85][S01]all encapsulating thing because there's no sort of opt-out here.[1315.35] [1315.35][S01]I mean, this is we, as all of humanity[1317.81] [1317.81][S01]collectively, we're either developing these technologies[1320.39] [1320.39][S01]or we're not.[1321.208] [1321.208][S02] Yeah, we are.[1322.5] [1322.5][S02]And I mean, I think that's an unavoidable truth[1325.76] [1325.76][S02]to some degree.[1326.52] [1326.52][S02]And that's why it's so, I think, important[1328.7] [1328.7][S02]to also invest in democracy.[1330.54] [1330.54][S02]And this is sort of a side point,[1332.04] [1332.04][S02]but it's the way we make regulation,[1334.44] [1334.44][S02]the methods whereby which we regulate technology actually[1337.4] [1337.4][S02]matter.[1338.04] [1338.04][S02]I do think that democracies have a better[1340.73] [1340.73][S02]way of discovering people's real views on this[1343.25] [1343.25][S02]and respecting and representing us[1345.44] [1345.44][S02]collectively than, for example, authoritarian systems have.[1349.15] [1349.15][S01] I do also want to talk about the role of private[1351.65] [1351.65][S01]companies in all of this, because I mean, of course,[1354.33] [1354.33][S01]we're having this conversation within the four walls of \"Google[1357.98] [1357.98][S01]DeepMind.\"[1358.53] [1358.53][S01]What do you think the role should[1360.32] [1360.32][S01]be in having private companies to shape this regulation?[1364.35] [1364.35][S02] I think ideally, if we do it right,[1366.63] [1366.63][S02]the way that all of these policy discussions work,[1369.77] [1369.77][S02]or what's sometimes called lobbying works,[1372.81] [1372.81][S02]if lobbying really works for society's best, it's a knowledge[1376.4] [1376.4][S02]exchange equation.[1378.0] [1378.0][S02]So we give knowledge around how this technology works.[1381.18] [1381.18][S02]And in exchange for that, we get influence over how[1384.17] [1384.17][S02]the regulation is shaped.[1385.7] [1385.7][S02]And that is the role I think private companies should[1388.13] [1388.13][S02]take to a large degree in all emerging technologies.[1391.8] [1391.8][S02]And that's not just true for AI.[1393.15] [1393.15][S02]It's true for life sciences, for example.[1396.15] [1396.15][S02]There is this knowledge asymmetry[1397.76] [1397.76][S02]where a lot of the knowledge sits in the private sector,[1400.52] [1400.52][S02]and we need to even it out.[1401.96] [1401.96][S02]And I don't mean to say, sometimes hear people say,[1404.59] [1404.59][S02]we need to educate policymakers.[1406.22] [1406.22][S02]I think that's wrong.[1407.21] [1407.21][S02]I think that we have to have some kind of mutual dialogue[1410.2] [1410.2][S02]here where we educate about the technology,[1412.55] [1412.55][S02]but policymakers educate us about democracy, institutions,[1415.67] [1415.67][S02]and values in ways that actually can help us think about, OK,[1418.97] [1418.97][S02]how do we shape this technology?[1421.01] [1421.01][S02]A good example is, we understood early on[1423.49] [1423.49][S02]that there was going to be requirements to figure out[1426.88] [1426.88][S02]how to deal with content that was synthetically produced.[1430.06] [1430.06][S02]And so those values, those political sort[1433.21] [1433.21][S02]of interests in some way also informed the work[1437.2] [1437.2][S02]that we have done on a very technical thing called SynthID.[1440.2] [1440.2][S02]So figuring out how watermarking works,[1442.52] [1442.52][S02]how we can make it possible to distinguish[1445.54] [1445.54][S02]real content from content that's been synthetically generated,[1449.36] [1449.36][S02]is something that you learn if you have this ongoing dialogue[1453.13] [1453.13][S02]between the political and the technical,[1454.88] [1454.88][S02]between the public and the private.[1456.44] [1456.44][S01] But I mean, is that a universal belief[1459.76] [1459.76][S01]among the tech industry?[1461.39] [1461.39][S01]Some people, I mean, there is sort[1463.0] [1463.0][S01]of a push for local expertise, internal expertise being[1468.31] [1468.31][S01]the ultimate thing to be considered.[1470.493] [1470.493][S02] I think the more[1471.91] [1471.91][S02]you can build, not just expertise,[1474.185] [1474.185][S02]but also understanding of this technology in the public sector,[1476.81] [1476.81][S02]the better it is.[1477.518] [1477.518][S02]So I'm not opposed to that, but I mean,[1480.007] [1480.007][S02]you put your finger on something that's important.[1482.09] [1482.09][S02]Of course, not everyone in technology sector[1484.0] [1484.0][S02]thinks the same way.[1485.05] [1485.05][S02]We have lots of different perspectives.[1487.01] [1487.01][S02]And it also depends a little bit on where your ideological sort[1490.51] [1490.51][S02]of home is.[1491.95] [1491.95][S02]An interesting example of a difference[1494.32] [1494.32][S02]that we've seen in the last 10 years or so[1496.36] [1496.36][S02]is the difference between internet policy and AI policy.[1500.03] [1500.03][S02]Because if you think back to the early days of the internet,[1502.69] [1502.69][S02]to 1996, the internet ethos amongst[1507.12] [1507.12][S02]people who discussed policy, and politics,[1508.87] [1508.87][S02]and the internet was highly libertarian,[1511.82] [1511.82][S02]extremely libertarian.[1513.17] [1513.17][S02]In 1996, John Perry Barlow publishes the \"Declaration[1517.35] [1517.35][S02]of Independence of Cyberspace,\" which is a fantastically[1520.74] [1520.74][S02]bombastic document.[1522.01] [1522.01][S02]It sort of says, you weary giants of steel and flesh,[1525.13] [1525.13][S02]leave us alone.[1526.51] [1526.51][S02]That's the basic tonality in the early internet policy debates.[1531.79] [1531.79][S02]In the AI policy debates, it's really different[1534.12] [1534.12][S02]because if you speak to young engineers[1535.8] [1535.8][S02]today, if you speak to anyone working on this, if you speak[1538.35] [1538.35][S02]to researchers, they say, no, this technology[1540.6] [1540.6][S02]is powerful enough that it needs to be regulated in some way.[1544.69] [1544.69][S02]And we have AI companies persistently saying[1548.04] [1548.04][S02]that this technology needs to be regulated.[1549.97] [1549.97][S02]Now, that begs another question.[1551.62] [1551.62][S02]When you say that, you should be able to answer how.[1554.28] [1554.28][S02]And I think that's a challenge, and we'll probably get to that.[1557.35] [1557.35][S02]But I do think that one of the things that you see[1559.77] [1559.77][S02]is that the ethos is different, and that's really important.[1562.57] [1562.57][S01] But there are some who[1563.987] [1563.987][S01]think that self-regulation should be the main lever that[1567.63] [1567.63][S01]gets used as we go forwards.[1568.797] [1568.797][S02] Self-regulation[1570.172] [1570.172][S02]can be good for two reasons.[1571.36] [1571.36][S02]One is because you want to keep sort of your ability[1574.41] [1574.41][S02]to do whatever you want.[1575.53] [1575.53][S02]So self-regulation becomes almost a defensive move.[1578.495] [1578.495][S02]You say, yes, yes, we'll self-regulate.[1580.12] [1580.12][S02]Another reason for self-regulating[1582.6] [1582.6][S02]can be that there is so much we don't[1584.73] [1584.73][S02]know that we need to make sure that we constantly[1587.91] [1587.91][S02]change the regulation.[1589.48] [1589.48][S02]Self-regulation is easier to change than legislation is.[1592.47] [1592.47][S02]But I don't think that the two are mutually exclusive.[1595.81] [1595.81][S02]Self-regulation with transparency[1597.69] [1597.69][S02]allows for legal review and auditing, for example.[1600.58] [1600.58][S02]So you can imagine different ways in which we as a company[1603.91] [1603.91][S02]could test our models and then review openly[1606.33] [1606.33][S02]how the models have been tested, like the AI Act says.[1609.61] [1609.61][S02]And then there is a possibility for legal and auditing[1613.68] [1613.68][S02]and review of that from outside of the company.[1616.18] [1616.18][S02]So you combine them in different ways.[1617.92] [1617.92][S01] I guess one of the other big counterarguments[1620.61] [1620.61][S01]is that this whole picture is really further colored[1624.96] [1624.96][S01]by the pursuit of profit, in the sense that companies who[1628.8] [1628.8][S01]have a hand in deciding regulations,[1631.26] [1631.26][S01]can you trust that they're doing so for public good,[1636.26] [1636.26][S01]or for competitive advantage?[1638.295] [1638.295][S02] And the profit motive[1639.92] [1639.92][S02]is always going to be something that we[1641.75] [1641.75][S02]should take into account.[1642.9] [1642.9][S02]But it shouldn't disqualify companies[1645.14] [1645.14][S02]from actually having some moral authority.[1647.97] [1647.97][S02]And I think the problem is that often the profit motive is used[1650.78] [1650.78][S02]as this blanket argument saying, you only[1652.52] [1652.52][S02]think about your profit.[1653.78] [1653.78][S02]And you can't do that, because if you only[1655.85] [1655.85][S02]think about your profit and mercilessly drive profit,[1658.62] [1658.62][S02]you're not going to be able to recruit the right talent.[1661.02] [1661.02][S02]People are not going to want to work on a company like that.[1663.52] [1663.52][S02]You're not going to be able to do deals[1665.532] [1665.532][S02]with others in the industry who want to be seen[1667.49] [1667.49][S02]as responsible and respectable.[1669.42] [1669.42][S02]There's a huge difference between a company that's[1671.75] [1671.75][S02]trying to maximize its profit quarter to quarter,[1674.67] [1674.67][S02]and a company that's trying to maximize its profits over 100[1677.45] [1677.45][S02]years.[1678.03] [1678.03][S02]And in order to do that, you actually[1679.79] [1679.79][S02]have to be a good player.[1680.91] [1680.91][S02]It's like any prisoner's dilemma that you reiterate.[1684.26] [1684.26][S02]The tit for tat rule.[1685.56] [1685.56][S02]What turns out to be most sustainable in a repeated game[1689.24] [1689.24][S02]is to be a trusted and somewhat good player.[1692.407] [1692.407][S02]I'm more worried about something else,[1693.99] [1693.99][S02]actually, if I can put on my worry hat.[1696.26] [1696.26][S02]And that is that I think currently,[1698.6] [1698.6][S02]investments in science from the private sector,[1701.42] [1701.42][S02]and this is true for a lot of different sectors,[1704.25] [1704.25][S02]are much, much higher than investments[1706.28] [1706.28][S02]from the public sector.[1707.66] [1707.66][S02]And I think that over the last couple of decades,[1710.79] [1710.79][S02]the prioritization of science as a public venture[1714.35] [1714.35][S02]has sadly fallen down a lot.[1717.39] [1717.39][S02]I actually would like to see more public investment[1719.96] [1719.96][S02]in science, more public work.[1721.9] [1721.9][S02]We've been, for example, deeply supportive[1723.65] [1723.65][S02]of the notion of a national AI resource in different shapes[1728.63] [1728.63][S02]or forms around the world, where we believe[1730.73] [1730.73][S02]that public investment in this technology[1732.74] [1732.74][S02]will actually give more public knowledge,[1735.45] [1735.45][S02]will give more insight into how the technology works,[1738.27] [1738.27][S02]and that will be a good thing.[1739.52] [1739.52][S01] And do you think then[1740.895] [1740.895][S01]with that as the background, there[1742.55] [1742.55][S01]is a risk that we'll end up in a future[1744.59] [1744.59][S01]where AI is in the hands of a very small number of companies?[1749.758] [1749.758][S02] There's always that risk.[1751.55] [1751.55][S02]And I think one of the things that we[1753.092] [1753.092][S02]see is that for many of the other markets or sectors that we[1756.19] [1756.19][S02]look at, we have this structure where[1758.26] [1758.26][S02]you have a few companies that are in some way or shape very[1762.91] [1762.91][S02]important in that sector.[1764.033] [1764.033][S02]You see it in pharma, you see it in telco, you see it in energy,[1766.7] [1766.7][S02]you see it in all of these different sectors.[1768.575] [1768.575][S02]And the question then is, if that is the case,[1772.16] [1772.16][S02]is that then harmful or is that OK?[1774.98] [1774.98][S02]Is it OK that that's sort of an effect[1776.8] [1776.8][S02]of industrial organization and economic pressures?[1779.8] [1779.8][S02]Is it something that holds back innovation?[1782.3] [1782.3][S02]Is it something that sort of reduces consumer welfare?[1785.39] [1785.39][S02]That's the question you have to ask.[1787.34] [1787.34][S02]I think I'm less worried about the market structure, where[1790.24] [1790.24][S02]you could say you have four or five really large pharma[1792.91] [1792.91][S02]companies or energy companies than I am about[1795.64] [1795.64][S02]how that impacts the overall change[1797.5] [1797.5][S02]and progress in those sectors.[1799.7] [1799.7][S01] OK, I want to get on to some really concrete examples[1802.54] [1802.54][S01]of regulation.[1803.27] [1803.27][S01]Those four different approaches exactly as you described.[1805.91] [1805.91][S01]Of course, the EU's AI Act came into force earlier this year.[1809.57] [1809.57][S01]And they've really gone for this risk-based approach.[1812.87] [1812.87][S01]So at the one end, the kind of unacceptable risks[1815.89] [1815.89][S01]are things like social credit scores.[1818.57] [1818.57][S01]And then down at the other end, the very low risk[1821.118] [1821.118][S01]stuff for things like chatbots.[1822.41] [1822.41][S01]Do you think this is broadly a good approach?[1825.045] [1825.045][S02] So I think it's[1826.42] [1826.42][S02]a good one to start with that.[1827.57] [1827.57][S02]I think it's actually when the commission originally[1829.737] [1829.737][S02]published its proposal, I think it was a really thoughtful way[1833.65] [1833.65][S02]to approach a problem that is very[1836.02] [1836.02][S02]much like the one we discussed.[1837.35] [1837.35][S02]What do you do with the technology[1838.767] [1838.767][S02]when it's hard to predict the harms?[1841.947] [1841.947][S02]And so in some way, I think it's a really good way[1844.03] [1844.03][S02]to think about things.[1844.947] [1844.947][S02]Where is the risk?[1846.77] [1846.77][S02]And then tune the regulation to the risk.[1850.63] [1850.63][S02]I think the downsides, of course,[1852.86] [1852.86][S02]as to how you assess risk and who assesses it.[1857.21] [1857.21][S02]It's been pointed out by people that if companies[1859.51] [1859.51][S02]get to assess their own risk entirely,[1861.62] [1861.62][S02]then that's going not to work, to your point[1863.8] [1863.8][S02]about self-regulation.[1864.792] [1864.792][S02]Now, that's not what the regulation says.[1866.5] [1866.5][S02]The regulation has other views about this.[1868.71] [1868.71][S02]Another argument against the risk-based regulation,[1872.62] [1872.62][S02]of course, is that you are not looking at all[1875.73] [1875.73][S02]at the upside or the benefits.[1878.02] [1878.02][S02]So you say, this is a high risk, but what if it's high reward?[1881.85] [1881.85][S02]Say we have a high risk application that[1884.58] [1884.58][S02]can be used to, we go back to the medical case, cure a disease[1888.72] [1888.72][S02]or sort of massively improve a situation.[1891.04] [1891.04][S02]Shouldn't we then at some point factor in the fact[1894.51] [1894.51][S02]that it's so high reward and say, OK, we're[1898.02] [1898.02][S02]going to regulate this, but we're also[1899.88] [1899.88][S02]going to give you a little bit of leeway[1901.83] [1901.83][S02]because we think that risk alone isn't the best metric here?[1905.16] [1905.16][S02]That criticism could be leveled against the system.[1908.85] [1908.85][S02]But I think that it's probably even harder[1912.21] [1912.21][S02]to assess reward than risk.[1914.1] [1914.1][S02]So it becomes even messier to do that.[1916.71] [1916.71][S02]And that is what you're seeing at work here,[1919.9] [1919.9][S02]if you scratch the surface, is a tension[1923.04] [1923.04][S02]between the American and the European approach,[1925.66] [1925.66][S02]because the American approach is the cost-benefit principle,[1928.947] [1928.947][S02]whereas you have the precautionary principle[1930.78] [1930.78][S02]in Europe.[1931.28] [1931.28][S02]And the precautionary principle says, let's look at the risk.[1934.02] [1934.02][S02]And the cost-benefit principle says,[1935.83] [1935.83][S02]let's look at the reward and the risk and see how they compare.[1939.3] [1939.3][S01] I do wonder a little bit[1940.8] [1940.8][S01]about that difference between the US and the EU approach,[1943.972] [1943.972][S01]because I guess now there's a reasonable chance[1945.93] [1945.93][S01]that those executive orders will be overturned.[1948.19] [1948.19][S01]Is there a chance that we could end up in a situation where[1951.18] [1951.18][S01]there is this big gap between the types of regulation that you[1954.177] [1954.177][S01]get in the US and the types of regulation that you have within[1956.76] [1956.76][S01]the EU, which then exacerbates the \"winner takes all\"[1960.3] [1960.3][S01]thing where really get this huge advantage in innovation[1963.54] [1963.54][S01]for American companies?[1964.907] [1964.907][S02] I think there is a definite risk[1966.99] [1966.99][S02]that regulation becomes determinative[1969.33] [1969.33][S02]in what kind of economic growth you get, what kind of welfare[1973.49] [1973.49][S02]you get.[1973.99] [1973.99][S02]And I do think that that's one of the things[1976.2] [1976.2][S02]that the European model will test.[1980.11] [1980.11][S02]And we will see if that kind of regulatory model[1982.76] [1982.76][S02]can carry the kind of innovation that Europe wants.[1986.22] [1986.22][S02]The hypothesis that the European regulator has[1989.75] [1989.75][S02]is that a clear playing field, sort of clear rules,[1994.25] [1994.25][S02]and a way to access the entire European market that's[1997.76] [1997.76][S02]been designed for the European AI Act[1999.77] [1999.77][S02]is going to boost innovation, and it's[2002.56] [2002.56][S02]going to boost investments, it's going[2004.21] [2004.21][S02]to boost the use of this technology[2005.74] [2005.74][S02]because people now know what to do.[2007.52] [2007.52][S02]And their hypothesis then, is that the US model will[2011.74] [2011.74][S02]create so much uncertainty.[2013.43] [2013.43][S02]And with US litigiousness and liability,[2016.85] [2016.85][S02]it's actually going to make it much slower[2019.27] [2019.27][S02]than if you had a comprehensive piece of legislation.[2021.83] [2021.83][S02]Those hypotheses will be tested out in real time.[2024.015] [2024.015][S01] Because I mean, there[2025.39] [2025.39][S01]are some situations where actually regulation has[2027.67] [2027.67][S01]accelerated innovation.[2029.03] [2029.03][S01]I mean, I'm thinking here about emissions on vehicles ending up[2033.52] [2033.52][S01]causing innovation in the electric vehicle market.[2035.98] [2035.98][S01]I mean, as a Formula One fan, the years that[2038.933] [2038.933][S01]are most exciting in terms of innovation[2040.6] [2040.6][S01]are the ones where new regulations come in.[2042.71] [2042.71][S01]And as you say, people have this framework[2044.89] [2044.89][S01]within which they know that they have to innovate.[2047.03] [2047.03][S02] Yeah, I think that can be absolutely true.[2049.53] [2049.53][S02]That depends on that regulation being very specific,[2052.04] [2052.04][S02]setting out a space that is viable.[2053.929] [2053.929][S02]And there are examples of this in technology regulation too.[2056.69] [2056.69][S02]I think that the Digital Millennium Copyright[2058.75] [2058.75][S02]Act, for example, which was put in place very early on,[2061.73] [2061.73][S02]had certain rules for what you had to do in order[2065.469] [2065.469][S02]to escape liability.[2067.07] [2067.07][S02]And one of the things was that you[2068.62] [2068.62][S02]had to have some way for people to claim their content[2071.08] [2071.08][S02]on the platform, which led to content ID, which[2074.193] [2074.193][S02]is one of the greatest innovations[2075.61] [2075.61][S02]that YouTube brought to the table.[2077.48] [2077.48][S02]And so you can see how that piece of innovation[2080.38] [2080.38][S02]was encouraged by the way that legislation was set up.[2083.21] [2083.21][S02]That's hard to do, but it's absolutely possible to do.[2086.042] [2086.042][S01] Do you think then, that we will see[2088.0] [2088.0][S01]similar types of innovation?[2089.54] [2089.54][S01]I mean, I'm thinking here about some of the generative AI stuff,[2093.76] [2093.76][S01]SynthID you mentioned earlier, which[2095.739] [2095.739][S01]has come about as a way to identify AI-generated content.[2101.52] [2101.52][S01]Do you think that the EU regulations[2104.79] [2104.79][S01]will force sort of wider adoption of those kind of ideas[2107.88] [2107.88][S01]and maybe even a wider agreement on standardized systems[2112.05] [2112.05][S01]for them?[2113.533] [2113.533][S02] I hope so.[2114.7] [2114.7][S02]I'm not entirely sure that we will see that,[2116.795] [2116.795][S02]because one of the things that you require in order[2118.92] [2118.92][S02]to get that innovative effect is that you[2121.8] [2121.8][S02]require a certain precision.[2123.37] [2123.37][S02]And currently, the European AI Act[2125.31] [2125.31][S02]is still being negotiated in a code of practice.[2128.26] [2128.26][S02]It's going to be implemented in all of the member states.[2130.69] [2130.69][S02]So we don't know yet if it will be clear enough[2133.5] [2133.5][S02]to drive that kind of innovation.[2135.1] [2135.1][S02]I think the innovation on SynthID and watermarking[2137.82] [2137.82][S02]comes from a global political dialogue, where you in the US,[2141.55] [2141.55][S02]in some of the United States, and in Europe, of course,[2144.04] [2144.04][S02]have sought for a means to figure out[2146.28] [2146.28][S02]how you can bring back some way of assessing the authority[2151.26] [2151.26][S02]or authoritativeness of content.[2153.4] [2153.4][S02]And I think that's a good example of how[2156.54] [2156.54][S02]regulation, or regulatory concerns, or policy concerns[2159.87] [2159.87][S02]can drive innovation.[2161.23] [2161.23][S01] I mean, I guess there's also the necessary,[2164.032] [2164.032][S01]but not sufficient argument here as well, which[2165.99] [2165.99][S01]is that it's all very well, sort of labeling, just as an example,[2169.12] [2169.12][S01]AI-generated content.[2170.56] [2170.56][S01]But if you're not doing it in real time, then[2173.04] [2173.04][S01]actually the harms can be caused--[2174.76] [2174.76][S01]you can't just retrospectively label things.[2177.2] [2177.2][S02] And the entire approach of labeling things[2179.7] [2179.7][S02]is interesting too, because we focus on the misinformation,[2182.19] [2182.19][S02]on the fake information.[2183.52] [2183.52][S02]And one of the things you can say[2184.983] [2184.983][S02]is that that's going to be really,[2186.4] [2186.4][S02]really, really hard because of the delay you mentioned[2188.7] [2188.7][S02]or because of the amount of information.[2190.51] [2190.51][S02]What we may want to do instead is to figure out,[2192.52] [2192.52][S02]is there another way for us to assign authority to content?[2195.16] [2195.16][S02]Can we actually say, not that this is misinformation[2198.0] [2198.0][S02]or this is synthetic, but this is really authoritative[2201.66] [2201.66][S02]because we have this flat information surface[2204.09] [2204.09][S02]that we're struggling with, and finding out what's actually what[2206.965] [2206.965][S02]or what's true in this flat information[2208.59] [2208.59][S02]surface is really hard.[2209.662] [2209.662][S02]Now, if you could build some peaks and valleys,[2211.62] [2211.62][S02]and if you could make this into an authoritative landscape,[2216.0] [2216.0][S02]then you could actually start to build in a way to sift out,[2220.43] [2220.43][S02]not just the fake, but also the highly valuable.[2223.41] [2223.41][S02]And if you look at the early days of newspapers,[2226.48] [2226.48][S02]that's actually what happened.[2227.73] [2227.73][S02]You started with a plethora of different newspapers[2231.02] [2231.02][S02]that were more or less libelous every single one of them.[2233.7] [2233.7][S02]Infamous scribblers, I think George Washington called them[2236.75] [2236.75][S02]in the US.[2237.51] [2237.51][S02]They would just make stuff up.[2240.36] [2240.36][S02]And then after a while, you saw the evolution[2243.65] [2243.65][S02]of publishing standards.[2245.22] [2245.22][S02]You saw some of the companies say,[2247.13] [2247.13][S02]this is printed in our newspaper,[2248.82] [2248.82][S02]so it was really relevant.[2250.23] [2250.23][S02]And you saw people building authority around content.[2253.59] [2253.59][S02]They weren't trying to attack the bad stuff.[2256.2] [2256.2][S02]They were trying to show you ways to the good stuff.[2259.14] [2259.14][S01] Yeah, I really like that idea.[2261.357] [2261.357][S01]Although, I also think that this is something that takes time.[2263.94] [2263.94][S01]And newspapers, as a really lovely example,[2268.08] [2268.08][S01]is something that developed over the course of decades rather[2271.19] [2271.19][S01]than the situation that we're in now where things are changing[2274.85] [2274.85][S01]in a lightning fast way.[2276.24] [2276.24][S01]Trust takes time to build up.[2279.05] [2279.05][S01]The EU actually also has this unacceptable risk category,[2283.34] [2283.34][S01]which I guess includes things like social credit scores[2286.37] [2286.37][S01]and lots of real time biometric identification.[2290.548] [2290.548][S02] And those are interesting[2292.34] [2292.34][S02]and we should discuss them.[2293.465] [2293.465][S02]I actually think it's quite useful for a society to say,[2296.31] [2296.31][S02]here are some uses that we're absolutely not going to condone.[2299.67] [2299.67][S02]Social scoring is one, for example.[2302.34] [2302.34][S02]I think that's good.[2303.39] [2303.39][S02]I think that is actually both helpful and healthy.[2307.53] [2307.53][S02]And I agree with that being a prohibited category.[2310.38] [2310.38][S02]And I think that part of the AI Act[2313.13] [2313.13][S02]is actually very reasonable because it points[2315.44] [2315.44][S02]to stuff that we do not want.[2317.07] [2317.07][S02]And it's interesting because the European Union has done[2319.88] [2319.88][S02]this in many different cases.[2321.27] [2321.27][S02]It actually has selected not to have[2324.53] [2324.53][S02]certain kinds of technology.[2326.01] [2326.01][S02]It did with the prohibited categories of AI,[2328.35] [2328.35][S02]but also historically, famously, it has done the same with GMOs.[2331.55] [2331.55][S01] How about within \"Google DeepMind, then?[2333.82] [2333.82][S01]Do you have certain applications that you consider off[2336.07] [2336.07][S01]limits here?[2336.65] [2336.65][S02] We have our AI principles.[2338.21] [2338.21][S02]So one example is that there are certain applications of weapons,[2340.63] [2340.63][S02]et cetera, that we wouldn't consider.[2342.35] [2342.35][S02]But generally go into the AI principles[2344.47] [2344.47][S02]and consulting them gives you a sense of how we try to balance[2347.08] [2347.08][S02]the different equities here.[2348.35] [2348.35][S01] I do wonder about in terms of particular projects,[2351.17] [2351.17][S01]how do you decide at \"Google DeepMind\"[2352.99] [2352.99][S01]which projects you will and won't get involved in?[2355.28] [2355.28][S01]How does that decision process happen?[2357.043] [2357.043][S02] There is a set of principles[2358.96] [2358.96][S02]against which we test ethically the different kinds of things[2362.23] [2362.23][S02]we do.[2362.9] [2362.9][S02]That is then taken to two different councils,[2365.77] [2365.77][S02]actually, that look, review, interview the engineers,[2369.47] [2369.47][S02]set out different conditions, and then give[2371.713] [2371.713][S02]a green light or a yellow light or a red light[2373.63] [2373.63][S02]and say generally, here's how you should think about this[2376.6] [2376.6][S02]from our ethical perspective.[2378.71] [2378.71][S02]So these ethical review boards is[2380.612] [2380.612][S02]something that we're not the only ones who have them,[2382.82] [2382.82][S02]there are others who have them too.[2384.278] [2384.278][S02]They are essentially a way of looking at how different[2386.77] [2386.77][S02]projects can be assessed, what we should do,[2389.137] [2389.137][S02]and if we think they should be, if we should proceed with them[2391.72] [2391.72][S02]or not.[2392.39] [2392.39][S01] If that's the frameworks that[2393.49] [2393.49][S01]exist at the moment, I think it's probably also[2395.448] [2395.448][S01]important to talk about frontier models.[2397.97] [2397.97][S01]So this idea of the most powerful AI models,[2402.115] [2402.115][S01]kind of at the very cutting-edge and the risk[2403.99] [2403.99][S01]that they might pose.[2404.99] [2404.99][S01]And so we got to talk to Demis to kick off this season.[2409.25] [2409.25][S01]And he said that he would recommend[2411.97] [2411.97][S01]kind of beefing up existing regulations at the moment[2414.31] [2414.31][S01]in the domains where we have AI, but also making sure[2420.28] [2420.28][S01]that you understand and test the frontier model simultaneously.[2423.26] [2423.26][S01]And then he said, start regulating around that maybe[2426.13] [2426.13][S01]in a couple of years time.[2427.293] [2427.293][S01]I mean, it's slightly unfair of me[2428.71] [2428.71][S01]to give you a quote from Demis and then[2430.465] [2430.465][S01]ask you to explain what he meant.[2431.84] [2431.84][S02] Yeah, but there you go.[2433.49] [2433.49][S02]Happy to.[2433.9] [2433.9][S02]Yes.[2434.23] [2434.23][S01] But there you go.[2435.438] [2435.438][S01]Why in a couple of years?[2436.73] [2436.73][S01]Why not now?[2437.72] [2437.72][S02] I think that the evidence base[2439.72] [2439.72][S02]we have is simply too slim.[2441.14] [2441.14][S02]I think what he is sort of hinting[2442.72] [2442.72][S02]at is that if we get into the habit of testing[2444.82] [2444.82][S02]these models now, when the capabilities are impressive,[2447.72] [2447.72][S02]but not dangerous, that will be a really good way for us to have[2451.32] [2451.32][S02]built habits, to have built regulation, to have built[2454.02] [2454.02][S02]institutions that can do this.[2456.42] [2456.42][S02]In the UK, for example, the AI Safety Institute[2458.88] [2458.88][S02]was put in place as a government initiative in the UK first,[2463.03] [2463.03][S02]but now there are AI safety institutes in many places,[2465.54] [2465.54][S02]to look at and understand more about foundation models[2469.47] [2469.47][S02]or really powerful frontier models.[2471.37] [2471.37][S02]And they've recruited scientists, recruited policy[2474.27] [2474.27][S02]people to really figure out how the government can learn more[2478.71] [2478.71][S02]about these models.[2480.04] [2480.04][S02]And then you might see different kinds of other capabilities[2482.79] [2482.79][S02]evolve through this particular testing work[2485.31] [2485.31][S02]or through this sort of analysis that we're doing.[2488.23] [2488.23][S02]And that, in turn, will help us to figure out what[2490.95] [2490.95][S02]does good regulation look like.[2492.68] [2492.68][S02]The challenge is the pace of evolution of these models.[2496.03] [2496.03][S02]And you can see that if you do a retrospective and say, OK, let's[2500.52] [2500.52][S02]assume that you were a regulator or a legislator,[2503.13] [2503.13][S02]and you're asked to regulate AI in 2015.[2507.27] [2507.27][S02]That's nine years ago.[2508.98] [2508.98][S02]What do you think that you would produce?[2511.3] [2511.3][S02]And I think that's--[2513.17] [2513.17][S02]if you do that, you'll see easily why he says, wait[2516.36] [2516.36][S02]a couple of years, because the knowledge we have,[2519.13] [2519.13][S02]the evidence we're getting, all of that[2521.55] [2521.55][S02]is going to be different in a couple of years, even,[2524.55] [2524.55][S02]I think to some degree, the unit of analysis.[2528.34] [2528.34][S02]One of the things that we have seen[2530.1] [2530.1][S02]is a huge focus on this notion of the model[2533.07] [2533.07][S02]as the object of regulation.[2535.44] [2535.44][S02]And that is, while it's a sort of a helpful first approximation[2539.61] [2539.61][S02]of where regulation should sit, one[2541.47] [2541.47][S02]of the things we're seeing now is[2543.33] [2543.33][S02]that as the effectiveness of the scaling laws[2546.0] [2546.0][S02]seem to be diminishing, people are building[2548.82] [2548.82][S02]several different models and then[2550.41] [2550.41][S02]having them work together in different ways.[2552.58] [2552.58][S02]So you get architectures that don't have a single model that[2557.16] [2557.16][S02]is feasible to regulate, but where[2559.74] [2559.74][S02]it's a collection, or a set, or a portfolio of models together[2562.79] [2562.79][S02]that produces the capability.[2564.66] [2564.66][S02]It's easy to see that those capabilities that[2566.78] [2566.78][S02]are produced by two models will be different.[2569.01] [2569.01][S02]It's a little bit like thinking about how would you[2571.61] [2571.61][S02]say that you want to regulate a symphony orchestra.[2574.7] [2574.7][S02]Do you regulate the guy with a triangle?[2576.99] [2576.99][S02]Do you regulate the conductor?[2579.12] [2579.12][S02]That seems to be pretty fair, but at some point,[2582.18] [2582.18][S02]the conductor also needs to make sure that everything else is[2585.08] [2585.08][S02]regulated.[2585.98] [2585.98][S02]And so I think to some degree, we're[2587.48] [2587.48][S02]moving into an architectural paradigm that may even[2590.81] [2590.81][S02]challenge the notion that these frontier models in the singular[2596.12] [2596.12][S02]are the right thing to regulate.[2598.05] [2598.05][S01] I mean, you said there a moment ago that we're[2601.01] [2601.01][S01]at the point now where these models are[2604.1] [2604.1][S01]impressive, but not harmful.[2606.12] [2606.12][S01]But then I think that there are some situations[2609.68] [2609.68][S01]in which large language models can be harmful already, right?[2612.66] [2612.66][S01]I mean, I'm thinking here in terms of amplification[2615.14] [2615.14][S01]of misinformation.[2616.14] [2616.14][S01]I'm thinking in terms of bias and disinformation.[2620.37] [2620.37][S01]I mean, is sitting back and waiting[2623.96] [2623.96][S01]until we have further information definitely[2626.45] [2626.45][S01]the right approach, or could we not move now[2630.2] [2630.2][S01]and be willing to be flexible as we go along?[2632.46] [2632.46][S02] Well, you're right and I'm wrong.[2633.63] [2633.63][S02]They can be harmful now.[2634.68] [2634.68][S02]But I think one thing that's important[2636.29] [2636.29][S02]is that I'm not recommending we sit back and wait.[2638.49] [2638.49][S02]One of the things that is also often stresses[2641.18] [2641.18][S02]is that we don't have the right benchmarks or the right tests[2643.82] [2643.82][S02]to figure out how we assess what the models can actually do.[2648.44] [2648.44][S02]And that means that we don't exactly[2649.94] [2649.94][S02]know what the possible harms are.[2652.11] [2652.11][S02]But this we can solve, and we can solve this[2654.92] [2654.92][S02]by attacking the problem now.[2656.7] [2656.7][S02]And that's why the scientific exploration and evaluation[2659.81] [2659.81][S02]of models is going to be incredibly important to invest[2662.6] [2662.6][S02]in and work with over the coming couple of years,[2665.43] [2665.43][S02]together with the public sector and third parties,[2668.81] [2668.81][S02]and together with other companies as well.[2671.97] [2671.97][S02]And that's why we set up the Frontier Model Forum, which[2674.72] [2674.72][S02]is an organization of different kinds of AI companies[2677.78] [2677.78][S02]coming together to try to build these evaluations[2680.65] [2680.65][S02]on a scientific basis.[2682.36] [2682.36][S01] So am I getting the picture here then,[2684.89] [2684.89][S01]that with these frontier models, with the real cutting-edge[2688.24] [2688.24][S01]of artificial intelligence, it's having[2690.403] [2690.403][S01]the discussion about the safety before they're[2692.32] [2692.32][S01]built, while they're built, and after they're built.[2695.15] [2695.15][S01]It's kind of every step of the process?[2696.923] [2696.923][S02] Yes.[2697.84] [2697.84][S02]And I do think that after is more important than it usually[2701.11] [2701.11][S02]is, because there's only so much pre-testing you[2704.26] [2704.26][S02]can do before you've actually built the artifact.[2706.34] [2706.34][S02]If you think about in regulation, you often say that,[2708.548] [2708.548][S02]should we have something be safe by design,[2711.02] [2711.02][S02]make sure that it's absolutely safe from the outset,[2713.66] [2713.66][S02]or should we have it be safe by testing?[2715.91] [2715.91][S02]So we test it afterwards and see if it's safe.[2718.22] [2718.22][S02]You do it differently for cars.[2719.66] [2719.66][S02]Cars you test and you build them safe.[2721.435] [2721.435][S02]You hope that the design is safe,[2722.81] [2722.81][S02]and then you test it a lot.[2723.95] [2723.95][S02]This is even more so.[2725.777] [2725.777][S02]What you want to do is you want to build it,[2727.61] [2727.61][S02]then you want to understand it, you want to test it,[2729.777] [2729.777][S02]you want to explore how it works.[2731.33] [2731.33][S02]You also want to explore how it works[2733.69] [2733.69][S02]in a sociotechnical context.[2735.17] [2735.17][S01] Well, I think, that's[2736.545] [2736.545][S01]the important point, the difference between the example[2739.208] [2739.208][S01]of cars, or I'm thinking pharmaceuticals as well,[2741.25] [2741.25][S01]COVID vaccine, for instance, where[2743.74] [2743.74][S01]you have the opportunity to test it before it[2746.08] [2746.08][S01]gets embedded in society.[2747.89] [2747.89][S01]You have a very particular controlled environment.[2751.48] [2751.48][S01]With some of these frontier models,[2753.04] [2753.04][S01]that might not necessarily always be possible.[2756.64] [2756.64][S02] Well, I think you can not--[2758.563] [2758.563][S02]so that's an interesting question.[2759.98] [2759.98][S02]Can you test it before you release it?[2762.1] [2762.1][S02]We do a ton of testing, and all AI companies do,[2765.1] [2765.1][S02]before we put anything out to the public[2767.59] [2767.59][S02]or make it generally available.[2769.495] [2769.495][S02]And that's one part of the processes[2770.995] [2770.995][S02]that you said Demis talks about, we[2772.75] [2772.75][S02]build the institutional habits of making sure[2775.42] [2775.42][S02]that we test, we red team with selected third parties,[2778.49] [2778.49][S02]we work with the AI safety institutes,[2780.44] [2780.44][S02]and we have this sort of period where we test it.[2783.2] [2783.2][S02]It's not as if we test it by throwing it out to market.[2785.812] [2785.812][S01] On that note, though, I mean,[2787.52] [2787.52][S01]there are people who have said that actually we[2789.73] [2789.73][S01]should be building kill switches into these things.[2791.87] [2791.87][S01]And that actually perhaps that should even[2793.62] [2793.62][S01]be part of regulatory control, is that an insistence on that.[2796.3] [2796.3][S01]What's your take on that idea?[2797.555] [2797.555][S02] I don't know what that would look like.[2799.93] [2799.93][S02]I would have to look at some kind of computer science[2802.41] [2802.41][S02]proof that shows me that a kill switch can be really effective.[2805.78] [2805.78][S02]Now, if we're talking about a system that's[2808.08] [2808.08][S02]sort of vastly more intelligent than us,[2810.33] [2810.33][S02]then I think it's probably hard to build a kill[2812.61] [2812.61][S02]switch into that system.[2814.722] [2814.722][S02]I can build a kill switch into my lamp.[2817.0] [2817.0][S02]It's the on/off button.[2818.12] [2818.12][S02]That's fine.[2818.62] [2818.62][S02]I can do that.[2819.48] [2819.48][S02]My lamp is likely to not outsmart me often,[2822.52] [2822.52][S02]and so I can do that.[2823.63] [2823.63][S02]But what do I do with the system that is not just highly complex,[2827.08] [2827.08][S02]but also distributed?[2828.588] [2828.588][S02]The notion of a kill switch goes back[2830.13] [2830.13][S02]to what we talked about before, that there's[2831.963] [2831.963][S02]a single model we can regulate.[2833.65] [2833.65][S02]Now, what if the model actually uses, say, 10, 20, 200,[2839.26] [2839.26][S02]200,000 different kinds of nodes in a network?[2842.89] [2842.89][S02]Where do you put the kill switch?[2845.25] [2845.25][S02]I mean, the way that people now try[2847.47] [2847.47][S02]to shut off the internet is by shutting it off at the border[2850.71] [2850.71][S02]and essentially saying, we will have no internet whatsoever,[2854.13] [2854.13][S02]because it's highly distributed.[2856.51] [2856.51][S02]A system like this, an AGI that is basically[2859.71] [2859.71][S02]a network of different kinds of models,[2861.7] [2861.7][S02]will be even more distributed and will have a natural tendency[2865.83] [2865.83][S02]to allocate or reallocate its tasks[2869.55] [2869.55][S02]or its capabilities across this network if any piece of it[2872.28] [2872.28][S02]is shut down.[2873.45] [2873.45][S01] If we do watch and wait,[2875.68] [2875.68][S01]I mean, are there any emerging capabilities[2878.22] [2878.22][S01]that you are particularly concerned about?[2880.83] [2880.83][S02] I think it's worth[2882.33] [2882.33][S02]looking more closely at persuasion and deception,[2885.76] [2885.76][S02]not because I think that they in themselves[2888.09] [2888.09][S02]are extremely dangerous, but because we need to understand[2890.65] [2890.65][S02]so much more about both persuasion and deception[2893.85] [2893.85][S02]and how they work if there is technology or AI in the mix.[2898.62] [2898.62][S01] But this does sound one area[2900.84] [2900.84][S01]in which there are some very clear sort of blanket rules that[2905.46] [2905.46][S01]might actually be useful.[2907.0] [2907.0][S01]A model should always let you know that it's a model.[2910.11] [2910.11][S01]You should never pretend to be a human.[2911.86] [2911.86][S01]Deception should never happen.[2913.11] [2913.11][S01]I mean, these sound sort of beginning[2916.01] [2916.01][S01]to get towards sort of regulatory ideas.[2919.898] [2919.898][S02] And the declaration[2921.44] [2921.44][S02]that you're interacting with an AI[2922.64] [2922.64][S02]is actually a part of many regulations[2924.29] [2924.29][S02]today, which I think is not a bad thing.[2926.593] [2926.593][S01] We started off by talking about the four[2928.76] [2928.76][S01]different approaches that you see globally towards regulation,[2931.62] [2931.62][S01]and the experiment that they are effectively[2934.04] [2934.04][S01]running in real time.[2936.162] [2936.162][S01]Do you think that we're going to get to a point[2938.12] [2938.12][S01]where international bodies will collaborate on AI regulation,[2941.7] [2941.7][S01]or do you think that we'll stay in this situation[2944.888] [2944.888][S01]where different regions of the world approach it differently?[2947.43] [2947.43][S02] There are different attempts[2949.347] [2949.347][S02]at international collaboration through the OECD,[2951.95] [2951.95][S02]which is not all of the countries in the world,[2954.32] [2954.32][S02]through the UN, for the G7.[2956.25] [2956.25][S02]So there are these nascent attempts[2958.16] [2958.16][S02]to try to figure out what are the overarching rules[2961.082] [2961.082][S02]that we want to make sure are true.[2962.54] [2962.54][S02]Things like, AI should benefit humanity.[2964.76] [2964.76][S02]Very, very general, but still some kind[2967.58] [2967.58][S02]of at least indication to what we think the policy tonality[2971.0] [2971.0][S02]should be.[2972.44] [2972.44][S02]I think that there will be challenges here,[2974.88] [2974.88][S02]and I think those challenges are exacerbated[2976.76] [2976.76][S02]by geopolitical pressures and tensions.[2978.95] [2978.95][S02]They're also exacerbated by the fact[2980.45] [2980.45][S02]that there is a lot of national competitiveness involved.[2983.7] [2983.7][S02]And so we're not in a period of human history[2987.86] [2987.86][S02]where multilateral institutions are super strong.[2990.48] [2990.48][S02]So while I'm very happy that we have[2993.553] [2993.553][S02]these international collaborations[2994.97] [2994.97][S02]and the early attempts, I'm not entirely sure[2997.91] [2997.91][S02]how we progress from the very general agreements[3001.3] [3001.3][S02]to, for example, limits on the use of AI in simple situations.[3005.75] [3005.75][S02]Say that you wanted to take the European AI Act[3008.02] [3008.02][S02]and say we should universally outlaw social scoring with AI.[3013.42] [3013.42][S02]Would that work or are there countries[3015.67] [3015.67][S02]that actually believe that social scoring is a net good?[3018.32] [3018.32][S02]Well, there are, and so we should think through,[3020.89] [3020.89][S02]where are the areas where international collaboration can[3023.56] [3023.56][S02]happen?[3024.38] [3024.38][S01] I mean, one place that has not yet[3027.972] [3027.972][S01]planted their flag about regulation at all is the UK.[3031.403] [3031.403][S01]Do you think that when it does, it[3032.82] [3032.82][S01]will end up being closer to the EU or to the US?[3037.71] [3037.71][S02] The sort of approach in the UK has been[3040.603] [3040.603][S02]advertised by the new Labour government as something where[3043.02] [3043.02][S02]they're going to put a UK bill out before the end of the year[3046.05] [3046.05][S02]for consultation.[3047.41] [3047.41][S02]And they're looking at a set of narrow regulations[3051.48] [3051.48][S02]that will put them closer to the US, I think,[3053.95] [3053.95][S02]than to the EU, which is interesting in many different[3057.2] [3057.2][S02]ways.[3057.7] [3057.7][S02]And then I think they'll continue to push.[3059.88] [3059.88][S02]And this I really think is important.[3061.542] [3061.542][S02]They'll continue to push for the sectoral application[3063.75] [3063.75][S02]of the rules, because I think there[3065.37] [3065.37][S02]is a lot of value in how a sector or a set of professionals[3069.09] [3069.09][S02]work with this.[3070.436] [3070.436][S02]I think it would be very, very valuable[3072.99] [3072.99][S02]to have a set of nurses or teachers[3074.52] [3074.52][S02]talk about what would make most sense to them[3076.83] [3076.83][S02]in their professional day to day when[3079.17] [3079.17][S02]it comes to the use of these technologies.[3081.13] [3081.13][S02]That kind of normative regulation[3083.94] [3083.94][S02]is actually really important.[3085.66] [3085.66][S02]And if you can encourage that on top[3087.57] [3087.57][S02]of whatever narrow regulation you believe is needed,[3090.25] [3090.25][S02]I think you can actually progress in[3092.22] [3092.22][S02]what is another objective of regulation,[3094.51] [3094.51][S02]and that is the diffusion of the technology in the economy.[3097.18] [3097.18][S02]One way to measure the effectiveness and the value[3099.9] [3099.9][S02]of regulation is actually to look[3101.85] [3101.85][S02]at how fast is this technology diffusing through the economy.[3105.395] [3105.395][S02]And that's something that I think[3106.77] [3106.77][S02]we don't discuss often enough, because technology diffusion[3109.74] [3109.74][S02]also unlocks benefits and also unlocks welfare.[3112.3] [3112.3][S02]And so as we look at different kinds of regulation,[3116.02] [3116.02][S02]that's actually one, not the only one,[3117.87] [3117.87][S02]but one interesting metric to think through.[3120.113] [3120.113][S01] I mean, I think what's[3121.53] [3121.53][S01]been really interesting about this conversation is[3123.613] [3123.613][S01]that I don't feel as though there are[3125.37] [3125.37][S01]any aspects of regulation where you're like, no,[3127.72] [3127.72][S01]that's definitely a bad idea.[3129.357] [3129.357][S01]I mean, it's still sort of feels like you're[3131.19] [3131.19][S01]open to all of the potential options on the table.[3133.45] [3133.45][S01]I do wonder, though, in terms of-- is that fair?[3135.485] [3135.485][S02] I have no real views.[3137.11] [3137.11][S02]Yes.[3137.61] [3137.61][S01] No, no, no, no, no.[3138.96] [3138.96][S02] No, no.[3139.36] [3139.36][S02]But I do think there are bad things that you can do.[3141.28] [3141.28][S02]And one of them is banning technology.[3142.9] [3142.9][S02]I think that's always wrong.[3144.19] [3144.19][S02]So there's a hard line.[3145.653] [3145.653][S02]You shouldn't ban technology because you[3147.32] [3147.32][S02]don't know how it works.[3148.77] [3148.77][S02]I also think it's probably wrong to try[3151.19] [3151.19][S02]to regulate too much at the data level[3153.11] [3153.11][S02]and say, these data sets must be absolutely perfect,[3156.228] [3156.228][S02]because I think that will be really, really hard to do.[3158.52] [3158.52][S02]There are some aspects of that in the EU AI Act[3161.317] [3161.317][S02]that I think will be really difficult,[3162.9] [3162.9][S02]but can be cleared up through government practice[3165.65] [3165.65][S02]and implementation in different ways.[3167.49] [3167.49][S02]But it is also, to be honest, a wide open field.[3171.57] [3171.57][S02]And we need to not foreclose any of the models we have,[3176.34] [3176.34][S02]but explore them equally and figure this out as we go along.[3179.793] [3179.793][S01] But then on the flip side of that,[3181.71] [3181.71][S01]what do you think is the most urgent thing that[3183.668] [3183.668][S01]requires attention?[3184.47] [3184.47][S01]I mean, the wait and see slowly, slowly, nice[3188.12] [3188.12][S01]and carefully is all good.[3189.93] [3189.93][S01]But where do you see that there's the most urgency?[3192.503] [3192.503][S02] I think building the institutions,[3194.67] [3194.67][S02]making sure we have the science that makes it possible[3196.97] [3196.97][S02]for us to understand this technology,[3198.63] [3198.63][S02]whether it's a single model or a network of models,[3201.03] [3201.03][S02]making sure that there are ways of testing and benchmarking[3203.66] [3203.66][S02]the system, not sitting and waiting,[3205.47] [3205.47][S02]but constantly building up the public knowledge about these[3209.79] [3209.79][S02]together with the private sector,[3211.38] [3211.38][S02]I think is super important.[3213.03] [3213.03][S02]I want a regulatory curiosity, and I[3214.94] [3214.94][S02]think that's actually what we have[3217.52] [3217.52][S02]in a lot of the different regions we described.[3219.59] [3219.59][S02]Regulatory curiosity, understanding the technology,[3222.012] [3222.012][S02]figuring out where are the levers that we could[3223.97] [3223.97][S02]pull if we needed to pull them.[3225.54] [3225.54][S02]What lever is better than another lever?[3227.28] [3227.28][S02]Is it better to regulate the data, the algorithm, the system,[3230.28] [3230.28][S02]the user, the deployment?[3231.81] [3231.81][S02]Where in the value chain should we be?[3233.75] [3233.75][S02]That regulatory curiosity is absolutely[3236.06] [3236.06][S02]essential and urgent because the technology[3238.25] [3238.25][S02]is developing so fast.[3239.727] [3239.727][S01] What a fascinating conversation.[3241.56] [3241.56][S01]Thank you so much.[3242.55] [3242.55][S01]That was really, really wonderful.[3244.06] [3244.06][S01]Thank you.[3244.56] [3244.56][S02] Thanks for having me.[3245.31] [3245.31][S02]Thank you.[3246.205] [3246.205][S01] That was a discussion[3247.58] [3247.58][S01]that probably left us with more questions than answers.[3250.47] [3250.47][S01]But I do think that one takeaway for me,[3253.41] [3253.41][S01]anyway, is the extraordinary complexity that[3256.97] [3256.97][S01]is involved in regulating this technology,[3259.58] [3259.58][S01]to allow for benefits while mitigating against harms.[3263.16] [3263.16][S01]And there really aren't simple solutions here.[3266.37] [3266.37][S01]There aren't any quick wins, no silver bullets,[3268.77] [3268.77][S01]no ideas that are even guaranteed[3270.92] [3270.92][S01]to continue to work across time.[3273.9] [3273.9][S01]Everything is a balance here in every direction.[3277.35] [3277.35][S01]But as hard as this will be to get right,[3280.89] [3280.89][S01]the one thing we can be sure of, as Nicklas[3283.19] [3283.19][S01]said, the only way to proceed without any risk of harms[3288.05] [3288.05][S01]is to not proceed at all.[3290.775] [3290.775][S01]You have been listening to \"Google DeepMind--[3292.65] [3292.65][S01]The Podcast\" with me, Professor Hannah Fry.[3295.1] [3295.1][S01]If you enjoyed that episode, then[3296.99] [3296.99][S01]do subscribe to our YouTube channel.[3299.1] [3299.1][S01]You can also find us on your favorite podcast platform.[3302.82] [3302.82][S01]And we have got plenty more episodes[3305.6] [3305.6][S01]on a whole range of topics to come, so do check those out too.[3309.63] [3309.63][S01]See you next time![3310.64] [3310.64][MUSIC PLAYING][3314.59]"}
{"file_name": "audio/val_000016.wav", "transcription": "[0.0][MUSIC PLAYING][3.458] [6.916][S01] Welcome back to \"Google DeepMind: The Podcast\"[9.76] [9.76][S01]with me, your host, Professor Hannah Fry.[12.13] [12.13][S01]In this episode, we get to speak to one[14.35] [14.35][S01]of the most legendary figures in the world of computer[17.23] [17.23][S01]science, Jeff Dean.[19.2] [19.2][S01]He was there in the late 1990s writing the code that would turn[22.81] [22.81][S01]Google from a small startup to the multinational company it is[27.52] [27.52][S01]today.[28.4] [28.4][S01]Jeff spearheaded TensorFlow, one of the programming tools[31.75] [31.75][S01]responsible for the democratization[33.67] [33.67][S01]of machine learning.[35.06] [35.06][S01]He's the one who pushed the boundaries[36.97] [36.97][S01]of artificial intelligence in the direction[39.22] [39.22][S01]of large-scale models.[41.24] [41.24][S01]And if that wasn't enough, he also co-founded[44.29] [44.29][S01]Google's AI research project, Google Brain,[47.42] [47.42][S01]and was one of the earliest pioneers of a new neural network[50.86] [50.86][S01]architecture called Transformers.[53.36] [53.36][S01]Not sure that one's going to catch on.[55.27] [55.27][S01]I think this is why people joke that Jeff Dean's resume just[58.84] [58.84][S01]lists the things that he hasn't done.[60.39] [60.39][S01]It's shorter that way.[61.9] [61.9][S01]More recently, Jeff, as Google's chief scientist,[64.819] [64.819][S01]has occupied one of the most important seats[66.89] [66.89][S01]around the table as the two great AI arms of Alphabet[70.31] [70.31][S01]have merged, DeepMind and Google Brain.[73.2] [73.2][S01]His latest baby, which he co-parents,[75.39] [75.39][S01]is known as Gemini, which takes large language models well[78.89] [78.89][S01]beyond language alone.[80.67] [80.67][S01]Gemini is a multimodal model which can understand text, code,[85.05] [85.05][S01]audio, images, and video.[87.36] [87.36][S01]It is AI through and through and almost certainly the direction[91.79] [91.79][S01]that Google Search itself is heading in.[94.265] [94.265][S01]Jeff, thank you so much for joining me.[95.89] [95.89][S02] Thank you for having me.[97.08] [97.08][S02]It's a delight to be here.[98.24] [98.24][S01] So, OK, 25 years, a quarter of a century at Google I[102.38] [102.38][S01]want to a little bit of what it was like back in the early days,[105.055] [105.055][S01]right, like in the '90s, when you first joined,[107.22] [107.22][S01]when Google wasn't the sort of slick organization that it is[109.9] [109.9][S01]now.[110.4] [110.4][S01]Was it all a lot of laptops with stickers on it and sort of[113.645] [113.645][S01]coding in flip-flops?[115.25] [115.25][S02] Sadly, it was pre-laptops.[116.87] [116.87][S01] Pre-laptops.[117.87] [117.87][S02] Well, mostly.[119.03] [119.03][S02]Yeah, we all had those giant CRT-based--[120.71] [120.71][S01] Yeah, of course.[121.25] [121.25][S02] --monitors, those pre-LCD monitors.[123.63] [123.63][S02]So they took up a lot of desk space.[126.02] [126.02][S01] Not very portable.[127.27] [127.27][S02] My desk was like a door on two sawhorses.[131.21] [131.21][S02]That was the-- and you could adjust it yourself[134.36] [134.36][S02]by getting under the desk and standing up with your back[136.82] [136.82][S02]to get it to the next higher setting.[138.86] [138.86][S01] Really?[139.76] [139.76][S02] Yeah.[141.057] [141.057][S01] Amazing.[141.89] [141.89][S02] And when I started, we[143.265] [143.265][S02]were in this small office area, actually not-- maybe three times[146.33] [146.33][S02]as big as this room.[148.117] [148.117][S01] The whole of Google.[149.45] [149.45][S02] The whole of Google on University Avenue[151.575] [151.575][S02]in Palo Alto above what's now a T-Mobile cell phone store.[156.2] [156.2][S02]And the really fun and exciting thing in those days[162.11] [162.11][S02]was we were a small company, but we[166.61] [166.61][S02]could see that people were using our service more and more[169.13] [169.13][S02]because we were providing good-quality, high-quality[172.28] [172.28][S02]search.[173.42] [173.42][S02]And you could see your traffic growing[175.82] [175.82][S02]day over day, week over week.[177.65] [177.65][S02]So we'd always be trying to not melt on[179.9] [179.9][S02]Tuesday at noon, which was the peak traffic hour of the week.[184.14] [184.14][S02]And that would require we deploy more computers quickly,[188.3] [188.3][S02]that we optimize our code to make it run faster,[191.37] [191.37][S02]that we come up with new and interesting innovations[194.45] [194.45][S02]for next month's index that make it able to serve more[197.9] [197.9][S02]users with the same hardware.[199.108] [199.108][S01] I can imagine it being very exciting.[201.15] [201.15][S01]Was there a moment when you guys realized--[203.03] [203.03][S01]was there a moment when you were like,[204.613] [204.613][S01]ah, OK, this is really going to be big?[206.63] [206.63][S02] I mean, I think you could[208.13] [208.13][S02]see that from the very earliest days when I was joining--[211.11] [211.11][S02]I joined the company because our traffic was growing really fast,[214.49] [214.49][S02]and we felt like by focusing on returning really high-quality[220.25] [220.25][S02]search results and doing that quickly and giving users what[224.66] [224.66][S02]they want--[225.3] [225.3][S02]we actually want to get people off our site[227.63] [227.63][S02]as quickly as possible to the information they need--[230.96] [230.96][S02]that was a winning proposition.[233.73] [233.73][S02]And users seemed to like our service.[235.83] [235.83][S02]And so it seemed reasonably promising,[240.21] [240.21][S02]I would say, from even the early days.[241.793] [241.793][S01] There's quite a big gap[243.252] [243.252][S01]between reasonably promising, though,[244.81] [244.81][S01]and sort of what it ended up being.[246.6] [246.6][S01]Has it been a surprise to you-- like, to all of you?[249.74] [249.74][S02] I mean, I think there's been a bunch of things[252.14] [252.14][S02]that we've branched out into that were obviously[255.308] [255.308][S02]hard to anticipate.[256.1] [256.1][S02]Autonomous vehicles-- it's hard to fathom[259.04] [259.04][S02]that when you're working on a search engine.[261.95] [261.95][S02]But I think the sort of gradual broadening out[265.67] [265.67][S02]of our portfolio of products to other kinds of information[269.54] [269.54][S02]makes a lot of sense.[270.72] [270.72][S02]So going from public web pages to helping users organize[275.27] [275.27][S02]their own email with Gmail, things[277.31] [277.31][S02]like that, those are sort of natural evolutions of things[280.58] [280.58][S02]that solve real problems people have.[284.48] [284.48][S02]And they get you in the state of, OK, well,[287.1] [287.1][S02]now we don't just have one product.[289.53] [289.53][S02]We have a handful of products that people[292.22] [292.22][S02]use fairly regularly.[293.345] [293.345][S01] And I sort of wonder.[294.72] [294.72][S01]Looking back through all that time,[296.178] [296.178][S01]do you think that Google was always a search company,[299.57] [299.57][S01]or do you think it was an AI company sort of pretending[302.48] [302.48][S01]to be a search company?[303.807] [303.807][S02] Yeah, I mean, I think a lot[305.39] [305.39][S02]of the problems we wanted to tackle as a company really[309.23] [309.23][S02]were sort of ones that would require[312.53] [312.53][S02]AI to really, truly solve.[314.22] [314.22][S02]And so along the way, in a long period of time, 25 years,[319.01] [319.01][S02]we've been progressively tackling some of those hard AI[323.09] [323.09][S02]problems and making progress on them[324.62] [324.62][S02]and then using the new techniques that are now[327.32] [327.32][S02]starting to work in the midst of search[329.9] [329.9][S02]and in the midst of all of our other products.[332.04] [332.04][S01] Do you think that Google will always[334.04] [334.04][S01]be a search company, or do you think-- is it[336.08] [336.08][S01]even a search company now?[337.32] [337.32][S01]Is it changing?[338.54] [338.54][S02] Well, one of the things[339.957] [339.957][S02]I really like about Google is our mission[341.93] [341.93][S02]is still incredibly relevant even 25 years later-- organize[348.022] [348.022][S02]the world's information and make it universally[349.98] [349.98][S02]accessible and useful.[351.66] [351.66][S02]And I feel like Gemini is really helping[354.18] [354.18][S02]us push in the direction of understanding[357.33] [357.33][S02]lots of different kinds of information--[359.53] [359.53][S02]so textual data, software code, which is kind of texty in nature[365.83] [365.83][S02]but very structured in certain ways,[368.08] [368.08][S02]but also all the other kinds of modalities of input[371.46] [371.46][S02]that humans are sort of fluent in.[373.57] [373.57][S02]We're naturally-- we read stuff, but we also[376.56] [376.56][S02]see stuff with our eyes and hear stuff with our ears.[379.95] [379.95][S02]And you want models to be able to sort of take in information[383.91] [383.91][S02]in its many forms and also produce information in text form[390.09] [390.09][S02]or maybe generate audio so that you[392.22] [392.22][S02]can have a conversation with the model[395.04] [395.04][S02]or produce imagery, if that's appropriate,[397.68] [397.68][S02]or annotate text with graphs or things like that.[400.84] [400.84][S02]So we're really trying to make a single model that[403.95] [403.95][S02]can take in all those modalities and produce all those modalities[406.77] [406.77][S02]and use that capability when it makes sense.[410.718] [410.718][S01] Can you remember when you first[412.51] [412.51][S01]came across neural networks?[414.41] [414.41][S02] Oh, yeah, actually.[416.35] [416.35][S02]So neural networks have had an interesting history.[420.044] [420.044][S02]AI is a quite old discipline, and the early phases of AI[424.78] [424.78][S02]were about how do we define rules about how things work.[429.11] [429.11][S02]So that was like the '50s, '60s, '70s, to some extent.[433.39] [433.39][S02]And then neural networks kind of came along in the '70s and had[438.61] [438.61][S02]a wave of excitement in the late '80s and early '90s.[442.36] [442.36][S02]And I was actually an undergrad in 1990[445.39] [445.39][S02]at University of Minnesota.[446.81] [446.81][S02]And I was taking a class in parallel processing, which[449.44] [449.44][S02]is this idea of how do you take problems and break them down[453.49] [453.49][S02]into pieces that can be done on different computers.[455.78] [455.78][S02]And then kind of in conjunction, all those computers[459.97] [459.97][S02]work together to solve a single problem.[462.305] [462.305][S01] I guess this is the point where the computing[464.68] [464.68][S01]power wasn't quite what we have today.[466.88] [466.88][S01]So how do you make computers work as a team?[468.8] [468.8][S02] Right, yeah.[469.758] [469.758][S02]And at that time, neural networks[472.49] [472.49][S02]was a sort of particular approach to machine learning[476.93] [476.93][S02]and AI that involved kind of very kind[481.79] [481.79][S02]of crude approximations of how we[484.49] [484.49][S02]think real human or other brains work with neurons.[487.57] [487.57][S02]So that's why they're called neural networks because they're[490.07] [490.07][S02]made up of artificial neurons.[492.74] [492.74][S02]And artificial neurons have connections to other neurons[496.04] [496.04][S02]below them.[496.89] [496.89][S02]And then they look at kind of the signals that come up[500.18] [500.18][S02]from those artificial neurons, and they[502.07] [502.07][S02]decide how interested are they in that particular pattern[506.75] [506.75][S02]of signals.[507.57] [507.57][S02]And then they decide, should they[509.66] [509.66][S02]be excited enough to send a signal further[512.87] [512.87][S02]up the neural network?[514.32] [514.32][S02]And so that's one artificial neuron.[516.09] [516.09][S02]And a neural network is made up of lots of layers[519.74] [519.74][S02]of lots of these neurons.[521.64] [521.64][S02]And so the higher layer neurons build on the representations[525.65] [525.65][S02]of the lower neurons.[527.3] [527.3][S02]And so if you're, for example, building a neural network[530.66] [530.66][S02]for image problems, the lowest neurons in the lowest layer[535.46] [535.46][S02]might learn features like, oh, it's a splotch of red or green,[540.0] [540.0][S02]or there's an edge at this orientation.[542.04] [542.04][S02]And then the next level up might learn, oh,[544.55] [544.55][S02]it's an edge with yellow on one side.[548.6] [548.6][S02]And then higher up, it might be, oh, it looks[551.51] [551.51][S02]like a nose or ears or a face.[555.17] [555.17][S02]And so by building these layered, learned abstractions,[560.09] [560.09][S02]these systems can actually develop[562.01] [562.01][S02]very powerful pattern-recognition[564.71] [564.71][S02]capabilities.[565.29] [565.29][S02]And that's what people were excited[566.748] [566.748][S02]about about neural networks in kind of 1985, 1990.[570.075] [570.075][S01] But we're talking teeny, teeny, tiny little--[572.45] [572.45][S02] Teeny networks, yeah.[573.783] [573.783][S02]So they could not recognize faces and cars and things.[578.96] [578.96][S02]They could recognize little patterns in artificial--[584.07] [584.07][S02]artificially generated patterns.[585.75] [585.75][S01] Yeah.[585.95] [585.95][S01]Like, you have a grid, and it can recognize maybe[587.992] [587.992][S01]a cross or something.[588.89] [588.89][S02] Or a handwritten digit.[590.307] [590.307][S02]Is it a seven or an eight?[591.518] [591.518][S01] That's fancy.[592.56] [592.56][S02] Yeah, that was fancy.[593.893] [593.893][S02]But that was kind of what they could do at that time.[596.6] [596.6][S02]But people were excited because they[598.37] [598.37][S02]could solve those kinds of problems[600.26] [600.26][S02]that other systems, based on purely logically specified[604.07] [604.07][S02]rules of what a seven means, weren't actually[607.25] [607.25][S02]able to do very well in a way that[609.17] [609.17][S02]generalized to all kinds of messy, handwritten sevens.[612.39] [612.39][S02]So I was kind of intrigued by that after my two lectures[615.86] [615.86][S02]on neural nets, and I decided I would[618.56] [618.56][S02]do a senior thesis, honors thesis,[622.52] [622.52][S02]on parallel training of neural networks[624.27] [624.27][S02]because I felt like we just need more compute.[626.73] [626.73][S02]What if we use the 32 processor machine in the department[630.14] [630.14][S02]and made a bigger system that we could train bigger neural nets?[634.58] [634.58][S02]So that was what I spent a couple of months--[637.43] [637.43][S02]three months on.[638.39] [638.39][S01] Did it work?[639.09] [639.09][S02] Yeah, yeah.[640.007] [640.007][S02]So anyway, I was very excited.[641.727] [641.727][S02]I was like, oh, 32 processors is going[643.31] [643.31][S02]to cause neural nets to really, really hum--[645.325] [645.325][S01] Take off.[646.2] [646.2][S02] --and sing.[647.48] [647.48][S02]Turns out I was wrong, naive undergrad me.[651.23] [651.23][S02]We needed about a million times as much processing power[653.99] [653.99][S02]to get them to really start to work well on real problems[656.78] [656.78][S02]that you might sort of care about.[658.382] [658.382][S01] Yeah.[659.09] [659.09][S02] But then, thanks to 20 years[661.43] [661.43][S02]of progress of Moore's law and much faster CPUs[665.45] [665.45][S02]and computational devices and stuff,[667.56] [667.56][S02]we actually then started to have practical systems that[671.48] [671.48][S02]had a million times as much compute as even[673.52] [673.52][S02]our fancy 32 processor machine.[676.68] [676.68][S02]And so I started to get interested in neural nets[680.51] [680.51][S02]again when Andrew Ng, who's a Stanford faculty member,[684.96] [684.96][S02]was consulting at Google one day a week.[686.79] [686.79][S02]And I bumped into him in one of our many micro kitchens.[689.918] [689.918][S02]I'm like, oh, what are you doing at Google?[691.71] [691.71][S02]He's like, well, I haven't really[693.085] [693.085][S02]figured it out yet because I just started consulting here.[695.79] [695.79][S02]But some of my students at Stanford[697.73] [697.73][S02]are getting good results on neural networks.[699.92] [699.92][S02]I'm like, oh, really?[701.22] [701.22][S02]Why don't we train really, really big neural networks?[703.56] [703.56][S02]So that was the genesis of our work[705.71] [705.71][S02]on neural networks at Google.[707.19] [707.19][S02]And then we formed a small team called the Google Brain[710.18] [710.18][S02]Team to start looking at how could we[713.21] [713.21][S02]train very large neural networks using Google's[716.06] [716.06][S02]computational resources.[717.74] [717.74][S02]And so we sort of built this software infrastructure that[723.14] [723.14][S02]enabled us to take a neural network description and then[726.5] [726.5][S02]break it down into pieces that would be done on different[730.13] [730.13][S02]computers, different members of this parallel team,[733.88] [733.88][S02]and then communicate amongst themselves in ways that they[737.51] [737.51][S02]needed to do in order to sort of tackle the overall problem[740.9] [740.9][S02]of how do you train a single neural network on 2,000[744.2] [744.2][S02]computers.[745.58] [745.58][S02]And so that was kind of the earliest software[747.86] [747.86][S02]we built for really scaling up neural network training.[751.49] [751.49][S02]And it enabled us to train models that were 50 to 100 times[755.33] [755.33][S02]larger than sort of existing neural networks.[758.2] [758.2][S01] 2011, right?[759.64] [759.64][S02] Yeah, this was like early 2012.[762.39] [762.39][S01] So this is before the big breakthroughs[765.86] [765.86][S01]in image recognition.[767.2] [767.2][S01]This is like way back.[768.53] [768.53][S01]And in many ways, you were doing then the same thing[771.08] [771.08][S01]that you were doing previously of just[772.97] [772.97][S01]kind of stitching computers together.[775.292] [775.292][S02] Just like my undergrad thesis.[777.0] [777.0][S01] Exactly.[777.295] [777.295][S01]But again--[777.8] [777.8][S02] I'm like, hey, we could do that[778.76] [778.76][S02]again but at big scale.[780.227] [780.227][S01] And yet, this time--[781.56] [781.56][S02] This time, it actually[782.48] [782.48][S02]worked because the computers were faster,[783.89] [783.89][S02]and we used a lot more of them.[785.14] [785.14][S01] Did it feel like a bit of a gamble,[787.098] [787.098][S01]though, back in 2011?[788.36] [788.36][S02] Oh, yeah.[789.52] [789.52][S02]The system that we built for training these neural networks[793.34] [793.34][S02]and trying different ways of breaking them apart,[795.71] [795.71][S02]I actually named it DistBelief because it's--[799.98] [799.98][S02]partly because people didn't think it was really[801.98] [801.98][S02]going to work and also because it was a distributed[805.58] [805.58][S02]system that could build these--[808.01] [808.01][S02]one of the things we wanted to train[809.51] [809.51][S02]was belief networks in addition to neural networks.[811.68] [811.68][S01] Oh, I love it.[812.27] [812.27][S01]I love it.[812.88] [812.88][S02] So DistBelief it was.[814.22] [814.22][S01] DistBelief.[815.178] [815.178][S01]Amazing.[815.7] [815.7][S01]So, while this was going on stateside,[818.82] [818.82][S01]this side of the Atlantic, it was the beginnings of DeepMind.[822.475] [822.475][S02] Yes.[823.1] [823.1][S01] And I know that you were[825.62] [825.62][S01]the person tasked with coming over and checking them out,[828.605] [828.605][S01]right?[829.105] [829.105][S01]Can you tell me about that story?[830.61] [830.61][S02] Yeah, so actually, Geoffrey Hinton, who's[835.04] [835.04][S02]a very well-known machine learning researcher,[837.51] [837.51][S02]spent a summer at Google in 2011, I think,[842.12] [842.12][S02]and we couldn't figure out how to classify him.[844.56] [844.56][S02]So he got classified as an intern, which is a little funny.[847.603] [847.603][S01] Most senior intern in the entire history[849.77] [849.77][S01]of the world.[850.5] [850.5][S02] And so he and I were working together.[853.11] [853.11][S02]And then somehow, we found out about DeepMind.[855.405] [855.405][S02]I think Geoffrey had known a little bit[857.03] [857.03][S02]about sort of the formation of the company.[859.08] [859.08][S02]And some other people also said, oh, yeah, there's[861.402] [861.402][S02]this company over in the UK doing interesting--[863.36] [863.36][S01] And it was teeny tiny at this time.[864.67] [864.67][S01]I mean, like--[865.253] [865.253][S02] Yeah, like probably 40 or 50 people or something.[869.18] [869.18][S02]And so we decided as a company that we would go check them out[873.53] [873.53][S02]as a potential acquisition.[875.96] [875.96][S02]And so I was in California.[878.49] [878.49][S02]Geoffrey was back in Toronto, where he was[881.03] [881.03][S02]a faculty member at that time.[883.7] [883.7][S02]And Geoff has a bad back, so he can't actually fly commercially[888.62] [888.62][S02]because he can't sit down.[890.19] [890.19][S02]He can only lay down or stand up.[893.15] [893.15][S02]Airlines don't like it when you stand up during takeoff.[897.47] [897.47][S02]So we had to figure out a solution, which[900.05] [900.05][S02]was to get a medical bed in a private plane.[902.97] [902.97][S02]And so a bunch of us took off at California, flew to Toronto,[908.66] [908.66][S02]scooped Geoffrey up from the tarmac,[910.83] [910.83][S02]put him in the medical bed, and then we all flew to the UK[916.19] [916.19][S02]and landed in some--[917.9] [917.9][S02]not one of the main airports.[919.62] [919.62][S02]It was at the edge of town.[923.04] [923.04][S02]And we all got in a big van, and we trooped off[925.13] [925.13][S02]to visit DeepMind, which was near Russell Square, I think.[929.42] [929.42][S02]And we were really tired from flying the previous night.[933.96] [933.96][S02]But then we got 13 straight 20-minute lectures in a row[939.977] [939.977][S02]of all the different things they were doing.[941.81] [941.81][S01] What, from the team at DeepMind?[943.11] [943.11][S02] Yeah, from the team.[943.625] [943.625][S01] Oh wow.[944.31] [944.31][S02] So we have some work, some--[945.74] [945.74][S01] While jet lagged.[946.71] [946.71][S02] While jet lagged.[947.93] [947.93][S01] It's like something from a sitcom.[949.847] [949.847][S02] Yeah, yeah, exactly.[951.45] [951.45][S02]So then we got some presentations on some[956.08] [956.08][S02]of the Atari work they were doing,[958.01] [958.01][S02]which was published later, on how do you use reinforcement[962.71] [962.71][S02]learning to learn to play old Atari 2600 games--[966.38] [966.38][S02]so games like \"Breakout\" or \"Pong\" or various other ones--[970.87] [970.87][S02]which was quite interesting.[972.58] [972.58][S02]And then--[973.18] [973.18][S01] Because you guys hadn't[973.84] [973.84][S01]been doing the reinforcement learning at that time.[975.74] [975.74][S02] Right.[976.448] [976.448][S02]We'd mostly been focusing on how do[978.07] [978.07][S02]you scale up large-scale supervised and unsupervised[981.7] [981.7][S02]learning.[982.6] [982.6][S02]And--[983.1] [983.1][S01] And the reinforcement learning[984.85] [984.85][S01]is much more motivated by rewards.[986.9] [986.9][S02] Yeah, so I think all of these techniques[990.19] [990.19][S02]are really useful, and they're often useful in combination.[992.96] [992.96][S02]So reinforcement learning you should[994.54] [994.54][S02]think of as you have some agent operating in an environment,[1000.51] [1000.51][S02]and at every step, there's a bunch of different moves[1004.26] [1004.26][S02]you could make or actions you could take.[1006.57] [1006.57][S02]So in the game of \"Go,\" for example,[1009.02] [1009.02][S02]you could play a stone in any of a whole bunch of different[1012.27] [1012.27][S02]positions.[1012.93] [1012.93][S02]In Atari, you could move your joystick up,[1015.54] [1015.54][S02]down, or left or right, or you could[1017.04] [1017.04][S02]push the left or right button.[1019.59] [1019.59][S02]And often, in these situations, you[1022.05] [1022.05][S02]don't get an immediate reward.[1024.06] [1024.06][S02]So in \"Go,\" you play a stone, and you don't know[1026.31] [1026.31][S02]if that's a good idea or not until the whole process[1030.72] [1030.72][S02]of the rest of the game plays out.[1032.64] [1032.64][S02]And one of the interesting things about reinforcement[1035.069] [1035.069][S02]learning is it's able to take kind[1037.44] [1037.44][S02]of long sequences of actions and then[1040.319] [1040.319][S02]attribute rewards or negative rewards[1043.89] [1043.89][S02]to the sequence of actions that you took in proportion to how[1048.45] [1048.45][S02]unexpected it was, based on when you made that move,[1052.15] [1052.15][S02]did you think it was a good idea.[1053.53] [1053.53][S02]And then you won.[1055.24] [1055.24][S02]So maybe you should increase your idea of that[1057.57] [1057.57][S02]was a good idea a little bit.[1058.84] [1058.84][S02]Or maybe you lost, and you should decrease your sense[1061.695] [1061.695][S02]that that was a good idea a little bit.[1064.26] [1064.26][S02]And so that's kind of the main idea[1066.39] [1066.39][S02]behind reinforcement learning.[1067.66] [1067.66][S02]And it's a quite effective technique,[1069.34] [1069.34][S02]especially in environments where it's very unclear immediately[1073.75] [1073.75][S02]whether that was a good idea.[1075.23] [1075.23][S01] Yeah, absolutely.[1076.53] [1076.53][S02] In contrast, supervised learning[1078.37] [1078.37][S02]is where you have an input and you have a ground truth output.[1082.94] [1082.94][S02]So the classic example is you have a bunch of images,[1086.17] [1086.17][S02]and each image has been labeled with one[1088.12] [1088.12][S02]of a bunch of categories.[1089.485] [1089.485][S02]So there's an image--[1091.48] [1091.48][S02]car.[1092.27] [1092.27][S02]There's another image-- ostrich.[1094.22] [1094.22][S02]Another image-- a pomegranate.[1098.08] [1098.08][LAUGHTER][1100.24] [1100.24][S02]If you have a rich set of categories.[1102.88] [1102.88][S01] Exactly.[1103.93] [1103.93][S01]Tell me.[1104.51] [1104.51][S01]When you were here at DeepMind, and you[1106.135] [1106.135][S01]decided that you would do the acquisition, was Demis nervous?[1109.79] [1109.79][S02] I don't know if he was nervous.[1111.54] [1111.54][S02]I mean, I think that I said, well, I've[1115.92] [1115.92][S02]seen all these nice presentations,[1117.4] [1117.4][S02]but can I look it a little bit of the code?[1119.61] [1119.61][S02]And so-- because I wanted to make[1121.53] [1121.53][S02]sure there was real code behind it[1123.12] [1123.12][S02]and see what the coding standards were like,[1127.002] [1127.002][S02]did people actually write comments,[1128.46] [1128.46][S02]and that kind of thing.[1129.418] [1129.418][S02]So Demis kind of was a little unsure about this.[1133.39] [1133.39][S02]I said, oh, it doesn't have to be super secret code.[1136.03] [1136.03][S02]Just pick some little bit of code and show it to me.[1138.6] [1138.6][S02]And so I went into an office with one of the engineers,[1141.69] [1141.69][S02]and we kind of sat down for 10 minutes.[1143.67] [1143.67][S02]And I said, OK, what does this code do?[1146.83] [1146.83][S02]And, oh, OK, that thing, what does that do?[1149.35] [1149.35][S02]And can you show me the implementation of that?[1152.25] [1152.25][S02]And I came out satisfied.[1153.875] [1153.875][S01] It was neat and tidy.[1155.25] [1155.25][S02] It was reasonably neat and tidy.[1157.042] [1157.042][S02]I mean, for a small company trying to move quickly,[1159.97] [1159.97][S02]it was kind of researchy code.[1162.72] [1162.72][S02]But it was clearly interesting and well-documented.[1168.89] [1168.89][S02]And so--[1169.39] [1169.39][S01] I've heard that when you do your code,[1171.473] [1171.473][S01]you put a little thing, which is LGTM.[1174.213] [1174.213][S02] Oh yeah, it looks good to me.[1175.88] [1175.88][S02]Yeah, I use that in real life, too, not just for code reviews.[1179.51] [1179.51][S01] So in these presentations then,[1181.4] [1181.4][S01]can you remember what your impression was?[1183.993] [1183.993][S02] Yeah, I mean, it seemed[1185.41] [1185.41][S02]like they were doing really interesting work, particularly[1188.65] [1188.65][S02]in the reinforcement learning side.[1190.67] [1190.67][S02]We were focused on scaling, so we were training models[1193.84] [1193.84][S02]that were much, much bigger than the ones DeepMind was[1197.47] [1197.47][S02]playing with at the time.[1198.97] [1198.97][S02]But they were learning to use reinforcement learning[1202.24] [1202.24][S02]to sort of solve kind of game play, which[1206.157] [1206.157][S02]is a nice, clean environment for reinforcement learning.[1208.49] [1208.49][S02]But it seemed like the combination of reinforcement[1211.63] [1211.63][S02]learning plus a lot of the scaling work[1213.58] [1213.58][S02]that we had been working on would be a really good one.[1216.4] [1216.4][S01] Because it's like I guess[1217.24] [1217.24][S01]you're sort of approaching a problem from two[1219.115] [1219.115][S01]different directions-- like, really tiny with reinforcement[1221.8] [1221.8][S01]learning, really small, like toy models and building up.[1224.24] [1224.24][S01]And then you're kind of at this very, very big scale[1226.57] [1226.57][S01]with this sort of rich understanding,[1230.21] [1230.21][S01]understanding in inverted commas.[1231.82] [1231.82][S02] Yeah.[1232.51] [1232.51][S01] But then it's sort of putting the two together[1234.58] [1234.58][S01]where things become really powerful.[1236.03] [1236.03][S02] Yeah, yeah, indeed.[1237.28] [1237.28][S02]And that was a lot of the motivation[1240.73] [1240.73][S02]behind the combination we did last year[1244.0] [1244.0][S02]of the kind of legacy DeepMind and legacy Brain[1248.56] [1248.56][S02]and other parts of Google research.[1250.52] [1250.52][S02]We decided we would just combine the units together and form[1253.75] [1253.75][S02]Google DeepMind.[1255.13] [1255.13][S01] Yeah, and Gemini as the--[1256.99] [1256.99][S02] And Gemini, which actually predated[1259.03] [1259.03][S02]the idea of combining but really was like,[1261.5] [1261.5][S02]hey, we should really all work together on these problems[1264.25] [1264.25][S02]because we're all kind of sniffing around[1267.43] [1267.43][S02]the same kind of general direction of trying[1271.0] [1271.0][S02]to train really high-quality, large-scale, multimodal models.[1274.76] [1274.76][S02]And it doesn't make sense to fragment our ideas[1277.75] [1277.75][S02]and not work together and fragment our compute resources[1280.51] [1280.51][S02]and so on.[1281.03] [1281.03][S02]We should really just put all this together,[1283.72] [1283.72][S02]build a combined team to go after this problem,[1287.45] [1287.45][S02]and that's what we did.[1288.74] [1288.74][S01] So why Gemini?[1289.84] [1292.418][S02] I actually named it.[1293.71] [1293.71][S01] Did you?[1294.16] [1294.16][S02] Yeah.[1294.54] [1294.54][S01] I mean, you're very-- after the--[1295.95] [1295.95][S02] I do like naming things.[1296.85] [1296.85][S01] DistBelief.[1297.09] [1297.09][S02] It's fun.[1297.25] [1297.25][S02]Yeah, DistBelief.[1297.57] [1297.57][S01] It's your--[1298.89] [1298.89][S02] So Gemini relates to twins,[1301.92] [1301.92][S02]and I felt it was a good name for the twins of legacy DeepMind[1306.57] [1306.57][S02]and legacy Brain kind of coming together[1309.15] [1309.15][S02]to really start working together on an ambitious, multimodal[1313.465] [1313.465][S02]project.[1313.965] [1313.965][S01] I guess also Gemini--[1315.585] [1315.585][S01]I'm just thinking of the space missions.[1318.473] [1318.473][S02] Yeah.[1319.14] [1319.14][S01] It's like a precursor to Apollo.[1320.84] [1320.84][S02] Yeah, a good thing about a name that[1322.798] [1322.798][S02]has multiple meanings is--[1324.21] [1324.21][S02]so that was another reason to pick the name.[1326.47] [1326.47][S02]It's sort of the precursor to ambitious space program[1331.8] [1331.8][S02]progress.[1332.405] [1332.405][S01] So I want to come on to the multimodal stuff.[1334.78] [1334.78][S02] OK.[1335.363] [1335.363][S01] Just before I do, I guess[1337.26] [1337.26][S01]one of the big reasons why this big change has happened[1341.07] [1341.07][S01]in the sort of public consciousness of chat bots[1343.2] [1343.2][S01]and large language models is in part because of some work[1346.68] [1346.68][S01]that came out of Google Brain with Transformers.[1349.87] [1349.87][S01]If you'll sort of forgive the pun,[1351.757] [1351.757][S01]can you tell us a little bit about that Transformer work[1354.09] [1354.09][S01]and how transformative it's been?[1355.81] [1355.81][S02] Sure, yeah.[1356.98] [1356.98][S02]So it turns out a lot of the problems[1359.58] [1359.58][S02]you want to deal with in language[1361.95] [1361.95][S02]and in a bunch of other domains are problems of sequences.[1365.71] [1365.71][S02]So if you think about autocomplete in Gmail--[1369.987] [1369.987][S02]you're typing a sentence, and can the system[1371.82] [1371.82][S02]help you by finishing your sentence[1375.36] [1375.36][S02]or your thought for you?[1377.43] [1377.43][S02]A lot of that relies on seeing part of a sequence[1380.37] [1380.37][S02]and then predicting the rest of it.[1382.35] [1382.35][S02]And essentially, that's what these large language[1385.08] [1385.08][S02]models are trained to do.[1387.07] [1387.07][S02]They're trained to take in data one word[1389.97] [1389.97][S02]or one piece of a word at a time and then predict[1392.91] [1392.91][S02]what is the next thing that will follow.[1395.24] [1395.24][S01] Like fancy autocomplete.[1396.74] [1396.74][S02] Like fancy autocomplete.[1398.38] [1398.38][S02]Yeah, it turns out to be useful.[1399.732] [1399.732][S02]You can model a lot of different problems this way as well.[1402.19] [1402.19][S02]So translation you can model as taking[1405.75] [1405.75][S02]in the English version of a sentence[1407.85] [1407.85][S02]and then training the model to then output[1410.28] [1410.28][S02]the French version of the sentence when you have[1413.85] [1413.85][S02]enough English-French sentence pairs to sort of train[1417.36] [1417.36][S02]as like a sequence.[1419.28] [1419.28][S02]You can also use this in healthcare settings,[1421.655] [1421.655][S02]like if you're trying to predict a patient in front[1426.18] [1426.18][S02]of you is reporting these symptoms[1427.62] [1427.62][S02]and they have these lab test results, and in the past,[1429.87] [1429.87][S02]they've had these things, you can model that whole thing[1432.81] [1432.81][S02]as a sequence.[1433.77] [1433.77][S02]And then you can predict what are the likely diagnoses that[1437.94] [1437.94][S02]would make sense if you have other de-identified data[1442.17] [1442.17][S02]that you can train on that has also kind of been organized[1445.5] [1445.5][S02]in these sequences.[1446.74] [1446.74][S02]And the way you can do that is you[1448.77] [1448.77][S02]just hide the rest of the sequence,[1450.84] [1450.84][S02]and you force the model to try to predict what happens next.[1454.412] [1454.412][S02]It's quite an interesting thing that it's[1456.12] [1456.12][S02]so applicable to language translation, healthcare[1459.93] [1459.93][S02]settings, DNA sequences, all kinds of things.[1462.915] [1462.915][S01] But it's about the bit that you're paying attention[1465.54] [1465.54][S01]to at any point in time.[1467.2] [1467.2][S02] Yeah, so the models that were successful[1470.83] [1470.83][S02]prior to the Transformer architecture[1473.17] [1473.17][S02]were what are called recurrent models, where[1477.22] [1477.22][S02]they have some internal state.[1479.05] [1479.05][S02]And every time they see a word, they[1481.72] [1481.72][S02]do some processing to update their internal state.[1484.36] [1484.36][S02]And then they go on to the next word,[1486.25] [1486.25][S02]and then they do that again.[1487.85] [1487.85][S02]So now, they move their state forward a bit[1491.2] [1491.2][S02]and update the state with respect to the next word[1494.74] [1494.74][S02]that they just saw.[1495.68] [1495.68][S02]And so you can imagine this as like a 12-word sentence.[1499.49] [1499.49][S02]You're doing that updating of the state 12 times.[1502.37] [1502.37][S02]But every step is dependent on the previous one.[1505.25] [1505.25][S02]And so that means it's actually quite hard to get it to run fast[1508.36] [1508.36][S02]because you have this-- what's called a sequential dependency[1511.51] [1511.51][S02]where step seven depends on step six,[1514.13] [1514.13][S02]step six depends on step five, and so on.[1517.39] [1517.39][S02]So one of the things a collection of researchers[1520.45] [1520.45][S02]within Google Research did is they came up[1524.2] [1524.2][S02]with a pretty interesting idea, which is instead of just having[1527.5] [1527.5][S02]a single state that we update at every word,[1530.54] [1530.54][S02]let's process all the words all at once,[1533.95] [1533.95][S02]and let's remember the state that we get when[1536.86] [1536.86][S02]we're processing every word.[1538.84] [1538.84][S02]And then when we're trying to predict a new word,[1541.28] [1541.28][S02]let's pay attention to all of the previous states[1545.11] [1545.11][S02]and figure out how to learn to pay attention[1547.72] [1547.72][S02]to the important parts--[1549.05] [1549.05][S02]so that's the learned attention mechanism in Transformer--[1553.09] [1553.09][S02]in order to predict the next word.[1554.66] [1554.66][S02]And for some words, you might need to pay attention[1556.96] [1556.96][S02]to the previous word a lot.[1558.91] [1558.91][S02]For some contexts, it's very important[1561.82] [1561.82][S02]to pay attention kind of a little bit[1563.38] [1563.38][S02]to a lot of the words in the context.[1566.5] [1566.5][S02]But the important thing about that is it really[1568.51] [1568.51][S02]can be done in parallel.[1569.78] [1569.78][S02]You can take in 1,000 words, compute the state for each one[1573.91] [1573.91][S02]of them in parallel, and then that makes it sort of 10 to 100[1579.7] [1579.7][S02]times more efficient in terms of scaling and performance than[1585.22] [1585.22][S02]the previous recurrent models.[1586.53] [1586.53][S02]And so that was why that was such a big advance.[1588.53] [1588.53][S01] But then I guess there's other things that[1590.74] [1590.74][S01]seem to emerge from this.[1591.782] [1591.782][S01]I mean, sort of a conceptual understanding or maybe sort[1594.34] [1594.34][S01]of abstraction that's possible just through sequencing[1597.4] [1597.4][S01]and language alone--[1598.555] [1598.555][S01]I mean, was that a surprise?[1600.722] [1600.722][S02] Yeah, I mean, I think some of the earliest work we did[1603.43] [1603.43][S02]on language modeling in the Google Brain team was really[1606.4] [1606.4][S02]about modeling words not as their surface form of like[1610.75] [1610.75][S02]H-E-L-L-O or C-O-W but really about a high-dimensional vector[1617.89] [1617.89][S02]that represents kind of the way in which that word is used.[1621.95] [1621.95][S02]We're used to thinking in two and three dimensions as humans.[1625.13] [1625.13][S02]But when you have 100 dimensions or 1,000 dimensions,[1628.21] [1628.21][S02]there's a lot of room in 1,000-dimensional space.[1631.28] [1631.28][S02]But when you have things that are nearby,[1633.37] [1633.37][S02]and you've trained the model in such a way[1635.17] [1635.17][S02]that cow and sheep and goat and pig are all near each other,[1641.26] [1641.26][S02]and they're very far apart from espresso machine--[1643.905] [1643.905][LAUGHTER][1645.22] [1645.22][S01] Although milk could be in between.[1647.525] [1647.525][S02] Milk would probably be nearer the cow but kind[1649.9] [1649.9][S02]of in between the two.[1650.9] [1650.9][S01] Yeah.[1651.608] [1651.608][S02] It would probably be kind of on that 100-dimensional[1654.91] [1654.91][S02]line in 100-dimensional space.[1658.69] [1658.69][S02]So this is kind of why these models have surprisingly[1662.38] [1662.38][S02]powerful capabilities, I think, is because they're representing[1666.25] [1666.25][S02]things with so many high dimensions[1668.08] [1668.08][S02]that they can actually really latch on[1670.21] [1670.21][S02]to many different facets of a word[1673.3] [1673.3][S02]or a sentence or a paragraph simultaneously because there's[1677.38] [1677.38][S02]so much room in their representation.[1679.305] [1679.305][S01] It's sort of extracted the grounding[1681.64] [1681.64][S01]that we ourselves have given language, I guess.[1683.75] [1683.75][S02] Yeah, I mean, when we hear a word,[1686.12] [1686.12][S02]we don't just think of the surface form of the word.[1688.52] [1688.52][S02]We think cow-- oh, that triggers a bunch of other things,[1691.66] [1691.66][S02]like milk or espresso machine or milking and calf and bull.[1699.14] [1699.14][S02]And one of the things we found with those early word[1702.73] [1702.73][S02]representations was that directions had meaning.[1705.41] [1705.41][S02]So if you think about present tenses of verbs, like walk,[1711.95] [1711.95][S02]you would go in the same direction in this[1713.96] [1713.96][S02]100-dimensional space to get from walk to walked as you would[1719.3] [1719.3][S02]get from run to ran as you would go from read to read.[1728.033] [1728.033][S01] Wow.[1728.7] [1728.7][S02] Yeah.[1729.26] [1729.26][S01] So it actually understands-- understands,[1731.195] [1731.195][S01]I keep using that word, and I don't mean it.[1733.17] [1733.17][S01]But there is some representation of tenses[1736.4] [1736.4][S01]within the structure of these--[1737.92] [1737.92][S02] Yeah, and it's just emerged[1739.55] [1739.55][S02]from the training process.[1741.36] [1741.36][S02]It's not something we told it to do.[1742.86] [1742.86][S02]It's just the training algorithm we used[1745.16] [1745.16][S02]and the fact that language has lots[1747.35] [1747.35][S02]of ways in which particular forms are used[1752.6] [1752.6][S02]caused that to emerge.[1753.6] [1753.6][S02]And you could also, for example, change[1755.6] [1755.6][S02]from male or female versions of words and vice versa.[1759.75] [1759.75][S02]So cow to bull is the same direction[1763.85] [1763.85][S02]as queen to king or man to woman, woman to man, and so on.[1769.1] [1769.1][S01] Amazing.[1770.06] [1770.06][S01]But this is still-- this is just with language[1772.04] [1772.04][S01]that we're talking about here.[1773.07] [1773.07][S02] Yeah.[1773.21] [1773.21][S01] And so, OK, tell me tell[1774.72] [1774.72][S01]how does the multimodal aspect of this change?[1780.2] [1780.2][S01]How does it make it different?[1781.82] [1781.82][S02] Yeah, because you're still[1787.59] [1787.59][S02]representing the input data in these high-dimensional spaces.[1791.513] [1791.513][S02]And it's really a matter of how do[1792.93] [1792.93][S02]you get from the pixels of an image,[1795.4] [1795.4][S02]say, into something where, ideally, you'd[1800.04] [1800.04][S02]like the multimodal model to have the same kind of thing[1804.42] [1804.42][S02]that we have.[1805.14] [1805.14][S02]When we see a cow, that triggers kind[1808.66] [1808.66][S02]of similar activations in our brain to reading the word cow,[1812.83] [1812.83][S02]to hearing a cow moo, right?[1817.405] [1817.405][S02]And you kind of want to train models[1819.19] [1819.19][S02]so that they have that joint meaning and representation,[1823.18] [1823.18][S02]regardless of the way they arrived at that input data.[1826.93] [1826.93][S02]So if they see a video of a cow walking through a field,[1830.53] [1830.53][S02]that should trigger a whole bunch[1832.09] [1832.09][S02]of things that are related to that[1833.62] [1833.62][S02]in the model, based on the activations[1836.14] [1836.14][S02]that the model has built over--[1839.8] [1839.8][S02]typically, these are very deep, layered models.[1842.53] [1842.53][S02]And so the lowest layers typically[1844.33] [1844.33][S02]have very simple representations.[1846.26] [1846.26][S02]And then the deeper, the higher layers in the model,[1849.38] [1849.38][S02]build on those representations and build[1851.41] [1851.41][S02]more interesting and complex combinations[1855.76] [1855.76][S02]of features and representations of be it words or images or--[1860.377] [1860.377][S01] So when you're saying multimodal[1862.21] [1862.21][S01]from the ground up, which is kind of a big phrase[1864.252] [1864.252][S01]that you hear about Gemini, it's not[1866.44] [1866.44][S01]that you've got the word section over here,[1868.735] [1868.735][S01]the pixel section over here, and you're translating[1870.86] [1870.86][S01]between one and the other.[1871.895] [1871.895][S02] Right.[1872.24] [1872.24][S01] But in the model itself,[1873.847] [1873.847][S01]those representations are there.[1875.18] [1875.18][S02] Yeah, very early in the models.[1876.81] [1876.81][S01] Does that make it harder at the beginning[1879.018] [1879.018][S01]when you're setting it up?[1880.2] [1880.2][S01]Does it make it more difficult to do?[1884.59] [1884.59][S02] Yeah, I mean, I think figuring out[1886.63] [1886.63][S02]how to integrate different modalities into the model[1889.45] [1889.45][S02]and how should you train a multimodal model is[1892.93] [1892.93][S02]more complex than a simpler, pure language[1896.38] [1896.38][S02]or pure character-based model.[1899.14] [1899.14][S02]But you get a lot of benefits from it[1901.36] [1901.36][S02]in that you get sometimes cross-modal transfer, where[1905.74] [1905.74][S02]now seeing visual stuff about cows[1910.06] [1910.06][S02]actually helps inform the language.[1912.67] [1912.67][S02]Maybe it'd had seen a bunch of descriptions[1914.83] [1914.83][S02]of cows in meadows or something, but now[1917.8] [1917.8][S02]it suddenly has seen images of that and videos of that.[1922.39] [1922.39][S02]And it's actually able to bring those representations together[1926.17] [1926.17][S02]in a way that makes kind of similar things[1929.98] [1929.98][S02]trigger inside the model, regardless[1932.14] [1932.14][S02]of whether you saw the word cow or kind of the image of a cow.[1936.688] [1936.688][S01] Give me an example of the type of situation[1938.98] [1938.98][S01]you see this being useful in in the future.[1943.51] [1943.51][S02] Well, I think it's already useful, which is good.[1948.28] [1948.28][S02]I mean, as one example, you want to be[1950.2] [1950.2][S02]able to take in an image of a handwritten, whiteboard,[1957.83] [1957.83][S02]worked-out math problem and say, did the student[1961.42] [1961.42][S02]get this problem right?[1962.72] [1962.72][S02]Right?[1963.27] [1963.27][S02]And so now, you need to really bring[1965.7] [1965.7][S02]in the multimodal capabilities in one example.[1968.92] [1968.92][S02]You need to actually do handwriting recognition,[1972.19] [1972.19][S02]understand from that, OK, it's a physics problem[1975.54] [1975.54][S02]that someone's written on the board,[1977.04] [1977.04][S02]and it's got maybe a picture of a skier going down a slope.[1980.2] [1980.2][S02]In one of the early Gemini tech reports,[1982.57] [1982.57][S02]we had this good example of a student who had worked[1986.25] [1986.25][S02]out a problem on a whiteboard.[1988.14] [1988.14][S02]And you could actually ask Gemini,[1993.21] [1993.21][S02]did the student get this problem right?[1995.05] [1995.05][S02]If not, where did they go wrong?[1997.12] [1997.12][S02]And can you explain how to solve the problem correctly?[1999.76] [1999.76][S02]And it was actually able to tell that the student had incorrectly[2003.32] [2003.32][S02]applied the formula for a skier going down a frictionless slope.[2009.29] [2009.29][S02]And instead, they used the hypotenuse[2011.51] [2011.51][S02]instead of the height.[2012.69] [2012.69][S02]And it said, oh, no, actually, you should have used this.[2015.27] [2015.27][S02]And here's the problem worked out.[2017.16] [2017.16][S02]And it did all that and recognized all the handwriting[2020.75] [2020.75][S02]and the fact that this was a physics problem.[2022.83] [2022.83][S02]This kind of physics knowledge that the model already[2025.25] [2025.25][S02]had sort of was the right thing to apply.[2028.372] [2028.372][S01] I mean, that's a really neat way that you could[2030.83] [2030.83][S01]use the existing model of Gemini in the existing[2033.47] [2033.47][S01]model of education, I guess.[2034.86] [2034.86][S02] Totally.[2035.12] [2035.12][S02]Yeah.[2035.51] [2035.51][S01] But I suppose, actually, these[2037.26] [2037.26][S01]are not kind of isolated systems from one another.[2040.23] [2040.23][S01]So in some ways, do you think that these multimodal models[2043.04] [2043.04][S01]will change the way that we have to do education, full stop?[2047.2] [2047.2][S02] I mean, I think the potential[2049.06] [2049.06][S02]for using AI tools to help education is really amazing.[2056.27] [2056.27][S02]And we're sort of just at the beginning of this journey[2058.659] [2058.659][S02]as a society, I think.[2060.04] [2060.04][S02]We know, for example, that educational outcomes of students[2064.57] [2064.57][S02]who get one-on-one tutoring from another person[2067.75] [2067.75][S02]are two standard deviations better than students[2071.59] [2071.59][S02]who have a traditional classroom setting of a teacher and 30[2075.639] [2075.639][S02]or so students.[2077.77] [2077.77][S02]So how could we get everyone to the point[2081.9] [2081.9][S02]where they feel like they have the benefit[2084.87] [2084.87][S02]of an educational tutor that's one on one,[2087.268] [2087.268][S02]understands what they know, understands what they[2089.31] [2089.31][S02]don't know, can help them learn in the way that they learn best?[2093.84] [2093.84][S02]That is the potential of AI in education.[2096.94] [2096.94][S02]And I think, really, we're not that far away from something[2101.46] [2101.46][S02]where you could point a Gemini model or a future Gemini model[2106.23] [2106.23][S02]at something, some piece of material, and say,[2108.91] [2108.91][S02]can you help me learn this?[2110.33] [2110.33][S02]Take the chapter six in your biology textbook or something.[2114.31] [2114.31][S02]And it's got a bunch of images.[2116.32] [2116.32][S02]It's got a bunch of text.[2118.29] [2118.29][S02]Maybe it's got a lecture video that you watched as well.[2121.77] [2121.77][S02]And then you can actually say, I really[2124.32] [2124.32][S02]don't understand this thing.[2125.68] [2125.68][S02]Can you help me understand it?[2126.94] [2126.94][S02]It can ask you questions.[2128.35] [2128.35][S02]You can ask it questions.[2131.372] [2131.372][S02]You can answer the questions.[2132.58] [2132.58][S02]It can assess are you right or wrong[2134.9] [2134.9][S02]and really guides you in your learning journey[2139.16] [2139.16][S02]because it's individualized.[2140.87] [2140.87][S02]And we should be able to get that to many, many people[2145.73] [2145.73][S02]around the world in not just English, in languages[2150.92] [2150.92][S02]spoken by--[2153.77] [2153.77][S02]hundreds and hundreds of languages all around the world.[2156.853] [2156.853][S01] I mean, so I take what[2158.27] [2158.27][S01]you said about the lots of different languages[2160.187] [2160.187][S01]and trying to make these as broadly available as possible.[2163.1] [2163.1][S01]But is there a danger of creating[2165.14] [2165.14][S01]a bit of a two-tier system here where, on the one hand,[2167.79] [2167.79][S01]people who have access to these tools, as you described,[2171.66] [2171.66][S01]get far better outcomes, accelerate[2173.57] [2173.57][S01]their own learning and their own productivity, and then[2176.51] [2176.51][S01]anybody who is not fortunate enough to have access[2178.64] [2178.64][S01]to the tools really struggles?[2180.87] [2180.87][S01]I mean, is that something that concerns you?[2183.04] [2183.04][S02] Yeah, I mean, I think there is definitely[2185.207] [2185.207][S02]a risk of creating two-tier systems.[2186.9] [2186.9][S02]I think what we should strive to do[2189.26] [2189.26][S02]is make these technologies as broadly[2191.99] [2191.99][S02]accessible and universally accessible, if we can,[2195.72] [2195.72][S02]for everyone and really try to lean into the strengths of what[2199.56] [2199.56][S02]that will do for society and to make it affordable or free[2205.11] [2205.11][S02]for people to take advantage of the capabilities for education.[2208.78] [2208.78][S02]For health care, I think, is another area that is hugely--[2213.45] [2213.45][S02]huge potential for AI to really make a big difference[2217.47] [2217.47][S02]in healthcare accessibility.[2219.582] [2219.582][S01] Go back to Gemini, if we can.[2221.29] [2221.29][S02] Sure.[2221.957] [2221.957][S01] OK, so I guess if you started off[2225.0] [2225.0][S01]with Google Search, factuality must[2227.1] [2227.1][S01]have been absolutely at the cornerstone of everything[2230.01] [2230.01][S01]that you cared about.[2230.92] [2230.92][S02] Yeah.[2231.587] [2231.587][S01] But Gemini, I mean, you work with it all the time.[2234.588] [2234.588][S01]I imagine you've seen it say some quite outlandish things.[2237.005] [2237.005][S02] Yeah.[2237.672] [2237.672][S01] How are you sort of squaring[2240.0] [2240.0][S01]that circle in your head of releasing,[2242.76] [2242.76][S01]perhaps, some of the need for absolute factuality[2245.82] [2245.82][S01]at all times?[2246.93] [2246.93][S02] Yeah, it's actually a tricky balance as a company[2249.63] [2249.63][S02]because we are, from our origins, a search-based company.[2254.35] [2254.35][S02]And as you say, providing accurate,[2257.22] [2257.22][S02]factual information is kind of the pinnacle[2260.22] [2260.22][S02]of a search-engine experience.[2262.5] [2262.5][S02]And I think we actually had built interesting large language[2265.71] [2265.71][S02]models internally that people enjoyed conversing with.[2271.68] [2271.68][S02]Actually, some of them were available[2275.52] [2275.52][S02]internally during the pandemic.[2277.25] [2277.25][S02]So people were all at home.[2278.59] [2278.59][S02]And you could actually see internal usage spike[2281.01] [2281.01][S02]during lunchtime because people were having conversations[2283.92] [2283.92][S02]with their virtual chatbot.[2285.87] [2285.87][S02]Because who else are you going to talk to[2288.12] [2288.12][S02]when you're home alone or whatever?[2290.82] [2290.82][S02]But these models are trained to predict plausible next tokens,[2297.88] [2297.88][S02]essentially.[2298.57] [2298.57][S02]So a token, you can think of it as a word or a piece of a word.[2302.5] [2302.5][S02]And so when you predict plausible next tokens,[2305.56] [2305.56][S02]that's a different thing than that is absolute truth, right?[2309.0] [2309.0][S02]It's a probabilistically plausible sentence.[2313.11] [2313.11][S02]And that's different than a fact.[2314.91] [2314.91][S02]And I think one of the things we realized over time[2318.03] [2318.03][S02]was these models can actually be quite useful even if they're not[2322.23] [2322.23][S02]100% factual.[2323.47] [2323.47][S02]And so I think realizing there's all these other use cases,[2326.85] [2326.85][S02]or can you summarize this slide deck in five bullets--[2331.33] [2331.33][S02]and yes, you could argue about is that fifth bullet[2334.22] [2334.22][S02]exactly right.[2335.03] [2335.03][S02]But it's still pretty useful to get[2336.85] [2336.85][S02]4.5 bullets that are factually accurate about the slide deck.[2341.08] [2341.08][S02]And we're striving to make it five factually accurate bullets.[2344.66] [2344.66][S02]But even without that, I think the utility of these models[2347.98] [2347.98][S02]is actually quite high.[2348.95] [2348.95][S01] Was it an uncomfortable realization?[2350.95] [2350.95][S01]Because, of course, other labs did[2352.12] [2352.12][S01]push out their models earlier.[2353.543] [2353.543][S02] Yeah, yeah.[2354.46] [2354.46][S01] Do you think that guys had an abundance of caution[2358.36] [2358.36][S01]because of this factual issue?[2360.14] [2360.14][S02] I mean, I think we had[2361.96] [2361.96][S02]a number of different concerns, factuality being one of them.[2365.17] [2365.17][S02]Toxicity and bias in the way the model is trained[2369.64] [2369.64][S02]and the outputs that it can produce[2371.83] [2371.83][S02]is an area where we want to make the model less[2376.48] [2376.48][S02]biased in a lot of ways.[2378.43] [2378.43][S02]And so there were a whole number of areas[2381.28] [2381.28][S02]where we wanted to sort of be relatively[2383.92] [2383.92][S02]cautious before releasing things to the general public.[2389.16] [2389.16][S02]And I think we've gotten a lot of those issues kind of sorted[2394.17] [2394.17][S02]out enough that we think the products we've put out[2396.87] [2396.87][S02]in this space are useful even though there's obviously[2400.32] [2400.32][S02]room for improvement in things like factuality[2403.11] [2403.11][S02]and in bias and other areas.[2406.74] [2406.74][S02]So I think that's taken a little bit of an adjustment[2409.5] [2409.5][S02]for people-- is strive for the best you can be but also realize[2414.72] [2414.72][S02]that by not releasing something, you're sort of holding back[2419.34] [2419.34][S02]something that could be useful for a lot of people[2422.22] [2422.22][S02]even with its sort of foibles.[2424.253] [2424.253][S01] But then with those foibles,[2425.92] [2425.92][S01]then, so in which direction do we go from here?[2428.62] [2428.62][S01]Do you think that--[2429.42] [2429.42][S01]I mean, it sort of seems to me that there's[2431.212] [2431.212][S01]been this real shift in the way that computing happens,[2435.17] [2435.17][S01]as it were.[2435.67] [2435.67][S01]You get a calculator, you put in the same sum twice,[2438.01] [2438.01][S01]you get the same answer twice.[2439.05] [2439.05][S02] Yeah.[2439.41] [2439.41][S01] Whereas we're now in an era[2441.18] [2441.18][S01]of probabilistic computing.[2443.35] [2443.35][S01]And so I wonder whether--[2445.95] [2445.95][S01]is it that the public has to come to terms with that[2448.33] [2448.33][S01]and sort of accept that we're in an era where[2452.555] [2452.555][S01]things are much more human-like in that they can make mistakes?[2455.18] [2455.18][S01]Or is it something that you think is fixable?[2458.27] [2458.27][S02] I think it's some of both, right?[2460.52] [2460.52][S02]I mean, I think there's a bunch of sort[2463.455] [2463.455][S02]of technical approaches to some of these problems[2466.22] [2466.22][S02]that will make the factuality area issues better.[2470.09] [2470.09][S02]One instance is if you think about the data[2473.37] [2473.37][S02]the model is trained on--[2475.16] [2475.16][S02]trillions of tokens of text and other data[2478.97] [2478.97][S02]that are then mixed together in this giant soup of billions[2483.26] [2483.26][S02]and billions of parameters--[2486.23] [2486.23][S02]I like to think of that as like you've seen a lot of stuff,[2490.05] [2490.05][S02]but you don't recall it very well.[2493.25] [2493.25][S02]Whereas if you take information that is in--[2496.73] [2496.73][S02]one of the things we've been pushing on in Gemini[2499.22] [2499.22][S02]is having a long context window.[2501.72] [2501.72][S02]So when you have a long bit of space[2504.38] [2504.38][S02]where you can put a lot of direct information[2506.87] [2506.87][S02]that you're trying to summarize or manipulate or compare[2511.49] [2511.49][S02]in various ways or extract information[2513.38] [2513.38][S02]from, that information in the context window, the model[2516.83] [2516.83][S02]actually has a much clearer view of, right?[2519.15] [2519.15][S02]It's got the actual text and the representations[2522.15] [2522.15][S02]of that text not tangled together[2525.168] [2525.168][S02]with everything else it's seen.[2526.46] [2526.46][S01] So this context window is sort of the bit[2528.668] [2528.668][S01]that the model can see as important at that moment.[2532.507] [2532.507][S02] Yeah, it can sort of reason[2534.09] [2534.09][S02]about that in more fidelity than other things[2537.45] [2537.45][S02]that it's seen in its training process.[2539.59] [2539.59][S02]So it can take five PDFs of scientific articles,[2544.123] [2544.123][S02]and then you can ask questions about it,[2545.79] [2545.79][S02]like can you please tell me common themes[2548.97] [2548.97][S02]across these articles?[2550.0] [2550.0][S02]And then it's actually able to do that because it[2552.48] [2552.48][S02]has its own representation of all[2556.44] [2556.44][S02]the contents of those articles.[2557.95] [2557.95][S02]And that's one of the reasons we've[2559.408] [2559.408][S02]been pushing a lot on very long context windows[2562.02] [2562.02][S02]for Gemini models is we think that's a really[2564.18] [2564.18][S02]useful capability for factuality,[2566.68] [2566.68][S02]for video summarization, for all kinds of things.[2570.12] [2570.12][S01] Is there a limit, though, to the context window?[2572.62] [2572.62][S01]Can you just push and push and push and push[2574.453] [2574.453][S01]until it's sort of an infinite context window?[2577.47] [2577.47][S02] That is an excellent question.[2579.39] [2579.39][S02]Currently, the computational aspects of the attention process[2584.49] [2584.49][S02]are quite expensive.[2585.63] [2585.63][S02]So the longer you try to make it, the more expensive it gets.[2589.232] [2589.232][S01] Expensive in terms of time but also computing--[2591.69] [2591.69][S02] Compute, time--[2592.47] [2592.47][S01] --and eventually money.[2593.928] [2593.928][S02] Time and money and compute and all kinds of things.[2597.45] [2597.45][S02]But we think it may be possible to sort of come up with[2602.34] [2602.34][S02]algorithmic improvements that enable you to go beyond[2605.31] [2605.31][S02]the 2-million token context window,[2608.337] [2608.337][S02]which is what we have now.[2609.42] [2609.42][S02]I mean, a million tokens is quite a lot.[2611.37] [2611.37][S02]A million tokens is about 600 pages of text.[2615.54] [2615.54][S02]So that's like most books, 20 articles.[2620.67] [2620.67][S02]It's an hour of video.[2621.713] [2621.713][S01] How about on the other side?[2623.38] [2623.38][S01]Because you said it was a little bit of both,[2624.89] [2624.89][S01]that perhaps people have to adjust their expectations.[2627.14] [2627.14][S02] Yeah, so I think these models are tools,[2631.47] [2631.47][S02]and people need to understand the capabilities of their tools[2635.56] [2635.56][S02]but also some of the ways in which you probably[2638.61] [2638.61][S02]don't want to use the tool.[2641.99] [2641.99][S02]So, I mean, I think it's a bit of a educational process[2648.1] [2648.1][S02]for people.[2651.13] [2651.13][S02]Don't just trust every fact that comes out of a language model[2655.66] [2655.66][S02]right off the bat.[2656.81] [2656.81][S02]You need to apply a bit of scrutiny to that.[2661.51] [2661.51][S02]Sort of like--[2663.256] [2663.256][S02]I think we've taught people these days[2666.22] [2666.22][S02]that if you see something online,[2668.9] [2668.9][S02]that doesn't necessarily make it true.[2670.81] [2670.81][S02]I think a similar degree of skepticism[2672.985] [2672.985][S02]for some kinds of things from language models[2674.86] [2674.86][S02]is probably also appropriate.[2676.72] [2676.72][S02]That skepticism may decrease over time as the models improve,[2681.73] [2681.73][S02]but it's good to take it with a healthy dose of,[2685.6] [2685.6][S02]oh, that might not actually be true.[2689.453] [2689.453][S01] Aside from the context[2690.87] [2690.87][S01]windows, are there ways that you can[2694.5] [2694.5][S01]yourself, when you're sort of writing in prompts, sort of[2697.69] [2697.69][S01]minimize the risk of ending up with something that's[2701.37] [2701.37][S01]a complete hallucination?[2702.64] [2702.64][S02] So one technique that Google researchers kind of[2707.61] [2707.61][S02]came up with is what's called chain of thought prompting.[2711.22] [2711.22][S02]So in the same way if you just give the model[2716.19] [2716.19][S02]a sort of interesting math problem,[2718.17] [2718.17][S02]and you say, OK, what's the answer, it may get it correct,[2724.44] [2724.44][S02]but it may not.[2725.9] [2725.9][S02]And if instead, you say, here's an interesting math problem,[2730.41] [2730.41][S02]can you please show your work step by step--[2732.747] [2732.747][S02]so if you remember back to your fourth grade math teacher--[2735.205] [2735.205][S01] Oh, I do.[2735.77] [2735.77][S02] --he or she was probably saying,[2737.562] [2737.562][S02]you should really show your work step by step[2739.92] [2739.92][S02]and then get to the final answer and then[2741.933] [2741.933][S02]write the final answer down.[2743.1] [2743.1][S02]And that's partly because that helps[2746.6] [2746.6][S02]you get through that multi-step thinking process of how[2750.23] [2750.23][S02]do you actually go from what's being asked to, OK,[2753.48] [2753.48][S02]I need to calculate this and calculate this based on that[2756.48] [2756.48][S02]and so on and finally get to the answer.[2758.85] [2758.85][S02]And It turns out that not only makes the model's output more[2765.66] [2765.66][S02]interpretable because it kind of tells you[2767.73] [2767.73][S02]what steps it's going through, but it also[2770.632] [2770.632][S02]makes it more likely to get the correct answer.[2772.59] [2772.59][S01] What if it's not a math problem, though?[2774.757] [2774.757][S02] Yeah, I mean, even in non-crisply[2777.87] [2777.87][S02]defined right answer things, domains,[2782.02] [2782.02][S02]this approach kind of works.[2783.78] [2783.78][S02]And there's a bit of subtlety.[2786.66] [2786.66][S02]And I think people need to actually learn[2789.45] [2789.45][S02]how to use these models.[2790.93] [2790.93][S02]And the way in which you prompt them[2793.17] [2793.17][S02]is actually a big differentiator in how high quality[2797.88] [2797.88][S02]the output is.[2798.69] [2798.69][S02]Like, if you say, summarize this,[2801.81] [2801.81][S02]that might lead to one outcome.[2803.31] [2803.31][S02]If you say, please summarize this and give me[2806.64] [2806.64][S02]five bullet points that highlight[2810.12] [2810.12][S02]the major important pieces of the article[2812.55] [2812.55][S02]and identify two cons that the author wrote down--[2817.78] [2817.78][S02]if you say that, that's a much clearer set[2820.213] [2820.213][S02]of instructions to what the model should[2821.88] [2821.88][S02]do than just summarize this.[2823.078] [2823.078][S01] So when we put these things together, then,[2825.37] [2825.37][S01]so sort of breaking down step-by-step processes[2828.77] [2828.77][S01]but also understanding more context[2831.91] [2831.91][S01]and the multimodal stuff, too, are we[2834.31] [2834.31][S01]moving towards a situation where these kind of multimodal models[2837.82] [2837.82][S01]will understand us as individuals and our preferences?[2840.97] [2840.97][S02] Yeah, I mean, I think--[2842.68] [2842.68][S02]what you really want, I think, is a sort[2845.62] [2845.62][S02]of very personal version of Gemini for you[2848.23] [2848.23][S02]that understands what it is you're trying to do right now[2851.87] [2851.87][S02]but also understands the context in which you're[2854.41] [2854.41][S02]trying to do that.[2855.76] [2855.76][S02]I'm vegetarian.[2856.76] [2856.76][S02]So if I'm asking Gemini about restaurant recommendations[2859.9] [2859.9][S02]in London, and it knew that I was a vegetarian,[2862.58] [2862.58][S02]it would recommend different things than if I was not.[2866.29] [2866.29][S02]And I think a general model that is serving[2869.95] [2869.95][S02]the needs of every person the same[2872.86] [2872.86][S02]is not going to be as good as one[2874.57] [2874.57][S02]that actually understands a lot about you and your context.[2879.29] [2879.29][S02]There are some kinds of queries you[2880.93] [2880.93][S02]might like to ask a model that you can't quite[2884.59] [2884.59][S02]do today with Gemini but you could imagine wanting to do.[2887.88] [2887.88][S02]Can you take the pictures I took on the hike last week[2891.16] [2891.16][S02]and make an illustrated storybook[2894.52] [2894.52][S02]for my kid's bedtime tonight?[2897.91] [2897.91][S02]And it would know where those pictures from on your hike were[2901.405] [2901.405][S02]and how to make an illustrated storybook that[2903.28] [2903.28][S02]would appeal to your child.[2904.787] [2904.787][S02]It would maybe know how old your child[2906.37] [2906.37][S02]was to make it age appropriate.[2909.08] [2909.08][S02]So I think-- you can't do that now,[2911.18] [2911.18][S02]but that could be something useful.[2914.36] [2914.36][S02]People would want-- you'd want people to opt in to that.[2917.43] [2917.43][S02]I think the more information you want the model to know and have[2922.55] [2922.55][S02]in context, I think the more you want to sort of have people[2927.72] [2927.72][S02]understand what is happening.[2930.915] [2930.915][S02]One of the things I think we'll be able to do[2932.79] [2932.79][S02]is not train a version of the model on that data[2936.22] [2936.22][S02]but just have the right information available in context[2939.27] [2939.27][S02]in order to sort of call upon it when generating responses.[2942.437] [2942.437][S02]And I think that would be pretty nice.[2944.02] [2944.02][S01] So then you've got like this general structure[2948.57] [2948.57][S01]that you can almost imprint your own context onto.[2952.29] [2952.29][S01]But then that's kind of private for you.[2953.98] [2953.98][S02] That's right.[2954.64] [2954.64][S01] Nice.[2954.96] [2954.96][S02] Yeah.[2955.17] [2955.17][S02]That seems like it'd be pretty good.[2956.46] [2956.46][S01] Yeah, that would be pretty good.[2958.33] [2958.33][S01]Are we limited here to just audio and visual and things[2962.987] [2962.987][S01]that you can see on a screen-- language, whatever?[2965.07] [2965.07][S01]Or do we ever expect that these kind of assistants[2968.49] [2968.49][S01]will come out of our computers, as it were?[2971.398] [2971.398][S02] Yeah, I mean, I think there's actually[2973.44] [2973.44][S02]a lot of different kinds of new modalities of data that aren't[2978.09] [2978.09][S02]sort of strictly human modalities[2980.13] [2980.13][S02]that we want these models to understand,[2982.15] [2982.15][S02]so lots of temperature readings across the Earth[2986.74] [2986.74][S02]to help with weather prediction or genetic sequences[2990.37] [2990.37][S02]or LiDAR data for autonomous vehicles or robotics[2994.78] [2994.78][S02]applications.[2996.02] [2996.02][S02]And then in on setting, you want these models to perhaps[3002.34] [3002.34][S02]be able to help with real-world robotics applications,[3006.52] [3006.52][S02]be able to talk to a robotic device,[3009.67] [3009.67][S02]give it sort of instructions in plain language.[3013.51] [3013.51][S02]Can you please go into the kitchen[3015.52] [3015.52][S02]and wipe the counter down and recycle the soda[3020.35] [3020.35][S02]can I left on the counter and then[3022.0] [3022.0][S02]bring me a bag of pistachios or something, right?[3025.665] [3027.227][S02]Robots have traditionally not been able to understand language[3029.81] [3029.81][S02]like that, but I think we're on the cusp of enabling[3032.84] [3032.84][S02]that kind of capability and then being able to have robots do[3036.89] [3036.89][S02]50 or 100 useful tasks, even in messy environments[3040.91] [3040.91][S02]like this room, rather than the traditional setting in which[3045.17] [3045.17][S02]robots have already been deployed in the world, which[3047.63] [3047.63][S02]is very controlled environments like factory, assembly line[3051.86] [3051.86][S02]kind of things, where they go from there to there.[3053.97] [3053.97][S02]And it's a very predictable thing.[3055.41] [3055.41][S01] We've been talking here[3056.15] [3056.15][S01]as assistants, these things as being sort of augmenting[3059.0] [3059.0][S01]human abilities in that way.[3060.48] [3060.48][S01]And I can see it in medical settings, in education settings.[3063.9] [3063.9][S01]But is there more that the multimodal aspect of this[3068.33] [3068.33][S01]offers us in terms of, I don't know,[3070.52] [3070.52][S01]like how we understand the world?[3072.2] [3072.2][S02] Yeah, I mean, I think what these models can[3074.45] [3074.45][S02]do now is often do a few steps of reasoning[3078.62] [3078.62][S02]to get from what you asked it to do in order[3081.23] [3081.23][S02]to accomplish something.[3082.6] [3082.6][S02]And I think as these models improve in capability,[3085.98] [3085.98][S02]you'll be able to sort of get models[3088.41] [3088.41][S02]to work with you to do much more complex tasks.[3092.55] [3092.55][S02]And it's sort of a difference between[3097.97] [3097.97][S02]can you order a bunch of chairs at the chair[3101.0] [3101.0][S02]rental place versus plan me a conference, right?[3104.935] [3104.935][S01] Mmm.[3106.64] [3106.64][S02] The latter is much higher level, much more complex.[3112.272] [3112.272][S02]The right model would kind of ask you a bunch of follow up[3115.46] [3115.46][S02]questions because there's ambiguity in there.[3117.39] [3117.39][S02]How many people are coming?[3118.64] [3118.64][S02]What's it about?[3121.01] [3121.01][S02]Just like a human.[3122.06] [3122.06][S01] What country are you in?[3122.51] [3122.51][S02] Yeah, what country are you in?[3123.72] [3123.72][S02]Where do you want to have it?[3125.36] [3125.36][S02]When?[3128.06] [3128.06][S02]And then we'd set off and actually[3130.67] [3130.67][S02]be able to accomplish a lot of the kind of things[3133.25] [3133.25][S02]that you might want done in order[3135.92] [3135.92][S02]to do that high-level goal.[3137.31] [3137.31][S01] But then if you have this sort[3139.22] [3139.22][S01]of conceptual link, sort of these conceptual links--[3142.823] [3142.823][S01]I'm going back to cow here, right?[3144.24] [3144.24][S01]And it understands pictures.[3146.39] [3146.39][S01]And it understands, I don't know, I guess, gravity,[3149.52] [3149.52][S01]having seen videos on the internet.[3153.193] [3153.193][S02] It probably watched introductory lectures[3155.36] [3155.36][S02]on physics.[3155.86] [3155.86][S01] Right?[3156.61] [3156.61][S01]Oh, wow.[3157.53] [3157.53][S01]OK, so it understand it from that perspective.[3159.533] [3159.533][S02] Yeah.[3160.2] [3160.2][S01] Yeah.[3160.908] [3160.908][S02] Also seeing a bunch of things falling.[3163.65] [3163.65][S01] OK, so then could you go in one day[3166.15] [3166.15][S01]and say, draw me the blueprint for a really efficient[3169.21] [3169.21][S01]aeroplane?[3170.23] [3170.23][S02] Yeah, I mean, I think one[3171.73] [3171.73][S02]of the things these models need to be partnered with[3176.45] [3176.45][S02]is some exploratory process.[3181.11] [3181.11][S02]And that exploratory process can come in the form of maybe[3185.33] [3185.33][S02]it doesn't need to give you an answer in 200 milliseconds.[3189.27] [3189.27][S02]Maybe you'd be happy with your airplane tomorrow, right?[3193.835] [3193.835][S02]And so I think at that point, then you have a lot more[3197.18] [3197.18][S02]freedom in how would you design systems[3199.16] [3199.16][S02]to be able to efficiently do things like that where they can[3202.46] [3202.46][S02]go off and try a few experiments maybe in a simulator[3206.66] [3206.66][S02]that they have access to, or maybe they[3208.64] [3208.64][S02]create a simulator for basic fluid dynamics or something.[3213.0] [3213.0][S02]And they try a bunch of designs.[3217.14] [3217.14][S02]Maybe they have some ideas about what airplane shapes make sense,[3221.54] [3221.54][S02]having seen a bunch of existing airplanes.[3224.81] [3224.81][S02]And so then they can kind of try to accomplish[3227.3] [3227.3][S02]what it is you asked.[3228.48] [3228.48][S02]Hopefully, they first asked you, well,[3230.61] [3230.61][S02]what characteristics do you want your airplane to have?[3233.105] [3233.105][S01] It was a paper airplane all along[3234.98] [3234.98][S01]is what I wanted.[3235.67] [3235.67][S02] Paper airplane.[3236.34] [3236.34][S02]Yeah, it's important to know if it's paper.[3238.41] [3238.41][S02]That reduces the cost a lot.[3242.07] [3242.07][S02]So I think those kinds of things will come eventually.[3247.3] [3247.3][S02]It's a little hard to tell exactly when[3249.57] [3249.57][S02]those capabilities-- that's a pretty complicated sort[3252.01] [3252.01][S02]of integration of what you want the reasoning in the model[3256.62] [3256.62][S02]to do, the knowledge it needs, what you're asking it to do,[3260.4] [3260.4][S02]and how you're asking it.[3261.85] [3261.85][S02]But we're already seeing pretty big advances in capabilities[3266.04] [3266.04][S02]of these models over 5-year, 10-year periods.[3269.68] [3269.68][S02]And so over a 5-year, 10-year period, that might be possible.[3274.44] [3274.44][S02]It might even be sooner than that for can you help me[3278.07] [3278.07][S02]design an airplane with these characteristics.[3280.23] [3280.23][S01] But I guess these are[3281.605] [3281.605][S01]like the early, early precursors to what we[3284.04] [3284.04][S01]might hope Apollo would be.[3286.02] [3286.02][S02] Yeah, exactly.[3287.14] [3287.14][S02]That's why it's Gemini.[3288.12] [3288.12][S01] That's why it's Gemini.[3289.68] [3289.68][S01]Amazing.[3290.37] [3290.37][S01]Jeff, thank you so much for joining me.[3292.18] [3292.18][S02] It's a pleasure to be here.[3293.08] [3293.08][S02]Thank you for having me.[3294.08] [3294.08][S01] In a lot of ways, I think Jeff's whole story[3296.49] [3296.49][S01]is one about scale.[3298.05] [3298.05][S01]For Google Search, it was about how do you[3300.09] [3300.09][S01]get more of the web, more users, faster queries.[3303.84] [3303.84][S01]For neural networks, it was about more computing power,[3306.97] [3306.97][S01]more machines.[3308.17] [3308.17][S01]And in the recent era of machine learning,[3310.39] [3310.39][S01]it's been about more and more and more data.[3314.19] [3314.19][S01]But something emerges from all of that--[3316.27] [3316.27][S01]a genuine conceptual model of the world, one[3319.38] [3319.38][S01]that is capable of abstraction and has already a proven ability[3323.7] [3323.7][S01]to enhance human productivity.[3326.44] [3326.44][S01]And it's telling that Jeff isn't finished there.[3329.08] [3329.08][S01]There is more to come, more sensors, more modes and, when[3332.79] [3332.79][S01]combined with the reinforcement learning tools that were born[3335.37] [3335.37][S01]in this building, maybe also more progress on the path[3339.36] [3339.36][S01]to AGI.[3341.49] [3341.49][S01]If you've enjoyed this episode, please[3343.47] [3343.47][S01]make sure that you subscribe to our podcast.[3345.39] [3345.39][S01]And if you have any feedback, or you[3347.04] [3347.04][S01]want to suggest a guest that you'd like to hear from, then[3349.8] [3349.8][S01]why not leave us a comment on YouTube?[3351.65] [3351.65][S01]Until next time.[3353.72]"}
{"file_name": "audio/val_000017.wav", "transcription": "[1.458][MUSIC PLAYING][4.86] [6.804][S01] Welcome to \"Google DeepMind-- the Podcast.\"[9.7] [9.7][S01]I'm Professor Hannah Fry.[11.01] [11.01][S01]Now, I want to start today, unusually,[13.09] [13.09][S01]perhaps with a clip from another podcast.[16.17] [16.17][S01]Listen to this.[16.83] [16.83][AUDIO PLAYBACK][17.497] [17.497][S01]- What's the overall message here?[19.06] [19.06][S01]Is it social commentary, artistic expression, or just[23.34] [23.34][S01]a really elaborate joke?[24.55] [24.55][S01]- That's the beauty of this piece, I think.[26.46] [26.46][S01]It defies easy categorization.[28.87] [28.87][S01]It exists in this liminal space between language[32.009] [32.009][S01]and non-language, between art and absurdity.[34.007] [34.007][END PLAYBACK][34.59] [34.59][S01] This is a very interesting discussion,[36.673] [36.673][S01]which, as you might have guessed, is AI generated.[39.37] [39.37][S01]But what is notable about this particular clip,[42.055] [42.055][S01]aside from the fact that neither of the two podcast hosts[44.43] [44.43][S01]have ever existed, is that their conversation,[47.52] [47.52][S01]a mini treatise on human nature and our relationship with art,[51.3] [51.3][S01]was generated from the most unusual of prompts.[55.26] [55.26][S01]The podcast itself was created by a new feature[58.26] [58.26][S01]called Audio Overview, part of NotebookLM,[61.38] [61.38][S01]a personalized AI research assistant from Google Labs.[64.68] [64.68][S01]Now, NotebookLM is powered by Gemini,[67.04] [67.04][S01]and it lets you upload your sources, anything from PDFs[71.47] [71.47][S01]to videos to generate insights, explanations and, of course,[75.77] [75.77][S01]podcasts.[76.79] [76.79][S01]We often think of AI as just crunching through data[79.9] [79.9][S01]and spitting out answers, but NotebookLM[82.72] [82.72][S01]draws on expertise from storytelling[85.36] [85.36][S01]to present information in an engaging way.[88.46] [88.46][S01]And we wanted to see what happens[90.19] [90.19][S01]when you ask NotebookLM to analyze what most people would[94.63] [94.63][S01]consider to be nonsense, a single document containing[98.17] [98.17][S01]just two words repeated a thousand times over--[101.8] [101.8][S01]\"cabbage\" and \"puddle.\"[104.12] [104.12][S01]And here is the result.[105.27] [105.27][AUDIO PLAYBACK][105.65] [105.65][S01]- So I have to admit, at first, I was like,[107.442] [107.442][S01]what is going on here?[110.87] [110.87][S01]But the more I think about it, [LAUGHS] the curious I get.[114.28] [114.28][S01]- It is fascinating, isn't it?[116.3] [116.3][S01]We're like, dealing with this one-piece puzzle.[118.37] [118.37][S01]- Right[118.91] [118.91][S01]- And we're trying to figure out, well,[119.93] [119.93][S01]what does this piece tell us?[121.54] [121.54][S01]What do you think?[122.39] [122.39][S01]What's your first impression?[123.89] [123.89][S01]- Honestly, it's almost like hypnotic or something.[128.03] [128.03][S01]If you were really staring into a puddle and all[130.36] [130.36][S01]you saw were these cabbages floating around--[132.355] [132.355][S01]- I can see it.[132.98] [132.98][S01]- --it's a little unsettling but also kind of funny.[135.222] [135.222][END PLAYBACK][135.805] [135.805][S01] Several minutes of intellectual analysis[138.04] [138.04][S01]packed to the brim with seemingly relevant ideas that[141.43] [141.43][S01]are nowhere in the original document.[144.61] [144.61][S01]It's actually quite impressive, really.[147.8] [147.8][S01]I am joined today by two people who are deeply involved[150.64] [150.64][S01]in writing NotebookLM story.[152.51] [152.51][S01]Joining us from San Francisco is Steven Johnson,[155.06] [155.06][S01]NotebookLM's editorial director and also a \"New York Times\"[158.8] [158.8][S01]best-selling author.[160.13] [160.13][S01]And in Mountain View, California,[161.81] [161.81][S01]Raiza Martin is a senior product manager for AI at Google Labs[165.46] [165.46][S01]who leads the team behind NotebookLM.[167.74] [167.74][S01]Welcome to the podcast, both of you.[169.84] [169.84][S01]Now, I want to start with the feature[171.76] [171.76][S01]that everybody's been talking about, this Audio Overview.[175.61] [175.61][S01]And well, I understand that you've got a little clip[178.36] [178.36][S01]that you want to play me.[180.02] [180.02][S03] Yes, let's play the clip.[181.92] [181.92][S03]I think you will enjoy this, Hannah.[183.12] [183.12][S01] OK, here we[183.47] [183.47][AUDIO PLAYBACK][184.137] [184.137][S01]- Welcome back, everyone.[185.34] [185.34][S01]Ready for another deep dive?[187.04] [187.04][S01]Today, we're shrinking down, way down.[190.71] [190.71][S01]- Microscopic, you might say?[192.11] [192.11][S01]- Exactly.[192.61] [192.61][S01]Think about those tiny little droplets of water,[195.732] [195.732][S01]you know, like the ones you see on a freshly washed car.[198.065] [198.065][S01]- Oh yeah.[198.84] [198.84][S01]- But imagine those droplets clinging to an airplane wing.[201.257] [201.257][S01] [LAUGHS] Oh my gosh.[202.59] [202.59][S01]- Or on a plant leaf.[203.6] [203.6][S01]- Right.[204.1] [204.1][S01]Being sprayed with pesticides.[206.39] [206.39][S01]The way those droplets behave--[207.92] [207.92][S01] Oh, you guys.[208.22] [208.22][S01]- --is actually incredibly important for all kinds[210.65] [210.65][S01]of things--[211.19] [211.19][S01]- It is?[211.76] [211.76][S01]- Making planes safer.[213.06] [213.06][S01]- More efficient farming.[214.5] [214.5][S01]- Even figuring out how rain forms.[216.92] [216.92][S01]- Wow, that's fascinating.[218.73] [218.73][S01]- We're diving into some serious research today.[220.73] [220.73][END PLAYBACK][221.313] [221.313][S01] That was my PhD, the first page of my PhD thesis.[224.67] [224.67][S01]Extraordinary.[225.86] [225.86][S01]I mean, frankly, there is no good stuff,[227.78] [227.78][S01]apart from heavy equations in math.[230.31] [230.31][S01]OK.[230.81] [230.81][S01]Lots of things to notice about that.[232.59] [232.59][S01]For starters, they made it sound much more[234.47] [234.47][S01]exciting than it actually is.[236.585] [236.585][S03] That's the point.[237.96] [237.96][LAUGHS][238.82] [238.82][S01] But also, though, the sort of back and forth.[241.51] [241.51][S01]I mean, the two voices, they were finishing[243.66] [243.66][S01]each other's sentences.[244.75] [244.75][S01]It felt very fluid, pardon the pun, very natural.[249.707] [249.707][S03] Imagine defending your dissertation now.[252.04] [252.04][S03]You could just play the podcast and leave it at that, I think,[255.93] [255.93][S03]if you'd only had that at your disposal back then.[258.22] [258.22][S01] Raiza, have you been surprised by people's reaction[260.1] [260.1][S01]to this?[260.6] [260.6][S01]Because it's had really quite serious uptake, hasn't it?[263.73] [263.73][S02] Yes, and I think the most surprising thing to me,[266.92] [266.92][S02]and, really, equally delightful, is how people are using it.[271.18] [271.18][S02]I think I imagined how they might.[273.28] [273.28][S02]But I think the beautiful thing about launching something[275.91] [275.91][S02]with this much sort of excitement around it[278.1] [278.1][S02]is you see a whole new universe of what everybody has been[281.85] [281.85][S02]trying, from things that are funny,[283.42] [283.42][S02]things that are entertaining, things that are inspiring[286.17] [286.17][S02]or really meaningful.[287.8] [287.8][S02]It's just been incredible.[289.32] [289.32][S02]I actually probably spend a good chunk[291.045] [291.045][S02]of my day, a third of my day, just listening to these.[293.295] [293.295][LAUGHS][293.64] [293.64][S03] [LAUGHS] Really?[294.3] [294.3][S01] You set up a Discord server, didn't you,[296.467] [296.467][S01]just to let people share stories about the ways[298.763] [298.763][S01]that they're using it.[299.68] [299.68][S01]What kind of things have come up?[301.81] [301.81][S03] So that was an interesting example,[303.992] [303.992][S03]playing your dissertation, because one of the things that I[306.45] [306.45][S03]think genuinely surprised us is people[308.43] [308.43][S03]would put their CVs and their resumes in there,[311.01] [311.01][S03]and it was almost like a little like hype machine.[314.2] [314.2][S03]If you were feeling down about yourself,[316.12] [316.12][S03]you would listen to like a 10-minute audio conversation[320.58] [320.58][S03]between two very enthusiastic hosts.[322.29] [322.29][S03]You're like, wow, Steven has really done a lot in his career.[325.91] [325.91][S03]It's very impressive.[327.14] [327.14][S03]But actually, a more serious version of that--[329.39] [329.39][S03]I mean, that's kind of fun and playful,[331.22] [331.22][S03]but people are using it, like, you can kind of workshop[335.02] [335.02][S03]things you're working on.[336.59] [336.59][S03]So you can upload a short story you're working on[339.01] [339.01][S03]and say, hey, give me some constructive criticism on this.[343.09] [343.09][S03]And you listen to people talking about your work[347.65] [347.65][S03]and they're very good at pulling out[350.32] [350.32][S03]the kind of interesting twists or focusing[352.48] [352.48][S03]on the characters that are particularly compelling or not.[356.29] [356.29][S03]And so it's a way of getting a little--[358.0] [358.0][S03]it's almost like a little focus group for stuff[360.37] [360.37][S03]that you're working on, which is really amazing.[362.95] [362.95][S01] I guess also hearing people actually[364.99] [364.99][S01]talk about it out loud adds that kind of extra layer of, I don't[368.56] [368.56][S01]know, objectivity almost, Raiza.[371.17] [371.17][S02] I would say it's been really surprising[376.75] [376.75][S02]because if we think about it, a lot of the content or content[380.11] [380.11][S02]generation, if you just render it in text, it's not new.[384.1] [384.1][S02]It's like if I upload my CV and then I[386.8] [386.8][S02]have an LLM spit out something that[388.72] [388.72][S02]says like, oh, here's Raiza's career,[390.62] [390.62][S02]write a summary of sorts--[391.97] [391.97][S02]maybe there's a few interesting tidbits[393.79] [393.79][S02]that it pulls out here and there--[395.57] [395.57][S02]that was novel two years ago, and everybody[398.17] [398.17][S02]was excited by that.[399.41] [399.41][S02]But I think adding that new layer or that new modality[403.09] [403.09][S02]of just very human-like voices, I[405.94] [405.94][S02]think it connects with people in a very different way.[411.28] [411.28][S02]Personally, I call this type of technology human-like, where[415.66] [415.66][S02]you recognize it as being very similar to you[418.72] [418.72][S02]and it resonates with you in a different way as a result.[421.76] [421.76][S02]And I think the first time I listened to my CV,[425.33] [425.33][S02]I knew what to expect.[426.68] [426.68][S02]But when I heard it, I still felt that bubble inside of me,[430.28] [430.28][S02]like the woo![431.47] [431.47][S02][LAUGHS] And I think that's the magic of new modalities.[435.37] [435.37][S03] I think the other point on this[437.5] [437.5][S03]is that human beings have been learning and exchanging[440.26] [440.26][S03]information through conversation for hundreds of thousands[443.77] [443.77][S03]of years.[444.53] [444.53][S03]We've been learning by reading structured text[447.04] [447.04][S03]on a page for 500 years, and structured text[451.06] [451.06][S03]on a screen for 30 years.[454.72] [454.72][S03]And so when you activate that sense of a genuine human-like[458.83] [458.83][S03]conversation, it's just a deep, ancient kind of ancestral part[464.05] [464.05][S03]of who we are that--[466.06] [466.06][S03]I think that's one of the reasons why it just[468.64] [468.64][S03]lights up people when they hear it for the first time.[470.96] [470.96][S01] Also interesting, I think[471.76] [471.76][S01]that you decided to have two hosts rather[473.71] [473.71][S01]than just one person sort of talking into space, as it were,[477.32] [477.32][S01]which--[477.82] [477.82][S01]I guess it speaks to the point that you're making, Steven.[480.58] [480.58][S03] Yeah, it's just a very different format.[483.2] [483.2][S03]If you just have one person, it feels like text to speech,[485.85] [485.85][S03]right?[486.35] [486.35][S03]We've heard text to speech before.[487.767] [487.767][S03]You're just like, the computer is turning the text that it just[490.45] [490.45][S03]wrote into something I can listen to, which is great.[494.12] [494.12][S03]And we're interested in trying to figure out ways we[496.57] [496.57][S03]can do that in other formats.[498.142] [498.142][S03]But to get the conversation right--[499.6] [499.6][S03]and we can dive into this in more detail--[501.47] [501.47][S03]there are all these subtle things[502.9] [502.9][S03]that you have to make work.[504.44] [504.44][S03]Nobody wants to listen to two robots talk to each other.[507.46] [507.46][S03]That will fail and be unlistenable after 30 seconds.[512.59] [512.59][S03]You have to master all these very subtle, weird things[515.77] [515.77][S03]that people do in conversation for it to work.[517.76] [517.76][S01] To make it human-like, exactly as you said.[520.052] [520.052][S01]Raiza, I want to come back to those features[521.95] [521.95][S01]a little bit later, to the Audio Overview,[524.42] [524.42][S01]because I also wanted to discuss the origins of NotebookLM.[528.58] [528.58][S01]How did it come about, Raiza?[530.353] [530.353][S02] For one, I think a lot of people[532.27] [532.27][S02]think that NotebookLM is new because of the Audio Overview[535.48] [535.48][S02]feature.[536.0] [536.0][S02]We had such a massive influx of people and people[538.51] [538.51][S02]were like, wow, what is this?[539.93] [539.93][S02]A brand-new thing from Google.[541.4] [541.4][S02]But actually, we've been working on NotebookLM for over a year.[544.49] [544.49][S02]We first announced it at Google I/O[547.24] [547.24][S02]last year, as Project Tailwind.[549.43] [549.43][S02]And before then, we actually had been incubating it inside[553.12] [553.12][S02]of Google Labs.[554.39] [554.39][S02]And it's actually how Steven and I met.[557.38] [557.38][S02]Steven was brought in.[559.288] [559.288][S02]What was your original title, Steven?[560.83] [560.83][S03] I was visiting scholar.[562.42] [562.42][S02] Yeah.[562.69] [562.69][LAUGHS][562.96] [562.96][S03] Yes.[563.9] [563.9][S03]Then I became editorial director.[566.157] [566.157][S02] That's right, he was promoted.[567.99] [567.99][S02]And at the time, Josh Woodward, who now leads Google Labs--[573.18] [573.18][S02]he's the vice president--[574.97] [574.97][S02]told me, he was like, I want you to build a new AI business.[578.85] [578.85][S02]And I thought to myself, what does it[580.47] [580.47][S02]take to actually do that?[581.83] [581.83][S02]But what I'll say is, one of my early inspirations[584.37] [584.37][S02]was just watching Steven work.[586.26] [586.26][S02]Honestly, just understanding how he does, what he does,[589.398] [589.398][S02]I was like, wow, that could be a real superpower if you[591.69] [591.69][S02]could give that to people.[592.92] [592.92][S03] It was a mix of Steven[594.503] [594.503][S03]is abnormal in his research habits,[597.75] [597.75][S03]but maybe we could turn this into a mainstream pursuit[600.0] [600.0][S03]somehow.[600.72] [600.72][S03]Yeah, it was interesting because I had had this long history[605.34] [605.34][S03]writing books, and Josh had read some of those books[608.97] [608.97][S03]and had read some things that I was writing about,[611.28] [611.28][S03]tools for thought, basically, like how do you use software[614.07] [614.07][S03]to help you think and help you develop your ideas and research?[617.05] [617.05][S03]This is the middle of 2022, so language models[619.32] [619.32][S03]were at the top of the list then.[620.94] [620.94][S03]And so he kind of reached out to me and said,[622.96] [622.96][S03]hey, any chance you would want to come to Google[625.68] [625.68][S03]and help build the tool that you have always[628.02] [628.02][S03]wanted to help people learn and organize their ideas,[632.07] [632.07][S03]now built on top of language models?[633.81] [633.81][S03]And what Raiza and I-- kind of right from the beginning--[636.28] [636.28][S03]I think I met Raiza day two at Google.[639.64] [639.64][S03]We were like, let's build something new.[641.51] [641.51][S01] This came about at a time[642.64] [642.64][S01]when large language models were at the top of the agenda.[645.015] [645.015][S01]In those early conversations, how[646.39] [646.39][S01]did you see this as being fundamentally different to just,[649.658] [649.658][S01]I don't know, like uploading a document on Gemini[651.7] [651.7][S01]and getting it to summarize it for you?[653.29] [653.29][S03] From the very beginning--[654.41] [654.41][S03]we call it source grounding.[655.85] [655.85][S03]That's the way we describe it.[657.1] [657.1][S03]You supply the source information[658.48] [658.48][S03]that you want to work with.[659.36] [659.36][S03]It might be the story you're writing.[660.77] [660.77][S03]It might be the book you're researching.[662.437] [662.437][S03]It might be your journals.[663.71] [663.71][S03]It might be the marketing documents you're working on.[665.96] [665.96][S03]And uploading that to the model then creates[668.53] [668.53][S03]a kind of personalized AI that is an expert in the information[672.52] [672.52][S03]that you care about.[673.64] [673.64][S03]And that was not-- no one was talking about that[676.36] [676.36][S03]in the middle of 2022.[678.17] [678.17][S03]So that was like the first thing we built, was like--[680.87] [680.87][S03]I mean, we uploaded part of one of my books,[683.32] [683.32][S03]and I could have this very crude conversation with the model that[687.46] [687.46][S03]was not at all like what you see now in text or with audio.[691.13] [691.13][S03]But you could get a little taste of what[693.67] [693.67][S03]it would be like to have all the ideas you were working with[697.77] [697.77][S03]instead of just talking to an open-ended model that just[700.17] [700.17][S03]had its general knowledge, actually have[701.64] [701.64][S03]that personalized knowledge.[702.807] [702.807][S03]And it was great because it also reduced hallucinations.[706.36] [706.36][S03]It made it more factual.[707.65] [707.65][S03]You could fact-check it.[709.17] [709.17][S03]You could go back and see the original source material.[711.73] [711.73][S03]That's a big part of the whole NotebookLM experience.[714.75] [714.75][S03]That was the beginning of it.[716.02] [716.02][S03]And everything we've done is built on that platform,[718.187] [718.187][S03]and Audio Overviews is just, OK, take that insight of,[720.75] [720.75][S03]I supply my sources and now, I turn it into something else.[723.46] [723.46][S03]In this case, it's an audio conversation.[725.64] [725.64][S01] Because I guess the real key difference[727.08] [727.08][S01]here is that it's very focused on the sources[729.0] [729.0][S01]that you're giving it and anything[731.08] [731.08][S01]that's connected to that rather than just as you[733.08] [733.08][S01]say, this general model.[734.7] [734.7][S02] Yeah, I think that--[736.71] [736.71][S02]I'll say, too, that what we've seen[739.53] [739.53][S02]is I think it's a little bit harder to get started[742.41] [742.41][S02]with this paradigm because it's so new.[744.63] [744.63][S02]The idea that one, you're talking to an AI, two,[748.065] [748.065][S02]you have to bring your own stuff.[749.44] [749.44][S02]So I think there's a little bit of a layer where[750.93] [750.93][S02]it's like, OK, you have to convince somebody[752.763] [752.763][S02]that it's worth doing.[753.88] [753.88][S02]But once you can get somebody over that hump,[756.22] [756.22][S02]it's just massively useful because--[758.5] [758.5][S02]I think about the work that I do every day,[760.61] [760.61][S02]the work Steven does every day, and many people[763.177] [763.177][S02]around the world that work on computers every day,[765.26] [765.26][S02]we are working with very specific sets of information,[769.43] [769.43][S02]shared contexts that we have with others.[772.21] [772.21][S02]We do research, we pull it in, we[774.1] [774.1][S02]want to extract our own insights from it.[776.48] [776.48][S02]I think that's what makes NotebookLM really special[778.87] [778.87][S02]and has made it special from the beginning.[780.77] [780.77][S01] So it does include these text elements too then,[782.84] [782.84][S01]because as you say, the podcast part[784.34] [784.34][S01]is the bit that's most notable.[787.055] [787.055][S02] That's right.[788.18] [788.18][S02]So the podcast thing is the most recent development[792.31] [792.31][S02]in NotebookLM, but we actually launched a year ago,[795.67] [795.67][S02]where it was primarily a chat feature.[798.53] [798.53][S02]So you're chatting with the system using your sources,[802.97] [802.97][S02]and it's always referencing back to exactly what pieces[806.98] [806.98][S02]of your content that it used.[808.238] [808.238][S01] So give me some more mundane examples[810.28] [810.28][S01]of how people are using this, like on a day-to-day level[812.613] [812.613][S01]then, Steven.[814.0] [814.0][S03] Yeah, so we actually[816.19] [816.19][S03]see a huge amount of usage of the product[818.08] [818.08][S03]just with the text features.[819.61] [819.61][S03]And suddenly, you have this amazing resource[823.03] [823.03][S03]that can answer any question about all-- hundreds of pages[826.3] [826.3][S03]of documents.[827.68] [827.68][S03]And in the text version, you get citations and everything.[830.787] [830.787][S03]It's a very scholarly thing, actually,[832.37] [832.37][S03]and you would appreciate it.[833.537] [833.537][S03]You get your answers back and every fact that the model says[837.34] [837.34][S03]has a little inline footnote, and you[839.452] [839.452][S03]can click directly on that footnote[840.91] [840.91][S03]and go and read the original passage.[843.17] [843.17][S03]Writers, journalists, obviously, are using it.[845.33] [845.33][S03]This comes a little bit out of my involvement for the project.[848.21] [848.21][S03]I have one notebook that has thousands and thousands[852.075] [852.075][S03]of quotes from books that I've read over the years,[854.2] [854.2][S03]plus a lot of the text of books that I've written.[856.7] [856.7][S03]And that notebook has basically like my brain[860.95] [860.95][S03]kind of captured in the AI.[862.85] [862.85][S03]And so whenever I work on anything, I[864.935] [864.935][S03]have a new idea for something, I'll go into that notebook[867.31] [867.31][S03]and be like, hey, what do you think about this idea?[870.825] [870.825][S03]And the AI will say, hey, Steven,[872.87] [872.87][S03]you read something related to that seven years ago.[875.78] [875.78][S03]What about this passage?[877.16] [877.16][S03]And so it's a true extension of my memory, so that kind[882.15] [882.15][S03]of stuff.[882.65] [882.65][S03]And the other thing, last thing I'll say,[883.51] [883.51][S03]is we're not training the model on this information.[886.1] [886.1][S03]So your information is secure, it's private.[888.048] [888.048][S03]It's not going to get into the general knowledge of the model[890.59] [890.59][S03]and be used by somebody else.[891.89] [891.89][S03]So you can put private information in there.[894.35] [894.35][S03]And when you put a couple of years of your journal[899.02] [899.02][S03]in a large-context model like this,[902.89] [902.89][S03]you can get these amazing insights[904.96] [904.96][S03]and you can turn them into audio overviews[907.09] [907.09][S03]and listen to two people talk about yourself.[909.44] [909.44][S03]Or you can just be like, what was I thinking about last May?[913.13] [913.13][S03]Give me an overview of all the stuff that was going on.[916.18] [916.18][S03]And 20 seconds later, you'll have this amazing kind[919.06] [919.06][S03]of document of your own life.[920.485] [920.485][S01] Rather than just recalling stuff,[922.36] [922.36][S01]they can actually be insightful, in terms of your own journals.[925.398] [925.398][S02] I would say yes, because I've[927.19] [927.19][S02]used it for that purpose.[928.94] [928.94][S02]And one of the things that I like to ask it after uploading--[932.72] [932.72][S02]I do these weekly journals --is I say, how much have I[936.01] [936.01][S02]changed over time?[937.84] [937.84][S02]And it's really remarkable.[940.55] [940.55][S02]It's been able to pull out for me, really interesting nuances[943.94] [943.94][S02]that I haven't been able to observe about myself.[946.94] [946.94][S02]It's been able to say things like hey,[949.46] [949.46][S02]you tend to associate a lot of negativity[951.68] [951.68][S02]with this particular topic.[953.91] [953.91][S02]You associate a lot of positivity with this topic.[956.4] [956.4][S02]And it's just really interesting because I think, to your earlier[959.45] [959.45][S02]question around the mundane use cases,[962.28] [962.28][S02]I think we see a lot more of those,[965.52] [965.52][S02]which is just people trying to take the work that they're[968.6] [968.6][S02]doing every day.[969.69] [969.69][S02]For example, sales teams use this a lot[971.93] [971.93][S02]to share knowledge with each other.[973.46] [973.46][S02]Makes a lot of sense.[974.43] [974.43][S02]There's a lot of technical, complex changing documentation,[977.61] [977.61][S02]so it's really nice to have an AI partner.[979.65] [979.65][S02]I think that's really different from how a lot of AI systems[983.42] [983.42][S02]work today, right?[984.9] [984.9][S02]I use everything.[987.24] [987.24][S02]I use everything that's out there,[988.71] [988.71][S02]and the prompts that I write are massive.[991.77] [991.77][S02]The first thing that I write is, you are a blah.[994.5] [994.5][S02]This is what we are doing.[996.03] [996.03][S02]Here are the documents that are relevant.[998.22] [998.22][S02]And I think for NotebookLM, this just shortcuts it.[1001.16] [1001.16][S02]It's just a project space.[1002.282] [1002.282][S02]It knows what you're talking about.[1003.74] [1003.74][S02]You can have a conversation forever.[1005.9] [1005.9][S02]It takes up to 25 million words.[1007.82] [1007.82][S02]It's just contextually quite massive.[1010.687] [1010.687][S03] I think one of the things that was interesting[1013.27] [1013.27][S03]and maybe a little bit distinctive about it[1015.062] [1015.062][S03]was so many of the questions about what makes this product[1018.1] [1018.1][S03]work or not work are not so much technological questions[1021.73] [1021.73][S03]as they are editorial stylistic questions.[1023.89] [1023.89][S03]Like, what is the right kind of answer when you get[1026.38] [1026.38][S03]an audio overview that works?[1030.579] [1030.579][S03]What's the style?[1031.52] [1031.52][S03]What's the house style for those conversations?[1034.0] [1034.0][S03]What level should they be pitched at?[1035.66] [1035.66][S03]And those are not technological questions.[1037.73] [1037.73][S03]Those are language questions.[1039.619] [1039.619][S03]And that's the crazy reality of the language-model age,[1043.383] [1043.383][S03]is that all these things that used to be just mostly[1045.55] [1045.55][S03]a question of, let's get the programming right now,[1048.14] [1048.14][S03]become more about the rhetoric of it all.[1050.973] [1050.973][S01] Well, actually, I want[1052.39] [1052.39][S01]to dig into some of the house style a little bit more,[1055.41] [1055.41][S01]I guess.[1055.91] [1055.91][S01]Why did you decide to go into Audio Overview?[1058.43] [1058.43][S01]What was it that inspired that?[1059.99] [1059.99][S01]I mean, there are already quite a lot of podcasts.[1062.17] [1062.17][S01]Let's be honest.[1063.58] [1063.58][S03] When Audio Overviews really began--[1066.02] [1066.02][S03]it was a great example of the lab's structure,[1069.35] [1069.35][S03]I think, really working well because it[1071.98] [1071.98][S03]was another small team inside of labs[1074.35] [1074.35][S03]that were just kind of focused on the audio version of this.[1077.565] [1080.71][S03]And part of the idea of it was not[1082.78] [1082.78][S03]so much to compete with podcasts but rather[1084.99] [1084.99][S03]that there was a whole universe of content[1086.74] [1086.74][S03]that you would never-- the economics of generating[1089.47] [1089.47][S03]a podcast for it would never make any sense.[1091.7] [1091.7][S03]But if you could generate one automatically,[1094.67] [1094.67][S03]you might have five people that would want to listen to it,[1097.132] [1097.132][S03]or one person who would want to listen to it, or 20 people,[1099.59] [1099.59][S03]but not 200,000.[1101.21] [1101.21][S03]And so we want to create a podcast[1104.47] [1104.47][S03]based on our team meetings from the last week[1106.69] [1106.69][S03]so we can review them.[1107.69] [1107.69][S03]That's not going to be a commercial business.[1110.05] [1110.05][S03]No one's going to ask you to host that,[1113.052] [1113.052][S03]but actually, it might be useful for that team.[1115.01] [1115.01][S03]And so they had started developing this thing,[1117.55] [1117.55][S03]and Raiza and I heard it probably[1120.96] [1120.96][S03]in March or April of this year.[1123.88] [1123.88][S03]And as everyone who's heard an audio overview,[1127.17] [1127.17][S03]initially, we're just like, wow, what did I just hear?[1129.97] [1129.97][S03]That was amazing.[1130.93] [1130.93][S03]But we realized pretty early on that part of our mission[1135.03] [1135.03][S03]with NotebookLM was to build a tool that[1137.22] [1137.22][S03]helps people understand things.[1138.7] [1138.7][S03]And suddenly, we were like, oh wait, people really[1141.09] [1141.09][S03]understand, and remember, and pay attention[1144.21] [1144.21][S03]when they hear something in the form of an engaging conversation[1147.75] [1147.75][S03]between two smart people.[1149.4] [1149.4][S03]We've released it internally to Googlers over the summer,[1153.34] [1153.34][S03]and that was, I think, when we started[1155.31] [1155.31][S03]to think this is going to be a hit,[1157.8] [1157.8][S03]because you could just see the delight that people had with it.[1161.1] [1161.1][S03]So while we were surprised that it[1163.86] [1163.86][S03]went quite as crazy as it did, we knew we were on to something.[1167.52] [1167.52][S01] Now, I remember the last season,[1170.01] [1170.01][S01]we got to hear a demo of WaveNet, which, of course, is[1172.62] [1172.62][S01]one of the first AI models to generate this human-like speech.[1176.26] [1176.26][S01]And it was quite impressive back then,[1179.54] [1179.54][S01]but I mean, presumably, there have[1181.22] [1181.22][S01]been technological advancements that have happened since, that[1183.915] [1183.915][S01]have been necessary to make something like Audio Overview[1186.29] [1186.29][S01]possible.[1187.32] [1187.32][S02] I think the underlying model for NotebookLM[1191.09] [1191.09][S02]is Gemini 1.5 Pro, and that just creates, really,[1196.76] [1196.76][S02]to me, incredible content.[1199.5] [1199.5][S02]The voice models, the audio model that we use, that,[1203.0] [1203.0][S02]by itself, is a breakthrough.[1204.795] [1204.795][S02]And I think that's what you're talking about,[1206.67] [1206.67][S02]which is the realism of the human voice,[1209.37] [1209.37][S02]the human-like voices that we hear,[1211.4] [1211.4][S02]and pair that with the approach that we've taken--[1215.04] [1215.04][S02]and Steven can speak more to this, too--[1217.46] [1217.46][S02]of editorializing the content to thinking about,[1220.19] [1220.19][S02]how do we create something really useful[1222.26] [1222.26][S02]and really fun for you that's engaging?[1224.85] [1224.85][S03] Yeah, that's a great segue actually,[1228.03] [1228.03][S03]to something I was going to say, which is about interestingness.[1231.32] [1231.32][S03]So Simon, who's one of the leads on the audio side,[1237.02] [1237.02][S03]he sometimes has a slogan for Audio Overviews, which[1240.57] [1240.57][S03]is make anything interesting.[1242.47] [1242.47][LAUGHS][1243.148] [1243.148][S03]So, like, whatever, make your dissertation interesting.[1245.44] [1245.44][S03]I'm sure it was interesting.[1246.69] [1246.69][S01] It wasn't.[1247.607] [1247.607][LAUGHS][1249.218] [1249.218][S03] And so it's a great example[1251.01] [1251.01][S03]of a convergence of three different technologies[1254.67] [1254.67][S03]or breakthroughs that make something magical happen.[1257.97] [1257.97][S03]Gemini itself, and it can do this with text as well,[1261.76] [1261.76][S03]is incredibly good at pulling out[1263.97] [1263.97][S03]interesting facts or ideas or stories[1266.51] [1266.51][S03]from the material you give it.[1267.76] [1267.76][S03]So I do this all the time.[1269.14] [1269.14][S03]I upload something new and say, tell me[1270.79] [1270.79][S03]the most interesting things from this, just in text.[1272.86] [1272.86][S03]Computers could never do that before.[1274.402] [1274.402][S03]You couldn't command F for interestingness.[1276.37] [1276.37][S03]This was not a search query you could do.[1278.125] [1278.125][S01] But how are you defining it even?[1280.0] [1280.0][S01]I mean, what does it mean?[1281.108] [1281.108][S03] I believe that it comes out[1282.9] [1282.9][S03]of the basic idea behind language models, which[1285.84] [1285.84][S03]is that they're predictive.[1287.1] [1287.1][S03]They're like, given this string of text,[1289.15] [1289.15][S03]I expect the next thing to happen.[1290.71] [1290.71][S03]And so what interestingness is kind of controlled surprise.[1295.733] [1295.733][S03]I thought this was going to be the case,[1297.4] [1297.4][S03]but actually, there's some new information here[1299.358] [1299.358][S03]that I wasn't expecting.[1300.61] [1300.61][S03]And so it makes sense, in a way, that the language models[1304.38] [1304.38][S03]would be good at this because their basic circuitry is[1307.95] [1307.95][S03]prediction.[1308.59] [1308.59][S03]And so they're looking through all this information.[1310.757] [1310.757][S03]Given their training data, what in this information[1312.99] [1312.99][S03]is novel or--[1313.867] [1313.867][S01] Surprising.[1314.825] [1314.825][S03] --defies their expectations.[1316.57] [1316.57][S03]So it's very good at that.[1317.44] [1317.44][S03]So that's an underlying Gemini thing, right?[1319.33] [1319.33][S03]And the hosts of the show are instructed[1321.72] [1321.72][S03]to find the interesting material and present it[1325.95] [1325.95][S03]to the user in an engaging way.[1328.03] [1328.03][S03]So that's one capability.[1329.71] [1329.71][S03]The second thing that is really cool about this[1332.4] [1332.4][S03]is that the instructions take the script that is generated,[1337.45] [1337.45][S03]and they add noise to the script.[1339.73] [1339.73][S03]So they add what are called disfluencies,[1342.1] [1342.1][S03]so all the stammers, and the \"likes,\"[1344.04] [1344.04][S03]and the interjections that humans actually have when they[1348.21] [1348.21][S03]speak.[1348.73] [1348.73][S03]And it turns out you need that because if you[1351.03] [1351.03][S03]don't have that noise, it sounds too robotic.[1354.653] [1354.653][S01] Mhm.[1355.32] [1355.32][S03] And then finally, there's[1357.028] [1357.028][S03]the audio voices themselves, and what[1358.71] [1358.71][S03]they do is all these subtle things, like in English,[1363.17] [1363.17][S03]speakers will raise their voice a little bit if they're not[1366.168] [1366.168][S03]sure about what they're saying.[1367.46] [1367.46][S03]Or for emphasis, they will slow down what they're saying.[1370.47] [1370.47][S03]All these things that we do natively,[1373.14] [1373.14][S03]we never even think about it.[1374.43] [1374.43][S03]But no computer could do that until now,[1376.86] [1376.86][S03]and that's the part of it that just like lights up.[1379.35] [1379.35][S03]And that's the underlying language vocal model,[1382.14] [1382.14][S03]audio model, that didn't exist a year ago.[1384.935] [1384.935][S01] It's the voice modulation, right?[1386.81] [1386.81][S01]It's like when you--[1389.39] [1389.39][S01]I remember years ago at the BBC, being taught[1394.25] [1394.25][S01]to make content sound engaging.[1396.86] [1396.86][S01]And they give you a copy of \"Winnie the Pooh\" to read.[1399.56] [1399.56][S01]And then they say, OK, read it as you would a newsreader[1402.89] [1402.89][S01]and you read it very, very flat.[1404.28] [1404.28][S01]And then they say, read it as you would to a child,[1406.07] [1406.07][S01]and you notice, exactly as you say,[1407.61] [1407.61][S01]Steven, that your voice goes up at certain points[1409.89] [1409.89][S01]and it goes down at other points.[1411.66] [1411.66][S01]The range that you have and the speeds completely changes.[1414.33] [1414.33][S01]But you've built all of those aspects into this.[1418.05] [1418.05][S01]I mean, how on Earth do you do that?[1420.655] [1420.655][S03] Yeah, we should make it very clear,[1422.78] [1422.78][S03]we do not build the vocal model.[1424.902] [1424.902][S01] You plural, you plural.[1426.36] [1426.36][S03] We have no idea how it was built,[1428.402] [1428.402][S03]and geniuses inside of Google built that,[1431.54] [1431.54][S03]and we inherited that technology.[1433.11] [1433.11][S03]And we have been running with it and showing[1434.63] [1434.63][S03]how it could be useful, but we did not build it.[1436.64] [1436.64][S03]One of the questions that people have is--[1441.02] [1441.02][S03]it's English only right now, and people[1443.27] [1443.27][S03]are very eager for it to come into different languages.[1446.55] [1446.55][S03]And we are very eager for that, too,[1448.655] [1448.655][S03]because we have a wonderful international audience.[1450.78] [1450.78][S03]But it's not something you can do easily[1452.51] [1452.51][S03]because the intonations and all those little conversational tics[1456.47] [1456.47][S03]are different in every language.[1458.16] [1458.16][S03]And so you can't just be like, change the words into Spanish[1461.15] [1461.15][S03]and press Play.[1462.76] [1462.76][S02] I was just going to add that DeepMind actually[1465.26] [1465.26][S02]has a recent blog post about the audio model,[1469.37] [1469.37][S02]and how it was built, and who built it, and all the research[1472.07] [1472.07][S02]papers underneath it.[1473.4] [1473.4][S02]I think if we could share that, we should.[1476.022] [1476.022][S01] Yeah.[1476.73] [1476.73][S01]No, absolutely.[1477.71] [1477.71][S01]I think one thing that's really noticeable playing around[1480.085] [1480.085][S01]with this is how it is very versatile across different types[1484.56] [1484.56][S01]of data that you give it.[1486.72] [1486.72][S01]And so the way that you're describing[1488.3] [1488.3][S01]this, Steven, is that you're sort[1489.68] [1489.68][S01]of coding in all of the disfluencies,[1492.5] [1492.5][S01]but how do you stop this thing from just[1495.14] [1495.14][S01]sounding like a bunch of cliches every single time?[1497.51] [1497.51][S01]Raiza.[1498.482] [1498.482][S02] I actually think it's[1499.94] [1499.94][S02]hard to get it not to sound like a bunch of cliches every time.[1503.61] [1503.61][S02]I think because of trying to standardize interestingness,[1509.25] [1509.25][S02]that's really, actually quite difficult.[1512.16] [1512.16][S02]And so interestingness tends to sound the same after hearing it[1515.78] [1515.78][S02]enough times.[1517.4] [1517.4][S02]And that's why we actually introduced the first improvement[1520.55] [1520.55][S02]to this particular launch, which is, we're letting users--[1524.4] [1524.4][S02]I call it pass a note to the hosts,[1526.46] [1526.46][S02]where you could slip them a little instruction on, hey, you[1529.35] [1529.35][S02]know what, maybe less of the cliche.[1531.51] [1531.51][S02]Go deeper on this topic.[1533.25] [1533.25][S02]And it will change the way that they talk about whatever content[1537.38] [1537.38][S02]you've given them.[1538.145] [1538.145][S01] Should I imagine this[1539.52] [1539.52][S01]as almost though you have different kind of dials?[1542.32] [1542.32][S01]Like, maybe you turn up the quirky dial,[1544.62] [1544.62][S01]and maybe you turn up the historical-fact dial.[1548.07] [1548.07][S01]Or how can I think of this?[1550.77] [1550.77][S03] Well, imagine one thing[1552.87] [1552.87][S03]that I'm very interested in.[1555.58] [1555.58][S03]What if you could give each of the hosts[1558.87] [1558.87][S03]a different kind of field of expertise?[1562.62] [1562.62][S03]Right now, they basically are kind of interchangeable.[1565.6] [1565.6][S03]They don't have defined perspectives on the world.[1568.38] [1568.38][S03]One takes the lead in the conversation,[1570.6] [1570.6][S03]and we switch back and forth randomly.[1573.81] [1573.81][S03]But what if you were like, OK, I'm a city planner,[1576.51] [1576.51][S03]and I'm working on this design for this new town square.[1580.2] [1580.2][S03]And I want one of them to be an environmental activist,[1582.96] [1582.96][S03]and I want one of them to be an economist.[1585.4] [1585.4][S03]And now, let's have a conversation[1587.79] [1587.79][S03]and let's have a debate.[1589.12] [1589.12][S03]And suddenly, they have different perspectives[1591.295] [1591.295][S03]because one of the things-- this is something[1593.17] [1593.17][S03]I've written about a lot in my books over the years[1595.83] [1595.83][S03]--is that people are more creative, make better decisions,[1599.97] [1599.97][S03]when they have a diverse pool of expertise helping[1603.72] [1603.72][S03]them make the choices or come up with the ideas they're[1606.09] [1606.09][S03]trying to do.[1606.73] [1606.73][S03]And that's also on our roadmap for 2025.[1611.002] [1611.002][S01] Will I actually be able to interact[1612.96] [1612.96][S01]with these hosts in the future?[1614.41] [1614.41][S01]Like, I don't know, interrupt them[1615.9] [1615.9][S01]and join their conversation?[1617.915] [1617.915][S03] Well, we actually[1619.29] [1619.29][S03]showed a version of this at I/O, the big Google Developers[1624.24] [1624.24][S03]Conference where we first rolled out this feature, announced it.[1627.22] [1627.22][S03]And they do their audio-podcast format,[1631.63] [1631.63][S03]and then Josh Woodward, the head of Labs, in the demo,[1634.78] [1634.78][S03]interrupts and says, hey, can you--[1636.39] [1636.39][S03]they're talking about physics.[1638.22] [1638.22][S03]And he's like, hey, can you use a basketball metaphor here,[1641.44] [1641.44][S03]because my son is listening?[1642.93] [1642.93][S03]And they're like, oh, great.[1644.56] [1644.56][S03]OK.[1645.3] [1645.3][S03]Someone called into the show, basically, and they're like,[1648.06] [1648.06][S03]let's do it in a basketball metaphor.[1651.07] [1651.07][S03]So that has been, publicly, part of what we wanted to do,[1655.72] [1655.72][S03]and you can imagine we're very eager to bring that to people.[1660.37] [1660.37][S01] I mean, you paint a really compelling picture.[1663.16] [1663.16][S01]I do also wonder, though, is there the danger[1665.46] [1665.46][S01]here that you could have--[1668.11] [1668.11][S01]picking up on a minor detail in the corpus of text,[1671.34] [1671.34][S01]and then make it into a much bigger thing than it necessarily[1674.42] [1674.42][S01]is?[1674.92] [1674.92][S01]I mean, we're still at the situation[1676.42] [1676.42][S01]where large language models can kind of hallucinate,[1679.78] [1679.78][S01]can not necessarily put the right emphasis[1681.66] [1681.66][S01]on different parts of what it's reporting.[1684.585] [1684.585][S03] The early days of three weeks[1686.46] [1686.46][S03]ago, when we were testing this customization passive note[1689.73] [1689.73][S03]from the producers feature that Raiza's talking about,[1691.98] [1691.98][S03]I uploaded an article I'd written a couple of years ago,[1695.04] [1695.04][S03]and I gave them the instructions to give me[1699.39] [1699.39][S03]relentless criticism of this piece in the style of an insult[1704.19] [1704.19][S03]comic at a roast because, again, they're kind of instructed[1709.26] [1709.26][S03]to be enthusiastic.[1710.56] [1710.56][S03]So I upload this piece, and it was cool.[1713.38] [1713.38][S03]They immediately were like, what is Johnson's problem?[1716.76] [1716.76][S03]Did he even do any research for this piece?[1719.19] [1719.19][S03]But they also kind of reached for a criticism of it that,[1723.78] [1723.78][S03]genuinely--[1724.62] [1724.62][S03]I'm not just saying this because I[1726.037] [1726.037][S03]wrote it and I'm defensive-- was kind of wrong.[1728.112] [1728.112][S03]It kind of misread it a little bit.[1729.57] [1729.57][S03]And I couldn't quite tell whether it[1731.07] [1731.07][S03]was because I'd instructed them to be so extreme[1733.56] [1733.56][S03]or whether they just-- it's almost like,[1735.872] [1735.872][S03]I keep saying this to people.[1737.08] [1737.08][S03]It's like they don't really hallucinate in the way[1739.29] [1739.29][S03]that the first-generation models do.[1741.43] [1741.43][S03]It's just that they sometimes get confused[1743.22] [1743.22][S03]or they misinterpret something in a way that humans do,[1746.37] [1746.37][S03]and their take is a little bit off.[1748.62] [1748.62][S01] Well, what about humor though?[1750.37] [1750.37][S01]I mean, we're talking about all of these different types[1752.25] [1752.25][S01]of examples.[1753.06] [1753.06][S01]Have they ever made you laugh?[1755.03] [1755.03][S02] Yes.[1755.78] [1755.78][S02]Yes.[1756.367] [1756.367][S03] Yes.[1757.2] [1757.2][S02] Actually, I will say[1758.617] [1758.617][S02]that they have made me laugh through the cleverness,[1763.35] [1763.35][S02]and the humor, and the exploration of other people,[1767.02] [1767.02][S02]because I myself--[1768.443] [1768.443][S02]I don't think I could have come up[1769.86] [1769.86][S02]with the funny cases on my own.[1771.76] [1771.76][S02]But just seeing what people have tried[1774.66] [1774.66][S02]in the outside world with the technology,[1777.98] [1777.98][S02]that's been really funny.[1779.64] [1779.64][S02]And somebody uploaded a document to NotebookLM,[1782.6] [1782.6][S02]and the document just said the words \"poop\" and \"fart\" in it.[1787.74] [1787.74][S02]And when I saw that that's what it was-- the person[1790.13] [1790.13][S02]posted it on Twitter.[1791.07] [1791.07][S02]They're like, that's all this is.[1792.445] [1792.445][S02]Listen to the podcast.[1793.92] [1793.92][S02]I was like oh dear.[1796.4] [1796.4][S02]What is this about to be?[1797.97] [1797.97][S02]But it was hilarious.[1799.38] [1799.38][S02]It was so good.[1800.43] [1800.43][S02]And the thing that makes it so funny[1803.27] [1803.27][S02]is that there were moments that were truly hilarious,[1806.3] [1806.3][S02]and then it would dip into, but what does it really mean?[1810.455] [1810.455][S02]And it would be thoughtful.[1811.58] [1811.58][S02]It would be bizarre.[1812.95] [1812.95][S02]It would be thought provoking.[1814.2] [1814.2][S02]And I'm like, am I really listening to this?[1816.502] [1816.502][S02]But I took it very seriously.[1817.71] [1817.71][S02]It was great.[1818.335] [1818.335][S01] Yeah, I guess in some ways,[1819.96] [1819.96][S01]though, that's sort of hilarious in the way[1822.56] [1822.56][S01]that the AI is kind of oblivious to how absurd the challenge is,[1827.96] [1827.96][S01]that it's been said.[1829.027] [1829.027][S02] I think on that one, they mentioned,[1831.11] [1831.11][S02]is somebody's trying to trick us into just saying a bunch[1833.84] [1833.84][S02]of \"poop\" and \"fart\"?[1835.73] [1835.73][S02]And I was like, I think so.[1838.35] [1838.35][S01] I do also think that the more traditional forms[1841.76] [1841.76][S01]of humor, so not just laughing at how oblivious the AI is,[1845.19] [1845.19][S01]but a lot of that seems to me like it's about the build up[1848.232] [1848.232][S01]and release of tension.[1849.19] [1849.19][S03] Yeah.[1849.51] [1849.51][S01] So, it's the kind of similar thing[1851.427] [1851.427][S01]about you're making a prediction of where you're[1853.485] [1853.485][S01]expecting a sentence to go, and then it[1855.11] [1855.11][S01]goes in a different direction.[1856.5] [1856.5][S01]Is this something that you think that it will[1858.74] [1858.74][S01]be able to do in the future?[1860.19] [1860.19][S01]Because I don't think it's particularly good at it now.[1862.792] [1862.792][S03] I actually had this sense[1864.5] [1864.5][S03]in the early days, the first couple of weeks,[1866.28] [1866.28][S03]really, that it was out--[1867.322] [1867.322][S03]I actually wrote about this briefly--[1869.58] [1869.58][S03]which was that they actually weren't very good at humor.[1872.34] [1872.34][S03]They had banter and they were playful,[1876.14] [1876.14][S03]but they didn't really like crack good jokes[1878.99] [1878.99][S03]or have genuinely funny things.[1880.53] [1880.53][S03]And then it turned out, as Raiza said,[1882.68] [1882.68][S03]that users were able to push them into being genuinely funny.[1886.77] [1886.77][S03]They had to be put in a funny situation, as it were.[1889.23] [1889.23][S03]Like, we've been given this poop-fart document.[1891.27] [1891.27][S03]Another one was a completely coherent-looking scientific[1895.91] [1895.91][S03]paper with charts, and graphs, and published in with footnotes[1899.268] [1899.268][S03]and everything, except that every word in the paper was[1901.56] [1901.56][S03]\"chicken,\" just \"chicken,\" \"chicken,\" \"chicken,\" \"chicken,\"[1903.37] [1903.37][S03]\"chicken,\" \"chicken.\"[1904.26] [1904.26][S03]And every footnote was \"chicken,\" \"chicken,\" \"chicken,\"[1905.92] [1905.92][S03]\"chicken.\"[1906.16] [1906.16][S03]All the charts were \"chicken.\"[1907.41] [1907.41][S03]And so they gave them that, and that was the first part where[1909.99] [1909.99][S03]I actually really laughed.[1911.197] [1911.197][S03]They were just like, what is even happening?[1913.03] [1913.03][S03]And they made some funny jokes.[1915.01] [1915.01][S03]And so it's like they have to be prodded[1918.15] [1918.15][S03]into it by an unusual situation, in a weird way.[1921.28] [1921.28][S01] You did mention something there, actually,[1923.53] [1923.53][S01]that I want to pick up on.[1924.613] [1924.613][S01]There are people who have made a criticism of this technology,[1927.36] [1927.36][S01]saying that it's a threat to the podcasting world, that you could[1930.45] [1930.45][S01]be flooding the podcasting world with lots[1932.97] [1932.97][S01]of generic, low-quality AI-generated podcasts.[1937.12] [1937.12][S01]Is there a response that you have to that?[1939.047] [1939.047][S02] What is most interesting and nuanced[1941.13] [1941.13][S02]about it is that what we've found[1943.38] [1943.38][S02]is that people are creating content[1945.18] [1945.18][S02]of things that probably don't have a podcast about it[1948.06] [1948.06][S02]to begin with.[1948.97] [1948.97][S02]It really is--[1950.02] [1950.02][S02]I don't want to say mundane.[1951.79] [1951.79][S02]But it really is things that nobody is going[1954.06] [1954.06][S02]to make a whole show about.[1956.007] [1956.007][S02]And I think that is interesting.[1957.34] [1957.34][S02]I think we're putting power in people's hands[1960.39] [1960.39][S02]to create content that they want that they ordinarily[1963.63] [1963.63][S02]wouldn't have access to.[1964.75] [1964.75][S02]The second piece of this around the low-quality content,[1970.71] [1970.71][S02]I would say that most of the content that I have heard,[1975.75] [1975.75][S02]ones on the internet, just people posting on the Discord,[1979.89] [1979.89][S02]the quality is quite high.[1981.55] [1981.55][S02]I think on the third note, all of the generations[1986.01] [1986.01][S02]from NotebookLM are also watermarked with SynthID,[1989.13] [1989.13][S02]and so we've taken a very responsible and cautious[1992.61] [1992.61][S02]approach to making sure that as we create the machinery,[1997.02] [1997.02][S02]or as we launch machinery, where you can create[2000.29] [2000.29][S02]audio outputs that are very human-like,[2002.96] [2002.96][S02]we want to make sure that we approach that with watermarking.[2005.505] [2005.505][S03] One of the other things[2007.13] [2007.13][S03]that's interesting here that I think[2008.63] [2008.63][S03]you're getting at a little bit in this line of questioning[2011.66] [2011.66][S03]is, we are personifying these people.[2016.82] [2016.82][S03]They do sound human, and we do all these things[2018.8] [2018.8][S03]to make them sound human.[2019.842] [2019.842][S03]And the interesting thing about this[2021.98] [2021.98][S03]is actually the philosophy that we've[2024.32] [2024.32][S03]had up until Audio Overviews with the product,[2028.1] [2028.1][S03]was in the text version of NotebookLM,[2030.5] [2030.5][S03]it actually does not try to sound particularly human.[2034.23] [2034.23][S03]It's very kind of factual, and it[2036.77] [2036.77][S03]doesn't try to be your friend on some level.[2038.792] [2038.792][S01] Yeah, it's quite cold almost.[2040.5] [2040.5][S03] Yeah, it's almost cold.[2042.125] [2042.125][S03]And that was kind of a bit of the idea of the house style[2046.31] [2046.31][S03]was that, but you can't do that with voice.[2049.469] [2049.469][S03]That's the thing that became very clear the second we first[2053.27] [2053.27][S03]heard these.[2053.85] [2053.85][S03]It's like, you can't say, convey this through a conversation,[2057.78] [2057.78][S03]but don't sound human.[2059.13] [2059.13][S03]Don't pretend to be a person.[2060.6] [2060.6][S03]There's no place where the human ear will tolerate that.[2064.177] [2064.177][S01] I do wonder about that, though,[2065.969] [2065.969][S01]because I mean, in that way, you are, as you say,[2068.489] [2068.489][S01]leaning in a different direction to--[2070.425] [2070.425][S01]I mean, lots of the other conversations[2072.05] [2072.05][S01]that I've had with Google DeepMind about how[2074.15] [2074.15][S01]you should try and avoid anthropomorphization.[2076.199] [2076.199][S01]You should avoid trying to think of them as \"they.\"[2078.429] [2078.429][S01]We've been describing the podcast host as \"they\"[2080.429] [2080.429][S01]the entire conversation.[2081.86] [2081.86][S01]I mean, are there dangers or concerns[2084.31] [2084.31][S01]that are associated with anthropomorphization[2086.53] [2086.53][S01]of these characters?[2087.87] [2087.87][S02] I think that by personifying them[2090.1] [2090.1][S02]to a certain extent in the way that we have, like[2092.35] [2092.35][S02]adding texture to the way that they describe things,[2095.9] [2095.9][S02]making them sound more human-like,[2097.63] [2097.63][S02]I think it's a way to make information easier to consume[2101.26] [2101.26][S02]and easier to make something more useful.[2104.51] [2104.51][S02]And I think that the reality is that we probably[2108.28] [2108.28][S02]shouldn't resist these types of approaches[2111.79] [2111.79][S02]if we believe that there is enough value associated[2114.48] [2114.48][S02]with them.[2114.98] [2114.98][S02]And I really do.[2116.27] [2116.27][S02]I really think that--[2117.7] [2117.7][S02]I've seen-- I don't know if you've seen on TikTok --all[2120.88] [2120.88][S02]of these people uploading their study materials,[2123.01] [2123.01][S02]and they're like, wow, I can study so much faster.[2126.59] [2126.59][S02]I think about the cases like that where I'm like,[2128.87] [2128.87][S02]are these people being harmed?[2130.64] [2130.64][S02]What is the actual danger?[2132.11] [2132.11][S02]And I'm not saying this to be like, well, clearly, right?[2134.93] [2134.93][S02]It's good for society.[2136.13] [2136.13][S02]But I really am thinking, what are they losing[2138.91] [2138.91][S02]as part of this experience?[2140.24] [2140.24][S02]And I think that it's less about the personification[2144.1] [2144.1][S02]or the anthropomorphization of the hosts themselves[2147.49] [2147.49][S02]and more about, OK, what did you lose by listening[2150.46] [2150.46][S02]instead of reading?[2151.58] [2151.58][S02]Maybe that's it.[2152.66] [2152.66][S03] Yeah, and that's a great point, Raiza.[2154.91] [2154.91][S03]And the other thing that I would add[2156.55] [2156.55][S03]to on that is it turns out to be a very powerful way to learn[2160.21] [2160.21][S03]and to understand is through dialogue,[2162.22] [2162.22][S03]and through asking follow-up questions,[2164.2] [2164.2][S03]and steering the focus towards the things[2167.26] [2167.26][S03]that you need to in a complex body of work.[2170.26] [2170.26][S03]But that kind of dialogue, if you[2172.9] [2172.9][S03]wanted to have a conversation about a book[2176.38] [2176.38][S03]and really engage with it, most people[2180.527] [2180.527][S03]don't have access to the author of the book.[2182.36] [2182.36][S03]Most people don't have access to an expert tutor that understands[2185.068] [2185.068][S03]the complexities of the book.[2186.64] [2186.64][S03]But now, with AI, those kinds of conversational explorations[2190.81] [2190.81][S03]are possible.[2192.16] [2192.16][S01] It's kind of a much more ancient way[2194.2] [2194.2][S01]to explore things, exactly as you describe.[2197.03] [2197.03][S01]I do wonder, though, I mean, you're talking about here,[2199.71] [2199.71][S01]people don't have access to the author,[2201.75] [2201.75][S01]but what's to stop somebody from uploading a book where,[2205.557] [2205.557][S01]actually, you really don't want them to have[2207.39] [2207.39][S01]a conversation with the author?[2208.682] [2208.682][S01]I'm thinking here like putting in \"Mein Kampf\"[2211.7] [2211.7][S01]or the \"Anarchist Cookbook.\"[2213.318] [2213.318][S03] Yeah, I mean, there's[2214.86] [2214.86][S03]a kind of underlying safety layer[2216.54] [2216.54][S03]that Google spent a lot of time working on,[2218.677] [2218.677][S03]DeepMind and spent a lot of time working on.[2220.51] [2220.51][S03]So if there are obviously offensive, dangerous things,[2223.26] [2223.26][S03]that, you can catch.[2225.96] [2225.96][S03]The trickier thing is, what happens in terms of politics?[2228.695] [2228.695][S03]So if you upload something that's[2230.07] [2230.07][S03]within the bounds of conventional political[2232.32] [2232.32][S03]discussion, but it may be more right wing or more left wing,[2234.9] [2234.9][S03]how should the host respond to that?[2236.83] [2236.83][S03]And so we specifically included instructions that say, listen,[2239.62] [2239.62][S03]if it feels political, then you should[2242.97] [2242.97][S03]adopt the attitude of hey, we're not taking sides in this.[2246.005] [2246.005][S03]We are just going to have a conversation about what[2248.13] [2248.13][S03]this document says, and we're not going to endorse it[2250.68] [2250.68][S03]or critique it in that way.[2252.01] [2252.01][S03]And we figured that was the best compromise[2254.58] [2254.58][S03]for those kind of complicated political stances.[2257.243] [2257.243][S02] I think there's also[2258.66] [2258.66][S02]the interesting sort of line, where, I think, there's[2262.85] [2262.85][S02]the safety concern and then I think[2264.81] [2264.81][S02]there's censorship concern.[2266.68] [2266.68][S02]And actually, in the early days, we ran into this a lot[2270.0] [2270.0][S02]before the safety filters were much more sophisticated, where--[2274.02] [2274.02][S02]people study difficult topics, people[2276.84] [2276.84][S02]study things that happened in history that have quite[2279.24] [2279.24][S02]a bit of violence, racism.[2282.06] [2282.06][S02]These are topics that are fraught.[2284.28] [2284.28][S02]But I think it would be wrong to create a tool that blocks[2289.5] [2289.5][S02]content generically without a thought around the intent[2293.31] [2293.31][S02]of the user so that we're not allowing users[2296.07] [2296.07][S02]to create harmful content, but at the same time, if--[2300.06] [2300.06][S02]most of our users, especially in the beginning,[2302.65] [2302.65][S02]were learners, educators.[2303.76] [2303.76][S02]Like if you're studying history, you[2306.84] [2306.84][S02]are definitely going to run into a safety filter.[2309.23] [2309.23][S03] Well, that was my problem.[2310.98] [2310.98][S03]The last book that I wrote was-- actually,[2312.87] [2312.87][S03]you mentioned \"The Anarchist Cookbook.\"[2314.847] [2314.847][S03]Part of it is about the history of anarchism[2316.68] [2316.68][S03]and the kind of roots of terrorism in the early anarchist[2319.5] [2319.5][S03]world.[2320.3] [2320.3][S03]And so I was using NotebookLM to help me research that book[2323.658] [2323.658][S03]because I was writing it.[2324.7] [2324.7][S03]And it was constantly like, I'm sorry,[2326.24] [2326.24][S03]I can't answer that question because you are obviously[2327.94] [2327.94][S03]a terrorist, Steven.[2328.78] [2328.78][S03]And I'm like, no, no.[2329.735] [2329.735][S01] You're definitely on a list[2331.36] [2331.36][S01]somewhere, Steven, aren't ya?[2332.87] [2332.87][LAUGHS][2333.37] [2333.37][S02] That's right.[2334.36] [2334.36][S03] Maybe I still have a job.[2336.068] [2336.068][LAUGHS][2336.7] [2336.7][S01] There is also this question about personal data.[2339.4] [2339.4][S01]I know that this is something that[2340.817] [2340.817][S01]has been really subject to a lot of discussion[2342.955] [2342.955][S01]with large language models and people uploading documents to it[2345.58] [2345.58][S01]and being concerned about it, kind[2346.997] [2346.997][S01]of feeding into the next generation of models.[2349.48] [2349.48][S01]So how do you make sure, in NotebookLM,[2352.24] [2352.24][S01]as you said, that the information that you upload[2355.72] [2355.72][S01]can be private and remain so?[2358.33] [2358.33][S03] Yeah, so this actually[2360.07] [2360.07][S03]is an opportunity to explain something[2361.37] [2361.37][S03]that I think is really important here,[2362.99] [2362.99][S03]which is the idea of the context window or the model.[2366.34] [2366.34][S03]So a context window is effectively[2368.68] [2368.68][S03]like the short-term memory of a language model.[2371.33] [2371.33][S03]The long term memory is like its training data,[2373.37] [2373.37][S03]like its general knowledge of the world.[2375.29] [2375.29][S03]And the context is the stuff you put in with your query[2379.09] [2379.09][S03]when you ask a question.[2380.65] [2380.65][S03]And anything in the context window is transitory.[2384.288] [2384.288][S03]The second you close your session, it disappears.[2386.33] [2386.33][S03]It gets wiped from the memory of the model.[2388.122] [2388.122][S03]What that also means is that's why it's private.[2390.94] [2390.94][S03]We're not training the model on your information.[2394.46] [2394.46][S03]All we're doing is putting it in the short-term memory[2396.71] [2396.71][S03]in the model, letting the model answer questions.[2399.17] [2399.17][S03]And then when you close the session,[2400.67] [2400.67][S03]it's like the model has completely forgotten anything[2402.61] [2402.61][S03]that you've given to it.[2403.79] [2403.79][S01] So in terms of the future of this--[2406.13] [2406.13][S01]I mean, this is still quite a young product.[2408.61] [2408.61][S01]What are the things are you hoping to include on it?[2411.165] [2411.165][S02] I think we've seen so much excitement[2413.29] [2413.29][S02]about the audio feature, so I think[2415.57] [2415.57][S02]we can definitely commit to that to being on the future roadmap.[2418.77] [2418.77][S02]I think it's alluded to more controls, more voices, more[2421.79] [2421.79][S02]personas, more languages.[2424.59] [2424.59][S02]I think that's just such an exciting horizon for us.[2426.992] [2426.992][S03] The one that I'm so[2428.45] [2428.45][S03]excited to think about, which we've just started[2431.21] [2431.21][S03]to scratch the surface of, is--[2433.04] [2433.04][S03]there's a lot of tools for asking questions and listening[2436.07] [2436.07][S03]to explanations of things, but what[2438.68] [2438.68][S03]about writing with these sources at your disposal?[2443.9] [2443.9][S03]How do we write in a source grounded environment?[2447.02] [2447.02][S03]And so just as a writer myself, I[2449.92] [2449.92][S03]think that that's going to be an amazing thing.[2452.43] [2452.43][S03]So we have some really, really cool things in the works.[2455.96] [2455.96][S01] I do also wonder about different modalities.[2458.43] [2458.43][S01]I mean, you've gone to audio, but, presumably, you[2461.84] [2461.84][S01]could go to video at some point too.[2464.12] [2464.12][S02] Yeah, and, actually, there's a fun idea[2467.33] [2467.33][S02]we have for video, which is like--[2468.96] [2468.96][S02]we're not talking about fully generative video yet,[2471.84] [2471.84][S02]but imagine if you could do even something really basic.[2474.75] [2474.75][S02]You upload these slide decks, they have charts,[2477.39] [2477.39][S02]they have diagrams, you have PDFs of papers.[2480.06] [2480.06][S02]Just take the content that's already there.[2482.16] [2482.16][S02]And NotebookLM is already incredible at this[2484.43] [2484.43][S02]because of our citations model.[2486.77] [2486.77][S02]The fact that we know exactly where every piece of the answer[2492.8] [2492.8][S02]comes from--[2494.15] [2494.15][S02]we use it to generate audio overviews,[2496.23] [2496.23][S02]we use it to generate textual answers.[2498.27] [2498.27][S02]I think it wouldn't be that big of a leap[2500.0] [2500.0][S02]to generate short videos using your own content.[2502.848] [2502.848][S01] I do really like, Steven,[2504.39] [2504.39][S01]how you're describing this often as the thing that you[2507.29] [2507.29][S01]use to make the podcast that nobody else would want to make.[2510.0] [2510.0][S01]But the point here, I guess, is that you're not[2514.4] [2514.4][S01]trying to replace all podcasts.[2516.18] [2516.18][S01]There are presumably things that you expect NotebookLM will never[2519.53] [2519.53][S01]be able to do.[2520.38] [2520.38][S03] Yeah, people, I think,[2522.44] [2522.44][S03]will generally always prefer to hear two actual humans talking[2525.86] [2525.86][S03]about a topic.[2526.62] [2526.62][S03]If there is economics or passion enough[2531.47] [2531.47][S03]to generate a podcast on a topic,[2533.13] [2533.13][S03]humans actually talking to each other will be the choice.[2536.16] [2536.16][S03]It's just turns out that there's this vast,[2539.37] [2539.37][S03]uncharted territory that just wasn't-- no one ever thought[2542.57] [2542.57][S03]about making a podcast based on the family trip to Alaska,[2547.74] [2547.74][S03][LAUGHS] because it just didn't make sense to rent a studio[2552.32] [2552.32][S03]to do that.[2553.11] [2553.11][S03]But now, you can just take everybody's journal entries[2555.62] [2555.62][S03]and photos and upload it to NotebookLM,[2557.54] [2557.54][S03]and you can have a podcast based on your family trip.[2559.86] [2559.86][S03]And so I think that's where it turns out[2562.35] [2562.35][S03]there's just all this untapped kind of blank space on the map[2566.84] [2566.84][S03]that we've just started to explore.[2570.062] [2570.062][S01] Do you think that there[2571.52] [2571.52][S01]are elements of, like, human-content creation[2574.88] [2574.88][S01]that are really hard to capture with AI,[2578.45] [2578.45][S01]or the AI will maybe never be able to capture?[2581.5] [2581.5][S03] Yeah, that's the thing[2583.37] [2583.37][S03]we're trying to figure out.[2585.38] [2585.38][S03]I mean, the one idea I think that I'm really interested in[2590.86] [2590.86][S03]is like, how capable are these models[2594.52] [2594.52][S03]at thinking and developing ideas that are really long form?[2599.8] [2599.8][S03]So book writing-- so when you're coming up with the idea[2603.31] [2603.31][S03]for a book, you're really thinking it's-- one,[2606.22] [2606.22][S03]it's an incredibly long-term process and you're thinking[2608.65] [2608.65][S03]about a presentation of information that's going to go[2611.92] [2611.92][S03]on for 300 pages.[2613.218] [2613.218][S03]It's going to involve all this complexity, all this narrative[2615.76] [2615.76][S03]complexity.[2616.73] [2616.73][S03]And you couldn't approach that all[2619.237] [2619.237][S03]with a language model right now.[2620.57] [2620.57][S03]You could work on little bits of it.[2622.24] [2622.24][S03]You could say, OK, I'm trying to set up this scene[2624.04] [2624.04][S03]or I'm trying to figure out what the narrative should be,[2626.415] [2626.415][S03]but you can't actually imagine the whole thing.[2630.49] [2630.49][S03]That, right now, is just a human-exclusive capability.[2634.22] [2634.22][S03]And I think it will be for a long time, and it may always be.[2637.94] [2637.94][S03]But who knows where we're going to end up?[2639.88] [2639.88][S01] Both the wood and the trees simultaneously.[2642.26] [2642.26][S03] Yeah.[2643.135] [2643.135][S03]Yeah, and I think they're the kind of seeds of that.[2647.63] [2647.63][S03]There's some promising signals, but people[2651.087] [2651.087][S03]who write books for a living, I think,[2652.67] [2652.67][S03]can feel confident that they will[2654.078] [2654.078][S03]continue to be able to do that.[2655.37] [2655.37][S01] Yeah, although writing books for a living[2657.578] [2657.578][S01]is one of the most torturous professions there is.[2659.91] [2659.91][LAUGHS][2660.81] [2660.81][S01]As someone who's trying to write one at the moment,[2662.935] [2662.935][S01]I want you guys to hurry up, please.[2664.97] [2664.97][S01]Well, thank you both for joining me.[2666.75] [2666.75][S01]That was a really, really fascinating discussion.[2669.06] [2669.06][S01]Appreciate it.[2670.07] [2670.07][S03] Thanks for having us.[2671.16] [2671.16][S02] Thank you.[2672.16] [2672.16][S02]Thanks for having us.[2673.56] [2673.56][S01] You know, I think there's actually[2675.477] [2675.477][S01]something quite heartwarming about the way[2677.64] [2677.64][S01]that NotebookLM has captured people's imagination,[2680.137] [2680.137][S01]because on the one hand, you've got this technology that[2682.47] [2682.47][S01]is operating at the absolute cutting edge of what[2685.5] [2685.5][S01]is possible with some of the most sophisticated AI models[2688.92] [2688.92][S01]out there.[2689.77] [2689.77][S01]And it's something that's designed[2691.71] [2691.71][S01]to deal with this very modern problem about how we are often[2695.76] [2695.76][S01]overwhelmed with having to process[2697.53] [2697.53][S01]these large amounts of, often, quite dense and maybe[2701.1] [2701.1][S01]quite-boring information.[2702.87] [2702.87][S01]And they've hit upon a solution that is so innately human, so[2709.08] [2709.08][S01]ancient and appealing, the idea of listening[2712.83] [2712.83][S01]in to a conversation between two excitable and interested people.[2717.61] [2717.61][S01]And, of course, the fastest way to make[2720.09] [2720.09][S01]a human prick up their ears and pay attention is through gossip.[2724.47] [2724.47][S01]And this is like sitting around a fire[2727.29] [2727.29][S01]while an AI uses that very trick to help[2730.95] [2730.95][S01]you digest 25 pages of a snorefest-lecture series.[2735.58] [2735.58][S01]I mean, put it this way.[2736.78] [2736.78][S01]If it can make my PhD thesis sound interesting,[2739.34] [2739.34][S01]then this has the potential to be quite a powerful tool.[2742.33] [2742.33][S01]You have been listening to \"Google DeepMind-- the Podcast,\"[2745.12] [2745.12][S01]with me, Professor Hannah Fry.[2746.68] [2746.68][S01]If you enjoyed that episode, then[2748.3] [2748.3][S01]do subscribe to our YouTube channel.[2750.64] [2750.64][S01]And you can also find us on your favorite podcast platform.[2753.95] [2753.95][S01]And of course, we have got plenty more episodes[2756.49] [2756.49][S01]on a whole range of topics to come, so do check those out too.[2760.94] [2760.94][S01]See you next time.[2762.25] [2762.25][MUSIC PLAYING][2765.3]"}
{"file_name": "audio/val_000018.wav", "transcription": "[0.0][S01] There was a really nice quote that I came across,[2.02] [2.02][S01]which is that almost no prerequisite[3.72] [3.72][S01]to any major invention was made with that invention in mind.[7.14] [7.14][S01]Can you imagine a point in the future[9.76] [9.76][S01]where you are letting agents loose in these environments[13.04] [13.04][S01]without specifying an objective for them?[15.52] [15.52][S03] I think we as humans,[17.42] [17.42][S03]we decide what's interesting.[19.14] [19.14][S03]I think there is even an example where[20.96] [20.96][S03]the entire evolution of mathematics[22.96] [22.96][S03]was guided by people deciding what's next, what's interesting,[26.78] [26.78][S03]what's not interesting, and just the problem being hard[30.098] [30.098][S03]doesn't mean that it's interesting at all, right?[32.14] [32.14][S02] I believe very strongly[33.2] [33.2][S02]that we need simulation, and I also[35.182] [35.182][S02]believe very strongly that we won't[36.64] [36.64][S02]be able to build a simulator of the real world any other way.[40.038] [40.038][S02]So when you combine those two things,[41.58] [41.58][S02]I think, yes, it is a big step for my version of AGI.[45.327] [45.327][MUSIC PLAYING][47.662] [51.163][S01] Welcome back to \"Google DeepMind--[53.08] [53.08][S01]The Podcast.\"[53.86] [53.86][S01]I'm Professor Hannah Fry.[55.72] [55.72][S01]Now the latest video generation models[58.2] [58.2][S01]have impressed the entire world.[59.8] [59.8][S01]They've created this near perfect imitation of reality.[64.0] [64.0][S01]But the limitations of video is that you are just[66.72] [66.72][S01]a viewer rather than a participant.[69.08] [69.08][S01]And that's not how humans experience the real world.[72.24] [72.24][S01]We instead can navigate environments we've never been to[75.36] [75.36][S01]and still have an expectation of what we're likely to encounter.[78.96] [78.96][S01]We can explore in every feasible direction, kind[82.16] [82.16][S01]of without limits, and interact with things that we chance upon[86.12] [86.12][S01]along the way.[87.52] [87.52][S01]And that is the next great frontier for this technology,[91.32] [91.32][S01]to move beyond generating a perfect recording of a scene[95.08] [95.08][S01]and towards building a dynamic simulation of a world[98.4] [98.4][S01]we can finally step into.[100.96] [100.96][S01]Enter Genie 3, a prototype world model[104.12] [104.12][S01]that can generate an unprecedented variety[107.32] [107.32][S01]of interactive environments.[109.4] [109.4][S01]It's already been described as a stepping stone towards AGI.[113.48] [113.48][S01]And with me today are two of its creators,[116.52] [116.52][S01]Shlomi Fruchter, research director,[118.46] [118.46][S01]and Jack Parker-Holder, research scientist.[120.69] [120.69][S01]Welcome to the podcast, both of you.[122.19] [122.19][S01]Can you sum this up in a sentence for me?[123.898] [123.898][S01]What is Genie 3?[124.98] [124.98][S02] It's a real time, interactive world model[127.73] [127.73][S02]that allows you to create diverse, visually interesting[130.97] [130.97][S02]worlds from a text prompt.[132.71] [132.71][S02]So there is no underlying game engine, no structure, no code.[137.27] [137.27][S02]It's just a neural network that's[139.09] [139.09][S02]predicting every single pixel in reaction[141.41] [141.41][S02]to inputs from the user and also the past.[144.45] [144.45][S02]And so the flexibility in the diversity[146.49] [146.49][S02]of things you can create in basically no time[149.21] [149.21][S02]is quite unprecedented.[150.91] [150.91][S01] You haven't had a whole army of artists sitting[153.45] [153.45][S01]in rooms constructing a world in order to be able to interact.[156.033] [156.033][S03] Yes, I think the point is that you can create[158.617] [158.617][S03]any world that you can imagine.[160.19] [160.19][S03]And that's not something that you can do with a game engine.[162.825] [162.825][S01] OK.[163.45] [163.45][S01]Let's have a look at it, because you've[164.45] [164.45][S01]got some demos for me, right?[165.89] [165.89][S03] Yeah.[166.807] [166.807][S03]So we have a few.[168.05] [168.05][S03]The first one, I think you might like it.[170.63] [170.63][S03]So it's basically playing a cat.[172.75] [172.75][S01] OK.[173.973] [173.973][S01]You've got me already.[174.89] [174.89][S03] A ginger cat.[176.73] [176.73][S01] Excellent.[177.792] [177.792][S03] Let me run it now.[179.25] [179.25][S03]We basically prompted it with being a cat.[181.71] [181.71][S01] So this is reacting to the inputs[184.21] [184.21][S01]that you're giving it.[185.35] [185.35][S03] Yes, exactly.[186.61] [186.61][S01] Is the light going to change as you go into the--[189.152] [189.152][S01]oh, look at that.[190.81] [190.81][S01]Look at that.[191.61] [191.61][S03] Yes.[192.485] [192.485][S03]So the model is basically trying to predict[194.77] [194.77][S03]what's going to happen next based on the sequence of inputs[198.29] [198.29][S03]that it gets, and it does it in real time.[200.155] [200.155][S01] So looking at this superficially,[202.03] [202.03][S01]I mean, this doesn't look that different from if I just[205.09] [205.09][S01]loaded up a video game, right?[206.99] [206.99][S01]How is it different?[208.15] [208.15][S03] Oh, I think the key difference is[209.45] [209.45][S03]that this is generated from text,[211.09] [211.09][S03]so every single pixel is generated from a model which[214.65] [214.65][S03]is just predicting the pixels.[216.467] [216.467][S01] I mean, there's also the detail that you're seeing[219.05] [219.05][S01]in this 3D environment.[220.51] [220.51][S01]It's also quite reminiscent of some of the stuff[222.77] [222.77][S01]that we're seeing with Veo if I didn't know that you[225.81] [225.81][S01]were interacting with it.[227.27] [227.27][S01]How is it different from that?[228.93] [228.93][S03] So when you create a video using Veo,[231.66] [231.66][S03]so you provide a prompt and then the model[233.41] [233.41][S03]is trying to figure out how to create[234.952] [234.952][S03]this entire video of, say, eight seconds from start to finish.[241.1] [241.1][S03]And once it's ready, then you cannot change how the camera[245.78] [245.78][S03]moves around, and definitely you cannot explore it much more than[250.54] [250.54][S03]just eight seconds.[251.332] [251.332][S01] Can you use an image to prompt this or is[253.54] [253.54][S01]it only text?[254.16] [254.16][S03] Yes, so we just found out[255.91] [255.91][S03]that we can actually use an image and videos to prompt[259.62] [259.62][S03]them all.[260.899] [260.899][S03]In this particular case, we found that we can actually[263.3] [263.3][S03]use paintings.[264.56] [264.56][S03]For example, this is \"Nighthawks\" by Edward Hopper.[266.927] [266.927][S01] Very famous picture.[268.26] [268.26][S03] Very famous picture, and from 1942.[271.56] [271.56][S03]And yeah, basically we asked Genie 3[274.9] [274.9][S03]to let us walk into the painting.[278.58] [278.58][S01] So this painting is of a very vivid image, a street[283.3] [283.3][S01]corner at nighttime.[285.16] [285.16][S01]You're looking in through the glass to see a man and a woman[289.06] [289.06][S01]leaning up against a bar, and then[291.5] [291.5][S01]someone serving drinks the other side of the counter.[293.92] [293.92][S01]It's got these rich greens.[295.92] [295.92][S01]The pavement underneath.[297.34] [297.34][S01]The way the light falls is really, really evocative.[300.58] [300.58][S01]But now that you're navigating this space and turning around,[304.46] [304.46][S01]I mean, this is really extraordinary.[307.64] [307.64][S01]It's like you've imagined what would be on either side.[309.973] [309.973][S01]Can you go around the back of the man and the woman?[312.14] [312.14][S01]I want to see the back of their heads.[313.78] [313.78][S01]And you can.[314.92] [314.92][S01]So it's like, got all dimensions of the original painting.[317.745] [317.745][S01]Let's turn around and see the rest of the street as well.[320.12] [320.12][S01]And in the distance, you can see these street lights just pinging[324.78] [324.78][S01]off into the distance as the road extends.[326.86] [326.86][S01]Some billboards above other shops and establishments[331.94] [331.94][S01]off into the distance.[333.26] [333.26][S01]And now we're looking back.[335.02] [335.02][S01]We've chosen to look back at the image of the original painting[338.1] [338.1][S01]to see it there.[339.26] [339.26][S01]That is really amazing.[340.92] [340.92][S01]It's all consistent, all completely consistent.[344.0] [344.0][S01]Have you got another one for us?[345.58] [345.58][S03] Yeah.[346.497] [346.497][S03]We have one that we're just kind of controlling the jet[349.5] [349.5][S03]ski around a few islands.[353.08] [353.08][S03]So let's see how it goes.[354.623] [354.623][S01] Tell me the original prompt.[356.29] [356.29][S03] So it's sailing a jet ski[358.04] [358.04][S03]through the waters around the islands of Kauai.[360.43] [360.43][S01] Ooh, sounds dreamy.[362.05] [362.05][S03] The waters have different ramps[364.35] [364.35][S03]that we can go up on.[365.85] [365.85][S03]Yeah.[366.35] [366.35][S01] OK.[366.81] [366.81][S03] Let's see.[367.49] [367.49][S01] All right, here we go.[368.49] [368.49][S01]OK.[368.99] [368.99][S01]So we are the POV of the person on the jet ski[373.33] [373.33][S01]We can see the hands in frame.[375.57] [375.57][S01]Both hands are consistent with each other, I should add.[378.43] [378.43][S01]The water is beautifully still.[380.41] [380.41][S01]You can see these islands in the background.[382.53] [382.53][S01]The sun is quite low in the sky and you[384.59] [384.59][S01]can see the reflection of its rays on the water.[388.31] [388.31][S01]Now, I see you're taking us up a ramp here.[390.21] [390.21][S03] Trying.[390.65] [390.65][S03]I'm a bit too slow.[391.61] [391.61][S03]I don't know.[392.425] [392.425][S01] What is it going to do on the way down?[394.55] [394.55][S03] Ah.[395.15] [395.15][S01] I mean--[396.11] [396.11][S01]oh, and it splashes when it hits the water.[398.41] [398.41][S03] Let's see what's-- let's look back.[399.77] [399.77][S01] And when you look round, when you look around[401.71] [401.71][S01]to the back, there's a trail in the water[403.85] [403.85][S01]exactly as you would expect from a real jet ski.[406.27] [406.27][S01]So are you seeing elements of where[408.03] [408.03][S01]it's understanding physics?[409.19] [409.19][S02] Definitely, yes.[410.69] [410.69][S02]There's some things we, I guess, refer to as emergent properties,[415.67] [415.67][S02]just from having general training[418.19] [418.19][S02]and seeing lots of different things.[420.51] [420.51][S02]When it sees a new scenario, it understands how smoke moves[423.51] [423.51][S02]or how water should flow.[425.15] [425.15][S02]But that's not maybe 100% accurate[427.59] [427.59][S02]in every single setting.[428.85] [428.85][S02]But it's got enough accuracy that you[430.51] [430.51][S02]do feel some sense of being in the scene,[432.61] [432.61][S02]and as humans, we can't obviously[434.27] [434.27][S02]spot these things that are wrong with it.[435.978] [435.978][S03] So I think, as Jack said,[437.728] [437.728][S03]there are definite limitations.[439.17] [439.17][S03]But on the other hand, after I've[441.23] [441.23][S03]been working on game engines in the past,[443.29] [443.29][S03]and I think we worked really hard to make[445.11] [445.11][S03]all of these kind of effects independently,[448.77] [448.77][S03]like the lens flare and the water simulation, everything.[451.213] [451.213][S03]And here we have basically a model[452.63] [452.63][S03]that can do all of that out of the box with some limitations,[455.81] [455.81][S03]of course.[456.31] [456.31][S01] Without even really trying.[457.65] [457.65][S02] And there's other things[459.483] [459.483][S02]that it can do that would be almost impossible to get[461.71] [461.71][S02]through other methods, right?[462.918] [462.918][S02]Like simulating other animals and people in the world.[467.25] [467.25][S02]I think that's something that's really exciting more[469.47] [469.47][S02]in the future as well, is to be able to interact[471.72] [471.72][S02]with other agents in the world.[473.452] [473.452][S01] Yeah.[474.16] [474.16][S01]I mean, this totally interactive environment[476.262] [476.262][S01]that ends up being consistent regardless of which direction[478.72] [478.72][S01]you push it in.[479.38] [479.38][S01]I mean, that's really extraordinary.[480.88] [480.88][S01]I mean, these demos demonstrate the proof of concept, I guess.[484.317] [484.317][S01]But how would you see this being used?[485.9] [485.9][S01]What are the kind of applications[486.8] [486.8][S01]that you'd be looking at?[487.78] [487.78][S03] One thing that we[488.64] [488.64][S03]are very excited about is using it[490.52] [490.52][S03]for actually the simulation environment for agents.[494.14] [494.14][S03]So, for example, you can imagine an agent that[497.6] [497.6][S03]wants to accomplish a goal.[499.84] [499.84][S03]And then we can put them in any environment that we can imagine,[504.02] [504.02][S03]maybe one that is more challenging for it.[507.513] [507.513][S03]And then it can explore the environment,[509.18] [509.18][S03]maybe try and accomplish a goal and learn from its mistakes.[511.7] [511.7][S03]Again, not without doing anything[513.08] [513.08][S03]in the real world, which is very expensive.[515.02] [515.02][S03]Another thing that we're excited about[516.76] [516.76][S03]is actually using those simulations for planning.[519.542] [519.542][S03]So if you have a robot or you have,[521.0] [521.0][S03]again, an agent that wants to accomplish a goal,[523.72] [523.72][S03]they can maybe do some rollouts in this simulation,[526.62] [526.62][S03]figure out what might happen.[528.1] [528.1][S03]For example, if they want to go across the road,[531.8] [531.8][S03]the agent can use the model to predict a few options.[535.3] [535.3][S03]Maybe there is a few scenarios.[537.7] [537.7][S03]Maybe a person going to cross its path,[539.718] [539.718][S03]maybe something else going to happen.[541.26] [541.26][S03]And then using those rollouts, it[543.28] [543.28][S03]can decide what's the next action it should take.[545.94] [545.94][S03]And then this is used for planning.[547.96] [547.96][S03]And beyond that, we just see a lot[549.56] [549.56][S03]of applications for education, for entertainment.[552.13] [552.13][S01] Just anchor this in a few examples for me.[554.38] [554.38][S01]So I mean, what's the idea here?[556.58] [556.58][S01]Are you going to, in a history lesson,[558.82] [558.82][S01]be able to create a world of Victorian England, for instance?[562.233] [562.233][S02] Exactly.[563.4] [563.4][S02]Yeah.[563.9] [563.9][S02]So imagine you are in front of a bunch of students.[567.235] [567.235][S02]They're obviously excited to learn about Victorian England,[569.9] [569.9][S02]but they've also got a lot of other distractions, things[571.88] [571.88][S02]that they're interested in as well.[573.36] [573.36][S02]Instead of just reading a textbook,[575.18] [575.18][S02]you can instead allow them to step into the world[577.38] [577.38][S02]and you can take them on a virtual tour,[579.24] [579.24][S02]in a sense, of what it would have been like to be there.[581.573] [581.573][S02]So for places that maybe are harder to access, maybe[584.96] [584.96][S02]for distant corners of the planet[586.52] [586.52][S02]or perspectives that you couldn't otherwise[588.65] [588.65][S02]get-- so being a jaguar, for example,[592.17] [592.17][S02]or other kind of animals, or being a shark,[595.036] [595.036][S02]or going back in the past.[596.382] [596.382][S02]These are things that you couldn't really[598.09] [598.09][S02]get any other way as an experience,[599.622] [599.622][S02]and it might make more visual learners in particular,[601.83] [601.83][S02]I think, resonate more with them.[603.323] [603.323][S01] And that's when you have the human who's[605.49] [605.49][S01]playing with the controls.[606.573] [606.573][S01]But as you say, if you then let an agent loose in this,[609.85] [609.85][S01]that opens up a whole other level of possibilities.[612.51] [612.51][S03] Yeah.[613.427] [613.427][S03]So an agent basically--[614.73] [614.73][S03]again, once you have a real or a very close[617.21] [617.21][S03]to real simulation of an environment,[619.35] [619.35][S03]then the agent can use it instead[621.53] [621.53][S03]of actually learning in the real world, which is very costly.[626.01] [626.01][S03]If an agent makes a mistake or if a robot makes[628.21] [628.21][S03]a mistake in the real world, it's much harder to fix.[634.29] [634.29][S03]Basically, this is a way for agents[636.49] [636.49][S03]to learn in a simulated environment[637.97] [637.97][S03]where we can control everything.[639.43] [639.43][S03]We can set up some environment that is maybe[641.37] [641.37][S03]more challenging or less predictable than what typically[645.49] [645.49][S03]the agent was trained to do.[647.49] [647.49][S03]And this way basically the agent can improve[650.49] [650.49][S03]in this safe simulation.[653.09] [653.09][S03]So we're very excited about that.[654.47] [654.47][S01] So let's say that you were maybe running a factory[657.053] [657.053][S01]and you wanted to install a robot with a particular task.[660.23] [660.23][S01]You can recreate the precise environment[662.05] [662.05][S01]it will find itself in and allow it to,[664.21] [664.21][S01]well, find its own mistakes.[666.27] [666.27][S02] Yeah.[666.37] [666.37][S02]I mean, it's a great example because this is already[668.537] [668.537][S02]something that we're close to being able to do already.[671.25] [671.25][S02]Robots are getting quite capable.[672.95] [672.95][S02]But I think what's even more exciting is things[674.912] [674.912][S02]that it's quite far away from being able to do,[676.87] [676.87][S02]that this could completely enable and unlock.[679.25] [679.25][S02]So having robots and embodied agents[682.13] [682.13][S02]actually in the real world.[683.93] [683.93][S02]The diversity of possible scenarios[686.17] [686.17][S02]is just quite hard to fathom, I think, for our current systems.[689.998] [689.998][S02]I think of this example of what would an embodied agent[692.29] [692.29][S02]do on Halloween, right?[693.723] [693.723][S02]Maybe one day a year it sees children running around[695.89] [695.89][S02]in costumes.[696.835] [696.835][S02]What would it do the first time it sees this?[698.71] [698.71][S02]It's quite a challenging scenario to prepare for.[701.453] [701.453][S02]And even if you've seen it before,[702.87] [702.87][S02]it might be different in the next year.[704.81] [704.81][S02]And so to really be able to simulate these rare events[707.26] [707.26][S02]and be able with text to describe any imaginable world,[710.812] [710.812][S02]to become robust to it, make sure the robots are safe[713.02] [713.02][S02]or the agents are safe.[714.78] [714.78][S02]They understand all these different things,[716.572] [716.572][S02]but they can also learn from their experience[718.447] [718.447][S02]as well, which we know is really important.[720.28] [720.28][S01] But then I also wonder,[721.16] [721.16][S01]I mean, all the examples that we've given so far[723.16] [723.16][S01]has been that experiencing the real world at the human level,[725.92] [725.92][S01]as it were, in terms of our size.[727.96] [727.96][S01]Could you shrink this down and create a simulated world[732.22] [732.22][S01]that was at the level of molecules or the human cell,[734.88] [734.88][S01]for instance?[735.44] [735.44][S03] Yeah, so we tried that[736.46] [736.46][S03]and we actually have a few examples we're moving around[739.1] [739.1][S03]like a blood vessel.[740.4] [740.4][S03]It's not necessarily always biologically accurate,[744.26] [744.26][S03]but this is not-- we don't think it's a fundamental limitation.[747.28] [747.28][S03]We could probably, if we had a more accurate simulation[752.82] [752.82][S03]that models can be trained on, then we see that in the future,[755.48] [755.48][S03]maybe we have other variants of models[757.22] [757.22][S03]that are able to specialize in this particular environment.[760.48] [760.48][S03]But we did try to focus more on the real world, the world[764.98] [764.98][S03]from the eyes of a person, because we[766.9] [766.9][S03]think that's the most widely applicable in terms[771.14] [771.14][S03]of the generality of the model.[773.38] [773.38][S01] Is this about just utilizing the existing[776.333] [776.333][S01]developments in artificial intelligence,[778.0] [778.0][S01]or is this also about stepping us[779.94] [779.94][S01]closer towards the goal of AGI?[781.232] [781.232][S02] We think this is definitely[783.19] [783.19][S02]a kind of a new kind of foundation model.[785.02] [785.02][S02]And that's why the breadth of applications[786.9] [786.9][S02]is so wide but also so nascent, because we've never really[792.42] [792.42][S02]had this kind of model before.[793.96] [793.96][S02]It blends and combines ideas from what[796.82] [796.82][S02]we've seen in language models and also video models,[799.46] [799.46][S02]with some of the techniques that we use.[801.68] [801.68][S02]And so it's kind of combining these different elements[804.02] [804.02][S02]in something quite new, which I think[806.58] [806.58][S02]is what's so exciting about it.[808.36] [808.36][S02]And this breakthrough that we've made[810.34] [810.34][S02]is that it might enable some completely new applications[812.86] [812.86][S02]that we didn't really have before.[814.58] [814.58][S02]And so it's still quite an early stage for this research,[817.28] [817.28][S02]but we're quite excited to see what the next few months bring.[820.16] [820.16][S01] Well, let me dig into that a bit deeper,[822.327] [822.327][S01]then, because I know that your background is much more in Veo[824.91] [824.91][S01]and the video side of things, but you were working on Genie 1[828.67] [828.67][S01]and Genie 2 before you and your team came on board.[833.27] [833.27][S01]What were you doing then?[834.41] [834.41][S01]What was the inspiration?[835.452] [835.452][S01]And how is it different from this iteration?[837.55] [837.55][S02] So before Genie,[839.05] [839.05][S02]I was working on open-ended learning.[841.77] [841.77][S02]So training agents in large simulated environments[844.59] [844.59][S02]where we could configure different components[846.75] [846.75][S02]of the world.[847.41] [847.41][S02]So we worked on the xLAM project,[849.15] [849.15][S02]and the idea there was to basically,[851.03] [851.03][S02]with procedural generation, generate[852.63] [852.63][S02]wide variety of environments that were still specified[855.35] [855.35][S02]in code and have the agent learn from these different experiences[858.47] [858.47][S02]and to become a generalist agent in simulation.[861.15] [861.15][S02]But ultimately, we were bottlenecked by the availability[863.51] [863.51][S02]of environments.[865.23] [865.23][S02]We were also, in my PhD, working a bit[867.11] [867.11][S02]on world models as well, but with[868.91] [868.91][S02]much more limited, constrained setting.[871.05] [871.05][S02]We were training world models from single environments,[873.59] [873.59][S02]typically quite low dimensional ones.[876.35] [876.35][S02]And the dream was really to combine these ideas,[878.548] [878.548][S02]to learn general world models that could be used,[880.59] [880.59][S02]simulators for any imaginable tasks,[882.61] [882.61][S02]and then train agents in them to solve completely new things[885.848] [885.848][S02]and have this kind of open-ended loop[887.39] [887.39][S02]where we generate new worlds and the agents[889.182] [889.182][S02]would learn from those.[890.55] [890.55][S02]We started with Genie 1 as a proof of concept, like,[893.15] [893.15][S02]could we do this at all?[894.17] [894.17][S02]Could we generate new worlds that were interactive at all?[897.13] [897.13][S02]And that was quite a major breakthrough.[899.11] [899.11][S02]And then with Genie 2, we scaled this to any 3D sort[901.79] [901.79][S02]of environment.[902.95] [902.95][S01] There were some things that emerged from--[906.03] [906.03][S01]emergent properties in that work.[907.435] [907.435][S01]Tell me a little bit about those.[908.81] [908.81][S01]Were they expected?[909.89] [909.89][S02] So for Genie 2, the question for us was,[912.63] [912.63][S02]can this idea scale?[913.75] [913.75][S02]Right?[914.25] [914.25][S02]Because Genie 1, it was quite a simple proof of concept.[918.422] [918.422][S02]It was, does this work at all?[920.69] [920.69][S02]Whereas Genie 2 was like, is this something[922.71] [922.71][S02]that could really scale to look more like what we see[925.95] [925.95][S02]in foundation models nowadays?[927.337] [927.337][S02]And we weren't sure if it would really work.[929.17] [929.17][S02]We weren't sure if it would be consistent for very long,[931.503] [931.503][S02]because Genie 1 only lasted a couple of seconds.[933.542] [933.542][S02]We weren't sure if it would work at higher resolution[935.75] [935.75][S02]because Genie 1 was 90p, which is very small images.[939.21] [939.21][S02]Genie 2 was 360p.[940.84] [940.84][S02]And then the type of diversity of environments[942.84] [942.84][S02]was a significant increase.[944.572] [944.572][S02]And so given all of that, we weren't sure[946.28] [946.28][S02]if it would be possible to have a single neural network that[948.78] [948.78][S02]could simulate anything within that domain.[952.16] [952.16][S02]And so when we actually got the model,[954.9] [954.9][S02]it was definitely an emergent property[956.68] [956.68][S02]for completely new worlds, and we used Imagen 3[959.4] [959.4][S02]to generate the starting frames back then.[962.28] [962.28][S02]It could do things simulate smoke[964.04] [964.04][S02]or, when you drove off the side of a cliff,[966.46] [966.46][S02]the car would have gravity, or if you landed in a puddle,[969.5] [969.5][S02]it splashed.[971.0] [971.0][S02]And it was quite surprising that this worked so well.[973.26] [973.26][S02]And that gave us the confidence that this next step with Genie 3[976.52] [976.52][S02]would be possible.[977.66] [977.66][S01] So Genie 1, 2D.[979.1] [979.1][S02] Yep.[980.1] [980.1][S01] Genie 2, 3D.[981.6] [981.6][S02] Mhm.[982.6] [982.6][S01] So with Genie 1, then, you're like,[984.4] [984.4][S01]right, we want to be able to create any sort of environment,[986.9] [986.9][S01]and you feed in--[987.892] [987.892][S01]I mean, it was platform games, wasn't it?[989.6] [989.6][S02] Yep.[990.08] [990.08][S01] Just tons and tons and tons and tons[992.08] [992.08][S01]of footage of platform games.[993.5] [993.5][S02] Exactly, yeah.[994.22] [994.22][S01] And then were there any emergent properties in that?[996.92] [996.92][S02] Well, the fact[997.52] [997.52][S02]that you could generate completely new ones, I think[999.92] [999.92][S02]was surprising.[1000.74] [1000.74][S02]I don't think that had really been shown before.[1002.86] [1002.86][S03] Even a painting, right?[1003.78] [1003.78][S02] Yeah.[1004.54] [1004.54][S02]Actually--[1004.8] [1004.8][S03] A drawing.[1005.44] [1005.44][S02] Yeah, I mean, you[1006.982] [1006.982][S02]know the work better than me.[1008.24] [1008.24][S02]We even had a picture of my dog in the park,[1010.6] [1010.6][S02]and you could move her left and right like a platform game.[1013.98] [1013.98][S02]But of course, she's not one.[1016.92] [1016.92][S02]We also had Jeff Clune was an advisor of the project.[1020.1] [1020.1][S02]His children did a bunch of drawings,[1022.04] [1022.04][S02]and we were able to animate those and move them[1024.44] [1024.44][S02]around like games.[1025.692] [1025.692][S02]And I think it's safe to say that was not[1027.4] [1027.4][S02]in the training data.[1028.3] [1028.3][S02]So I'm happy to call that an emergent property.[1032.44] [1032.44][S02]It looked quite different to what it was trained on.[1034.758] [1034.758][S02]And thanks, Shlomi, for reminding me,[1036.3] [1036.3][S02]because it's been a couple of years.[1037.938] [1037.938][S01] And then the step between Genie 1 and Genie 2 was[1040.48] [1040.48][S01]making it 3D.[1041.619] [1041.619][S02] Yeah, so it was increasing the diversity[1043.24] [1043.24][S02]of the things it could do.[1044.599] [1044.599][S02]So compared to Genie 1, which was just 2D platform games,[1048.98] [1048.98][S02]it was 2D games, but as well as 3D environments as well[1051.92] [1051.92][S02]in the same model.[1053.16] [1053.16][S02]It was also higher resolution, and there[1054.868] [1054.868][S02]was a lot more consistency so it would last a bit longer when[1057.41] [1057.41][S02]you interacted with it.[1058.75] [1058.75][S02]So it was really a step up in capability[1060.45] [1060.45][S02]in a few different dimensions, which[1062.33] [1062.33][S02]required a much more concentrated effort[1065.01] [1065.01][S02]to get that to be possible.[1066.77] [1066.77][S02]And we weren't sure if it would actually work at all.[1069.05] [1069.05][S02]So it was more of a proof of concept[1070.97] [1070.97][S02]at a slightly larger scale.[1072.69] [1072.69][S02]And that gave us confidence that what we have now[1075.01] [1075.01][S02]would be achievable.[1076.21] [1076.21][S01] So you were working on building an environment[1078.627] [1078.627][S01]that you could interact with, and meanwhile, in parallel, you[1081.57] [1081.57][S01]were working on video generation.[1083.23] [1083.23][S03] Yes.[1084.105] [1084.105][S03]So my background is actually in 3D game engines.[1086.91] [1086.91][S03]And a very long time ago, I used to work on simulations.[1090.49] [1090.49][S03]That was kind of where I started working on AI.[1094.297] [1094.297][S03]Back then we didn't use even AI.[1095.63] [1095.63][S03]We called it ML.[1097.49] [1097.49][S03]But in the last few years, I really[1100.17] [1100.17][S03]got more excited as the technology[1103.17] [1103.17][S03]evolved and then worked on image models, video models.[1106.81] [1106.81][S03]And yeah, in the last two years, video models, as we all know,[1110.13] [1110.13][S03]got to new levels of realism.[1112.51] [1112.51][S03]And I remember looking at one of the--[1114.33] [1114.33][S03]I think it was the Imagen video models and just saying,[1117.81] [1117.81][S03]how is it possible that there is a full simulation of the world[1121.09] [1121.09][S03]just in this model?[1122.15] [1122.15][S03]It's just mind blowing, if you think about the level of realism[1125.57] [1125.57][S03]that those models even initially achieved compared to simulation[1130.05] [1130.05][S03]using 3D graphics methods.[1133.09] [1133.09][S03]And then with Veo, what we've tried to do[1136.17] [1136.17][S03]is basically to build the best possible video model.[1140.41] [1140.41][S03]And when I saw the results, I started thinking, OK,[1143.95] [1143.95][S03]what happens if-- can we do that in real time?[1146.85] [1146.85][S03]And I was obviously following the work by Jack and team,[1150.93] [1150.93][S03]and I think that was also very inspiring.[1153.27] [1153.27][S03]And we just said, OK, we have to go to the next level.[1156.83] [1156.83][S01] So what were the elements from Veo[1158.85] [1158.85][S01]that you wanted to combine or learn from for this project?[1163.358] [1163.358][S01]Because it's not just the visual aesthetic.[1165.15] [1165.15][S03] The quality and the realism[1166.983] [1166.983][S03]is very something that we really invested in.[1169.45] [1169.45][S03]And I think there is a level of realism[1171.25] [1171.25][S03]where I think we kind of got to it with Veo 2 that--[1176.9] [1176.9][S03]the physics are not perfect, but it's good enough[1179.82] [1179.82][S03]to start to be useful so we can actually[1181.98] [1181.98][S03]create some scenes that are indistinguishable[1184.54] [1184.54][S03]from real footage.[1187.12] [1187.12][S03]Not everything.[1188.0] [1188.0][S03]Not all the time.[1188.72] [1188.72][S03]But it's starting to be there, right?[1190.48] [1190.48][S03]With Veo 3, we also added audio and other stuff.[1193.52] [1193.52][S03]So taking it to the next step of basically making[1196.82] [1196.82][S03]this interactive was the obvious next step.[1199.84] [1199.84][S03]But technically it's quite challenging, especially[1203.78] [1203.78][S03]terms of how fast we should create the next frames.[1210.46] [1210.46][S03]That's what was one of the core challenges[1212.38] [1212.38][S03]for the project for Genie 3.[1215.06] [1215.06][S03]And I think the approach that we have overall[1217.9] [1217.9][S03]is to try and understand how those models train,[1221.32] [1221.32][S03]how they're being trained, how they learn.[1225.36] [1225.36][S03]And the same kind of principles that helped us scale and improve[1230.84] [1230.84][S03]Veo, we found them to be useful for Genie 3 as well.[1233.253] [1233.253][S01] It seems to me--[1234.42] [1234.42][S01]OK.[1235.3] [1235.3][S01]As an outsider, it seems to me that the objective[1238.58] [1238.58][S01]of Genie and Veo are, I mean, even though they[1241.5] [1241.5][S01]look visually quite similar, the objectives of them[1243.82] [1243.82][S01]are quite different.[1244.78] [1244.78][S01]Veo, you're trying to create this very realistic,[1247.32] [1247.32][S01]non-interactive environment, whereas with Genie you[1250.62] [1250.62][S01]need to make this consistent, explorable world that you[1255.02] [1255.02][S01]can move around in.[1256.4] [1256.4][S01]Do you have to start from scratch[1258.38] [1258.38][S01]or is there kind of swaps that you can make?[1261.378] [1261.378][S03] So I think in a way,[1262.92] [1262.92][S03]if you think about the video model, you can tell it,[1265.36] [1265.36][S03]OK, I want to walk around maybe a volcano[1269.18] [1269.18][S03]or I want to-- the camera to move that way or another.[1272.04] [1272.04][S03]So what it does is basically looking at this entire video[1275.18] [1275.18][S03]and tries to create a coherent eight second or whatever[1279.82] [1279.82][S03]long video, and it can change the past and the future[1282.7] [1282.7][S03]at the same time.[1283.48] [1283.48][S03]I think that's the property of video models[1285.8] [1285.8][S03]that is very different from Genie 3.[1287.387] [1287.387][S02] Because it spits out the end result.[1289.72] [1289.72][S03] Exactly.[1290.762] [1290.762][S03]And you can think of it like as a painting[1292.91] [1292.91][S03]that you can change all the time everything on the canvas.[1297.27] [1297.27][S03]And while it's much, in a way, easier[1300.95] [1300.95][S03]than doing it in what we call autoregressive[1303.55] [1303.55][S03]or just extending one frame at a time.[1306.57] [1306.57][S03]So I think that's the fundamental difference.[1308.65] [1308.65][S03]There's still a lot of similar aspects.[1312.237] [1312.237][S03]For example, the way that we have eventually[1314.07] [1314.07][S03]to take some text input and convert it[1315.95] [1315.95][S03]to some kind of visual output.[1318.05] [1318.05][S03]So that's somewhat similar.[1319.61] [1319.61][S03]So I would say definitely we don't start from scratch.[1323.91] [1323.91][S03]We build on top of, I think, a lot of work in this space,[1326.73] [1326.73][S03]but there are definitely some novel properties[1329.07] [1329.07][S03]that we had to figure out.[1330.153] [1330.153][S01] Well, that's super interesting, actually,[1332.362] [1332.362][S01]the idea that time is the key component in this,[1334.73] [1334.73][S01]that you have to understand the past[1336.23] [1336.23][S01]and march forward to the future.[1337.85] [1337.85][S01]That's where the autoregressive stuff comes in, right?[1340.15] [1340.15][S02] Exactly.[1340.65] [1340.65][S02]So essentially every frame you see[1342.67] [1342.67][S02]is generated from scratch at that point in time.[1346.71] [1346.71][S02]So things that happen later in the interaction[1349.83] [1349.83][S02]aren't known yet, and things that happen at the beginning[1352.99] [1352.99][S02]have to all be remembered by the model.[1355.67] [1355.67][S02]So essentially, if we're doing 24 frames per second,[1358.67] [1358.67][S02]it's like doing image generation 24 times per second,[1362.99] [1362.99][S02]each one completely generating from scratch,[1365.63] [1365.63][S02]given all of the past and the actions[1367.55] [1367.55][S02]of the agent or human player.[1370.057] [1370.057][S01] There's an analogy here, though.[1371.89] [1371.89][S01]I mean, that's sort of the way that language models work,[1374.31] [1374.31][S01]right?[1374.68] [1374.68][S03] Yes, exactly.[1375.93] [1375.93][S03]So I think it's really a good example.[1377.57] [1377.57][S03]So we know that language models are basically[1379.75] [1379.75][S03]being trained to predict the next token or word.[1382.13] [1382.13][S03]So they look at text and try to guess, OK,[1384.73] [1384.73][S03]what's the distribution of words or tokens that happen to follow?[1389.27] [1389.27][S03]And what we have with autoregressive word models,[1392.11] [1392.11][S03]we actually have a similar problem.[1394.37] [1394.37][S03]We want to basically predict the next observation, which[1398.87] [1398.87][S03]is pretty much visual, the next frame, given[1401.51] [1401.51][S03]what was already seen.[1403.77] [1403.77][S03]And the nice thing, I think, in this kind of parallel to LLMs[1408.27] [1408.27][S03]is that LLMs learn from that very simple task,[1411.64] [1411.64][S03]a very rich, potentially, representation of the world[1415.16] [1415.16][S03]or how people think, how people solve problems, for example.[1420.0] [1420.0][S03]I think the reason that word models,[1421.74] [1421.74][S03]especially autoregressive word models,[1423.38] [1423.38][S03]are kind of exciting for us is because maybe[1426.64] [1426.64][S03]through that task, that is pretty much[1428.68] [1428.68][S03]simple task that anyone can understand,[1431.5] [1431.5][S03]they have to learn the dynamics of the world.[1434.26] [1434.26][S03]And I think it's nice, if you really think about it,[1438.2] [1438.2][S03]it's kind of a superset of intelligence.[1440.96] [1440.96][S03]Because through the world, if you say,[1443.02] [1443.02][S03]OK, now I'm playing chess with some grandmaster[1445.28] [1445.28][S03]and it was the next visual or the next frame would actually[1449.64] [1449.64][S03]be their next maybe move.[1452.3] [1452.3][S03]So of course, our model is not capable of doing that,[1455.22] [1455.22][S03]but at the limit, it kind of goes very, very far.[1457.558] [1457.558][S01] But it's these ideas of understanding[1459.6] [1459.6][S01]the past, the context, and being able to predict the next move[1462.76] [1462.76][S01]in the future, which is--[1464.16] [1464.16][S03] Yes.[1464.56] [1464.56][S02] Yeah.[1464.78] [1464.78][S02]It's also quite powerful, because it[1466.28] [1466.28][S02]means you can start in the same location[1468.24] [1468.24][S02]and do many very different things.[1470.72] [1470.72][S02]So from an agent's perspective, there[1473.08] [1473.08][S02]could even be quite a simple task,[1474.74] [1474.74][S02]but you want to get really good at it.[1476.48] [1476.48][S02]And so you can simulate various different scenarios.[1479.86] [1479.86][S02]This is very analogous to reinforcement learning paradigm,[1482.36] [1482.36][S02]where you have this end reset function that brings you back[1485.92] [1485.92][S02]to the same state.[1486.773] [1486.773][S02]And then you want to get more experience from there.[1488.94] [1488.94][S01] Well, let's inch through this then.[1490.898] [1490.898][S01]So the very first frame.[1492.88] [1492.88][S01]You can have an image, like you did with the painting,[1496.64] [1496.64][S01]but you can also have a text input, right?[1498.61] [1498.61][S03] Yes, of course.[1499.5] [1499.5][S01] I mean, I know you asked me[1501.2] [1501.2][S01]earlier to describe somewhere I'd been,[1503.72] [1503.72][S01]and I got to go to a hunter's lodge in Siberia once.[1506.42] [1506.42][S01]So it was the most impressive thing I could come up with.[1509.928] [1509.928][S01]And this is what I sent to you.[1511.22] [1511.22][S01]I went to hunter's lodge, sat on a reindeer skin in the woods[1514.96] [1514.96][S01]outside of Yakutsk in Siberia, drank vodka,[1517.52] [1517.52][S01]and ate frozen calf liver.[1519.117] [1519.117][S03] Yes, so--[1520.2] [1520.2][S01] Tell me what you did with that.[1521.22] [1521.22][S03] So first, we were surprised with the prompt,[1523.762] [1523.762][S03]but then we just tried to--[1525.0] [1525.0][S03]we put it into the system, and the system is basically able[1528.81] [1528.81][S03]to add a few more details to the prompt that you provided.[1534.75] [1534.75][S03]But still it does follow the key elements that you provided.[1539.85] [1539.85][S03]So for example, I think it's reasonably--[1543.09] [1543.09][S03]I'm not sure if it's a reindeer, but it's definitely--[1546.97] [1546.97][S03]so, and then here you can see the table.[1549.477] [1549.477][S01] It's really amazing.[1550.81] [1550.81][S01]The light, again.[1551.91] [1551.91][S01]You guys are very--[1552.97] [1552.97][S01]you love making the light late afternoon,[1555.37] [1555.37][S01]a beautiful golden hour.[1557.05] [1557.05][S02] You also missed the calf liver.[1559.175] [1559.175][S02]And it's frozen, too.[1560.37] [1560.37][S02]Frozen calf liver.[1561.51] [1561.51][S02]It's not even just calf liver.[1563.65] [1563.65][S01] That first frame is generated in the same way[1566.29] [1566.29][S01]as you might find in Veo.[1568.253] [1568.253][S03] Yeah.[1569.17] [1569.17][S03]So the model doesn't treat it like in any special way.[1571.55] [1571.55][S03]It's just like you give it a text,[1572.967] [1572.967][S03]and it just starts outputting frames[1577.05] [1577.05][S03]and it doesn't do any preparation before it.[1579.51] [1579.51][S03]It just throws you into the world[1581.13] [1581.13][S03]and you can go wherever you want.[1582.63] [1582.63][S01] And then it's from that first frame[1583.93] [1583.93][S01]that the prediction backwards and forwards.[1585.67] [1585.67][S02] Exactly, yeah.[1586.23] [1586.23][S02]So it's like text to first frame,[1588.19] [1588.19][S02]then first action to next frame and so on[1590.77] [1590.77][S02]and so forth from that point onwards, basically.[1593.05] [1593.05][S01] And how critical is the exact wording of the prompt[1595.85] [1595.85][S01]here?[1596.35] [1596.35][S01]I mean, can you get better images, better worlds[1598.53] [1598.53][S01]with better prompts?[1599.79] [1599.79][S02] Yeah, I think that's definitely true.[1601.19] [1601.19][S02]There's an art to prompting, I think,[1603.09] [1603.09][S02]all of these modern models, and some people are better at it[1606.37] [1606.37][S02]than others.[1606.87] [1606.87][S02]And fortunately, we have some people[1608.37] [1608.37][S02]who are much better than me at this.[1610.47] [1610.47][S03] This one worked pretty much out of the box.[1612.97] [1612.97][S02] This one, yeah.[1614.428] [1614.428][S02]Often they work pretty well, especially when[1616.41] [1616.41][S02]you have very vivid description, like the table with the vodka[1619.93] [1619.93][S02]and the frozen calf's liver.[1622.17] [1622.17][S02]But sometimes you can try something and not quite capture[1627.85] [1627.85][S02]exactly what you wanted first time,[1629.35] [1629.35][S02]and then you can iterate a little bit on the prompt[1631.01] [1631.01][S02]and then get something that's much more like what you wanted.[1633.35] [1633.35][S01] And does that require[1634.725] [1634.725][S01]you to regenerate the world?[1636.59] [1636.59][S01]Or given that it's marching forwards,[1638.51] [1638.51][S01]can you add things in on the fly?[1640.07] [1640.07][S03] So we do have ways[1641.528] [1641.528][S03]to add things on the fly, what we call promptable world events.[1644.94] [1644.94][S03]And this is just something that you can say, OK,[1647.36] [1647.36][S03]now I want, for example, a balloon to fly in[1651.82] [1651.82][S03]or some other character to show up.[1653.883] [1653.883][S03]This is something we're very excited about because it also[1656.3] [1656.3][S03]allows the environment to be more interesting for people,[1659.52] [1659.52][S03]but also, as we mentioned, more relevant for training[1663.82] [1663.82][S03]agents in simulation, because then we[1666.34] [1666.34][S03]can throw in something that happens in the world[1668.98] [1668.98][S03]that they have to adapt to.[1670.718] [1670.718][S01] So in that example, maybe[1672.26] [1672.26][S01]have a reindeer coming through.[1673.66] [1673.66][S03] Right.[1674.618] [1674.618][S03]We can have a reindeer, or we can[1676.14] [1676.14][S03]have another person is walking into the scene[1679.38] [1679.38][S03]and then we have to-- if it's an agent,[1682.1] [1682.1][S03]basically, it can respond to it, and if it's just[1685.18] [1685.18][S03]for entertainment purposes, then it's[1687.137] [1687.137][S03]much more interesting than just walking in a world[1689.22] [1689.22][S03]that nothing happens.[1690.095] [1690.095][S01] And this is all a key advantage[1691.887] [1691.887][S01]of the autoregressive part, I guess,[1693.46] [1693.46][S01]that you do have control over the future.[1695.52] [1695.52][S02] Exactly.[1695.98] [1695.98][S02]Yeah.[1696.48] [1696.48][S02]So you can just inject things on the fly[1699.38] [1699.38][S02]as you're generating things in real time.[1701.203] [1701.203][S01] The thing that I find extraordinary about this[1703.62] [1703.62][S01]is the consistency.[1704.94] [1704.94][S01]Right?[1705.44] [1705.44][S01]The memory of the system.[1707.42] [1707.42][S01]You turn away and then you turn back,[1709.04] [1709.04][S01]and it's exactly how you left it.[1710.76] [1710.76][S01]But presumably, if you're turning in a direction[1713.0] [1713.0][S01]you haven't yet turned in, then it's[1714.5] [1714.5][S01]sort of a stochastic process.[1716.58] [1716.58][S01]So how do you balance it, that it's sometimes a memory[1720.74] [1720.74][S01]and sometimes generated statistically?[1723.862] [1723.862][S02] Well, it's kind of a mixture of things.[1726.32] [1726.32][S02]So you can, in the text prompt, specify[1728.482] [1728.482][S02]things that are out of sight, which[1729.94] [1729.94][S02]I think is quite powerful compared to image prompting.[1732.96] [1732.96][S02]Because you can say, on the right is x, y, z,[1736.86] [1736.86][S02]and then when you actually look round to the right,[1739.04] [1739.04][S02]it's there when you actually play or interact in the world.[1742.885] [1742.885][S02]But then there is also an element[1744.26] [1744.26][S02]of the model using its world knowledge to generate things.[1746.68] [1746.68][S02]So in the Hopper example with the artwork,[1750.38] [1750.38][S02]it does generate the street that we haven't seen before.[1753.978] [1753.978][S02]And you can't be exactly certain what it's going to generate.[1756.52] [1756.52][S02]The model uses its own sort of intuition[1758.94] [1758.94][S02]of what should be there.[1760.54] [1760.54][S01] And that's intuition based on, I mean,[1762.73] [1762.73][S01]having watched, I mean, incredible amounts[1765.39] [1765.39][S01]of video footage in advance.[1767.093] [1767.093][S03] Yeah.[1768.01] [1768.01][S03]So basically the model is trying to just generate[1771.67] [1771.67][S03]a sequence of frames that is representative of the world.[1775.31] [1775.31][S03]So if it already saw some part or generated[1778.35] [1778.35][S03]some part of the world, then the right thing for the model to do[1781.15] [1781.15][S03]would be to make--[1782.47] [1782.47][S03]to recall this memory and use it.[1784.99] [1784.99][S03]So if I look back to where I've already been,[1787.99] [1787.99][S03]then the right thing for the model[1789.79] [1789.79][S03]would just to use the same thing.[1791.41] [1791.41][S03]But when you look to a new area from the model's perspective,[1794.17] [1794.17][S03]again, it can allow itself to generate something new[1796.69] [1796.69][S03]because it hasn't been seen.[1798.17] [1798.17][S03]So the model doesn't really treat it[1799.67] [1799.67][S03]in a fundamentally different way.[1802.49] [1802.49][S03]So the model, it learns to basically balance[1806.59] [1806.59][S03]the two aspects.[1808.17] [1808.17][S03]And again, all comes back to anchoring all of the generation[1812.55] [1812.55][S03]to the prompt or what the user provided.[1815.273] [1815.273][S03]That's where the information to the generation comes from.[1817.69] [1817.69][S01] But then I guess going back[1818.39] [1818.39][S01]to the language analogy that we were talking about earlier.[1820.89] [1820.89][S01]It's not that surprising that once[1822.445] [1822.445][S01]a statement has been established in a conversation,[1824.57] [1824.57][S01]that it remains consistent when you refer back to it.[1826.61] [1826.61][S02] Exactly, yeah.[1828.027] [1828.027][S02]And we've seen with language models, this ability[1830.47] [1830.47][S02]to have memory and consistency has been something that's[1834.51] [1834.51][S02]improved a lot recently, especially with the latest[1837.03] [1837.03][S02]Gemini models now versus two years ago.[1838.973] [1838.973][S03] Yeah.[1839.89] [1839.89][S03]I think what's interesting here is that while,[1842.11] [1842.11][S03]if you think about the number of the size of the memory[1846.31] [1846.31][S03]or the level of detail.[1847.61] [1847.61][S03]So if we think about I'm talking to Gemini,[1852.27] [1852.27][S03]and after a few sentences, it might refer to something[1854.99] [1854.99][S03]I've said before.[1855.97] [1855.97][S03]That's great.[1856.81] [1856.81][S03]But the number of details that we have in those visual words[1861.83] [1861.83][S03]is just like, it's staggering, right?[1863.49] [1863.49][S03]The quality of the memory, considering how much detail[1867.63] [1867.63][S03]and how much information it has to actually remember.[1870.61] [1870.61][S03]Yeah.[1871.11] [1871.11][S01] Well, how do you do that?[1872.17] [1872.17][S01]Does this require you having a sort of 3D representation[1874.91] [1874.91][S01]of the world that you're in?[1876.482] [1876.482][S02] We don't use anything[1878.19] [1878.19][S02]like that for this version of the model.[1880.56] [1880.56][S02]It's largely just learnt as an emergent property just from this[1884.817] [1884.817][S02]autoregressive prediction.[1885.9] [1885.9][S01] You guys with your emergent properties.[1887.34] [1887.34][S01]My goodness me.[1888.2] [1888.2][S02] Yeah.[1889.242] [1889.242][S03] We are very good students of the Beatles.[1891.68] [1891.68][S02] Yeah, I think it's similar to,[1893.68] [1893.68][S02]if you're predicting the next frame,[1895.18] [1895.18][S02]you just have to learn to remember these critical things[1897.52] [1897.52][S02]in the past.[1898.68] [1898.68][S02]And obviously the model has some representations[1900.68] [1900.68][S02]that prioritize the important details,[1903.4] [1903.4][S02]but it really is just predicting the next frame.[1905.46] [1905.46][S01] And is this analogous, then, to the fact[1907.627] [1907.627][S01]that language models have this conceptual understanding[1910.2] [1910.2][S01]and can describe similar things in--[1913.36] [1913.36][S01]the same things in different ways?[1914.943] [1914.943][S01]Is it the transformer architecture[1916.36] [1916.36][S01]here that's allowing you to do that?[1917.86] [1917.86][S03] Yes.[1918.735] [1918.735][S03]So the architecture is--[1919.76] [1919.76][S03]I think almost everything today is a transformer.[1922.34] [1922.34][S03]So yes, this is also a transformer.[1925.56] [1925.56][S03]And looking back, basically the model, we[1928.52] [1928.52][S03]will see what was already generated,[1931.68] [1931.68][S03]including in the user provided inputs, and based on that,[1935.32] [1935.32][S03]makes the prediction.[1936.54] [1936.54][S03]So I think the interesting question about explicit 3D is[1941.28] [1941.28][S03]that the model probably had to learn some representation,[1945.46] [1945.46][S03]but it's just not an explicit representation.[1948.3] [1948.3][S03]So we see that the ability of the model to understand 3D[1952.04] [1952.04][S03]environments is very strong.[1953.96] [1953.96][S03]And I think the most, to me, an emergent, if you want,[1958.2] [1958.2][S03]capability, is that it actually works[1959.96] [1959.96][S03]where it was not trained for.[1961.26] [1961.26][S03]For example, in taking a painting,[1964.14] [1964.14][S03]an oil painting in 1942 and actually making it into some[1968.84] [1968.84][S03]kind of a 3D environment.[1969.96] [1969.96][S03]That's pretty much out of its distribution.[1972.86] [1972.86][S01] Absolutely.[1973.88] [1973.88][S01]I want to go back to this idea of understanding physics.[1977.96] [1977.96][S01]Jack, if you've got this world that you can put an agent in,[1982.28] [1982.28][S01]does that mean that you can test how well it knows physics?[1985.46] [1985.46][S01]I mean, could you get a hammer and a feather, for instance,[1988.16] [1988.16][S01]and allow it to drop it at the same time?[1990.018] [1990.018][S02] You definitely could do that.[1992.06] [1992.06][S02]I think that would probably be close to the frontier[1995.0] [1995.0][S02]of the model's capabilities at this point.[1997.77] [1997.77][S02]We've seen it's quite good at visual things and things that[2001.09] [2001.09][S02]are more general concepts.[2002.31] [2002.31][S02]So you can imagine water occurs in quite[2005.09] [2005.09][S02]a lot of different scenarios it's seen before.[2008.21] [2008.21][S02]Gravity has probably occurred in many of those,[2010.87] [2010.87][S02]but probably not these exact objects.[2013.47] [2013.47][S02]But I think if you were to specify in the text prompt,[2016.11] [2016.11][S02]like in the world, there is a feather that[2018.81] [2018.81][S02]has less gravity and a hammer that is heavier,[2022.19] [2022.19][S02]maybe then it would work.[2023.51] [2023.51][S03] Yeah.[2024.427] [2024.427][S03]I think there are a lot of--[2025.65] [2025.65][S03]those models are inherently visual.[2027.943] [2027.943][S03]They don't know anything about the world[2029.61] [2029.61][S03]except for what we can see.[2032.01] [2032.01][S03]And that's a limitation, I would say.[2033.95] [2033.95][S03]Even for video models, some things[2036.53] [2036.53][S03]sometimes don't make sense.[2038.05] [2038.05][S03]And the model has to guess if something is heavy, for example.[2041.63] [2041.63][S03]There is no real way for a model to look at an image[2044.81] [2044.81][S03]and say, OK, that's how much this thing weigh.[2047.95] [2047.95][S03]So it kind of makes up the weight[2050.17] [2050.17][S03]and then try to simulate what would have happened.[2052.79] [2052.79][S03]And sometimes it breaks.[2054.69] [2054.69][S03]And we know that from video models, even the best video[2057.61] [2057.61][S03]models.[2058.31] [2058.31][S03]And I think what we have is basically we[2060.969] [2060.969][S03]solve an even harder problem because we have to, as we said,[2064.75] [2064.75][S03]not just-- we can't fix the past.[2066.77] [2066.77][S03]We have to go--[2067.732] [2067.732][S03]once the model generates something,[2069.19] [2069.19][S03]it has to roll with it.[2071.17] [2071.17][S03]And I think we have really good progress in terms of simulation,[2076.01] [2076.01][S03]like fluid dynamics.[2077.969] [2077.969][S03]But maybe some other aspects of the world physically[2081.25] [2081.25][S03]would not be accurate because of those limitations.[2083.78] [2083.78][S01] Well, tell me some of the work[2085.53] [2085.53][S01]that you've been doing with agents so far, because actually[2087.988] [2087.988][S01]on a previous episode, we got to talk about SIMA, the Scalable[2091.81] [2091.81][S01]Instructable Multiworld Agent.[2093.802] [2093.802][S02] Exactly.[2094.969] [2094.969][S01] Which they were putting into existing computer[2097.53] [2097.53][S01]game environments.[2098.77] [2098.77][S01]But I mean, you can now take SIMA[2102.01] [2102.01][S01]and put it into these generated environments too.[2104.47] [2104.47][S02] Exactly.[2105.637] [2105.637][S02]So the cool thing with this is that we've got--[2107.65] [2107.65][S02]at Google DeepMind, we have these agents[2109.45] [2109.45][S02]that are trained to be general.[2111.41] [2111.41][S02]The M is multiworld, as you just said.[2114.82] [2114.82][S02]And so what we're able to do is take all the worlds we generate[2117.58] [2117.58][S02]and then test whether these agents and the latest[2119.7] [2119.7][S02]versions of them are able to already use them[2122.42] [2122.42][S02]for agent training or collecting experiences or evaluations[2126.62] [2126.62][S02]as they are.[2127.42] [2127.42][S02]So you can say to it something like, navigate to the robot[2131.94] [2131.94][S02]over there, and then you can give it an image from the world[2136.44] [2136.44][S02]and it can take the first action.[2138.16] [2138.16][S02]And then from that point onwards,[2139.56] [2139.56][S02]it's interacting in the world through actions.[2142.1] [2142.1][S02]But the Genie 3 model does not know what the goal is.[2145.46] [2145.46][S02]So that makes it like a genuine simulation,[2148.56] [2148.56][S02]rather than if you were to tell Genie 3 that the SIMA agent is[2151.7] [2151.7][S02]trying to achieve this goal.[2153.46] [2153.46][S02]It might make the experience not authentic,[2155.58] [2155.58][S02]because it might make it happen in an incorrect way.[2157.888] [2157.888][S01] For it.[2158.68] [2158.68][S02] Yeah, exactly.[2159.52] [2159.52][S01] I guess if you've got a robot in the real world,[2161.32] [2161.32][S01]the world isn't helping the robot.[2163.19] [2163.19][S02] Exactly.[2163.28] [2163.28][S03] Exactly.[2164.322] [2164.322][S03]You can't say, OK, the robot has to go to fetch the red cube,[2167.09] [2167.09][S03]and then it looks to the left and there is a red cube.[2169.34] [2169.34][S03]It's a bit making it.[2170.215] [2170.215][S02] Yeah, and this[2171.632] [2171.632][S02]is a problem that we've encountered in other scenarios,[2174.16] [2174.16][S02]but you don't really get this problem[2175.74] [2175.74][S02]if you have this separation between the agent[2177.5] [2177.5][S02]and the environment.[2178.38] [2178.38][S02]And that's the really nice thing of working with agent teams that[2181.62] [2181.62][S02]are focused on building really capable agents,[2183.56] [2183.56][S02]but then having them access our environment as[2185.82] [2185.82][S02]if it's any other environment.[2187.38] [2187.38][S02]So the new worlds that are created by Genie 3[2190.5] [2190.5][S02]look the same as the existing worlds[2192.78] [2192.78][S02]that the SIMA agent was trained on.[2194.993] [2194.993][S01] But then what if there isn't a red cube?[2197.16] [2197.16][S01]What if it searches around forever[2198.66] [2198.66][S01]and there is no red cube?[2199.802] [2199.802][S03] So yeah, I think that's part of the things[2202.26] [2202.26][S03]that we're looking into is how we[2203.9] [2203.9][S03]can add more details as we go.[2206.76] [2206.76][S03]For example, if you put the agent in some room and then[2210.46] [2210.46][S03]it has to open maybe a drawer and find something in there.[2213.28] [2213.28][S03]So we want to be able to inject kind of events into the world[2220.9] [2220.9][S03]and control it.[2221.72] [2221.72][S03]So I think the interesting frontier here[2225.58] [2225.58][S03]is how you can make the world look very realistic[2229.62] [2229.62][S03]and also control what's happening[2232.31] [2232.31][S03]in the world in a way that still makes sense.[2234.847] [2234.847][S03]So I think what we've seen is that with our prompt about world[2237.43] [2237.43][S03]events, we can add things that happen in the world.[2240.29] [2240.29][S03]But if you just want something to pop into the world,[2242.81] [2242.81][S03]then it's not necessarily a plausible thing.[2245.41] [2245.41][S03]If you look at--[2246.367] [2246.367][S03]I don't know, if you're in the desert,[2247.95] [2247.95][S03]and then you're going to ask it, OK, now I[2249.7] [2249.7][S03]want to see an elephant, then where is this elephant[2253.098] [2253.098][S03]going to come from?[2253.89] [2253.89][S03]So maybe it's going to come from the side[2255.07] [2255.07][S03]or when you look to the left.[2256.29] [2256.29][S03]So I think there is something really interesting in what[2258.75] [2258.75][S03]does it mean to change a world, because the model[2262.59] [2262.59][S03]has some assumptions.[2263.65] [2263.65][S03]And when it comes to agents, this[2265.43] [2265.43][S03]is definitely an important capability[2267.91] [2267.91][S03]to be able to inject this new event into the world.[2270.61] [2270.61][S01] So you've done this already.[2272.51] [2272.51][S02] Yeah, we've got signs of life for this.[2274.968] [2274.968][S02]I think we haven't really say we've[2276.67] [2276.67][S02]got a full agent training loop where[2278.75] [2278.75][S02]we're already doing some large scale[2280.878] [2280.878][S02]training in these environments.[2282.17] [2282.17][S02]But what we can already do is test the agents in them[2284.83] [2284.83][S02]and see how they do.[2286.01] [2286.01][S02]And I think it's quite remarkable[2287.47] [2287.47][S02]that, given these things weren't developed together,[2290.65] [2290.65][S02]we just dropped the agent in and it can already do things.[2293.657] [2293.657][S02]And you can imagine all the different things[2295.49] [2295.49][S02]you could now use this for.[2296.71] [2296.71][S01] Color that in for me.[2297.57] [2297.57][S01]Give me an example.[2298.53] [2298.53][S03] If you have a factory[2300.59] [2300.59][S03]where you have some robots and you[2302.15] [2302.15][S03]want to introduce maybe a new--[2304.19] [2304.19][S03]maybe a very boring example, but a new machine, right?[2307.57] [2307.57][S03]And that wasn't there.[2308.65] [2308.65][S03]Or you changed somehow the structure of the building[2312.09] [2312.09][S03]and you want to test the robot before you actually[2314.482] [2314.482][S03]put them in the new building.[2315.69] [2315.69][S03]So again, this is like you can simulate the world that[2319.67] [2319.67][S03]basically is a variant of what maybe the agent has[2323.87] [2323.87][S03]seen in the past and see if it breaks.[2325.67] [2325.67][S03]Right?[2326.47] [2326.47][S03]Everything can happen in simulated environments[2328.99] [2328.99][S03]and not necessarily break your new machine.[2330.85] [2330.85][S03]So that's one example, I would say.[2333.265] [2333.265][S01] Find the unintended consequences.[2335.14] [2335.14][S03] Yeah, yeah.[2335.91] [2335.91][S03]And the evaluation of the model.[2337.49] [2337.49][S03]So that's even not training the agent,[2339.95] [2339.95][S03]just testing how well it adapts, maybe,[2341.95] [2341.95][S03]to a new variation of an environment.[2344.655] [2344.655][S01] All of these examples[2346.03] [2346.03][S01]you've given so far are where the agent has a specified[2349.04] [2349.04][S01]objective, which I know is the point of SIMA thus far.[2352.96] [2352.96][S01]But what about if you had agents that didn't have an objective?[2357.018] [2357.018][S01]There was a really nice quote that I came across,[2359.06] [2359.06][S01]which is that almost no prerequisite[2360.8] [2360.8][S01]to any major invention was made with that invention in mind.[2364.96] [2364.96][S01]Can you imagine a point in the future[2367.68] [2367.68][S01]where you are letting agents loose in these environments[2370.96] [2370.96][S01]without specifying an objective for them?[2373.203] [2373.203][S02] Yeah, exactly.[2374.62] [2374.62][S02]So the quote comes from \"Why Greatness Cannot Be Planned,\"[2377.74] [2377.74][S02]which is from Ken Stanley and Joel Lehman.[2379.9] [2379.9][S02]It's a great book.[2381.48] [2381.48][S02]And the general idea there is that searching[2384.0] [2384.0][S02]for interestingness might actually[2386.12] [2386.12][S02]lead to things that are more useful for practical goals[2389.383] [2389.383][S02]than if you just directly optimize for the practical goals[2391.8] [2391.8][S02]themselves.[2392.73] [2392.73][S02]And clearly, the bigger the domain[2395.8] [2395.8][S02]and the space for discovery, the more interesting things[2398.6] [2398.6][S02]could happen.[2399.74] [2399.74][S02]So they had this really nice example[2402.4] [2402.4][S02]quite a while ago with this paper called Picbreeder, where[2405.44] [2405.44][S02]essentially they allowed people to select images and combine[2410.08] [2410.08][S02]them to create new images that were mutations of those two.[2413.56] [2413.56][S02]People weren't directly optimizing for specific end[2416.24] [2416.24][S02]goals, but by just choosing what they found interesting,[2419.8] [2419.8][S02]they ended up discovering really cool, structured pictures,[2423.7] [2423.7][S02]like a skull or a butterfly, that weren't obvious[2426.56] [2426.56][S02]how you would reach that from the starting points.[2429.12] [2429.12][S02]And some of the stepping stones along the way[2431.2] [2431.2][S02]didn't really look much like the final goal,[2433.04] [2433.04][S02]and they wouldn't have been things you obviously[2435.04] [2435.04][S02]would have chosen to go for if you had those goals in mind.[2438.32] [2438.32][S02]And there are lots of examples of this in the real world.[2442.06] [2442.06][S02]If you are, for example, trying to reach the moon,[2444.66] [2444.66][S02]you wouldn't build a bigger ladder.[2446.28] [2446.28][S02]So optimizing along one dimension with maybe[2450.0] [2450.0][S02]a greedy, myopic approach doesn't always[2452.48] [2452.48][S02]lead you to make these big leaps.[2453.98] [2453.98][S01] Well, I mean, evolution itself[2455.73] [2455.73][S01]is like the classic example of iteration without objective.[2460.52] [2460.52][S02] Yeah, and we see this a lot in research.[2463.02] [2463.02][S03] Yeah.[2463.937] [2463.937][S03]My perspective on that is that I think we as humans,[2467.55] [2467.55][S03]we decide what's interesting.[2469.27] [2469.27][S03]I think there is even an example where[2471.05] [2471.05][S03]the entire evolution of mathematics[2473.39] [2473.39][S03]was guided by people deciding what's next, what's interesting,[2476.83] [2476.83][S03]what's not interesting, and just the problem being hard[2480.21] [2480.21][S03]doesn't mean that it's interesting at all.[2482.762] [2482.762][S03]And I think in a way, when we think[2484.89] [2484.89][S03]about generating new things in science,[2487.71] [2487.71][S03]it's like the goal or the-- it doesn't have to be a goal,[2491.31] [2491.31][S03]but there is some aspects of maybe beauty or interest that[2495.45] [2495.45][S03]is coming from us, and models might maybe[2498.89] [2498.89][S03]learn to simulate that.[2500.315] [2500.315][S03]But I think it's really important[2501.69] [2501.69][S03]to remember that we ultimately decide[2504.09] [2504.09][S03]what's interesting as our preferences as people.[2508.75] [2508.75][S01] I mean, you haven't had the move 37, as it were,[2511.25] [2511.25][S01]in this method.[2512.25] [2512.25][S03] Even in this case, the goal of Go or the game[2517.73] [2517.73][S03]was designed in a certain way that people[2519.93] [2519.93][S03]find it interesting to play, otherwise--[2522.32] [2522.32][S03]and it's a very ancient game.[2524.03] [2524.03][S03]So I'm just saying that the setup, even when there is--[2528.13] [2528.13][S03]or getting to the moon, this goal[2530.17] [2530.17][S03]was kind of maybe made by people.[2532.99] [2532.99][S03]So I'm just saying, I think there is still[2535.37] [2535.37][S03]broader constraints set by us, in a way.[2540.81] [2540.81][S03]If a machine comes up with a problem to solve,[2545.062] [2545.062][S03]we still have to say, is this an interesting problem?[2547.27] [2547.27][S03]Because otherwise we're just like, OK,[2548.59] [2548.59][S03]I don't care about that.[2549.77] [2549.77][S03]So I think there is still an aesthetic part to anything[2552.29] [2552.29][S03]that's open-ended to me.[2554.15] [2554.15][S02] Yeah, I think[2554.49] [2554.49][S02]there was this quote from Dennis a while ago about the levels[2557.17] [2557.17][S02]of creativity, and it was like, interpolation is one.[2559.97] [2559.97][S02]Like, you see a new cat and you can identify it as a cat.[2562.61] [2562.61][S02]Extrapolation was one where it's like, given the rules of Go,[2565.59] [2565.59][S02]can you discover a new move like move 37?[2567.93] [2567.93][S02]And then the third level is generating completely[2570.41] [2570.41][S02]new things.[2571.11] [2571.11][S02]So, could you actually invent Go, is what he said.[2573.43] [2573.43][S02]And we actually had this as a motivation for the Genie project[2576.15] [2576.15][S02]at the beginning.[2576.858] [2576.858][S02]It was like, can you create completely new things?[2580.29] [2580.29][S02]And I think we're starting to see that happen.[2582.49] [2582.49][S02]Since it's a completely new kind of model, someone on the team[2586.06] [2586.06][S02]will do something like create a certain kind of world,[2589.227] [2589.227][S02]and then immediately other members of the team are like,[2591.56] [2591.56][S02]that's really interesting.[2593.02] [2593.02][S02]And then they start evolving that idea themselves.[2596.58] [2596.58][S02]And then we post it on social media[2598.96] [2598.96][S02]and we see the reaction to some things,[2600.62] [2600.62][S02]and then we know that's interesting.[2602.12] [2602.12][S02]So then we create new things that way.[2604.08] [2604.08][S02]And that's just with a very limited access to the model.[2607.48] [2607.48][S02]So clearly you can see if we open this up a bit more[2610.74] [2610.74][S02]in the future, that it could lead[2612.66] [2612.66][S02]to some sort of open-ended creativity that way.[2615.382] [2615.382][S03] It's kind of like but still the evolution.[2617.84] [2617.84][S03]It's like an evolution with the criteria[2619.42] [2619.42][S03]that people find interesting.[2620.36] [2620.36][S02] Yeah, exactly.[2621.777] [2621.777][S02]There's people in the loop guiding the interestingness.[2624.5] [2624.5][S01] Well, then, OK, could you take--[2625.66] [2625.66][S01]I'm just thinking back to the conversation[2626.98] [2626.98][S01]that we had with Dave Silver earlier in the series[2629.063] [2629.063][S01]where he was saying, actually, there's[2631.34] [2631.34][S01]one way you remove humans from the equation[2633.34] [2633.34][S01]and actually you get even more surprising results, potentially.[2637.74] [2637.74][S01]OK, just play along with me here for a moment.[2640.3] [2640.3][S01]But could you get to a point where you just[2642.58] [2642.58][S01]simulate the first, I don't know, single-celled organism[2645.62] [2645.62][S01]and allow it to evolve inside of Genie[2649.34] [2649.34][S01]and actually watch the process of evolution[2651.62] [2651.62][S01]happen in a virtual environment?[2654.387] [2654.387][S02] That's a great question.[2656.22] [2656.22][S02]And that's the dream of an A life, open-ended evolution[2659.58] [2659.58][S02]community.[2661.62] [2661.62][S02]I think maybe the worlds that we create[2664.18] [2664.18][S02]are not fully rich enough, but I definitely[2666.46] [2666.46][S02]think that we're getting along that path.[2669.16] [2669.16][S02]So open-ended evolution and A life,[2671.123] [2671.123][S02]they've been designing worlds that[2672.54] [2672.54][S02]could facilitate this kind of thing, typically in code.[2675.86] [2675.86][S02]And so this could be an alternative approach[2679.18] [2679.18][S02]to getting maybe richer real world simulations.[2683.34] [2683.34][S02]And so in theory, I mean, we've made quite a lot of progress[2686.98] [2686.98][S02]pretty fast.[2687.92] [2687.92][S02]But if you've got the simulation to be fully like the real world,[2690.96] [2690.96][S02]and it had the kind of objectives and constraints[2693.3] [2693.3][S02]that lead to these kind of evolutionary steps,[2696.78] [2696.78][S02]then it's definitely plausible.[2699.04] [2699.04][S02]But I can't say it's definitely there yet.[2701.38] [2701.38][S03] It's not a direct answer,[2703.13] [2703.13][S03]but I actually tried.[2705.39] [2705.39][S03]I think there is maybe a very basic example of Game of Life.[2711.327] [2711.327][S01] John Conway's.[2712.41] [2712.41][S03] Yeah.[2713.327] [2713.327][S03]It has four rules, and then I actually[2715.19] [2715.19][S03]tried using Veo to simulate it.[2717.97] [2717.97][S03]You give it an image and it doesn't work.[2720.71] [2720.71][S03]Like, it does look like it's evolving.[2723.35] [2723.35][S03]And if you don't know the rules, you[2725.07] [2725.07][S03]would look like, yeah, it looks reasonable.[2727.323] [2727.323][S03]Different pixels light up, but it doesn't follow[2731.83] [2731.83][S03]the four rules of the game.[2733.89] [2733.89][S03]I think this is a good example for what our current models are[2738.47] [2738.47][S03]able to do, and what they are less-- still[2741.27] [2741.27][S03]limited in their ability to follow specific rules.[2744.35] [2744.35][S03]To actually evolve maybe life forms,[2747.01] [2747.01][S03]I think you need much more ability to do both,[2750.63] [2750.63][S03]to also simulate the physical world,[2752.41] [2752.41][S03]but also follow some basic rules of physics[2756.43] [2756.43][S03]in a very accurate way.[2757.803] [2757.803][S01] Constrained way.[2758.97] [2758.97][S03] Yes, and constrained way.[2759.95] [2759.95][S03]And I think we're not-- like, we see some glimpse of it,[2762.29] [2762.29][S03]but it's definitely very far from being able to have[2765.67] [2765.67][S03]evolution on a GPU.[2767.092] [2767.092][S01] Well, thank you for going philosophical with me[2769.55] [2769.55][S01]for a moment there.[2770.85] [2770.85][S01]I enjoyed that.[2771.77] [2771.77][S01]Let's come back down to Earth with a thump,[2773.97] [2773.97][S01]though, because, I mean, there are safety implications[2776.79] [2776.79][S01]with this.[2777.29] [2777.29][S01]What are your main concerns?[2778.572] [2778.572][S02] I think there's[2780.03] [2780.03][S02]different levels of concern.[2781.61] [2781.61][S02]There's the known things, and they're quite obvious.[2786.567] [2786.567][S02]I mean, things like violence, maybe we[2788.15] [2788.15][S02]wouldn't want to occur in the world in new ways.[2792.95] [2792.95][S02]And that's something that we can already start addressing.[2796.51] [2796.51][S02]But there's also maybe some more gray areas[2799.63] [2799.63][S02]where we're not actually really sure how[2802.582] [2802.582][S02]we feel about these things, like historical settings,[2804.79] [2804.79][S02]for instance.[2806.05] [2806.05][S02]Some of them may be unsavory for subtle reasons,[2810.35] [2810.35][S02]and those are just things that we as a team[2812.59] [2812.59][S02]can see quite clearly.[2814.595] [2814.595][S02]But I think there's probably also some things[2816.47] [2816.47][S02]that we haven't considered.[2817.73] [2817.73][S02]And we'd rather get those things right[2820.12] [2820.12][S02]by limiting our early access and getting feedback,[2823.857] [2823.857][S02]which we're already doing.[2824.94] [2824.94][S02]And we've already learned a lot from the folks[2826.52] [2826.52][S02]that we brought in a few weeks ago and the folks[2828.56] [2828.56][S02]that we're still interacting with.[2830.54] [2830.54][S01] Like what?[2831.72] [2831.72][S02] We found a lot of new use cases[2833.36] [2833.36][S02]that I didn't think of.[2834.46] [2834.46][S02]Vocational training could actually be quite impactful.[2837.44] [2837.44][S02]Lots of people can't go into the role[2839.8] [2839.8][S02]of things like firefighting, for instance.[2842.583] [2842.583][S02]What does it feel like to actually[2844.0] [2844.0][S02]be there without having a visceral experience?[2848.06] [2848.06][S02]It's probably something that you would benefit a lot from being[2850.92] [2850.92][S02]able to simulate in advance.[2852.34] [2852.34][S02]Even if it's not perfectly correct from a simulation[2854.56] [2854.56][S02]perspective, there may be elements to it[2857.12] [2857.12][S02]that just getting a sense of what[2859.04] [2859.04][S02]it's like to be situated in that specific circumstance,[2862.588] [2862.588][S02]that it's quite nice to be able to simulate in advance.[2864.88] [2864.88][S03] Minus the heat.[2865.98] [2865.98][S02] Yeah.[2867.022] [2867.022][S02]Minus the heat and the smoke.[2868.68] [2868.68][S01] And the genuine jeopardy.[2870.26] [2870.26][S02] Yes.[2871.28] [2871.28][S01] But actually, you raised another interesting point[2873.863] [2873.863][S01]there though, because you said even if it's not[2876.04] [2876.04][S01]perfectly realistic.[2877.76] [2877.76][S01]Is there also another danger about[2879.52] [2879.52][S01]this gap between what's simulated and what's real?[2883.04] [2883.04][S01]How do you make that as small as possible?[2885.22] [2885.22][S01]This is something-- the sim to real gap[2886.36] [2886.36][S01]is something we spoke about in this podcast lots of times[2888.24] [2888.24][S01]before.[2888.86] [2888.86][S01]Let's say you've got your example[2890.235] [2890.235][S01]where you've got a robot in a factory and it's moving around.[2893.052] [2893.052][S03] In terms of reliability.[2894.76] [2894.76][S01] It's interacting.[2895.66] [2895.66][S01]You can't directly take that and map it into the real world,[2898.22] [2898.22][S01]right?[2899.12] [2899.12][S03] Yeah.[2900.037] [2900.037][S03]So I think over time, we'll see more control,[2902.96] [2902.96][S03]basically being able to take maybe a real environment,[2906.04] [2906.04][S03]map it into the model so the model can base its generation[2912.0] [2912.0][S03]on the real environment.[2913.18] [2913.18][S03]And we see it to an extent, whether we're[2915.36] [2915.36][S03]starting from an image or starting from a video.[2918.2] [2918.2][S03]And now the question is, would it be perfectly the same[2922.0] [2922.0][S03]as the real world?[2922.82] [2922.82][S03]Probably not.[2923.54] [2923.54][S03]I don't think it's even well defined what does it mean.[2926.0] [2926.0][S03]But I think the gap is definitely narrowing.[2929.3] [2929.3][S03]So we can take environments.[2930.72] [2930.72][S03]In the past, we know a lot of the RL environments[2933.56] [2933.56][S03]look very far from anything that's[2936.17] [2936.17][S03]photorealistic or real world, but now we can go even closer.[2940.31] [2940.31][S03]But definitely there still remains a gap,[2942.27] [2942.27][S03]and we will have to see what are the implications.[2945.03] [2945.03][S03]And we're still not using it for any real world[2947.33] [2947.33][S03]deployment, of course.[2948.51] [2948.51][S03]Yeah.[2949.09] [2949.09][S02] Yeah, I think it's[2949.81] [2949.81][S02]kind of an iterative approach.[2951.06] [2951.06][S02]I don't think we're saying, right, now, we've got Genie 3,[2953.59] [2953.59][S02]we've solved simulation for any possible embodied task.[2956.69] [2956.69][S02]But I think what we can do is combine it[2959.65] [2959.65][S02]with other techniques.[2961.17] [2961.17][S02]So we still would train our agents in the same way[2963.81] [2963.81][S02]that we already did without this,[2965.27] [2965.27][S02]and we'd use it to augment the training process.[2968.37] [2968.37][S02]And then another element to it is[2970.81] [2970.81][S02]it's really important that it has some diversity.[2973.85] [2973.85][S02]So if it's always wrong in the same way,[2976.29] [2976.29][S02]then agents might learn to exploit that inaccuracy.[2979.395] [2979.395][S02]Whereas if the model can generate[2980.77] [2980.77][S02]quite diverse different worlds, then what we can do[2984.77] [2984.77][S02]is really test the breadth of agents' capabilities[2987.45] [2987.45][S02]and make sure that there's no scenario where[2989.33] [2989.33][S02]it does something really wrong.[2990.813] [2990.813][S02]That might actually be a strength.[2992.23] [2992.23][S02]So same in sim to real.[2994.31] [2994.31][S02]We want to do domain randomization.[2996.07] [2996.07][S02]Maybe by having a model that is a generative model, being[2999.25] [2999.25][S02]able to search the space of possibilities[3001.93] [3001.93][S02]and check that all of them, the agents[3004.29] [3004.29][S02]do something sensible might be a good thing.[3007.09] [3007.09][S02]But that doesn't mean you want to completely train it[3009.423] [3009.423][S02]that that is the real world.[3010.59] [3010.59][S02]Maybe you want to use it make it more adversarially robust[3014.41] [3014.41][S02]rather than learn specifics.[3016.582] [3016.582][S01] Let me make sure I understand that, then.[3018.79] [3018.79][S01]So if it's wrong but wrong in unpredictable ways,[3021.75] [3021.75][S01]then actually that might end up making the agent more[3024.25] [3024.25][S01]robust in the long run.[3025.85] [3025.85][S02] Yeah.[3026.17] [3026.17][S02]So what you want to do-- it's similar to domain[3028.13] [3028.13][S02]randomization-- is you want to make sure[3029.797] [3029.797][S02]that there's no plausible scenario where the agent could[3032.41] [3032.41][S02]do something really unsafe.[3034.65] [3034.65][S02]It's quite a different objective than if you[3036.73] [3036.73][S02]had one specifically incorrect scenario and then[3039.53] [3039.53][S02]you told the agent exactly how to behave from that one.[3042.81] [3042.81][S02]Instead, it's make it so that in any possible future world,[3046.71] [3046.71][S02]the agent should be able to do something sensible.[3048.93] [3048.93][S01] That's so interesting, then,[3050.597] [3050.597][S01]because I was sort of imagining that you were trying[3053.22] [3053.22][S01]to nudge this towards it being more realistic to more[3057.62] [3057.62][S01]towards getting more reliable outcomes towards closing[3060.94] [3060.94][S01]that gap of sim to real.[3062.24] [3062.24][S01]But I mean, the way you're describing[3063.782] [3063.782][S01]this is that not necessarily.[3066.212] [3066.212][S03] I think the question is,[3067.92] [3067.92][S03]what do you mean by reliable?[3069.52] [3069.52][S03]Because reliable, to me, is that is probably[3073.5] [3073.5][S03]a lot about following the instructions that we[3075.9] [3075.9][S03]provide the model.[3077.2] [3077.2][S03]So if we want the model to simulate a specific environments[3081.06] [3081.06][S03]and we describe it with a lot of detail,[3083.385] [3083.385][S03]we want the model to follow that.[3084.76] [3084.76][S03]If there is something not so plausible in this description,[3090.0] [3090.0][S03]the models should still follow that.[3091.7] [3091.7][S03]I think the challenge is that sometimes we[3094.94] [3094.94][S03]as people, or for various reasons,[3097.04] [3097.04][S03]we're interested in the less plausible scenarios.[3100.1] [3100.1][S03]We're just looking at the tree in the middle of-- for example,[3103.64] [3103.64][S03]in some of the examples we saw today,[3106.18] [3106.18][S03]you want some vodka and the calf liver.[3109.42] [3109.42][S03]All right?[3109.92] [3109.92][S03]So that's not a very--[3111.46] [3111.46][S03]it's not if you just sample from all[3113.3] [3113.3][S03]of the possible tables in Siberia, probably[3115.98] [3115.98][S03]that's not in the middle of the distribution.[3118.68] [3118.68][S03]So I think that the reliability to me[3121.3] [3121.3][S03]comes mostly from following the description that we provide[3124.86] [3124.86][S03]the model with and simulating the world in a way that[3127.58] [3127.58][S03]would be close to that.[3128.76] [3128.76][S02] I think that's a really good point,[3129.72] [3129.72][S02]actually.[3130.22] [3130.22][S02]I think it's in underspecified environments you want diversity,[3133.02] [3133.02][S02]because you want to be able to adapt[3134.78] [3134.78][S02]to anything within the plausible distribution.[3137.437] [3137.437][S02]But if you have a very well specified environment,[3139.52] [3139.52][S02]then you want it to be accurate.[3141.867] [3141.867][S02]And I think we're kind of seeing improvement[3143.7] [3143.7][S02]on both those dimensions, but we're probably not fully[3146.54] [3146.54][S02]there yet.[3147.638] [3147.638][S01] Let me go back to that AGI question--[3149.68] [3152.196][S01]if I may.[3153.86] [3153.86][S01]The end question that everyone always wants to ask.[3156.34] [3156.34][S01]Do you think that this is a step towards it, Jack?[3159.56] [3159.56][S02] I think AGI is something itself, which[3161.98] [3161.98][S02]is relatively subjective and people[3164.82] [3164.82][S02]have different interpretations of what you mean by AGI.[3167.9] [3167.9][S02]So I think it would be quite maybe grandiose to say[3171.83] [3171.83][S02]our model is the key thing in the whole field that[3174.39] [3174.39][S02]will enable AGI.[3175.75] [3175.75][S02]But I think for me, an AGI needs to be embodied[3178.63] [3178.63][S02]and be able to act in the physical world.[3180.64] [3180.64][S02]That's what really excites me.[3181.89] [3181.89][S02]I think that could really improve people's quality of life[3185.83] [3185.83][S02]in any demographic anywhere in the world.[3189.09] [3189.09][S02]And so with that framing, I definitely[3191.35] [3191.35][S02]think this is an important tool.[3192.83] [3192.83][S02]I can't see how an embodied AGI or AGI that is embodied[3196.71] [3196.71][S02]would be able to operate any scenario in the world[3199.75] [3199.75][S02]without being able to simulate it,[3201.33] [3201.33][S02]to gather experience and learn from its own experience.[3203.97] [3203.97][S02]Because that's the paradigm that we've used in other settings[3206.55] [3206.55][S02]to get superhuman capabilities or even[3208.83] [3208.83][S02]just robust capabilities.[3210.79] [3210.79][S02]And so I believe very strongly that we need simulation,[3214.47] [3214.47][S02]and I also believe very strongly that we[3216.27] [3216.27][S02]won't be able to build a simulator of the real world[3218.87] [3218.87][S02]any other way.[3219.908] [3219.908][S02]So when you combine those two things,[3221.45] [3221.45][S02]I think, yes, it is a big step for my version of AGI.[3224.993] [3224.993][S03] Yeah.[3225.91] [3225.91][S03]And I think it's a really good answer.[3228.19] [3228.19][S03]On top of that, I would just say that our current generation[3231.23] [3231.23][S03]of AI is limited to connect the digital world.[3236.67] [3236.67][S03]For AI to be useful for us, definitely it[3238.63] [3238.63][S03]has to have some kind of real world interaction.[3242.11] [3242.11][S03]So I think, again, it's a small step towards that,[3245.35] [3245.35][S03]towards that embodied AI.[3248.27] [3248.27][S03]And there are definitely a lot of gaps to get there,[3251.15] [3251.15][S03]so I think we need much better signals for, for example, robots[3255.31] [3255.31][S03]get while they walk through the world.[3257.03] [3257.03][S03]They need to get some physical response.[3259.61] [3259.61][S03]It's not enough just to have a visual input and output.[3263.91] [3263.91][S03]So I just think that this is definitely[3265.83] [3265.83][S03]a step towards that vision, yeah.[3267.21] [3267.21][S01] Because there is still a lot that this can't do.[3269.01] [3269.01][S01]I mean, the additional sensors being one of them.[3271.73] [3271.73][S01]But also doesn't handle people that well at the moment, Jack?[3274.272] [3274.272][S02] Exactly, and I[3275.688] [3275.688][S02]think that's really the key thing is I both think[3277.87] [3277.87][S02]this is one of the most promising technologies[3280.31] [3280.31][S02]to achieve sociable and socially aware robots,[3284.11] [3284.11][S02]and embodied agents, but also, I think[3285.95] [3285.95][S02]that's the biggest limitation, probably,[3288.28] [3288.28][S02]of the current iteration of the model[3290.32] [3290.32][S02]is that it doesn't do this perfectly.[3291.88] [3291.88][S02]Right?[3292.38] [3292.38][S02]Because our standards have raised,[3293.78] [3293.78][S02]and that's the thing that we now think[3295.12] [3295.12][S02]is something that isn't good enough.[3296.72] [3296.72][S02]But I think it's really critical that we do have that, right?[3299.82] [3299.82][S02]Because even if our robots and embodied agents[3302.56] [3302.56][S02]fully understand physics, I think[3304.52] [3304.52][S02]physics are fairly consistent around the world[3306.7] [3306.7][S02]but people are not, right?[3308.26] [3308.26][S02]And we want these agents, robots,[3311.24] [3311.24][S02]whatever form factor they come in,[3312.76] [3312.76][S02]to be able to really augment humans[3314.52] [3314.52][S02]and work with humans to make our quality of life better.[3317.8] [3317.8][S02]And so they need to understand how humans think,[3320.54] [3320.54][S02]work, interact, and be able to work with us on things.[3324.575] [3324.575][S02]And so that's what I think we're really[3326.2] [3326.2][S02]excited about, as one of the things that[3327.867] [3327.867][S02]might be enabled by our model.[3329.56] [3329.56][S03] Yeah, I think there are definitely[3331.685] [3331.685][S03]a lot of limitations that are in terms[3333.64] [3333.64][S03]of the quality of the generation.[3336.12] [3336.12][S03]But what I mostly--[3338.16] [3338.16][S03]I'm very excited about this, the pace.[3340.44] [3340.44][S03]If you think about it, we had Genie 2, Veo 2 in December,[3344.88] [3344.88][S03]and we definitely feel the pace, the impact on it[3348.8] [3348.8][S03]in our personal lives.[3350.24] [3350.24][S03]But the field is just moving fast.[3352.42] [3352.42][S03]And if we remember just, what, less than two years ago, we[3355.64] [3355.64][S03]had images generated with six fingers[3359.6] [3359.6][S03]and that was a big thing, and nobody[3361.16] [3361.16][S03]is speaking about that anymore.[3362.84] [3362.84][S03]So I don't see why we won't be able to generate people[3365.8] [3365.8][S03]in much higher fidelity and with everything that follows.[3369.662] [3369.662][S01] Is the goal here, then,[3371.12] [3371.12][S01]to have a foundational model?[3372.68] [3372.68][S01]To essentially do for simulated worlds what LLMs[3376.64] [3376.64][S01]have done for language?[3377.843] [3377.843][S02] Yeah, exactly.[3379.26] [3379.26][S02]I think you put it better than probably than I could.[3381.96] [3381.96][S02]I think this is really a step change as a foundation[3384.52] [3384.52][S02]model in terms of the breadth and generality and capabilities.[3387.94] [3387.94][S02]And I think this is probably similar to what Shlomi alluded[3390.56] [3390.56][S02]to as we've seen with images recently,[3392.38] [3392.38][S02]where there were things obviously[3393.84] [3393.84][S02]like the fingers to them now being, I mean,[3396.62] [3396.62][S02]pretty incredible.[3397.56] [3397.56][S02]We saw the same thing with video maybe in the past year[3400.04] [3400.04][S02]where, once we had something like Veo 2,[3402.14] [3402.14][S02]it's looking pretty amazing at this point.[3405.35] [3405.35][S02]And we saw this with language models[3406.85] [3406.85][S02]maybe three or four years ago, where[3408.53] [3408.53][S02]they started to get really capable,[3410.29] [3410.29][S02]and we wanted to get to that point[3411.73] [3411.73][S02]for this new kind of foundation model, sort[3413.97] [3413.97][S02]of an autoregressive world model.[3415.97] [3415.97][S02]Now we're there, there's a whole host[3417.73] [3417.73][S02]of different potential things that could be used for[3420.37] [3420.37][S02]and have impact on, and we're still fairly early in that right[3423.45] [3423.45][S02]now.[3423.95] [3423.95][S01] But are there elements of simulation too?[3426.62] [3426.62][S01]Do you think that you will ever be able to use this kind of idea[3431.17] [3431.17][S01]to recreate a lived experience, rather than just a visual one?[3435.17] [3435.17][S03] So we have many, many senses[3437.25] [3437.25][S03]that we're not even aware of.[3438.93] [3438.93][S03]So, for example, a proprioceptive perception[3443.13] [3443.13][S03]that we are basically we feel where[3445.29] [3445.29][S03]we are and we have this kind of notion[3447.49] [3447.49][S03]of where we are in the world.[3449.89] [3449.89][S03]When we think about actually putting people in a simulation[3453.33] [3453.33][S03]to really feel kind of immersed in it,[3455.75] [3455.75][S03]I think this is a huge part of it,[3457.43] [3457.43][S03]and basically the constraint to visual and maybe audio[3460.09] [3460.09][S03]is still too much of a constraint.[3463.09] [3463.09][S03]I think that there is definitely a potential for that,[3466.51] [3466.51][S03]but it goes through multiple technologies[3469.61] [3469.61][S03]that we have to build to actually get there.[3472.87] [3472.87][S03]So before that, I expect people to be[3475.17] [3475.17][S03]able to interact with photorealistic environments,[3479.39] [3479.39][S03]but still through some kind of an interface that will[3482.81] [3482.81][S03]be like a hybrid interface.[3485.57] [3485.57][S03]Like maybe they can feel some sensation[3487.65] [3487.65][S03]for maybe gloves or something.[3489.303] [3489.303][S02] I think I would also[3490.97] [3490.97][S02]say that there is definitely something[3492.553] [3492.553][S02]about being interactive in real time that[3494.69] [3494.69][S02]does make a big difference for the experience.[3496.89] [3496.89][S02]We've had members of the team say that they visited childhood[3500.01] [3500.01][S02]locations, for example, and did actually get a sense for it[3503.69] [3503.69][S02]that you couldn't really get from an image or a video.[3506.21] [3506.21][S02]So there is already some sort of degree of experience[3508.93] [3508.93][S02]that you can gain from this kind of model already.[3512.17] [3512.17][S02]And obviously, we're working hard[3513.69] [3513.69][S02]to make it an even more capable model in the future,[3517.27] [3517.27][S02]so maybe that will extend.[3518.733] [3518.733][S03] Yeah.[3519.65] [3519.65][S03]Actually, it reminds me of a project[3521.15] [3521.15][S03]that we had earlier in the year that the team at Google[3527.5] [3527.5][S03]used Veo, actually, to help people[3530.94] [3530.94][S03]with early onset of dementia to go back to their childhood[3534.78] [3534.78][S03]memories and reconstruct them.[3537.2] [3537.2][S03]So I can imagine that that might be,[3539.32] [3539.32][S03]for example, a potentially therapeutic tool[3542.82] [3542.82][S03]as well, that it can not only look at the video,[3545.64] [3545.64][S03]but maybe actually relive or remember some things[3549.82] [3549.82][S03]from their childhood.[3550.92] [3550.92][S03]So I think that even before, we don't[3552.78] [3552.78][S03]need to go very far for things to have[3555.34] [3555.34][S03]positive impact on the world.[3557.02] [3557.02][S01] Amazing.[3558.1] [3558.1][S01]That was absolutely fascinating.[3559.52] [3559.52][S01]Thank you so much.[3560.44] [3560.44][S02] Thanks for having us.[3561.28] [3561.28][S03] Thanks for having us, Hannah.[3562.58] [3562.58][S01] I think the most impressive part of this[3564.747] [3564.747][S01]is not what you're looking at on the screen.[3567.04] [3567.04][S01]It's how that is generated.[3569.64] [3569.64][S01]It's this change that this model represents,[3572.18] [3572.18][S01]from creating realistic images or videos of the real world,[3575.54] [3575.54][S01]as though they were frozen moments in time,[3579.34] [3579.34][S01]into something that can actually handle time in the way[3582.5] [3582.5][S01]that we experience it, with an arrow that's[3584.82] [3584.82][S01]pointing in only one direction, where effect follows cause[3588.94] [3588.94][S01]to build this consistent, forward moving world[3593.42] [3593.42][S01]where the present is a direct result of the past.[3597.5] [3597.5][S01]And that is why I think that this[3599.34] [3599.34][S01]is an early hint of something that's much bigger.[3602.48] [3602.48][S01]This is not just a new way to design games[3604.5] [3604.5][S01]or beautiful environments.[3606.26] [3606.26][S01]This here is the bedrock for machines[3609.3] [3609.3][S01]that can genuinely plan and reason about our world.[3614.905] [3614.905][S01]You have been listening to \"Google DeepMind--[3616.78] [3616.78][S01]The Podcast\" with me, Professor Hannah Fry.[3618.973] [3618.973][S01]Now we're going to take a little bit of a pause over the summer,[3621.64] [3621.64][S01]but we're going to be back with more episodes in the autumn[3624.38] [3624.38][S01]from Google HQ in California.[3626.86] [3626.86][S01]And in the meantime, do take a look[3628.34] [3628.34][S01]at our extensive back catalog, which covers everything[3631.22] [3631.22][S01]from tools for creators to AI for drug discovery.[3634.68] [3634.68][S01]See you soon.[3636.45]"}
{"file_name": "audio/val_000019.wav", "transcription": "[0.0][S04] Hello, and welcome to the Google DeepMind Podcast[2.6] [2.6][S04]with me, Professor Hannah Fry.[3.99] [3.99][S04]Now, those of you who are watching[5.42] [5.42][S04]will notice that we are not in the studio today,[7.68] [7.68][S04]and that is because we are backstage at the AI[10.49] [10.49][S04]for Science Forum, a very special event that[12.98] [12.98][S04]has been co-hosted between the Royal Society and Google[16.37] [16.37][S04]DeepMind.[17.07] [17.07][S04]And as part of the day, I had the opportunity[19.58] [19.58][S04]to speak to Demis Hassabis on stage, to interview him,[22.97] [22.97][S04]and we thought that you might like to hear that conversation.[26.1] [26.1][S04]So we have included it here as a special episode.[29.01] [29.01][S04]And then even more special than that,[31.29] [31.29][S04]as though one Nobel laureate wasn't enough,[34.47] [34.47][S04]we have three further Nobel laureates[37.31] [37.31][S04]joining me on stage for an extended panel for you to enjoy.[41.19] [41.19][MUSIC PLAYING][42.16] [47.99][S04]OK, now for the next session, it is always[52.25] [52.25][S04]an absolute treat for me and often a mind boggling[54.98] [54.98][S04]experience to get to interview your co-host for today,[60.3] [60.3][S04]sir Demis Hassabis.[61.81] [61.81][S04]Now, Demis is a computer scientist.[63.79] [63.79][S04]He's an Artificial Intelligence researcher.[65.74] [65.74][S04]He's an entrepreneur who co-founded DeepMind[68.13] [68.13][S04]in 2010, where he is still CEO.[72.03] [72.03][S04]Now, I don't know how much you know about Demis' background.[74.53] [74.53][S04]Because as well as the Nobel Prize, and the knighthood,[76.74] [76.74][S04]and being a fellow of the Royal Society, and AlphaFold,[79.11] [79.11][S04]and AlphaGo, and having 150,000 citations,[82.23] [82.23][S04]he also had a previous life as a very successful neuroscientist,[86.32] [86.32][S04]chess champion, and video games designer.[89.04] [89.04][S04]I have to say, I have a theory that, actually, Demis[91.98] [91.98][S04]worked out the secrets to AGI a little while ago[94.92] [94.92][S04]and has been hiding it in his basement and then just, like,[98.37] [98.37][S04]slowly eking out all of these massive breakthroughs[100.74] [100.74][S04]one-by-one to the rest of us.[102.057] [102.057][S04]But one thing is for sure, our conversations[103.89] [103.89][S04]are always genuinely fascinating.[105.55] [105.55][S04]So please, welcome Demis Hassabis.[106.98] [106.98][S04]Thank you.[107.65] [107.65][APPLAUSE][109.111] [111.06][S04]Thank you, Demis.[111.9] [111.9][S04]Thank you so much.[113.64] [113.64][S04]OK, I heard that you didn't know that the Nobel Prize was coming.[117.71] [117.71][S04]Is that true?[118.63] [118.63][S03] Yes, no--[119.23] [119.23][S04] Really, really?[119.98] [119.98][S03] Yeah.[120.255] [120.255][S04] How did you find out?[121.63] [121.63][S03] Well, it was quite a funny story, actually,[124.088] [124.088][S03]because we had heard some rumor that our fold had[126.43] [126.43][S03]been sort of nominated, but you never[129.19] [129.19][S03]expect something like that.[130.4] [130.4][S03]And in the morning, I was just getting on with my normal work.[132.8] [132.8][S03]Actually, my wife was working at home, as well,[134.758] [134.758][S03]and it got to about 10:30.[136.1] [136.1][S03]And we thought, oh, obviously, it's[137.56] [137.56][S03]not happened this year, because we hadn't heard.[139.56] [139.56][S03]And then, suddenly, my wife's computer[141.79] [141.79][S03]started ringing on her Skype, I think it was.[144.56] [144.56][S03]And I said, what's that annoying noise?[146.43] [146.43][S03]And then it turned out to be a call from Sweden,[148.43] [148.43][S03]and they were desperately trying to get hold of my number,[151.04] [151.04][S03]and they didn't have John's number either.[152.79] [152.79][S03]So everyone was in a kind of panic,[154.25] [154.25][S03]but it all added to the drama of the moment.[156.55] [156.55][S03]Just 10 minutes before--[157.78] [157.78][S04] You had john's number?[158.21] [158.21][S03] Yes, exactly.[158.81] [158.81][S03]Just before it was all announced.[159.825] [159.825][S04] Absolutely amazing, and I also[161.71] [161.71][S04]saw that you celebrated by having a poker[164.11] [164.11][S04]night with some chess champions, including Magnus[167.38] [167.38][S04]Carlsen and Hikaru, as well.[169.31] [169.31][S04]I want to know, who won?[170.635] [170.635][S04]What's your bluffing strategy?[171.885] [171.885][LAUGHING][172.588] [172.588][S03] Well, it turned out-- yeah, so that happened.[175.13] [175.13][S03]I think it was on the Wednesday, and it turned out[177.213] [177.213][S03]that there was a big, huge chess tournament going on in London.[182.77] [182.77][S03]And one of my chess friends, old chess friends[184.78] [184.78][S03]from when I was young, was hosting a poker and chess[187.39] [187.39][S03]evening the day after.[189.29] [189.29][S03]So I thought it was the perfect way to celebrate,[191.63] [191.63][S03]but I don't recommend a home poker game[195.94] [195.94][S03]with a couple of world poker champions[197.56] [197.56][S03]and a couple of ex-world chess champions[199.67] [199.67][S03]if you want to win some money.[201.95] [201.95][S03]But that's my idea of fun, and it was actually[204.902] [204.902][S03]an amazing night.[205.61] [205.61][S03]I know it sounds pretty nerdy, but it was actually my idea[209.83] [209.83][S03]of heaven to celebrate it.[211.64] [211.64][S04] Not at all.[211.81] [211.81][S04]It sounds like--[212.33] [212.33][S03] Yeah, I can't tell you the bluffing strategy.[213.7] [213.7][S03]Otherwise, I'll never beat Magnus again.[215.025] [215.025][LAUGHING][215.525] [215.525][S03]Yeah.[216.048] [216.048][S04] Well, this is true.[217.34] [217.34][S04]This is true.[218.59] [218.59][S04]I mean, aside from the Nobel Prize,[220.81] [220.81][S04]a sentence that no one has ever said, before that,[225.71] [225.71][S04]in advance of that, you also were named as a citation[229.21] [229.21][S04]laureates, because the AlphaFold work has now been cited over[232.72] [232.72][S04]28,000 times, which is incredible.[235.11] [235.11][S04]I mean, this is really not very long ago that it came through.[238.76] [238.76][S04]Are there standout applications for you that really resonate?[242.473] [242.473][S03] Well, I mean, it is mind boggling, really,[244.89] [244.89][S03]and it's everything we hoped would happen[247.34] [247.34][S03]by putting AlphaFold out there, and open sourcing[249.77] [249.77][S03]it, and putting it out with the community,[251.52] [251.52][S03]with our fantastic partners at EMBL-EBI, like Janet, and Ewan,[255.56] [255.56][S03]and the audience as well, to get it out there to everyone.[259.73] [259.73][S03]I mean, there's so many applications to mention,[263.43] [263.43][S03]but I think maybe I can mention two or three of my favorites[266.54] [266.54][S03]were kind of determining the structure of the nuclear pore[271.31] [271.31][S03]complex, one of the most important and largest proteins[274.04] [274.04][S03]in the human body.[275.228] [275.228][S03]And it's really important, because it[276.77] [276.77][S03]governs the molecules and nutrients that go in[279.17] [279.17][S03]and out of your cell nucleus.[280.74] [280.74][S03]And, actually, several teams used AlphaFold predictions,[284.51] [284.51][S03]as well as their own experimental data,[286.61] [286.61][S03]to finally piece that very, very complex protein[291.08] [291.08][S03]structure together.[292.48] [292.48][S03]So that was amazing.[293.89] [293.89][S03]Another piece of work that I really like[295.74] [295.74][S03]is [INAUDIBLE] work from the [? Broad, ?][298.32] [298.32][S03]developing a molecular syringe to deliver drug payloads[301.8] [301.8][S03]to hard parts of the body to reach,[304.2] [304.2][S03]and he also used AlphaFold to help[305.88] [305.88][S03]modify how that molecular syringe was designed.[309.58] [309.58][S03]And then maybe, finally, my other favorite project[311.94] [311.94][S03]is from the University of Portsmouth, John McKeon's group,[316.77] [316.77][S03]designing plastic eating enzymes using AlphaFold.[321.91] [321.91][S03]But I think it's just a small sample[324.21] [324.21][S03]of all the incredible work researchers are doing with it.[326.77] [326.77][S04] But these kind of projects, I[327.74] [327.74][S04]mean, these are your favorites, right?[329.323] [329.323][S04]The ones where it's not just sort of the technical idea[332.1] [332.1][S04]itself, it's the potential impact[333.54] [333.54][S04]that it can have further downstream?[334.55] [334.55][S03] Yeah, that's right.[336.008] [336.008][S03]I mean, that's why I was always intrigued by the protein folding[341.28] [341.28][S03]problem or protein structure prediction problem was because I[345.33] [345.33][S03]felt-- we sometimes call it a deepmind, a root node problem.[348.468] [348.468][S03]What we mean by that is, if you think[350.01] [350.01][S03]of the whole tree of knowledge, there[351.88] [351.88][S03]are certain problems, where if there are root node problems,[354.74] [354.74][S03]if you unlock them, if you discover a solution to them,[357.05] [357.05][S03]it would unlock a whole new branch or avenue of discovery.[361.01] [361.01][S03]And I always felt that determining a protein structure[364.57] [364.57][S03]in this way would do that, lead to disease understanding,[368.23] [368.23][S03]drug design, and much more, and that[370.383] [370.383][S03]seems to be what's happened.[371.55] [371.55][S04] OK, so not to be greedy, but what's next?[375.43] [375.43][S04]Is there another AlphaFold?[376.67] [376.67][S04]I mean, GNoME is a personal favorite of mine.[378.77] [378.77][S04]Maybe tell us a little bit about that project.[379.94] [379.94][S03] Yes, so GNoME, look, there's so many areas,[382.19] [382.19][S03]and actually, James mentioned it in the morning,[383.86] [383.86][S03]and Pushmeet talked about it as well,[385.402] [385.402][S03]that we're touching on nearly every area of science.[389.38] [389.38][S03]And GNoME is one of my favorite projects, which[391.42] [391.42][S03]is on material design, and I think[393.49] [393.49][S03]material design has some of the same characteristics that we[396.97] [396.97][S03]look for in a problem that's suitable for AI.[398.99] [398.99][S03]It's massive combinatorial space.[401.38] [401.38][S03]You need to try and build a model that[403.06] [403.06][S03]understands the physics and the chemistry[405.16] [405.16][S03]of the natural phenomena.[408.26] [408.26][S03]And then, if you have a model like that,[409.97] [409.97][S03]you can then maybe use it to do a very efficient search[413.27] [413.27][S03]through that search space, combinatorial space,[416.07] [416.07][S03]and then find an optimal solution.[417.81] [417.81][S03]And in materials, I think that would[419.63] [419.63][S03]be also just as groundbreaking.[421.17] [421.17][S03]You could imagine designing new batteries or maybe, one day,[424.208] [424.208][S03]discovering a room temperature superconductor.[426.125] [426.125][S03]It has always been one of my dreams.[428.99] [428.99][S03]So I think we're, obviously, at the early stages of that.[432.42] [432.42][S03]I would characterize it as sort of AlphaFold one level maybe,[435.865] [435.865][S03]and we've got to get to AlphaFold[437.24] [437.24][S03]two level of predictions.[439.89] [439.89][S03]But we can see a clear path from there,[441.66] [441.66][S03]and GNoME is the beginning of that work.[443.49] [443.49][S03]We published it in--[444.57] [444.57][S03]I think it was in Science, last year,[446.52] [446.52][S03]and we discovered 200,000 new crystals that no one had ever[451.55] [451.55][S03]seen before.[452.43] [452.43][S03]So I think that shows some of the potential of AI in things,[456.95] [456.95][S03]like material design.[457.86] [457.86][S03]And then maybe the other thing I'm excited about[459.86] [459.86][S03]is AI applied to mathematics and, perhaps,[462.98] [462.98][S03]solving one of the great conjectures, maybe[465.8] [465.8][S03]one of the Millennium prize problems,[468.11] [468.11][S03]and using AI as a big part of that solution.[474.03] [474.03][S04] So that idea, though, about GNoME,[476.12] [476.12][S04]also about AlphaFold, is you're sort of shortcutting,[479.87] [479.87][S04]or you're making the sort of synthetic world[483.8] [483.8][S04]really serve your advantage.[485.678] [485.678][S04]So you're not having to do everything[487.22] [487.22][S04]through experimentation.[488.51] [488.51][S04]I know that you and Paul Nurse have[490.04] [490.04][S04]been talking about a virtual cell for a long while.[493.697] [493.697][S04]Tell us a little bit about that.[495.03] [495.03][S03] Well, Paul has been one of my mentors.[499.07] [499.07][S03]He's very generously mentored me in the life sciences[501.5] [501.5][S03]for more than 25 years now or something,[504.09] [504.09][S03]and it's been incredible, sort of inspiring,[507.45] [507.45][S03]talking to Paul very regularly about these topics.[510.44] [510.44][S03]And, unusually, I think Paul for a life scientist, a biologist,[514.47] [514.47][S03]he's thought a lot about biology as an information system.[517.02] [517.02][S03]He's written a lot of interesting research papers[519.62] [519.62][S03]on that.[520.23] [520.23][S03]So we've always had this discussion in the background,[523.74] [523.74][S03]and I've always wondered-- every five years or so, I've thought,[526.73] [526.73][S03]do we have enough technologies, enough[528.7] [528.7][S03]know how to actually really attempt this Mount[532.565] [532.565][S03]Everest of a problem or trying to build[534.19] [534.19][S03]a virtual cell, a simulation of a cell, basically,[536.57] [536.57][S03]that would be predictive of something that's[538.96] [538.96][S03]really going to happen?[541.09] [541.09][S03]And every five years, I thought, we don't really[543.22] [543.22][S03]have enough technologies yet.[544.7] [544.7][S03]But, finally, now, I think the answer is, yes, we probably[548.35] [548.35][S03]do have enough know how and techniques that we could attempt[552.61] [552.61][S03]this seriously now.[553.94] [553.94][S03]And maybe, in the next five to 10 years,[555.92] [555.92][S03]we could eventually build up a picture[558.52] [558.52][S03]of a virtual cell, perhaps, a yeast cell to start off with,[561.62] [561.62][S03]as Paul's always been working on as the model organism for this.[567.34] [567.34][S03]And my way of thinking about it is you[569.32] [569.32][S03]can think of AlphaFold2 as essentially[571.69] [571.69][S03]solving the static picture of what a protein looks like.[574.52] [574.52][S03]But of course, we know biology is a dynamic system.[576.993] [576.993][S03]That's where all the interesting things happen in biology,[579.41] [579.41][S03]and AlphaFold3 is our first step towards trying[583.18] [583.18][S03]to model those interactions.[584.75] [584.75][S03]So AlphaFold3 can model pairwise interactions[587.59] [587.59][S03]between proteins and proteins, proteins and RNA, proteins,[590.09] [590.09][S03]and DNA.[591.28] [591.28][S03]And then maybe the next step up from that[593.08] [593.08][S03]would be to model a whole pathway,[594.86] [594.86][S03]and then, eventually, maybe we can make it to an entire cell.[599.26] [599.26][S04] Amazing.[600.37] [600.37][S04]I mean, I imagine that the advent of quantum computing[603.94] [603.94][S04]will make a difference now that it's possible to do simulations[606.673] [606.673][S04]down at the molecular level.[607.84] [607.84][S03] Yeah, so look, quantum computing[610.3] [610.3][S03]is very exciting.[611.23] [611.23][S03]It itself is accelerating, and James also[615.1] [615.1][S03]mentioned earlier this morning about[616.6] [616.6][S03]the interesting cross-pollination[620.62] [620.62][S03]that's happening between AI and quantum computing.[622.82] [622.82][S03]Actually, we collaborate a lot with our quantum computing[625.715] [625.715][S03]group at Google, which is one of the world's best quantum[628.09] [628.09][S03]computing teams, on things, like error correction codes[631.0] [631.0][S03]and things like that.[632.59] [632.59][S03]And of course, one of the uses of a quantum computer[635.74] [635.74][S03]would be to simulate quantum systems, like molecules,[640.33] [640.33][S03]and atoms, and things like that, and compounds,[643.14] [643.14][S03]and then produce a lot of synthetic data potentially.[646.97] [646.97][S03]But, interestingly, I also have a slightly controversial take,[650.43] [650.43][S03]which is--[651.32] [651.32][S03]and I've talked to some of the world's top quantum computer[655.22] [655.22][S03]scientists on this, is that, actually, I[657.47] [657.47][S03]believe that classical Turing machines, classical computers,[661.59] [661.59][S03]are capable of a lot more than we previously thought.[665.58] [665.58][S03]And I think that's what the work that we've been doing[668.12] [668.12][S03]has shown, both with AlphaFold, but also[670.43] [670.43][S03]previously with AlphaGo, our program that[673.28] [673.28][S03]beat the world champion at the game of Go, which[675.62] [675.62][S03]just as an indication of the complexity[677.69] [677.69][S03]Go, much more complicated than chess[679.58] [679.58][S03]and has more possible board positions[682.82] [682.82][S03]than there are atoms in the universe, 10 to the power 170.[685.71] [685.71][S03]So what that means is you can't possibly[687.56] [687.56][S03]brute force a solution to find the best[689.23] [689.23][S03]move in a particular position.[690.48] [690.48][S03]You have to do something much cleverer.[692.24] [692.24][S03]So both protein folding also has enormous numbers[695.57] [695.57][S03]of possibilities, if you were to do it naively and just try[698.51] [698.51][S03]every combination, which is the naive way of doing it.[701.65] [701.65][S03]But yeah, if you do things, like you[703.66] [703.66][S03]do a huge amount of pre-compute to build a model of the system[707.86] [707.86][S03]before you ask it the question that you're interested in,[711.19] [711.19][S03]what move should I play in this go position,[713.48] [713.48][S03]how does this novel protein fold, it turns out[716.92] [716.92][S03]you can actually come back with a near optimal move in Go[719.98] [719.98][S03]in a few seconds or fold a protein in a matter of minutes,[723.88] [723.88][S03]which you might have expected in protein space,[727.09] [727.09][S03]at least, that you might need a quantum computer for or quantum[730.21] [730.21][S03]algorithm for.[731.12] [731.12][S03]And it turns out, you don't.[732.32] [732.32][S03]So I actually think we should take this very seriously,[735.28] [735.28][S03]that classical systems, if you use them in the right way,[739.13] [739.13][S03]may be capable of modeling a lot more complex systems,[743.03] [743.03][S03]perhaps, even counterintuitively,[744.98] [744.98][S03]quantum systems.[745.7] [745.7][S03]Because, normally, you talk about needing quantum computers[748.24] [748.24][S03]to model any kind of classical system,[751.523] [751.523][S03]but it may be that classical systems can[753.19] [753.19][S03]model quantum systems.[754.49] [754.49][S03]And I've tested this out with some of the people,[756.94] [756.94][S03]like Professor Zeilinger, who won the Nobel Prize in physics[759.81] [759.81][S03]recently for his pioneering quantum computing work,[762.64] [762.64][S03]and he thought it was very interesting.[764.41] [764.41][S03]And David Deutsch, also, one of my scientific heroes,[767.97] [767.97][S03]who's basically invented quantum computing, said it was crazy,[771.97] [771.97][S03]but the right sort of crazy, which coming from him,[774.64] [774.64][S03]I take as a compliment and as also a sign[777.93] [777.93][S03]to maybe pursue it further.[779.055] [779.055][S04] So let me make sure I understand this then.[781.347] [781.347][S04]So, I guess, the classical computers[783.24] [783.24][S04]are thought of as deterministic machines,[785.53] [785.53][S04]and you're talking here about probabilistic questions.[788.305] [788.305][S04]But you can use the deterministic ones.[789.93] [789.93][S03] Yeah, and the idea[791.37] [791.37][S03]is that quantum systems are anything you try[793.74] [793.74][S03]to mold, any natural phenomena.[795.07] [795.07][S03]If you start trying to model it by every possibility[798.12] [798.12][S03]that it could take, that system, then you quickly[800.76] [800.76][S03]run out of computation in a classical system.[803.23] [803.23][S03]You need too many bits to model it,[805.29] [805.29][S03]but that's just not the way that you would attempt to do it[807.79] [807.79][S03]with a classical system.[808.79] [808.79][S03]You would build a model, first, that learns--[812.22] [812.22][S03]my conjecture on it would be that any natural phenomena[817.3] [817.3][S03]tends to have structure.[818.72] [818.72][S03]And if it has structure, you could potentially[821.17] [821.17][S03]learn it with a classical machine learning system,[824.17] [824.17][S03]learn an efficient model of that,[825.8] [825.8][S03]and then use that to search the possibilities[829.51] [829.51][S03]in an efficient way.[830.75] [830.75][S03]So I think that might get around some of the inefficiencies[834.94] [834.94][S03]from a naive way of doing this.[837.16] [837.16][S03]I mean, it's a pretty big claim, so I'm making it in a soft way.[841.31] [841.31][S03]But it's a kind of hobby of mine to look at this area,[844.782] [844.782][S03]and it's something, I think, that could be quite promising.[847.24] [847.24][S04] Another hobby of yours[848.68] [848.68][S04]that has now spun off into its own business, Isomorphic,[853.84] [853.84][S04]because, I mean, this is work using AlphaFold[856.15] [856.15][S04]to apply to drug discovery.[857.467] [857.467][S03] Yes.[858.3] [858.3][S04] You've got some pretty prestigious partnerships[861.1] [861.1][S04]now as well.[862.3] [862.3][S04]Can you tell us what Isomorphic is focusing on at the moment?[865.17] [865.17][S03] Yeah, so Isomorphic is our spin out[867.94] [867.94][S03]to try and revolutionize drug discovery from first principles[871.73] [871.73][S03]using AI from the ground up, trying to re-imagine the drug[876.74] [876.74][S03]discovery process.[878.07] [878.07][S03]And I had this in mind when we were doing AlphaFold.[880.53] [880.53][S03]So I think it's one of the most, obviously, good use cases of AI[886.46] [886.46][S03]is to cure diseases.[887.94] [887.94][S03]I mean, what better use of AI could there be?[890.72] [890.72][S03]So that has always been one of the number one things[893.48] [893.48][S03]I wanted to do with AI once it got mature enough.[896.69] [896.69][S03]So AlphaFold, of course, is a great tool[899.57] [899.57][S03]for fundamental research and fundamental biology research.[902.34] [902.34][S03]Over two million researchers around the world[904.88] [904.88][S03]have used AlphaFold and the structures we put out there,[907.67] [907.67][S03]but it's also a practical use to help with drug discovery.[911.37] [911.37][S03]Of course, knowing the structure of a protein[913.25] [913.25][S03]is only one small part of the whole drug discovery process.[917.36] [917.36][S03]So we spun out isomorphic after we did AlphaFold two[919.85] [919.85][S03]to build on that work and extend and build new machine learning[924.23] [924.23][S03]systems into adjacent areas, things,[926.87] [926.87][S03]like designing chemical compounds and drug compounds,[931.97] [931.97][S03]testing for toxicity and predicting things,[934.2] [934.2][S03]like ADME properties, important properties[936.24] [936.24][S03]that you need for drugs to work in the body[938.31] [938.31][S03]and minimize side effects and things like that.[940.57] [940.57][S03]And we're building-- you can think of it as building up[942.9] [942.9][S03]further AlphaFold-like models in these adjacent areas.[946.473] [946.473][S03]And then, eventually, we'll stick them all together,[948.64] [948.64][S03]and we hope that one day, in the near future, actually,[952.24] [952.24][S03]we'll reduce the time down from years, maybe even decade,[955.66] [955.66][S03]to design a drug down to months or perhaps even weeks.[959.46] [959.46][S03]And that would revolutionize, I think,[961.06] [961.06][S03]the drug discovery process.[962.015] [962.015][S04] This really does seem[963.21] [963.21][S04]like one of those endeavors, where you're not[965.085] [965.085][S04]having to make a choice between doing something that benefits[967.68] [967.68][S04]humanity and doing something that can be[969.81] [969.81][S04]profitable and sustain itself.[971.8] [971.8][S04]How much do you look for those kind of projects at the outset,[975.34] [975.34][S04]or how much of it is that you do sort of blue sky research[977.865] [977.865][S04]and then hope that it'll turn out that way in the end?[980.115] [980.115][S03] I think we do both.[983.49] [983.49][S03]We're research led as a group, always have been,[986.83] [986.83][S03]so we try to do what's the right next step for researching AGI,[992.69] [992.69][S03]Artificial General Intelligence, or something, like AlphaFold,[996.64] [996.64][S03]where you really want to solve the scientific challenge.[1000.04] [1000.04][S03]But then, in the back of my mind, I'm also pretty practical.[1003.49] [1003.49][S03]So I want to solve things that will also[1005.85] [1005.85][S03]have a practical, positive impact on the world[1008.46] [1008.46][S03]pretty directly.[1010.38] [1010.38][S03]So if you can find projects that do both,[1012.88] [1012.88][S03]then that's the holy grail, really,[1014.8] [1014.8][S03]and something that really worthwhile,[1016.86] [1016.86][S03]investing a lot of time and effort.[1018.54] [1018.54][S03]And AlphaFold was exactly like that.[1020.92] [1020.92][S03]So I was fascinated by the problem itself.[1023.64] [1023.64][S03]As Janet mentioned earlier, the more you get into proteins,[1027.16] [1027.16][S03]they're exquisite bionanomachines.[1030.25] [1030.25][S03]I mean, they're unbelievably beautiful.[1032.572] [1032.572][S03]I totally agree with Janet on that,[1034.03] [1034.03][S03]and you sort of fall in love with them[1035.613] [1035.613][S03]when you start working on them, and it's just unbelievable[1039.66] [1039.66][S03]what nature's engineered.[1041.55] [1041.55][S03]But then I always had in mind that would be a social good,[1047.079] [1047.079][S03]but then Isomorphic can do both, right?[1049.06] [1049.06][S03]I think we can cure many diseases with the help of AI,[1054.09] [1054.09][S03]but also, it should be an incredibly valuable company.[1057.34] [1057.34][S03]And that, of course, if that turns out to be the case,[1059.92] [1059.92][S03]that will give us more money to invest in fundamental research.[1063.85] [1063.85][S03]So it's all a virtuous cycle.[1065.68] [1065.68][S04] With all of these different projects[1067.68] [1067.68][S04]going on in fundamental research,[1068.95] [1068.95][S04]I mean, do you think that we are, at this moment, where we're[1071.492] [1071.492][S04]going to have sort of a vertical takeoff, as it were,[1074.02] [1074.02][S04]on progress?[1075.01] [1075.01][S03] I think we are on the cusp of that.[1077.41] [1077.41][S03]I really do feel like we're on the brink of a new golden era[1081.3] [1081.3][S03]of discovery, like the whole of today's symposium is called.[1086.82] [1086.82][S03]And I think what we need is a lot more[1089.61] [1089.61][S03]interdisciplinary science, so using AI, bringing[1093.0] [1093.0][S03]in the right way.[1093.92] [1093.92][S03]We're asking the right questions with domain experts,[1096.81] [1096.81][S03]and I think it's almost limitless,[1099.18] [1099.18][S03]what its applications could be.[1100.54] [1100.54][S03]And of course, AI itself is a scientific discipline,[1102.97] [1102.97][S03]is improving all the time.[1104.29] [1104.29][S03]So there's applying today's technologies directly[1108.4] [1108.4][S03]to the other fields, and then there's also[1110.98] [1110.98][S03]continuing to improve AI itself, and that as well[1114.79] [1114.79][S03]is a sort of exponential improvement.[1116.84] [1116.84][S03]So there's a lot of, I think, progress[1120.34] [1120.34][S03]to be made in just the next few years.[1122.33] [1122.33][S04] I think interdisciplinarity[1124.03] [1124.03][S04]has definitely been one of the big themes of today,[1126.41] [1126.41][S04]but I think that the other themes are about what science[1129.16] [1129.16][S04]looks like in the future.[1130.46] [1130.46][S04]What does it mean to do science in an era of AI?[1133.16] [1133.16][S04]And I know that you're a sort of big advocate[1135.34] [1135.34][S04]for the scientific method, but just talk[1138.24] [1138.24][S04]to me a little bit about that.[1139.49] [1139.49][S04]What does that look like as we kind of move forward?[1141.657] [1141.657][S03] Well, look, I think the scientific method[1144.22] [1144.22][S03]is arguably maybe the greatest idea humans have ever had[1147.83] [1147.83][S03]and, I think, underpins, obviously, all[1150.03] [1150.03][S03]of science, but then also because of that,[1152.41] [1152.41][S03]science and technology, modern civilization.[1155.36] [1155.36][S03]And I think more than ever, we need to anchor around[1159.94] [1159.94][S03]that method in today's world, and I think, especially[1164.68] [1164.68][S03]with something as powerful and potentially transformative[1167.96] [1167.96][S03]as AI, I think it's important we use more the scientific method[1172.22] [1172.22][S03]than, perhaps, lean on normal sort of technology,[1175.95] [1175.95][S03]A/B testing out in the wild.[1177.57] [1177.57][S03]That is often the case with new technologies,[1180.9] [1180.9][S03]and I feel we should treat this more as a scientific endeavor,[1184.1] [1184.1][S03]if possible, although, it obviously[1185.57] [1185.57][S03]has all the implications that breakthrough[1188.51] [1188.51][S03]technologies normally have in terms of the speed of adoption[1191.36] [1191.36][S03]and the speed of change.[1192.9] [1192.9][S03]So it's an interesting situation that we're in,[1196.22] [1196.22][S03]and I think we need to use the scientific method to do things,[1199.19] [1199.19][S03]like better understand what these systems do, build[1202.61] [1202.61][S03]benchmarks and rigorous evaluations[1204.86] [1204.86][S03]for understanding the limits of the capabilities,[1207.9] [1207.9][S03]interpretability.[1209.36] [1209.36][S03]Actually, I think we should be using neuroscience techniques,[1212.24] [1212.24][S03]which are, obviously, built for understanding real brains,[1215.06] [1215.06][S03]to model virtual brains.[1217.91] [1217.91][S03]I sometimes call it virtual brain analysis.[1220.05] [1220.05][S03]What's the equivalent of an fMRI of these neural networks, right?[1224.01] [1224.01][S03]And I think there's a lot we can learn, and bring across,[1227.53] [1227.53][S03]and translate from the natural sciences[1231.01] [1231.01][S03]to what is basically an engineering science.[1233.055] [1233.055][S03]I call it an engineering science,[1234.43] [1234.43][S03]because unlike the natural sciences,[1237.41] [1237.41][S03]you have to build the artifact of interest, first,[1240.413] [1240.413][S03]and then, once you have it, you can then[1242.08] [1242.08][S03]use the scientific method to reduce it down and understand[1245.32] [1245.32][S03]its components.[1247.18] [1247.18][S03]So it's a whole challenge.[1248.38] [1248.38][S03]That's for sure, and these systems, in my opinion,[1251.175] [1251.175][S03]are as complex as the natural phenomena we normally[1253.3] [1253.3][S03]want to study.[1253.94] [1253.94][S03]So no one should think that it would be easier[1256.24] [1256.24][S03]to understand one of these artificial systems[1258.28] [1258.28][S03]than it would be to understand a natural system.[1260.36] [1260.36][S03]I think they're just as complex in many ways.[1263.02] [1263.02][S03]So we've got a lot of work in front of us,[1264.77] [1264.77][S03]and I think this is something that industry, academia,[1268.7] [1268.7][S03]and civil society needs to come together to understand better,[1272.29] [1272.29][S03]including how to deploy these technologies.[1274.34] [1274.34][S04] Where does a scientist's intuition[1276.257] [1276.257][S04]fit in all of this?[1277.095] [1277.095][S03] Well, I think scientist's intuition[1279.94] [1279.94][S03]and creativity is critical.[1282.49] [1282.49][S03]I think, right now, the AI systems are just tools.[1285.38] [1285.38][S03]I think they're great for finding correlations,[1287.83] [1287.83][S03]and patterns, and structures in data, but for the moment,[1291.38] [1291.38][S03]they're not able to come up with their own hypotheses[1294.1] [1294.1][S03]or their own questions.[1295.503] [1295.503][S03]And I think, as all the scientists in the room know,[1297.67] [1297.67][S03]I think that's the hardest thing about science[1299.587] [1299.587][S03]is asking the right question, and that[1303.16] [1303.16][S03]has to still come from human scientists.[1305.327] [1305.327][S03]And for the foreseeable future, I think that will be the case.[1307.91] [1307.91][S03]So I think it could be-- that's why[1309.97] [1309.97][S03]I'm very excited about the use of these AI systems[1312.97] [1312.97][S03]as maybe the ultimate tools to help us accelerate[1315.79] [1315.79][S03]scientific discovery.[1316.665] [1316.665][S04] I guess on that topic of humans, I mean,[1318.832] [1318.832][S04]there has been some recent research saying[1321.13] [1321.13][S04]that the progress of scientific discovery[1324.46] [1324.46][S04]has slowed down over recent years.[1326.39] [1326.39][S04]What's your take on that, and what[1328.36] [1328.36][S04]could be done to change that?[1330.74] [1330.74][S03] Well, look, I think things--[1333.55] [1333.55][S03]I've seen studies like that, too,[1335.06] [1335.06][S03]and I think there's interesting conjectures as to why that is.[1337.99] [1337.99][S03]Science has become a bigger endeavor.[1340.017] [1340.017][S03]You need to have bigger teams, more expensive equipment,[1342.35] [1342.35][S03]and so on.[1343.23] [1343.23][S03]So that leads to some slowdown, and perhaps, the questions[1348.38] [1348.38][S03]we're now tackling are ever more complex.[1351.23] [1351.23][S03]So that's also tricky.[1353.6] [1353.6][S03]I think what I would recommend is,[1356.06] [1356.06][S03]again, I think, a lot of the advances in the next 10 years[1358.61] [1358.61][S03]are going to be interdisciplinary work, sort[1361.78] [1361.78][S03]of bringing together experts from two or more fields,[1365.63] [1365.63][S03]and then that making the big advances in between those areas.[1370.223] [1370.223][S03]And, actually, that's the story of DeepMind, which originally, I[1372.89] [1372.89][S03]think, was a combination of neuroscience ideas[1375.68] [1375.68][S03]with machine learning ideas.[1377.1] [1377.1][S03]It's also the story of AlphaFold,[1378.51] [1378.51][S03]which is the combination of biologists and chemists[1381.68] [1381.68][S03]on our team, along with machine learning experts and engineers.[1385.68] [1385.68][S03]So I've always found that's the place[1387.26] [1387.26][S03]to get the most advances most quickly,[1391.35] [1391.35][S03]and actually, we're announcing today, a $20 million[1395.6] [1395.6][S03]fund from Google.org to fund this kind[1398.45] [1398.45][S03]of interdisciplinary work in academia.[1400.74] [1400.74][S03]And I hope that other funders join that effort,[1403.873] [1403.873][S03]and I think that's what we need is[1405.29] [1405.29][S03]to train a new generation of PhDs and postdocs[1408.74] [1408.74][S03]in these kind of combination of these different areas.[1411.25] [1411.25][S04] What an extraordinary piece of news[1413.208] [1413.208][S04]to finish on, $20 million for research[1415.85] [1415.85][S04]in interdisciplinary teams.[1417.09] [1417.09][S04]Demis, thank you so much, indeed.[1420.157] [1420.157][S04]Thank you very much.[1420.99] [1420.99][APPLAUSE][1421.58] [1421.58][S04]OK, I think you go down to join the other laureates,[1425.84] [1425.84][S04]while we out the stage.[1428.63] [1428.63][S04]Now, I should tell you, actually,[1431.3] [1431.3][S04]the laureates, the word, laureates, by the way,[1434.01] [1434.01][S04]comes from the laurel wreath, which[1436.85] [1436.85][S04]was given to victors in ancient Greece as a sign of honor.[1439.79] [1439.79][S04]And every year, there are five separate prizes[1442.4] [1442.4][S04]that are awarded to those who are considered to have conferred[1446.84] [1446.84][S04]the greatest benefit to humankind courtesy of Alfred[1450.23] [1450.23][S04]Nobel.[1451.25] [1451.25][S04]There is a rumor that there is no Nobel Prize for mathematics,[1455.96] [1455.96][S04]because Alfred Nobel's great love ran away[1458.565] [1458.565][S04]with a mathematician.[1459.44] [1459.44][S04]Have we heard this?[1461.13] [1461.13][S04]I mean, we are charming.[1462.46] [1462.46][S04]What can I say?[1463.83] [1463.83][S04]Now, so joining me on the stage, we[1466.53] [1466.53][S04]have three further recipients of this esteemed prize,[1470.23] [1470.23][S04]John Jumper.[1470.73] [1470.73][S04]We've not yet had the pleasure of having John[1472.605] [1472.605][S04]Jumper on the stage thus far.[1474.07] [1474.07][S04]He is a director at Google DeepMind, where[1476.52] [1476.52][S04]he led the team that built AlphaFold and continues[1480.09] [1480.09][S04]to work on new methods to apply machine learning[1482.61] [1482.61][S04]to protein biology.[1483.82] [1483.82][S04]John has won numerous awards for his work,[1486.76] [1486.76][S04]including the Lasker Award, the Breakthrough[1489.63] [1489.63][S04]Prize in Life Sciences, the Canada International Award,[1493.08] [1493.08][S04]and of course, this year's Nobel Prize in chemistry.[1495.69] [1495.69][S04]We are also joined by Sir Paul Nurse,[1497.89] [1497.89][S04]who is CEO of the Francis Crick Institute[1499.86] [1499.86][S04]and winner of the 2001 prize in physiology of medicine[1503.7] [1503.7][S04]for his work also on proteins, specifically the protein[1506.22] [1506.22][S04]molecules that control the division of cells in the cell[1508.95] [1508.95][S04]cycle.[1509.82] [1509.82][S04]And, finally, Jennifer Doudna, who earlier shared her lessons[1513.9] [1513.9][S04]from CRISPR with James, she won the Nobel Prize in chemistry[1517.66] [1517.66][S04]in 2020.[1518.9] [1518.9][S04]Please, join me in welcoming our four Nobel laureates[1522.13] [1522.13][S04]to the stage.[1522.77] [1522.77][S04]Thank you.[1523.84] [1523.84][S04]OK, so I know that, often, you get[1526.45] [1526.45][S04]asked about where were when you heard,[1528.212] [1528.212][S04]what were you doing when you heard.[1529.67] [1529.67][S04]I want to ask you a slightly different question.[1531.94] [1531.94][S04]I want to ask you about whether there was a moment when[1534.88] [1534.88][S04]you realized that the work that you were doing[1537.16] [1537.16][S04]was genuinely groundbreaking.[1539.167] [1539.167][S04]Was there sort of a moment when you realized[1541.0] [1541.0][S04]the significance of it?[1542.27] [1542.27][S04]John?[1542.99] [1542.99][S06] I think there were kind of two moments for me.[1548.87] [1548.87][S06]One was actually watching Twitter[1551.41] [1551.41][S06]after-- whenever you release some work into the world,[1553.97] [1553.97][S06]you anxiously refresh Twitter, and we[1556.99] [1556.99][S06]stick to the word AlphaFold, and just see what tweets popped up.[1559.82] [1559.82][S06]And I remember seeing, when the database[1563.92] [1563.92][S06]became available, so many astounded grad students saying--[1567.95] [1567.95][S06]I think one was, how did they get my structure,[1570.68] [1570.68][S06]it hasn't been published, how in the world,[1574.1] [1574.1][S06]and just had absolute astounding.[1576.335] [1576.335][S06]Like, the number of people that said either[1580.327] [1580.327][S06]we predicted their structure, this person down the hall,[1582.66] [1582.66][S06]we predicted their structure, so I thought they were absolutely[1584.72] [1584.72][S06]blown away.[1585.22] [1585.22][S06]And I think the second moment was[1587.09] [1587.09][S06]there was a special issue of Science related[1588.978] [1588.978][S06]to the structure of the nuclear pore, the largest[1591.02] [1591.02][S06]collection of proteins in the human cell, protein chains,[1594.6] [1594.6][S06]and there had been, like, four papers in a special issue[1598.94] [1598.94][S06]of Science, and three out of the four[1600.71] [1600.71][S06]had made huge usage of AlphaFold.[1602.738] [1602.738][S06]Something like more than 100 mentions[1604.28] [1604.28][S06]of the word AlphaFold in Science,[1606.06] [1606.06][S06]and we had nothing to do with it.[1607.547] [1607.547][S06]We didn't know it was happening.[1608.88] [1608.88][S06]It just appeared one day, and that[1610.297] [1610.297][S06]was the moment when I really knew that people were doing[1613.49] [1613.49][S06]science worthy of appearing in Science, one of the most[1616.25] [1616.25][S06]prestigious journals in the world, on top of our tools[1619.495] [1619.495][S06]and without us.[1620.12] [1620.12][S06]And the moment at which people start making these discoveries[1623.48] [1623.48][S06]on top of what we've built, that's[1625.04] [1625.04][S06]when you really make it as someone who makes tools.[1627.267] [1627.267][S04] Jennifer, did you have a moment?[1629.1] [1629.1][S05] Well, I think for us,[1630.683] [1630.683][S05]yeah, maybe also two moments.[1632.16] [1632.16][S05]One was in the Fall of probably 2011, when[1638.36] [1638.36][S05]we had started a collaboration with Emmanuelle Charpentier[1641.09] [1641.09][S05]to work on CRISPR, which is a bacterial immune system,[1644.94] [1644.94][S05]and we wondered, how does it work?[1646.68] [1646.68][S05]And we had together figured out that it's an RNA guided[1652.1] [1652.1][S05]system that targets DNA for cutting,[1655.68] [1655.68][S05]for cleavage, and furthermore, that we could actually program[1659.57] [1659.57][S05]the system, because we understood[1661.13] [1661.13][S05]the chemistry of how it worked.[1662.97] [1662.97][S05]And it was really one of those aha moments.[1667.46] [1667.46][S05]Personally, I was just astounded that bacteria had figured out[1670.94] [1670.94][S05]how to do this and that, furthermore, we now understood[1674.63] [1674.63][S05]how to harness that system to manipulate DNA in new ways.[1679.65] [1679.65][S05]So that was one, and then the second one[1681.95] [1681.95][S05]was sort of akin to what you said,[1683.58] [1683.58][S05]John, where I think it was then fast forwarding about a year[1687.28] [1687.28][S05]or so.[1687.78] [1687.78][S05]We published that work in the Summer of 2012, and by the Fall,[1691.9] [1691.9][S05]I was starting to get emails from people[1693.9] [1693.9][S05]all over the world who had read the paper[1696.39] [1696.39][S05]and said, \"Oh, my gosh, this is so exciting.[1698.92] [1698.92][S05]I'm now using it to test genes in Drosophila.[1702.49] [1702.49][S05]I'm starting to test genes in zebrafish.[1705.66] [1705.66][S05]I'm excited about what I'm doing in human cells.\"[1708.55] [1708.55][S05]I was getting messages from people that I often barely knew[1711.6] [1711.6][S05]who were just very, very excited,[1713.08] [1713.08][S05]and you could start to feel the momentum that[1715.08] [1715.08][S05]was building in the field.[1716.945] [1716.945][S09] Cool.[1718.02] [1718.02][S09]Well, I'm the oldie here, so it's 1985, I'm afraid,[1722.97] [1722.97][S09]and I work on yeast, as you've already been told,[1726.79] [1726.79][S09]and my lab had worked out the genes that[1730.89] [1730.89][S09]controlled the cell cycle.[1733.39] [1733.39][S09]That is the process that leads to the reproduction of one cell[1736.38] [1736.38][S09]into two fundamental to growth and development in all living[1740.76] [1740.76][S09]organisms, and we'd identified that gene.[1744.01] [1744.01][S09]But the blunt thing is that, who cares about yeast?[1747.13] [1747.13][S09]I mean, being perfectly honest, I care about yeast,[1750.17] [1750.17][S09]but most of the world does not care about yeast.[1752.89] [1752.89][S09]So I found our work wasn't really being[1755.68] [1755.68][S09]taken much notice of.[1757.3] [1757.3][S09]So I thought, do humans have the same gene?[1760.82] [1760.82][S09]Now, this is long before the human genome sequence and so on.[1764.39] [1764.39][S09]So we did a crazy experiment.[1767.09] [1767.09][S09]I want to just tell you how crazy it was.[1769.01] [1769.01][S09]We had a mutant in yeast that couldn't grow,[1771.46] [1771.46][S09]because it was defective in this gene.[1773.81] [1773.81][S09]We took the first human cDNA library.[1777.05] [1777.05][S09]It's genES that had ever been made.[1780.13] [1780.13][S09]We didn't make it.[1781.13] [1781.13][S09]We got it a few months afterwards and fundamentally[1784.63] [1784.63][S09]sprinkled the genes onto the defective yeast,[1788.62] [1788.62][S09]arguing that, if there was a gene in humans that[1791.98] [1791.98][S09]could do the same job as the defective gene in yeast,[1795.71] [1795.71][S09]if the yeast cell took it up, if it could be expressed,[1799.55] [1799.55][S09]if it worked, then those cells would grow and divide,[1803.15] [1803.15][S09]and we could get the gene back and show that humans had it.[1808.32] [1808.32][S09]That experiment had no right to work.[1810.6] [1810.6][S09]Yeast, humans probably diverged 1,500 million years ago,[1817.11] [1817.11][S09]and we were demanding that it still worked after 1,500 million[1821.42] [1821.42][S09]years, and you know, it did work.[1824.54] [1824.54][S09]When we got that result--[1825.812] [1825.812][S09]and by the way, there were a couple[1827.27] [1827.27][S09]of months when we had to sequence the genes.[1829.103] [1829.103][S09]It took that long.[1830.01] [1830.01][S09]Then I used to think, I'm just going[1832.43] [1832.43][S09]to go home and believe this worked.[1834.86] [1834.86][S09]Tomorrow, I will go into work, and we[1837.59] [1837.59][S09]will have shown it didn't work, but it did work.[1840.9] [1840.9][S09]And that's when I thought, perhaps, it[1842.57] [1842.57][S09]might be recognized, because it wasn't yeast.[1844.56] [1844.56][S09]It was humans.[1845.34] [1845.34][S04] What an amazing collection of stories.[1847.423] [1847.423][S04]Now, I know that all of you must be bursting with questions[1851.54] [1851.54][S04]for our panel, so I might take some questions[1855.08] [1855.08][S04]from the audience.[1855.93] [1855.93][S04]Do we have mics running around?[1857.28] [1857.28][S04]Oh, there's one there.[1857.97] [1857.97][S04]I might take them in a little cluster.[1859.29] [1859.29][S04]Is there anyone else who would like to pop them in?[1861.48] [1861.48][S04]OK, so we have one here and another one here as well.[1864.03] [1864.03][S04]Roger?[1864.53] [1864.53][S01] This has been such an impressive day,[1868.18] [1868.18][S01]and Thank you for closing it out like this.[1871.09] [1871.09][S01]I've got a very playful question, which is,[1874.08] [1874.08][S01]if you could take yourself back in time and met your 18-year-old[1879.03] [1879.03][S01]or your 21-year-old self, and you could say something that[1883.38] [1883.38][S01]would reassure that person that you were making good life[1887.07] [1887.07][S01]choices, what would you say?[1890.26] [1890.26][S01]And you're not allowed to say, you're[1892.65] [1892.65][S01]going to win a Nobel Prize or anything to do with that.[1896.56] [1896.56][S01]So what would you say to your 18-year-old,[1901.66] [1901.66][S01]20-year-old former self?[1903.852] [1903.852][S04] OK, well, we'll collect another question.[1906.06] [1906.06][S04]I think Roger Highfield just here had a question too.[1909.63] [1909.63][S08] Roger Highfield, the Science Museum.[1911.86] [1911.86][S08]Just a comment, we've seen images[1914.73] [1914.73][S08]of Sycamore, the very cool looking Google quantum computer.[1918.37] [1918.37][S08]Can we, please, get some sexy looking AI hardware[1921.69] [1921.69][S08]that we can collect at the Science Museum, please?[1923.81] [1923.81][LAUGHING][1925.147] [1925.147][S08]Actually, a more serious point is-- and you've kind of alluded[1927.73] [1927.73][S08]to it, Demis, is this thing about how[1931.18] [1931.18][S08]AI is great at giving you answers,[1932.83] [1932.83][S08]but it doesn't give-- it's not very good at giving you[1935.08] [1935.08][S08]mechanistic insights.[1936.26] [1936.26][S08]You've got billions of parameters.[1937.97] [1937.97][S08]They've got no physical meaning.[1940.22] [1940.22][S08]How much of a barrier is this to the public trusting AI?[1945.59] [1945.59][S08]And I've talked to John about this as well.[1947.87] [1947.87][S08]Where are we in actually getting an AI[1950.68] [1950.68][S08]that can give us the laws of biology,[1953.18] [1953.18][S08]just like we've got the laws of physics,[1955.07] [1955.07][S08]so we get true mechanistic insights?[1957.18] [1957.18][S04] Fantastic.[1958.36] [1958.36][S04]Two very good questions there.[1959.66] [1959.66][S04]We'll start with the first one.[1960.952] [1960.952][S04]Demis, we go to you first.[1962.24] [1962.24][S04]What would you what would you say to your 18-year-old self?[1964.22] [1964.22][S03] Well, it's a little bit complicated for me.[1966.8] [1966.8][S03]I mean, I did actually have this plan when I was 18.[1969.91] [1969.91][LAUGHING][1971.02] [1971.02][S03]What is amazing is it worked out.[1973.31] [1973.31][S03]But I would have told myself--[1975.222] [1975.222][S03]I mean, it was a chess player in me is I always plan, like, many,[1977.93] [1977.93][S03]many years ahead.[1979.072] [1979.072][S03]That's what happens if you play chess from the age of four.[1981.53] [1981.53][S03]You end up like this.[1982.49] [1982.49][S03]But probably what I would have said to myself[1985.27] [1985.27][S03]is just enjoy the journey a bit more, because it will work out.[1989.86] [1989.86][S03]Because at the time, I was like, how[1991.36] [1991.36][S03]is this ever going to work out?[1992.652] [1992.652][S03]I had all these dreams, and then that's probably what[1995.05] [1995.05][S03]I'd recommend to my 18-year-old self.[1996.832] [1996.832][S04] John?[1997.54] [1997.54][S06] I think two things, actually.[1999.665] [1999.665][S06]Amusingly, I was thinking, don't worry, Carolyn will marry you.[2002.29] [2002.29][S06]But--[2002.79] [2002.79][LAUGHING][2003.99] [2006.85][S06]It's the gradient descent of life.[2008.77] [2008.77][S06]Do the right thing right now has worked out really well,[2012.06] [2012.06][S06]and be open to the interesting things that will open up[2014.945] [2014.945][S06]for you, and then, I think, we're[2016.32] [2016.32][S06]in one of the golden ages of biology, and AI, and biology.[2020.702] [2020.702][S06]And it's really fun to live through that.[2022.41] [2022.41][S06]So I think, like, don't be afraid to be[2024.78] [2024.78][S06]locally optimal, basically, the opposite of Demis' advice.[2027.605] [2027.605][LAUGHING][2028.806] [2030.193][S04] How about Jennifer?[2031.485] [2031.485][S04]Are you gradient descent, or are you[2033.81] [2033.81][S04]planning strategically in advance?[2035.445] [2035.445][S05] Probably closer to John.[2038.37] [2038.37][S05]Yeah, I think I would tell my 18-year-old self to follow[2042.85] [2042.85][S05]my passion and never, ever give up,[2045.74] [2045.74][S05]and don't listen to people that are naysayers.[2048.35] [2048.35][S05]That's really important.[2051.53] [2051.53][S09] I came from a non-academic background.[2054.59] [2054.59][S09]I couldn't believe that you could get paid for just[2059.17] [2059.17][S09]following your curiosity.[2061.33] [2061.33][S09]I still don't believe it now, 40 years later, 50 years later,[2067.48] [2067.48][S09]actually, more than that, 55 years later.[2070.78] [2070.78][S04] Just hold on to believing that we can,[2073.245] [2073.245][S04]OK, as president of the Royal Society?[2076.03] [2076.03][S04]You're good.[2077.08] [2077.08][S04]The second question from Roger there[2079.06] [2079.06][S04]about intuition from AI systems?[2081.135] [2081.135][S03] Yeah, look, I think, actually, this[2083.26] [2083.26][S03]is pretty interesting.[2084.23] [2084.23][S03]I'm not so worried about it as other people[2086.022] [2086.022][S03]are, because I think we're in a moment in time.[2088.58] [2088.58][S03]So what I said earlier, actually, my talk[2090.85] [2090.85][S03]was what I believe, which is that AI is an engineering[2095.25] [2095.25][S03]science.[2095.75] [2095.75][S03]So what it means is you have to build the artifact, first,[2098.67] [2098.67][S03]worthy of study, and then you can break it down[2101.28] [2101.28][S03]with a scientific method.[2102.87] [2102.87][S03]So what you've seen in the last five,[2104.92] [2104.92][S03]10 years is the building of the artifacts that are even worth[2108.42] [2108.42][S03]putting any effort into study, until you have things,[2111.24] [2111.24][S03]like today's transformer models and AlphaFold's, AlphaGo's.[2115.355] [2115.355][S03]The earlier systems were probably[2116.73] [2116.73][S03]not worth really putting the effort in to study,[2118.53] [2118.53][S03]because they were not sophisticated enough.[2120.49] [2120.49][S03]Now, they are.[2121.39] [2121.39][S03]So now, people are seriously studying the current systems.[2124.57] [2124.57][S03]On top of that, you've got the additional benefit[2127.11] [2127.11][S03]of these systems improving themselves, and at the minimum,[2133.15] [2133.15][S03]I think, we'll be in a situation,[2134.58] [2134.58][S03]where, firstly, the systems might[2136.29] [2136.29][S03]be able to explain themselves in language, or mathematics,[2139.41] [2139.41][S03]or code.[2140.17] [2140.17][S03]So we're getting close to that as you could say to a system,[2142.785] [2142.785][S03]OK, you've understood this.[2144.61] [2144.61][S03]Now, explain that in a mathematical equation[2147.39] [2147.39][S03]to the extent that it can be.[2148.76] [2148.76][S03]And I'm not sure biology can be explained,[2150.51] [2150.51][S03]like the laws of physics, by the way.[2152.08] [2152.08][S03]I think it is a lot messier.[2153.285] [2153.285][S03]It's more to do with interaction.[2154.66] [2154.66][S03]So I think a simulation would be more appropriate that you probe[2157.62] [2157.62][S03]than, like, Newton's laws of motion or something like that,[2160.695] [2160.695][S03]and I don't think biology could be reduced to that.[2162.82] [2162.82][S03]It's too complex.[2164.4] [2164.4][S03]And then the other thing is what I also said earlier[2166.98] [2166.98][S03]about applying neuroscience techniques, analysis[2169.05] [2169.05][S03]techniques to these artificial neural networks,[2173.59] [2173.59][S03]and we should be able to get, at least,[2175.47] [2175.47][S03]the same level of insights into them[2177.3] [2177.3][S03]as we do with natural brains.[2179.53] [2179.53][S03]So if you combine that together, we[2181.53] [2181.53][S03]should get pretty far already, let alone[2183.78] [2183.78][S03]with further engineering efforts that we're[2185.61] [2185.61][S03]going to put on top to decompose these systems.[2188.17] [2188.17][S03]So I think, in the next five years, we'll be out of this era[2190.89] [2190.89][S03]that we're currently in of kind of black boxes.[2193.95] [2193.95][S04] Fascinating.[2194.95] [2194.95][S04]OK, further questions?[2195.867] [2195.867][S04]All right, we've got a couple here.[2197.325] [2197.325][S04]So let's go one, two.[2198.287] [2198.287][S04]And then, if you can keep your hands up, actually,[2200.37] [2200.37][S04]so I can see, we'll go we'll go one, two,[2202.68] [2202.68][S04]and then we'll go over there.[2204.48] [2204.48][S04]Read them out in chunks, if that's OK.[2206.35] [2206.35][S04]Go for it.[2207.893] [2207.893][S02] Thank you.[2209.31] [2209.31][S02]Danny Newman Griffiths from University of Sheffield.[2211.477] [2211.477][S02]Just following up on the topic earlier[2213.66] [2213.66][S02]that came up about the social sciences in AI,[2217.39] [2217.39][S02]the idea was raised about potential for AI[2219.91] [2219.91][S02]to transform social sciences.[2221.3] [2221.3][S02]I want to ask you about flipping that around and thinking about,[2224.74] [2224.74][S02]in your understanding, how can you[2226.75] [2226.75][S02]see the work of social science as helping[2228.94] [2228.94][S02]to transform these next features of AI[2231.82] [2231.82][S02]and building the ways to transform that engineering[2235.33] [2235.33][S02]science into the interdisciplinarity[2237.34] [2237.34][S02]that we've been talking about, where AI tools are being[2239.95] [2239.95][S02]combined with really deep understanding of application[2243.13] [2243.13][S02]areas and contexts?[2244.79] [2244.79][S04] Fascinating question.[2246.19] [2246.19][S04]Your neighbor next to you.[2247.95] [2247.95][S01] A very simple question,[2250.6] [2250.6][S01]is attention still all we need?[2254.572] [2254.572][S04] Go for it.[2255.489] [2255.489][S07] Michael Chang from the National Eye Institute[2257.989] [2257.989][S07]in the US.[2258.53] [2258.53][S07]Thank you very much.[2259.58] [2259.58][S07]This was an awesome conference.[2261.25] [2261.25][S07]I have a-- the theme here is AI to do better science,[2264.86] [2264.86][S07]and I think one of the common things that we do is,[2268.34] [2268.34][S07]well, we can just teach scientists more about AI,[2271.15] [2271.15][S07]and they can use it.[2272.68] [2272.68][S07]My question is flipping that around.[2274.43] [2274.43][S07]Do you think there are things that AIs do not do,[2278.11] [2278.11][S07]that scientists should be getting more training on[2281.05] [2281.05][S07]to maximize their added value and what those would be?[2283.785] [2283.785][S07]Just love your perspectives on that.[2285.285] [2285.285][S04] Fascinating.[2286.4] [2286.4][S04]OK, all right, so we'll start off with--[2288.34] [2288.34][S04]I might start with you, Paul, actually,[2289.55] [2289.55][S04]because you were the one who was talking[2290.26] [2290.26][S04]about social sciences earlier.[2291.51] [2291.51][S04]So yeah, AI and social sciences, how much[2293.89] [2293.89][S04]do we have to think about that sort of interface between humans[2296.547] [2296.547][S04]and the machines?[2297.255] [2297.255][S09] Well, first of all,[2299.57] [2299.57][S09]I think we need to focus more on the social sciences[2302.68] [2302.68][S09]to help us scientists.[2306.97] [2306.97][S09]We have to appreciate, though, that this is pretty complicated,[2311.18] [2311.18][S09]the human interactions.[2313.21] [2313.21][S09]So I'm not sure how much it's going[2316.99] [2316.99][S09]to help us at the beginning, not sure about that,[2320.78] [2320.78][S09]but we really ought to be accommodating it,[2323.56] [2323.56][S09]ought to be thinking about how it might.[2326.47] [2326.47][S09]I could imagine certain social sciences problems,[2329.72] [2329.72][S09]and I suspect you're a social scientist,[2332.57] [2332.57][S09]to do with transport and things of this sort.[2335.52] [2335.52][S09]So I think there are some mechanical things, trying[2338.09] [2338.09][S09]to understand how human beings interact and work.[2341.42] [2341.42][S09]Well, you have to ask Carol, was it?[2344.751] [2344.751][S06] Yes, sure.[2347.17] [2347.17][S09] John, yes, exactly.[2348.63] [2348.63][S06] Oh, Caroline?[2349.78] [2349.78][S09] Yes, that Carol.[2351.113] [2351.113][LAUGHING][2352.04] [2352.04][S09]In other words-- sorry, I thought it was straightforward.[2354.78] [2354.78][S09]Getting emotion, understanding emotion--[2356.87] [2356.87][S06] You got my wife's name wrong.[2357.08] [2357.08][S09] --is going to be difficult.[2357.67] [2357.67][LAUGHING][2358.435] [2358.435][S03] And he was looking at me.[2360.143] [2360.143][S03]It was doubly confusing.[2362.365] [2362.365][S09] Over to you, John.[2364.99] [2364.99][S06] I mean, should I answer this,[2366.74] [2366.74][S06]or should I go to the--[2367.71] [2367.71][S04] Yeah, go to the attention, go to the attention.[2370.168] [2370.168][S06] I think attention is all you[2372.14] [2372.14][S06]need is an oversimplification.[2374.095] [2374.095][S06]I think one of the things that's really interesting[2376.22] [2376.22][S06]is AlphaFold isn't just, like, grab a transformer[2378.65] [2378.65][S06]off the shelf from the transformer store[2380.48] [2380.48][S06]and apply it to protein structure prediction.[2382.695] [2382.695][S06]It took a couple of years, because there's a lot of work,[2385.07] [2385.07][S06]and there's a lot of new ideas, where attention is a component,[2389.118] [2389.118][S06]but we have what we would call evil former.[2390.91] [2390.91][S06]We have new ideas on top.[2392.53] [2392.53][S06]We have all sorts of--[2394.995] [2394.995][S06]Demis runs an incredible AI research[2396.93] [2396.93][S06]org, which goes into work every day and doesn't just say, well,[2399.653] [2399.653][S06]we've got all we need.[2400.57] [2400.57][S06]There's a tension, right?[2401.612] [2401.612][S06]We all are doing this work, and it's really,[2403.78] [2403.78][S06]I think, people underestimate how much[2405.99] [2405.99][S06]really novel and exciting work is going on in AI[2409.11] [2409.11][S06]right now that are making these systems transformatively better,[2412.3] [2412.3][S06]both unlocking new data sources and learning a lot more[2414.84] [2414.84][S06]from the existing data sources.[2416.26] [2416.26][S06]AlphaFold is a story of having the same data as everyone else[2419.13] [2419.13][S06]and learning a tremendously more about protein structure from it.[2423.81] [2423.81][S06]So I think we'll continue to see these dividends from AI[2426.42] [2426.42][S06]research.[2427.005] [2427.005][S06]We're going to continue to see exciting new things,[2429.13] [2429.13][S06]and we may always call something a tension.[2432.04] [2432.04][S06]But it reminds me of this old computer scientist joke.[2434.65] [2434.65][S06]I don't know what the scientific computing language of the future[2436.95] [2436.95][S06]will look like, but it will be called Fortran, right?[2439.158] [2439.158][LAUGHING][2439.71] [2439.71][S06]So we tend to keep the labels of ideas and then update within it.[2445.317] [2445.317][S06]And I think we shouldn't underestimate[2446.9] [2446.9][S06]how much really wonderful, and exciting, and clever research is[2449.72] [2449.72][S06]going on these days.[2450.665] [2450.665][S03] Maybe I can just add quickly to that.[2452.873] [2452.873][S03]I agree with that.[2453.86] [2453.86][S03]And, actually, the transformer architecture,[2458.36] [2458.36][S03]which is what was invented with that paper, is being amazing[2462.93] [2462.93][S03]and, I think, will underpin, be one of the main components[2465.68] [2465.68][S03]of a future AGI system.[2467.61] [2467.61][S03]But my prediction is it's not going to be enough on its own.[2470.42] [2470.42][S03]I think we're going to need a couple[2471.92] [2471.92][S03]of other big breakthroughs like that in addition,[2475.16] [2475.16][S03]and they're still to come.[2476.67] [2476.67][S04] Jennifer, I might come to you just[2478.97] [2478.97][S04]to pick up on our third question there about the gaps in what AI[2483.71] [2483.71][S04]can do, and we need to really focus[2485.93] [2485.93][S04]on the skill set from humans.[2487.523] [2487.523][S05] Yeah, and thanks for that question,[2489.69] [2489.69][S05]Michael.[2490.19] [2490.19][S05]So I guess what I've been thinking about with regard[2493.13] [2493.13][S05]to that is there's been a lot of mention today about data[2496.64] [2496.64][S05]and the kind of data that are necessary for training models[2500.09] [2500.09][S05]like this.[2501.0] [2501.0][S05]And one of the challenges in biology, of course,[2503.858] [2503.858][S05]is the quality of data, but it's also[2505.4] [2505.4][S05]the quantity of data and the fact[2508.1] [2508.1][S05]that, typically, one needs a lot of data[2510.62] [2510.62][S05]that's of high quality to train models, at least, currently.[2514.4] [2514.4][S05]So what I'd like to see AI do for us scientists is to educate[2519.5] [2519.5][S05]us about how to collect data, perhaps, sparsely, but smartly,[2525.12] [2525.12][S05]so that your sparse data is broad enough[2528.65] [2528.65][S05]that it actually does provide the right platform for training.[2532.362] [2532.362][S05]And I think that's something we don't, at least[2534.32] [2534.32][S05]as an experimentalist, right now, we[2535.97] [2535.97][S05]don't think about that when we design experiments,[2538.08] [2538.08][S05]but we could.[2538.64] [2538.64][S04] I guess that kind of comes back to the asking[2540.65] [2540.65][S04]the right questions thing that we were talking about.[2542.19] [2542.19][S05] It does.[2542.6] [2542.6][S04] What is the [INAUDIBLE]?[2543.92] [2543.92][S06] That's one of the key questions.[2544.8] [2544.8][S04] Yeah, OK, we probably have time for another.[2547.68] [2547.68][S04]All right, we've got red jumper at the back.[2549.81] [2549.81][S04]You can go be number one, and then I[2551.75] [2551.75][S04]need to think strategically about microphones.[2553.888] [2553.888][S04]OK, if we go, there's a little cluster over here.[2555.93] [2555.93][S04]So we'll go 2, 3, 4 further back.[2558.56] [2558.56][S04]Yes, just-- there we go.[2561.21] [2561.21][S04]We found it.[2561.71] [2561.71][S04]OK, go for it, number one.[2562.89] [2562.89][S11] Hi, it's Wendy Hall, University of Southampton.[2565.728] [2565.728][S11]Someone who's been working with social scientists[2567.77] [2567.77][S11]for a long time, the big issue there,[2569.61] [2569.61][S11]collecting the data and the issues around privacy.[2573.47] [2573.47][S11]It makes it so much harder than the science you do.[2576.812] [2576.812][S11]The more we can do with that, the better.[2578.52] [2578.52][S11]My question, though, as someone who's[2580.91] [2580.91][S11]just been suffering from a cold, as many people, please,[2584.52] [2584.52][S11]can AI sort out the common cold?[2586.54] [2586.54][LAUGHING][2587.73] [2588.23][S04] Guys, what have you been doing?[2590.78] [2590.78][S04]OK, we've got three very closely together here.[2593.645] [2593.645][S04]Yeah, go ahead.[2594.27] [2594.27][S10] Hi, Thomas Crampton.[2596.372] [2596.372][S10]There's been a lot of talk, obviously, about great science,[2598.83] [2598.83][S10]a lot of fantastic science discussed.[2601.745] [2601.745][S10]What about the people outside of this room, though,[2603.87] [2603.87][S10]who don't understand science and might be suspicious of it?[2607.95] [2607.95][S10]How much concern do you have that society might reject[2610.755] [2610.755][S10]a lot of these great breakthroughs that[2612.38] [2612.38][S10]are coming to the world, and what[2613.755] [2613.755][S10]do you think we should do to address that?[2615.77] [2615.77][S04] Great question.[2617.12] [2617.12][S04]Next to you?[2618.12] [2618.12][S01] Hi, I'm [INAUDIBLE] from the African Institute[2621.35] [2621.35][S01]for Mathematical Sciences.[2623.76] [2623.76][S01]Thanks very much for a wonderful discussion.[2627.0] [2627.0][S01]I have two questions, if you permit.[2629.46] [2629.46][S01]The first is that, when people talk about AGI,[2633.1] [2633.1][S01]I think the common assumption is that the target is[2637.56] [2637.56][S01]human intelligence.[2639.07] [2639.07][S01]But human intelligence may be suboptimal,[2641.85] [2641.85][S01]because this is the result of historically contingent[2645.81] [2645.81][S01]evolutionary process.[2647.23] [2647.23][S01]So to what extent do you think AGI[2650.19] [2650.19][S01]is going to be something supra optimal with respect[2654.06] [2654.06][S01]to human intelligence?[2656.13] [2656.13][S01]The second is about inclusion.[2660.33] [2660.33][S01]Africa, by 2050, would have the largest population[2666.75] [2666.75][S01]of young adults in the world.[2668.52] [2668.52][S01]This population would be serving the world.[2671.29] [2671.29][S01]It will have to in order for humanity to continue its ascent.[2677.37] [2677.37][S01]To what extent is the community ensuring[2679.72] [2679.72][S01]that Africans are included?[2681.94] [2681.94][S01]I should mention that Google has been amazing,[2685.82] [2685.82][S01]and Google DeepMind has been an amazing supporter[2689.35] [2689.35][S01]of capacity building for young Africans in AI.[2692.77] [2692.77][S01]As a whole, what else is the community doing?[2696.32] [2696.32][S01]Thank you.[2697.07] [2697.07][S04] Great, and then, I think,[2698.02] [2698.02][S04]there's one bit immediately behind you as well.[2699.978] [2699.978][S04]Thank you.[2702.797] [2702.797][S01] Thank you.[2703.63] [2703.63][S01][INAUDIBLE] Digital Science.[2705.65] [2705.65][S01]I'm a recovering academic, a nanochemist by training,[2709.57] [2709.57][S01]and in chemistry, technological developments in things,[2713.26] [2713.26][S01]like microscopy, have always opened doors[2715.51] [2715.51][S01]to many, many other areas of research.[2718.28] [2718.28][S01]AI is very much the same, but I think[2720.49] [2720.49][S01]we've seen from a lot of the speakers and the discussions[2723.22] [2723.22][S01]today that a lot of the developments[2725.2] [2725.2][S01]come about through industry rather than academia.[2729.08] [2729.08][S01]So, Paul, I know you talked about research cultures[2731.62] [2731.62][S01]and how maybe-- some of our other speakers[2733.48] [2733.48][S01]talked about how maybe the way that we reward[2736.18] [2736.18][S01]successful research, at the moment,[2737.89] [2737.89][S01]isn't conducive to making space for innovations in the same way.[2742.65] [2742.65][S01]How do we change a global culture[2745.22] [2745.22][S01]when nobody wants to be the first to do that,[2749.46] [2749.46][S01]because they may be losing out?[2750.85] [2750.85][S04] OK, right.[2752.37] [2752.37][S04]Four extremely good questions, we[2754.43] [2754.43][S04]have precisely four minutes to answer them, everybody, OK?[2757.49] [2757.49][S04]All right, so Paul, we'll start with you.[2759.72] [2759.72][S04]Why isn't there a cure for the common cold yet?[2762.552] [2762.552][S09] Well, it's difficult, isn't it?[2766.37] [2766.37][S09]I'll go to the--[2769.43] [2769.43][S04] Of course, you absolutely can.[2771.64] [2771.64][S09] Keeping the public on board,[2774.18] [2774.18][S09]this is really critical.[2775.38] [2775.38][S09]It's crucial.[2776.04] [2776.04][S09]We really need to concentrate on it.[2777.66] [2777.66][S09]It is, of course, not the first time.[2779.85] [2779.85][S09]Nearly always, when there's been a new technology and changes,[2783.48] [2783.48][S09]there has been concern, and I think I mentioned it earlier.[2788.4] [2788.4][S09]We have to talk to the right people.[2790.17] [2790.17][S09]We have to talk to the public, and too often, these discussions[2793.88] [2793.88][S09]get hijacked by interest groups and people who[2797.88] [2797.88][S09]say they talk for the public when they talk often[2803.04] [2803.04][S09]for their own particular interests[2805.56] [2805.56][S09]or their own particular passion that they have.[2808.9] [2808.9][S09]So we have to work out ways in which[2810.9] [2810.9][S09]we can discuss with the public and discuss in a sensible way,[2815.65] [2815.65][S09]and I mentioned deliberative democracy.[2817.78] [2817.78][S09]It's expensive to do, but I think it's[2819.6] [2819.6][S09]a really important thing to do.[2822.18] [2822.18][S09]Because, bluntly, if you don't take the public with you,[2826.39] [2826.39][S09]we won't be able to see all the merits that can come out[2830.01] [2830.01][S09]of this, merits in terms of understanding[2832.68] [2832.68][S09]the world better around us, merits[2834.78] [2834.78][S09]in terms of actually using those discoveries for the public good.[2839.05] [2839.05][S09]We have to engage, and we have to convince the public[2841.913] [2841.913][S09]that these things are right.[2843.08] [2843.08][S04] I think I want to actually just add[2845.37] [2845.37][S04]on a little bit to Susie's question[2847.537] [2847.537][S04]there, which I think also incorporates[2849.12] [2849.12][S04]part of your question.[2850.98] [2850.98][S04]Because for all of you, we were talking[2852.9] [2852.9][S04]a lot today about measures of success and incentives[2855.42] [2855.42][S04]for scientists, and of course, all of you[2857.34] [2857.34][S04]have the ultimate prize of being Nobel Laureates.[2861.37] [2861.37][S04]How has your view of success changed,[2864.54] [2864.54][S04]and how should we kind of influence future generations[2869.07] [2869.07][S04]across the world of scientists to make sure that we're[2871.95] [2871.95][S04]getting the right outcomes?[2873.53] [2873.53][S04]John, do you want to go first?[2874.78] [2874.78][S06] I think that's--[2876.951] [2876.951][S06]I think one of the things that's really been informing me[2880.44] [2880.44][S06]over my scientific career is just[2882.15] [2882.15][S06]the power and the fun of working on a team, doing science,[2885.58] [2885.58][S06]and I was able to do it at Google DeepMind.[2887.8] [2887.8][S06]I did it for a time before my PhD,[2891.3] [2891.3][S06]and then the PhDs are a little bit lonelier, right?[2893.64] [2893.64][S06]You're working on a particular individual thing.[2895.81] [2895.81][S06]And I think, really, the power of working within a team[2899.503] [2899.503][S06]is really, really important.[2900.67] [2900.67][S06]I think that provides also its own motivation.[2902.89] [2902.89][S06]Science is about loads of failure[2905.01] [2905.01][S06]and occasional dramatic success, and we're all up here[2908.34] [2908.34][S06]from those dramatic successes.[2910.12] [2910.12][S06]But I think that--[2911.28] [2911.28][S06]and maybe Demis has it down to a repeatable science, but--[2914.16] [2914.16][LAUGHING][2914.92] [2914.92][S06]I think that, really, encouraging people[2919.96] [2919.96][S06]to work in teams, work together, and I[2921.633] [2921.633][S06]think that helps provide the kind of motivation that[2923.8] [2923.8][S06]leads to better science, that leads to more fun science.[2926.78] [2926.78][S06]If you don't have fun doing it, you won't do it.[2928.84] [2928.84][S06]All the great scientists seem to be having fun at it.[2930.91] [2930.91][S04] Jennifer?[2931.525] [2931.525][S05] Couldn't agree more.[2933.067] [2933.067][S05]And I have to say that, when I think back on my career so far,[2937.31] [2937.31][S05]I really feel great joy and pride in the students that I've[2942.43] [2942.43][S05]trained and the work that they're now doing,[2945.17] [2945.17][S05]and it's just there's something incredibly[2947.74] [2947.74][S05]satisfying about that.[2948.74] [2948.74][S05]And, actually, can I speak to one of the other questions?[2950.895] [2950.895][S04] Please.[2951.687] [2951.687][S05] You asked about the question about Africa[2954.16] [2954.16][S05]and involving African scientists and young people who[2958.36] [2958.36][S05]want to do science.[2959.63] [2959.63][S05]So I'm really proud that the Innovative Genomics Institute[2962.11] [2962.11][S05]has an ongoing effort right now in Kenya, where now, I think,[2966.108] [2966.108][S05]we've done three years running.[2967.4] [2967.4][S05]We've sent a team to different parts of Kenya,[2971.66] [2971.66][S05]where they've been working with scientists there[2974.36] [2974.36][S05]to work with them and really help them understand CRISPR.[2979.56] [2979.56][S05]It's really been really motivating[2981.86] [2981.86][S05]to see some of the videos that come back,[2983.9] [2983.9][S05]where these local scientists then go back[2986.45] [2986.45][S05]to their communities, and they work with students,[2988.55] [2988.55][S05]and they get excited, and they start[2990.05] [2990.05][S05]doing interesting, creative science[2993.59] [2993.59][S05]in their own laboratories.[2995.46] [2995.46][S05]So I'd like to see more of that.[2997.32] [2997.32][S05]I think there's an extraordinary opportunity there,[2999.93] [2999.93][S05]and I'm also very excited about the things[3001.93] [3001.93][S05]that I hear about Google doing.[3003.68] [3003.68][S05]So I think there's a big opportunity for all of us[3006.4] [3006.4][S05]to work together on that.[3007.62] [3007.62][S04] More and more people are[3009.12] [3009.12][S04]being paid to follow their curiosity, Paul.[3011.667] [3011.667][S09] You almost said what I was going to say.[3014.0] [3014.0][S04] Oh, sorry.[3014.355] [3014.355][LAUGHING][3014.89] [3014.89][S09] Yeah, but I'll amplify it slightly.[3017.54] [3017.54][S09]We live in this world of big data,[3020.08] [3020.08][S09]and sometimes, we have low standards,[3022.72] [3022.72][S09]that it's enough just to report lots of big data.[3026.74] [3026.74][S09]There are certain journals, very high profile journals,[3029.94] [3029.94][S09]that seem to do nothing, but actually[3032.07] [3032.07][S09]report lots and lots of data.[3033.97] [3033.97][S09]And I think it's worth giving some attention to creativity[3038.19] [3038.19][S09]in the world of big data, because creativity[3042.15] [3042.15][S09]can get lost in this scenario, when in fact, there[3046.56] [3046.56][S09]is enormous opportunities with big data[3049.75] [3049.75][S09]if we actually take a creative approach.[3052.96] [3052.96][S09]And we need to think a bit--[3054.55] [3054.55][S09]one, what is exactly creativity?[3057.1] [3057.1][S09]And I could go on about that, but I won't,[3059.28] [3059.28][S09]because it's near the end.[3060.7] [3060.7][S09]But we need to encourage in our colleagues,[3064.24] [3064.24][S09]in our students, creative thinking,[3067.29] [3067.29][S09]and that is a bit different from collecting[3069.6] [3069.6][S09]lots and lots of data.[3071.26] [3071.26][S09]But lots and lots of data really will[3073.62] [3073.62][S09]deliver if they take a creative approach to it.[3076.21] [3076.21][S04] Demis?[3077.43] [3077.43][S03] Well, so many things to pick up on,[3078.93] [3078.93][S03]and there's not much time.[3080.013] [3080.013][S03]But, I mean, yeah, creativity would[3081.568] [3081.568][S03]be a very interesting thing to have a whole panel[3083.61] [3083.61][S03]discussion on it.[3084.27] [3084.27][S03]Maybe the next time, we do this.[3085.603] [3085.603][S03]But maybe just to pick up on a couple of questions,[3088.22] [3088.22][S03]I think in terms of encouraging the next generation, for me,[3095.17] [3095.17][S03]one of my heroes was Feynman.[3096.578] [3096.578][S03]But it wasn't just his physics books,[3098.12] [3098.12][S03]which were very famous, right?[3099.37] [3099.37][S03]But, actually, it was his layperson's books[3101.62] [3101.62][S03]that inspired me to get into science,[3103.19] [3103.19][S03]and I really think all school kids should read them,[3105.65] [3105.65][S03]so Surely You're Joking, Mr. Feynman,[3107.56] [3107.56][S03]and The Great Joy In Finding Things Out.[3110.093] [3110.093][S03]Because I think those books more than any other I've read,[3112.51] [3112.51][S03]and maybe there are others like that,[3114.052] [3114.052][S03]show how exhilarating it is to be at the frontier of knowledge[3118.03] [3118.03][S03]and what that means.[3118.9] [3118.9][S03]I'm feeling getting goosebumps even just talking about that,[3121.4] [3121.4][S03]and it instilled that in me when I was--[3123.325] [3123.325][S03]I can't remember what age, like 10, or 11, or something.[3125.87] [3125.87][S03]And I think it would be great for school kids[3127.745] [3127.745][S03]to be exposed to that, how incredible and fun doing science[3131.74] [3131.74][S03]should be, and I think Feynman was[3133.193] [3133.193][S03]one of those people who had a lot of fun[3134.86] [3134.86][S03]doing his incredible science and as a great role model for that.[3138.203] [3138.203][S04] Thank you very much to all of you, very much indeed.[3140.87] [3140.87][S04]Thank you.[3141.37] [3141.37][S04]Well, as always with Demis, there[3143.38] [3143.38][S04]was a dizzying array of topics that were covered there.[3147.4] [3147.4][S04]We had room temperature superconductors.[3150.25] [3150.25][S04]We had nuclear pore complex.[3152.2] [3152.2][S04]We had molecular syringes.[3154.18] [3154.18][S04]There was plastic eating enzymes, drug design,[3157.69] [3157.69][S04]even poker bluffing strategies.[3160.5] [3160.5][S04]But I think for me, the most standout[3163.23] [3163.23][S04]moment from the conversation today[3165.81] [3165.81][S04]is about what it means to be a scientist in the era[3169.89] [3169.89][S04]of artificial intelligence, because this event was co-hosted[3174.15] [3174.15][S04]by the Royal Society, the Royal Society, which was founded[3177.3] [3177.3][S04]during the era of great lone geniuses,[3181.08] [3181.08][S04]standing on the shoulders of giants, as the phrase goes.[3184.45] [3184.45][S04]But I think that maybe we need to start[3186.99] [3186.99][S04]accepting that those days of lone geniuses[3189.84] [3189.84][S04]really are behind us.[3192.13] [3192.13][S04]Because I think, if we are going to address the biggest[3194.85] [3194.85][S04]challenges that society faces, climate change, energy, disease,[3199.93] [3199.93][S04]our understanding of the universe,[3201.7] [3201.7][S04]then what we really need is big, talented teams[3205.9] [3205.9][S04]of scientists who are working across traditional discipline[3209.14] [3209.14][S04]boundaries.[3210.17] [3210.17][S04]Politics has never needed science more,[3212.99] [3212.99][S04]but I think neither has the world.[3215.78] [3215.78][S04]And I think we need this collaboration[3217.72] [3217.72][S04]between our public institutions, between governments, our health[3221.26] [3221.26][S04]providers, and also the private sector,[3224.3] [3224.3][S04]because innovation is a pursuit for people, but also by people.[3231.43] [3231.43][S04]You've been listening to Google DeepMind, The Podcast,[3234.22] [3234.22][S04]with me, Professor Hannah Fry.[3235.96] [3235.96][S04]If you enjoyed that episode, do subscribe[3238.21] [3238.21][S04]to our YouTube channel.[3239.48] [3239.48][S04]You can also find us on your favorite podcast platform.[3242.99] [3242.99][S04]And we have got plenty more episodes[3245.08] [3245.08][S04]on a whole range of topics to come, so do check those out too.[3248.81] [3248.81][S04]See you next time.[3249.91] [3249.91][MUSIC PLAYING][3252.36]"}
{"file_name": "audio/val_000020.wav", "transcription": "[0.0][MUSIC PLAYING][3.311] [6.127][S02] Welcome to \"Google DeepMind, the Podcast\" with me,[8.71] [8.71][S02]your host, Professor Hannah Fry.[10.38] [10.38][S02]Now, when we first started thinking[12.18] [12.18][S02]about making this podcast way back in 2017,[15.82] [15.82][S02]DeepMind was this relatively small, focused AI research lab.[20.35] [20.35][S02]They'd just been bought by Google[21.9] [21.9][S02]and given the freedom to do their own quirky research[24.93] [24.93][S02]projects from the safe distance of London.[28.02] [28.02][S02]How things have changed.[29.65] [29.65][S02]Because since the last season, Google[31.26] [31.26][S02]has reconfigured its entire structure,[33.61] [33.61][S02]putting AI and the team at DeepMind[35.85] [35.85][S02]at the core of its strategy.[37.95] [37.95][S02]Google DeepMind has continued its quest[40.5] [40.5][S02]to endow AI with human-level intelligence,[44.16] [44.16][S02]known as artificial general intelligence, or AGI.[47.34] [47.34][S02]It has introduced a family of powerful new AI models called[51.15] [51.15][S02]Gemini, as well as an AI agent called[53.76] [53.76][S02]Project Astra that can process audio, video, image, and code.[58.39] [58.39][S02]The lab is also making huge leaps[61.74] [61.74][S02]in applying AI to a host of scientific domains,[64.76] [64.76][S02]including a brand new third version of AlphaFold,[67.73] [67.73][S02]which can predict the structures of all of the molecules[70.33] [70.33][S02]that you will find in the human body, not just proteins.[73.39] [73.39][S02]And in 2021, they spun off a new company, Isomorphic Labs,[77.84] [77.84][S02]to get down to the business of discovering[80.05] [80.05][S02]new drugs to treat diseases.[82.37] [82.37][S02]Google DeepMind is also working on powerful AI agents that[85.84] [85.84][S02]can learn to perform tasks by themselves using reinforcement[89.8] [89.8][S02]learning, and continuing that legacy of AlphaGo's[93.19] [93.19][S02]famous victory over a human in the game of Go.[96.59] [96.59][S02]Now, of course, you'll all have been following this podcast[99.927] [99.927][S02]since the beginning.[100.76] [100.76][S02]You'll all be familiar with the stories[102.94] [102.94][S02]behind all of those changes.[104.72] [104.72][S02]But just in case you are coming to us fresh, welcome.[108.2] [108.2][S02]You can find our first award-winning previous seasons[111.778] [111.778][S02]on Google DeepMind's YouTube channel,[113.32] [113.32][S02]or wherever you get your podcasts.[115.13] [115.13][S02]They also, those episodes go into detail[117.58] [117.58][S02]about a lot of the themes that we're[119.53] [119.53][S02]going to hear come up over and over again from the people[122.2] [122.2][S02]here, like reinforcement learning, deep learning,[125.03] [125.03][S02]large language models and, so on.[126.59] [126.59][S02]So have a listen.[128.229] [128.229][S02]They are really good, even if we do say so ourselves.[131.63] [131.63][S02]Now, all of the newfound attention[134.53] [134.53][S02]on AI since the last series does mean that there are quite a few[138.85] [138.85][S02]more podcasts out there for you to choose from.[141.35] [141.35][S02]But on this podcast, in just the same way as we always have,[144.86] [144.86][S02]we want to offer you something a little bit different.[147.11] [147.11][S02]We want to take you right to the heart of where these ideas are[151.12] [151.12][S02]coming from to introduce you to the people who[154.57] [154.57][S02]are leading the design of our collective future--[158.18] [158.18][S02]no hype, no spin, just compelling discussions[161.62] [161.62][S02]and grand scientific ambition.[164.3] [164.3][S02]So with all of that in mind, I am here[166.81] [166.81][S02]with the DeepMind co-founder and now CEO[169.66] [169.66][S02]of Google DeepMind, Demis Hassabis.[172.21] [172.21][S02]So with all of that in mind, do I have to call you Sir Demis[176.15] [176.15][S02]now?[176.65] [176.65][S01] No, absolutely not.[178.575] [178.575][S02] OK.[179.2] [179.2][S02]Well, Demis, welcome to the podcast.[180.59] [180.59][S01] Thank you.[181.39] [181.39][S02] Thank you very much for being here.[183.348] [183.348][S02]OK, I want to know, is your job easier or harder[186.37] [186.37][S02]now that there has been this explosion in public interest?[191.14] [191.14][S01] I think it's double edged.[193.04] [193.04][S01]I think it's harder because there's[195.13] [195.13][S01]just so much scrutiny, focus, and actually quite[197.38] [197.38][S01]a lot of noise in the whole field.[200.09] [200.09][S01]I actually preferred it when it was less people,[202.63] [202.63][S01]and maybe a little bit more focused on the science.[206.23] [206.23][S01]But it's also good because it shows that the technology is[209.86] [209.86][S01]ready to impact the real world in many different ways,[213.16] [213.16][S01]and impact people's everyday lives in positive ways.[216.29] [216.29][S01]So I think it's exciting, too.[217.882] [217.882][S02] Have you been surprised[219.34] [219.34][S02]by how quickly this has caught the public's imagination?[221.673] [221.673][S02]I mean, I guess you would have expected that eventually people[224.68] [224.68][S02]would have got on board.[225.68] [225.68][S01] Yes, exactly.[226.888] [226.888][S01]So at some point, those of us who've[230.47] [230.47][S01]been working on it like us for many years now, even decades,[234.56] [234.56][S01]so I guess at some point the general public[237.19] [237.19][S01]would wake up to that fact.[238.58] [238.58][S01]And effectively, everyone's starting[240.76] [240.76][S01]to realize how important AI is going to be.[242.99] [242.99][S01]But it's been quite surreal still[244.54] [244.54][S01]to see that actually come to fruition, and for that[247.45] [247.45][S01]to happen.[248.47] [248.47][S01]And I guess it is the advent of the chat bots and language[251.32] [251.32][S01]models because everyone, of course, uses language.[254.24] [254.24][S01]Everyone can understand language.[255.8] [255.8][S01]So it's an easy way for the general public[258.04] [258.04][S01]to understand and maybe measure where AI has got to.[262.24] [262.24][S02] I heard you describe these chat bots[264.34] [264.34][S02]as though they were unreasonably effective, which I really like.[267.903] [267.903][S02]And actually, later in the podcast[269.32] [269.32][S02]we are going to be discussing transformers, which[271.39] [271.39][S02]was the big breakthrough, I guess-- the big advance[274.87] [274.87][S02]that gave us those tools.[277.07] [277.07][S02]But tell me first, what do you mean by unreasonably effective?[281.145] [281.145][S01] What I mean by it[282.52] [282.52][S01]is I suppose if one were to wind back 5, 10 years ago,[286.27] [286.27][S01]and you were to say the way we're[289.27] [289.27][S01]going to go about this is build these amazing architectures,[292.48] [292.48][S01]and then scale from there, and not necessarily crack[295.81] [295.81][S01]specific things like concepts or abstractions.[299.878] [299.878][S01]These are a lot of debates we would have 5,[301.67] [301.67][S01]10 years ago is do you need a special way of doing[304.09] [304.09][S01]abstractions?[305.19] [305.19][S01]The brain certainly seems to do that.[307.35] [307.35][S01]But yet somehow, the systems, if you give them enough data--[310.39] [310.39][S01]i.e.[310.89] [310.89][S01]The whole internet-- then they do[313.47] [313.47][S01]seem to learn this and generalize[315.3] [315.3][S01]from those examples-- not just rote memorize,[318.31] [318.31][S01]but actually somewhat understand what they're processing.[323.76] [323.76][S01]And it's a little bit unreasonably effective[325.95] [325.95][S01]in the sense that I don't think anyone[327.84] [327.84][S01]would have thought that it would work as well as it has done,[330.96] [330.96][S01]say, five years ago.[332.02] [332.02][S02] Yeah.[332.728] [332.728][S02]I suppose it is a surprise that things[334.47] [334.47][S02]like conceptual understanding and abstraction[337.29] [337.29][S02]have emerged rather than been--[338.715] [338.715][S01] Yes, and we would have been--[340.59] [340.59][S01]probably we discussed last time things like concepts[343.41] [343.41][S01]and grounding--[345.374] [345.374][S01]grounding language in real world experience,[348.1] [348.1][S01]maybe in simulations or as robots embodied intelligence,[352.08] [352.08][S01]would have been necessary to really understand[355.44] [355.44][S01]the world around us.[356.92] [356.92][S01]And of course, these systems are not there yet.[360.13] [360.13][S01]They make lots of mistakes.[361.33] [361.33][S01]They don't really have a proper model of the world,[364.41] [364.41][S01]but they've got a lot further than one might expect just[368.26] [368.26][S01]by learning from language.[370.03] [370.03][S02] I guess we probably should actually[371.23] [371.23][S02]say what grounding is for those who haven't listened[373.397] [373.397][S02]to series 1 and series 2.[374.53] [374.53][S02]Because this was a big thing.[375.74] [375.74][S02]I mean, we were talking about this a lot.[377.81] [377.81][S02]So do you want to just give us an overview[379.18] [379.18][S02]of what grounding is?[380.055] [380.055][S01] Grounding is when--[383.08] [383.08][S01]one of the reasons the systems that were built[385.24] [385.24][S01]in the '80s and '90s, the classical AI systems built[388.09] [388.09][S01]at places like MIT, they were big logic systems.[391.04] [391.04][S01]So you can imagine them as huge databases of words[393.37] [393.37][S01]connected to other words.[394.76] [394.76][S01]And the problem was you could say something, a dog has legs,[398.99] [398.99][S01]and that would be in the database.[400.46] [400.46][S01]But the problem was, as soon as you showed it[402.335] [402.335][S01]a picture of a dog, it had no idea that collection of pixels[405.07] [405.07][S01]was referring to that symbol.[406.735] [406.735][S01]And that's the grounding problem.[408.11] [408.11][S01]So you have this symbolic representation,[410.09] [410.09][S01]this abstract representation, but what does it really[412.3] [412.3][S01]mean in the real world--[413.93] [413.93][S01]in the messy real world?[415.34] [415.34][S01]And then, of course, they tried to fix that,[417.62] [417.62][S01]but you never get that quite right.[419.77] [419.77][S01]And instead of that, of course, today's systems,[422.03] [422.03][S01]they're directly learning from the data.[424.31] [424.31][S01]So in a way, they're forming that connection[426.46] [426.46][S01]from the beginning.[427.4] [427.4][S01]But the interesting thing was that if you learn just[429.79] [429.79][S01]from language, in theory, there should[432.16] [432.16][S01]be missing a lot of the grounding that you need.[434.53] [434.53][S01]But it turns out that a lot of it is inferrable somehow.[437.3] [437.3][S02] Why, in theory?[438.49] [438.49][S01] Well, because where[439.33] [439.33][S01]is that grounding coming from?[440.39] [440.39][S01]These systems, at least the first large language models--[442.81] [442.81][S02] Don't exist in the real world.[443.56] [443.56][S01] --don't exist in the real world.[444.74] [444.74][S01]They're not connected to simulators.[446.36] [446.36][S01]They're not connected to robots.[447.98] [447.98][S01]They don't have any access to even--[450.323] [450.323][S01]they weren't multimodal to begin with, either.[452.24] [452.24][S01]They don't have access to the visuals or anything else.[455.78] [455.78][S01]It's just purely they live in language space.[457.91] [457.91][S01]So they're learning in an abstract domain,[461.27] [461.27][S01]so it's pretty surprising they can then infer some things[464.17] [464.17][S01]about the real world from that.[465.673] [465.673][S02] Which makes sense if the grounding gets in by people[468.34] [468.34][S02]interacting with the system and saying[469.75] [469.75][S02]that's a rubbish answer, that's a good answer.[471.5] [471.5][S01] Yes.[472.333] [472.333][S01]So for sure, part of that, if the[473.95] [473.95][S01]question that they're getting wrong,[475.52] [475.52][S01]the early versions of this, was due to grounding missing--[479.35] [479.35][S01]actually, the real world dogs bark in this way[481.57] [481.57][S01]or whatever it is-- and it's answering it incorrectly,[484.28] [484.28][S01]then that feedback will correct it.[486.2] [486.2][S01]And part of that feedback is from our own grounded knowledge.[489.77] [489.77][S01]So some grounding is seeping in like that for sure.[492.4] [492.4][S02] I remember seeing a really nice example[494.525] [494.525][S02]about crossing the English Channel versus walking[496.87] [496.87][S02]across the English Channel.[498.22] [498.22][S01] Exactly, those kinds of things.[499.49] [499.49][S01]And if it answered wrong, you would tell it it's wrong.[501.782] [501.782][S01]And then it would have to slightly figure out[503.77] [503.77][S01]that you can't walk across the Channel.[506.285] [506.285][S02] So some of these properties[507.91] [507.91][S02]that have emerged that weren't necessarily[510.1] [510.1][S02]expected to be, I want to ask you a little bit about hype.[513.44] [513.44][S02]Do you think that where we are right now,[516.77] [516.77][S02]how things are at this moment, is overhyped or underhyped?[521.679] [521.679][S02]Or is it just hyped, perhaps, in the wrong direction?[524.125] [524.125][S01] Yeah, I think it's more the latter.[526.25] [526.25][S01]So I would say that in the near term, it's hyped too much.[531.87] [531.87][S01]So I think people are claiming can do all sorts of things it[534.37] [534.37][S01]can't.[534.87] [534.87][S01]There's all sorts of startups and VC money chasing crazy ideas[539.05] [539.05][S01]that are just not ready.[541.48] [541.48][S01]On the other hand, I think it's still underhyped.[543.522] [543.522][S02] Coming from you, Demis--[545.022] [545.022][S01] Yes, I know, I know, I know.[547.01] [547.01][S02] AI in 2010.[548.11] [548.11][S01] Exactly, exactly.[549.485] [549.485][S01]But I think it's still underhyped or perhaps[552.96] [552.96][S01]underappreciated still even now what's[555.63] [555.63][S01]going to happen when we get to AGI and post-AGI.[558.66] [558.66][S01]I still don't feel like that's people[561.342] [561.342][S01]are quite understood how enormous that's going to be,[563.55] [563.55][S01]and therefore, the responsibility of that.[566.35] [566.35][S01]So it's both, really.[568.0] [568.0][S01]I think it's a little bit overhyped[569.58] [569.58][S01]in the near term at the moment.[572.56] [572.56][S01]We're going through that cycle.[574.207] [574.207][S02] I guess, though, so in terms of all of these[576.54] [576.54][S02]potential startups, and VC funding,[578.22] [578.22][S02]and so on, you who have lived and breathed this stuff[580.68] [580.68][S02]for, as you say, decades, are very well placed to spot which[584.88] [584.88][S02]ones are realistic goals and which ones aren't.[587.53] [587.53][S02]But for other people, how can they distinguish between[591.45] [591.45][S02]what's real and what isn't?[593.71] [593.71][S01] Yeah, well look, I think you need to look at--[597.63] [597.63][S01]obviously you've got to do your technical due diligence, have[601.29] [601.29][S01]some understanding of the technology,[602.91] [602.91][S01]and the latest trends.[605.24] [605.24][S01]I think also look at, perhaps, the background of the people[608.33] [608.33][S01]saying it, how technical they are.[610.08] [610.08][S01]Have they just arrived in AI last year from somewhere else?[612.898] [612.898][S01]I don't know.[613.44] [613.44][S01]They were doing crypto last year.[614.87] [614.87][S01]These might be some clues that perhaps they're[618.41] [618.41][S01]jumping on a bandwagon.[619.73] [619.73][S01]And it doesn't mean to say, of course, they[621.56] [621.56][S01]could still have some good ideas, and many will do.[624.06] [624.06][S01]But it's a bit more lottery ticket like, shall we say.[627.75] [627.75][S01]And I think that always happens when there's a ton of attention[632.077] [632.077][S01]suddenly on a place, and obviously, then the money[634.16] [634.16][S01]follows that.[635.73] [635.73][S01]And everyone feels like they're missing out.[637.83] [637.83][S01]And that creates a kind of opportunistic,[641.79] [641.79][S01]shall we say, environment, which is a little bit[643.79] [643.79][S01]opposite to those of us who've been in for decades[646.49] [646.49][S01]in a deep technology, deep science way, which is ideally[650.21] [650.21][S01]the way I think we need to carry on going as we get closer[653.21] [653.21][S01]to AGI.[653.94] [653.94][S02] Yeah.[654.36] [654.36][S02]And I guess one of the big things that we're[655.64] [655.64][S02]going to talk about in this series[657.057] [657.057][S02]is Gemini, which really comes from that very deep science[660.47] [660.47][S02]approach, I guess.[662.945] [662.945][S02]In what ways is Gemini different from the other large language[666.15] [666.15][S02]models that are released by other labs?[668.02] [668.02][S01] So from the beginning with Gemini,[670.12] [670.12][S01]we wanted it to be multi-modal from the start so it could[674.01] [674.01][S01]process not just language, but also audio, video, image, code--[679.09] [679.09][S01]any modality, really.[680.91] [680.91][S01]And the reason we wanted to do that was firstly,[683.8] [683.8][S01]we think that's the way to get these systems to actually[686.19] [686.19][S01]understand the world around them and build better world models.[689.68] [689.68][S01]So actually still going back to our grounding[691.8] [691.8][S01]question earlier, still building grounding in,[695.79] [695.79][S01]but piggybacking on top of language this time.[699.06] [699.06][S01]And so that's important.[700.597] [700.597][S01]And we also had this vision in the end[702.18] [702.18][S01]of having a universal assistant, and we[705.21] [705.21][S01]prototyped something called Astro,[706.66] [706.66][S01]which I'm sure we'll talk about, which[708.54] [708.54][S01]understands not just what you're typing, but actually the context[711.9] [711.9][S01]you're in.[712.45] [712.45][S01]And if you think about something like a personal assistant[714.87] [714.87][S01]or digital assistant, it will be much more useful the more[718.17] [718.17][S01]context it understood about what you're asking it for[720.84] [720.84][S01]or the situation that you're in.[722.92] [722.92][S01]So we always thought that would be a much more useful type[727.29] [727.29][S01]of system, and so we built multi-modality[729.69] [729.69][S01]in from the start.[730.54] [730.54][S01]So that was one thing, natively multi-modal.[732.67] [732.67][S01]And then at the time, that was the only model doing that.[735.88] [735.88][S01]So now the other models are trying to catch up.[738.27] [738.27][S01]And then the other big innovations[739.71] [739.71][S01]we had are on memory.[740.8] [740.8][S01]So long context.[742.21] [742.21][S01]So actually holding in mind 1 million-- or 2 million now--[746.43] [746.43][S01]tokens, you can think of them as more or less like words,[749.19] [749.19][S01]in mind.[749.89] [749.89][S01]So you can give it \"War and Peace,\" or even a whole--[753.21] [753.21][S01]because it's multi-modal-- a whole video now, a whole film,[756.396] [756.396][S01]or a lecture, and then get it to answer questions or find you[759.33] [759.33][S01]things within that video stream.[761.17] [761.17][S02] OK Project Astra, that's[762.67] [762.67][S02]the new universal AI agent, the one that can[766.26] [766.26][S02]take in video and audio data.[769.53] [769.53][S02]At Google.[770.14] [770.14][S02]I/O, I think you used the example of how Astra[772.89] [772.89][S02]could help you remember where you left[774.72] [774.72][S02]your glasses, for instance.[776.79] [776.79][S02]So I wonder, though, about the lineage[780.03] [780.03][S02]of this stuff because is this just a fancy, advanced version[786.12] [786.12][S02]of those old Google glasses?[787.405] [787.405][S01] So, of course, Google have a long history[789.78] [789.78][S01]of developing glass-type devices actually back to,[793.23] [793.23][S01]I think, 2012 or something.[794.53] [794.53][S01]So they were way ahead of the curve.[796.03] [796.03][S01]But maybe it was just missing this kind of technology[799.42] [799.42][S01]So you could actually understand-- a smart agent,[801.583] [801.583][S01]or a smart assistant that could actually[803.25] [803.25][S01]understand what it's seeing.[804.79] [804.79][S01]And so we're very excited about that digital assistant[808.99] [808.99][S01]to go around with you and understand the world around you.[811.78] [811.78][S01]So it seems a really-- when you use it,[814.3] [814.3][S01]it feels a really natural use case.[816.605] [816.605][S02] OK.[817.23] [817.23][S02]I want to rewind a tiny bit to the start of Gemini[822.06] [822.06][S02]because it came from two separate parts[825.03] [825.03][S02]of the organization.[826.0] [826.0][S01] Yes.[826.833] [826.833][S01]So we-- actually, last year we combined our two research[830.88] [830.88][S01]divisions at Alphabet.[832.69] [832.69][S01]So obviously, the old DeepMind, and then Google Brain into one[837.04] [837.04][S01]we call it super unit, bringing all the talent together--[840.3] [840.3][S01]that amazing talent we have across the company,[842.59] [842.59][S01]across the whole of Google, into one unified unit.[846.42] [846.42][S01]And what it meant was that we combined all the best[850.56] [850.56][S01]knowledge that we had from all the research[853.89] [853.89][S01]we were doing, but especially on language models.[855.94] [855.94][S01]So we had Chinchilla, and Gopher, and things like that,[858.73] [858.73][S01]and they were building things like PaLM, and LaMDA,[861.12] [861.12][S01]and early language models.[862.28] [862.28][S01]And they had different strengths and weaknesses,[864.28] [864.28][S01]and we pulled them all together into what became Gemini[867.24] [867.24][S01]as the first Lighthouse project that the combined group would[871.62] [871.62][S01]output.[872.193] [872.193][S01]And then the other important thing, of course,[874.11] [874.11][S01]was bringing together all the compute,[876.51] [876.51][S01]as well, so that we could, do these really massive training[879.72] [879.72][S01]runs and actually pull the compute resources together.[882.79] [882.79][S01]So it's been great.[883.787] [883.787][S02] I guess, in a lot of ways,[885.37] [885.37][S02]the focus of Google Brain and DeepMind was slightly different.[888.63] [888.63][S02]Is that fair to say?[889.57] [889.57][S01] Yeah.[890.445] [890.445][S01]So I think it was.[892.47] [892.47][S01]I mean, we were obviously focused, both of us,[894.63] [894.63][S01]on the frontiers of AI, and there[896.43] [896.43][S01]was a lot of collaborations already on a individual research[899.55] [899.55][S01]level, but maybe not on a strategic level.[902.17] [902.17][S01]Obviously, now the combined group, Google DeepMind,[904.6] [904.6][S01]I describe it as we're the engine room of Google now.[907.91] [907.91][S01]But it's worked really well.[909.2] [909.2][S01]I think there were a lot more similarities, actually,[911.408] [911.408][S01]in the way we were working than there were differences,[914.17] [914.17][S01]and we've continued to keep and double down our strengths[918.34] [918.34][S01]on things fundamental research.[920.24] [920.24][S01]So where does the next transformer architecture[922.83] [922.83][S01]come from?[923.33] [923.33][S01]We want to invent that.[924.288] [924.288][S01]Obviously, Google Brain invented the previous one.[926.927] [926.927][S01]We combined it with deep reinforcement[928.51] [928.51][S01]learning that we pioneered, and I still[931.21] [931.21][S01]think more innovations are going to be needed.[933.29] [933.29][S01]And I would back us to do that just as we've[935.38] [935.38][S01]done in the past 10 years collectively,[938.09] [938.09][S01]both Brain and DeepMind.[939.59] [939.59][S01]So it's been exciting.[941.82] [941.82][S02] I want to come back to that merge in a moment.[944.29] [944.29][S02]But I think just sticking on Gemini[946.21] [946.21][S02]for a second, how good is it?[948.29] [948.29][S02]How does it compare to other models?[950.21] [950.21][S01] Yeah, well, I think some of the benchmarks[952.9] [952.9][S01]are not--[953.743] [953.743][S01]the problem is that we need more--[955.16] [955.16][S01]I think this is one thing the whole field needs is[956.89] [956.89][S01]much better benchmarks.[957.805] [957.805][S02] Yeah.[958.513] [958.513][S02]How do you decide?[959.302] [959.302][S01] Well, there are some well-known benchmarks,[961.76] [961.76][S01]academic ones.[962.87] [962.87][S01]But they're getting saturated now,[964.682] [964.682][S01]and they don't really differentiate[966.14] [966.14][S01]between the nuances, between the different top models.[969.93] [969.93][S01]I would say there's three models that[971.63] [971.63][S01]are at the top, the frontier.[974.28] [974.28][S01]So it's Gemini from us, OpenAI's, GPT, of course,[978.717] [978.717][S01]and then Anthropic with their Claude models.[980.55] [980.55][S01]And then obviously, there's a bunch of other good models,[983.0] [983.0][S01]too, that people like Meta, and Mistral, and others built,[986.368] [986.368][S01]and they're differently good at different things.[988.41] [988.41][S01]It depends what you want--[990.05] [990.05][S01]coding, perhaps that's Claude.[991.55] [991.55][S01]And reasoning, maybe that's GPT.[993.35] [993.35][S01]And then memory stuff, long context,[995.51] [995.51][S01]and multimodal understanding, that would be Gemini.[998.09] [998.09][S01]Of course, we're continuing to-- all[999.62] [999.62][S01]of us are improving our models all the time.[1001.7] [1001.7][S01]So given where we started from, which Gemini as a project[1006.76] [1006.76][S01]only existed for a year, obviously,[1009.28] [1009.28][S01]based on some of our other projects,[1011.06] [1011.06][S01]I think our trajectory is very good.[1013.64] [1013.64][S01]So when we talk next time, we should hopefully[1016.57] [1016.57][S01]be right at the forefront.[1018.672] [1018.672][S02] Because there is still a way to go.[1020.63] [1020.63][S02]I mean, there are still some things[1022.088] [1022.088][S02]that these models aren't very good at.[1023.72] [1023.72][S01] Yes, for sure.[1024.68] [1024.68][S01]And actually, that's the big debate right now.[1026.75] [1026.75][S01]So this last set of things emerged[1029.829] [1029.829][S01]from the technologies that were invented five, six years ago.[1034.16] [1034.16][S01]The question is, they're still missing a ton of things--[1036.589] [1036.589][S01]so their factuality, they hallucinate, as we know.[1040.089] [1040.089][S01]They also not good at planning yet.[1042.579] [1042.579][S02] Planning in what sense?[1044.27] [1044.27][S01] Well, long term planning.[1046.099] [1046.099][S01]So they can't problem solve.[1048.55] [1048.55][S01]Something long term, you give it an objective,[1051.05] [1051.05][S01]they can't really do actions in the world for you.[1053.66] [1053.66][S01]So they're very much like passive Q&A systems.[1056.75] [1056.75][S01]You put the energy in by asking the question,[1058.84] [1058.84][S01]and then they give you some kind of response.[1061.99] [1061.99][S01]But they're not able to solve a problem for you.[1066.823] [1066.823][S01]You can't say something like, if you wanted it[1068.74] [1068.74][S01]as a digital assistant, you might want to say something[1071.298] [1071.298][S01]like, book me that holiday in Italy, and all the restaurants,[1073.84] [1073.84][S01]and the museums, and whatever, and it knows what you like,[1077.07] [1077.07][S01]but then it goes out and books the flights[1078.82] [1078.82][S01]and all of that for you.[1080.0] [1080.0][S01]So it can't do any of that.[1082.07] [1082.07][S01]But I think that's the next era--[1084.07] [1084.07][S01]these more agent-based systems, we[1086.02] [1086.02][S01]would call them, or agentic systems[1087.79] [1087.79][S01]that have agent-like behavior.[1090.288] [1090.288][S01]But of course, that's what we're expert in.[1092.08] [1092.08][S01]That's what we used to build with all our game agents--[1094.522] [1094.522][S01]AlphaGo and all of the other things[1095.98] [1095.98][S01]we've talked in about in the past.[1097.73] [1097.73][S01]So a lot of what we're doing is marrying[1100.66] [1100.66][S01]that work that we're, I guess, famous[1103.33] [1103.33][S01]for with the new large multimodal models.[1108.26] [1108.26][S01]And I think that's going to be the next generation of systems.[1112.52] [1112.52][S01]You can think of it as combining AlphaGo with Gemini.[1114.767] [1114.767][S02] Yeah, because I guess AlphaGo was very, very[1117.1] [1117.1][S02]good at planning.[1117.74] [1117.74][S01] Yes, it was very good at planning.[1118.82] [1118.82][S01]Of course, only in the domain, though, of games.[1121.07] [1121.07][S01]And so we need to generalize that[1123.49] [1123.49][S01]into the general domain of everyday workloads and language.[1128.897] [1128.897][S02] You mentioned a minute ago[1130.48] [1130.48][S02]how Google DeepMind is now the engine room of Google.[1134.27] [1134.27][S02]I mean, that is quite a big shift[1136.09] [1136.09][S02]since I was last here in the last couple of years ago.[1138.34] [1138.34][S02]Is Google taking quite a big gamble on you?[1140.455] [1140.455][S01] Well, I guess so.[1141.83] [1141.83][S01]I mean, I think Google have always understood[1145.0] [1145.0][S01]the importance of AI.[1147.01] [1147.01][S01]Sundar, when he took over as CEO,[1148.9] [1148.9][S01]said that Google was an AI-first company.[1152.45] [1152.45][S01]And we discussed that very early on in his tenure,[1155.12] [1155.12][S01]and he saw the potential in AI as the next big paradigm[1157.99] [1157.99][S01]shift after mobile and internet, but bigger than those things.[1160.992] [1160.992][S01]But then I think maybe in the last year or two,[1162.95] [1162.95][S01]we've really started living what that means--[1165.77] [1165.77][S01]not just from a research perspective,[1167.36] [1167.36][S01]but also from products and other things.[1169.97] [1169.97][S01]So it's very exciting, but I think[1172.87] [1172.87][S01]it's the right bet for us to coordinate all of our talents[1176.71] [1176.71][S01]together, and then push as hard as possible.[1179.983] [1179.983][S02] And then how about the other way around?[1182.15] [1182.15][S02]Because I guess from DeepMind, having that very strong research[1185.71] [1185.71][S02]and science focus, does becoming the engine room for Google now[1190.33] [1190.33][S02]mean that you have to care much more about commercial interests[1193.78] [1193.78][S02]rather than the purer stuff that--[1195.367] [1195.367][S01] Yeah, well, we do definitely[1197.2] [1197.2][S01]have to worry more about, and it's[1200.35] [1200.35][S01]in our remit now, the commercial interests.[1202.45] [1202.45][S01]But actually, there's a couple things say about that.[1205.4] [1205.4][S01]First of all, we're continuing on with our science work[1208.27] [1208.27][S01]in AlphaFolds, and you just saw AlphaFold 3 come out.[1211.45] [1211.45][S01]And we're doubling down on our investments there.[1214.42] [1214.42][S01]That's, I think, a unique thing that we do at Google DeepMind[1217.9] [1217.9][S01]now.[1219.61] [1219.61][S01]And even our competitors point at those things[1221.68] [1221.68][S01]as universal goods, if you like, that come out of AI.[1225.942] [1225.942][S01]And that's going really well.[1227.15] [1227.15][S01]And we spun out isomorphic to do drug discovery.[1229.82] [1229.82][S01]So it's very exciting, and that's all going really well.[1232.88] [1232.88][S01]And so we're going to continue to do that.[1234.76] [1234.76][S01]And then was all our work on climate and all of these things.[1238.022] [1238.022][S01]But then, we're quite a large team,[1239.48] [1239.48][S01]so we can do more than one things at once.[1241.23] [1241.23][S01]We're also building our large models, Gemini and et cetera,[1244.13] [1244.13][S01]and then we have a product team that we're building out[1246.82] [1246.82][S01]that is going to bring all this amazing technology to all[1251.86] [1251.86][S01]of the surfaces that Google has.[1253.64] [1253.64][S01]So it's an incredible privilege, in a way,[1256.6] [1256.6][S01]to have that there to plug in all of our stuff.[1259.94] [1259.94][S01]And we invent something, it immediately[1261.67] [1261.67][S01]can become useful to a billion people.[1263.84] [1263.84][S01]And so that's really motivating.[1266.33] [1266.33][S01]And actually, the other thing is there's a lot more convergence[1271.16] [1271.16][S01]now between the technology we need to develop for a product[1274.67] [1274.67][S01]to have AI in it and what you would do for pure AGI research[1278.33] [1278.33][S01]purposes.[1279.06] [1279.06][S01]So there's not really-- five years ago, you'd[1281.57] [1281.57][S01]have had to build some special case AI for a product.[1284.93] [1284.93][S01]Now, you can branch off your main research[1287.412] [1287.412][S01]and, of course, you still need to do some things that[1289.62] [1289.62][S01]are product specific, but maybe it's only 10% of the work.[1292.38] [1292.38][S01]So there's actually not that tension[1294.14] [1294.14][S01]anymore between what you would develop for an AI product[1298.23] [1298.23][S01]and what you would develop for trying to build AGI.[1300.75] [1300.75][S01]It's 90%, I would say, the same research program.[1305.88] [1305.88][S01]And then finally, of course, if you do products,[1308.4] [1308.4][S01]and you get them out into the world,[1309.9] [1309.9][S01]you learn a lot from that.[1311.66] [1311.66][S01]And people using it, and you learn[1313.4] [1313.4][S01]a lot about, oh, your internal metrics[1315.32] [1315.32][S01]don't quite match what people are saying,[1317.423] [1317.423][S01]so then you can update that.[1318.59] [1318.59][S01]And that's really helpful for your research.[1320.872] [1320.872][S02] Absolutely.[1321.83] [1321.83][S02]Well, OK, we are going to talk a lot more in this podcast[1325.34] [1325.34][S02]about those breakthroughs that have come from applying AI[1328.64] [1328.64][S02]to science, but I want to ask you about that tension[1331.73] [1331.73][S02]that there is between knowing when the right moment is[1335.33] [1335.33][S02]to release something to the public.[1337.11] [1337.11][S02]Because internally at DeepMind, those tools[1340.86] [1340.86][S02]like large language models were being used for research[1344.09] [1344.09][S02]rather than being seen as a potentially commercial thing.[1347.442] [1347.442][S01] Yeah, that's right.[1348.9] [1348.9][S01]So as you know, we've always taken responsibility incredibly[1352.43] [1352.43][S01]seriously here, and safety, right from the beginning,[1356.15] [1356.15][S01]way back when we started in 2010 and before that.[1358.95] [1358.95][S01]And Google then adopted some of our, basically, ethics[1362.745] [1362.745][S01]charter effectively into their AI principles.[1364.62] [1364.62][S01]So we've always been well aligned with the whole of Google[1367.037] [1367.037][S01]and wanting to be responsible about deploying this as one[1369.74] [1369.74][S01]of the leaders in this space.[1371.81] [1371.81][S01]And so it's been interesting now starting[1374.45] [1374.45][S01]to ship real products with Gen AI in them.[1376.98] [1376.98][S01]Actually there's a lot of learning that is going on,[1380.097] [1380.097][S01]and we're learning fast, which is good[1381.68] [1381.68][S01]because we're at relatively low stakes[1383.263] [1383.263][S01]here with the current technology.[1384.9] [1384.9][S01]So it's not that powerful yet.[1386.49] [1386.49][S01]But as it gets more powerful, we have to be more careful.[1389.63] [1389.63][S01]And that's just learning about the product teams[1392.95] [1392.95][S01]and other groups learning about how to test Gen AI technologies.[1397.5] [1397.5][S01]It's different from a normal piece of technology[1399.5] [1399.5][S01]because it doesn't always do the same thing.[1402.36] [1402.36][S01]It's almost like testing an open world game.[1404.922] [1404.922][S01]It's almost infinite what you can try and do with it,[1407.13] [1407.13][S01]so it's interesting to figure out how do you[1409.73] [1409.73][S01]do the red teaming on it.[1411.0] [1411.0][S02] So red teaming, in this case,[1412.708] [1412.708][S02]being where you're competing against yourselves?[1415.005] [1415.005][S01] Yeah.[1415.88] [1415.88][S01]So red teaming is when you set up a specific separate team[1419.27] [1419.27][S01]from the team that's developed the technology[1421.88] [1421.88][S01]to stress test it and try and break it in any way possible.[1425.4] [1425.4][S01]You actually need to use tools to automate that because nobody[1429.05] [1429.05][S01]can red team-- even if you had thousands of people doing it,[1431.902] [1431.902][S01]that's not enough compared to billions of users[1433.86] [1433.86][S01]when you put it out there.[1434.88] [1434.88][S01]They're going to try all sorts of things.[1436.588] [1436.588][S01]So it's kind of interesting to take that learning,[1440.03] [1440.03][S01]and then improve our processes so that our future launches will[1444.98] [1444.98][S01]be as smooth as possible.[1446.37] [1446.37][S01]And I think we got to do it in stages where there's[1448.82] [1448.82][S01]an experimental phase, then a closed beta, and then launch--[1453.805] [1453.805][S01]a little bit, again, like we used[1455.18] [1455.18][S01]to launch our games back in the day,[1458.09] [1458.09][S01]and learn at each step of the way.[1460.73] [1460.73][S01]And then the other thing we've got to do,[1462.74] [1462.74][S01]and I think we need to do more on, is use AI itself to help us[1467.15] [1467.15][S01]internally with red teaming and actually[1470.33] [1470.33][S01]spotting some errors automatically or triaging[1473.51] [1473.51][S01]that so that, then, our developers and human testers[1477.83] [1477.83][S01]can actually focus on those hard cases.[1480.078] [1480.078][S02] You said something really interesting[1482.12] [1482.12][S02]there about how you're just in a much more probabilistic space[1486.07] [1486.07][S02]here.[1486.57] [1486.57][S02]And then, if there's even a very small chance of something[1489.86] [1489.86][S02]happening, if you have enough tries,[1491.73] [1491.73][S02]eventually, something will go wrong.[1493.23] [1493.23][S02]And I guess there have been a couple of mistakes[1495.23] [1495.23][S02]that-- public mistakes.[1497.04] [1497.04][S01] Yeah, so that's why I think that,[1500.4] [1500.4][S01]as I mentioned, that product teams are just getting[1502.88] [1502.88][S01]used to the sorts of testing.[1504.39] [1504.39][S01]They tested these things, but they[1506.93] [1506.93][S01]have this stochastic nature, probabilistic nature.[1509.37] [1509.37][S01]So in fact, a lot of cases where if it[1512.107] [1512.107][S01]was a normal piece of software, you could say I've[1514.19] [1514.19][S01]tested 99.999% of things, so then extrapolates.[1517.88] [1517.88][S01]So then it's enough because there's[1519.62] [1519.62][S01]no way of exposing the flaw that it has if it has one.[1523.07] [1523.07][S01]But that's not the case with these generative systems.[1525.6] [1525.6][S01]They can do all sorts of things that[1527.36] [1527.36][S01]are a little bit left field, or out of the box,[1530.37] [1530.37][S01]out of distribution, in a way, from what you've seen before[1532.95] [1532.95][S01]if someone clever or adversarial decides to-- it's almost[1536.15] [1536.15][S01]like a hacker decides to test push it in some way.[1539.913] [1539.913][S01]And it could even be--[1540.83] [1540.83][S01]I mean, it's so combinatorial, it could even[1542.48] [1542.48][S01]be with all the things that you've happened[1543.89] [1543.89][S01]to have said before to it.[1545.1] [1545.1][S01]And then it's in some kind of peculiar state[1548.25] [1548.25][S01]which then-- or it's got its memories filled up[1550.25] [1550.25][S01]with this particular thing, and then[1551.75] [1551.75][S01]that's why it outputs something.[1553.59] [1553.59][S01]So there's a lot of complexity there, but it's not infinite.[1558.87] [1558.87][S01]So there's ways to deal with it.[1561.21] [1561.21][S01]But it's just a lot more nuanced than launching[1564.63] [1564.63][S01]normal technology.[1565.383] [1565.383][S02] I remember you saying,[1566.8] [1566.8][S02]I think it was in the first time I interviewed you[1569.82] [1569.82][S02]about how, actually, you have to think that this is a completely[1573.66] [1573.66][S02]different way of computing.[1575.95] [1575.95][S02]You have to move away from the things that we completely[1578.73] [1578.73][S02]understand-- the deterministic stuff--[1580.68] [1580.68][S02]into this much more messy, probabilistic error-ridden[1587.04] [1587.04][S02]place, as well as your testers.[1588.57] [1588.57][S02]Do you think the public slightly has[1590.07] [1590.07][S02]to shift its mindset on the type of computing that we're doing?[1593.23] [1593.23][S01] Yeah, I think so,[1594.75] [1594.75][S01]and maybe that's another thing, interestingly,[1597.51] [1597.51][S01]that we're thinking about is actually putting out[1602.16] [1602.16][S01]a kind of principles document or something before you release[1605.22] [1605.22][S01]something to show what is the expectation from this system.[1609.03] [1609.03][S01]What's it designed for?[1610.465] [1610.465][S01]What's it useful for?[1611.34] [1611.34][S01]What can't it do?[1613.05] [1613.05][S01]And I think there is some sort of education there needed of,[1618.42] [1618.42][S01]you'll be able to find it useful if you do these things with it,[1621.412] [1621.412][S01]but don't try and use it for these other things[1623.37] [1623.37][S01]because it won't work.[1624.43] [1624.43][S01]And I think that that's something[1628.2] [1628.2][S01]that we need to get better at clarifying as a field,[1632.01] [1632.01][S01]and then probably users need to get more experienced on.[1635.61] [1635.61][S01]And actually, this interesting.[1637.36] [1637.36][S01]This is probably why chatbots themselves came a little bit out[1640.08] [1640.08][S01]of the blue.[1641.11] [1641.11][S01]Even obviously ChatGPT, but even to OpenAI, it surprised them.[1644.25] [1644.25][S01]And we had our own chat bots, and Google had theirs.[1646.907] [1646.907][S01]And one of the things was we were looking at them,[1648.99] [1648.99][S01]and we were looking at all the flaws they still had,[1651.76] [1651.76][S01]and they still do.[1652.63] [1652.63][S01]And it's like, well, it's getting these things wrong,[1654.838] [1654.838][S01]and it sometimes hallucinates, and blah, blah, blah.[1657.25] [1657.25][S01]And there's so many things.[1658.81] [1658.81][S01]But then what we didn't realize is, actually,[1660.93] [1660.93][S01]there's still a lot of very good use cases for that even now that[1664.44] [1664.44][S01]people find very valuable-- summarizing documents,[1667.2] [1667.2][S01]and really long things, or writing--[1668.715] [1668.715][S02] Awkward emails?[1669.84] [1669.84][S01] --awkward emails, or mundane forms[1672.45] [1672.45][S01]to be filled in.[1673.98] [1673.98][S01]And there's all these use cases which, actually,[1676.44] [1676.44][S01]people don't mind if there's some small errors.[1680.05] [1680.05][S01]They can fix them easily, and saves a huge amount of time.[1682.9] [1682.9][S01]And I guess that was the surprising thing.[1684.79] [1684.79][S01]They discovered-- people discovered when you put it[1687.21] [1687.21][S01]in the hands of everyone, there were actually these valuable use[1690.6] [1690.6][S01]cases, even though the systems were flawed in all of these ways[1694.84] [1694.84][S01]we know.[1695.64] [1695.64][S02] Well, OK, so I think that sort of takes me[1697.47] [1697.47][S02]on to the next question I want to ask,[1699.053] [1699.053][S02]which is about open source.[1700.42] [1700.42][S02]Because when things are in the hands of people,[1703.24] [1703.24][S02]as you mentioned, really extraordinary things can happen.[1706.36] [1706.36][S02]And I know that DeepMind in the past[1707.94] [1707.94][S02]has open sourced lots of its research projects,[1710.5] [1710.5][S02]but it feels like that's slightly[1713.64] [1713.64][S02]changing now as we go forward.[1714.915] [1714.915][S02]So just tell me what your stance is on open source.[1717.04] [1717.04][S01] Yeah.[1717.28] [1717.28][S01]Well, look, we're huge supporters[1718.77] [1718.77][S01]of open source and open science, as you know.[1720.96] [1720.96][S01]I mean, we've given away and published[1723.24] [1723.24][S01]almost everything we've done, collectively, including like,[1727.53] [1727.53][S01]things like transformers, and AlphaGo.[1729.46] [1729.46][S01]We published all these things in \"Nature\" and \"Science.\"[1731.793] [1731.793][S01]AlphaFold was open source, as we covered last time.[1735.18] [1735.18][S01]And these are all good choices, and you're absolutely right.[1737.68] [1737.68][S01]That's the reason that all works is[1739.17] [1739.17][S01]because that's the way technology and science[1742.74] [1742.74][S01]advances as quickly as possible, by sharing information.[1746.31] [1746.31][S01]So almost always, that's a universal good[1748.85] [1748.85][S01]to do it like that, and that's how science works.[1751.73] [1751.73][S01]The only exception is when you--[1754.82] [1754.82][S01]and AGI and powerful AI does fall[1758.63] [1758.63][S01]into this-- is when you have a dual purpose technology.[1762.6] [1762.6][S01]And so then, the problem is that you[1765.08] [1765.08][S01]want to enable all the good use cases and all[1768.17] [1768.17][S01]the genuine scientists who are acting in good faith and so on,[1770.91] [1770.91][S01]technologists, to build on the ideas, critique[1773.6] [1773.6][S01]the ideas, and so on.[1774.63] [1774.63][S01]That's the way society advances the quickest.[1778.11] [1778.11][S01]But the problem is how do you restrict access[1781.4] [1781.4][S01]at the same time for bad actors who would take the same systems,[1785.22] [1785.22][S01]repurpose them for bad ends, misuse them--[1787.85] [1787.85][S01]weapon systems, who knows what?[1789.69] [1789.69][S01]And those general purpose systems[1792.65] [1792.65][S01]can be repurposed like that.[1794.4] [1794.4][S01]And it's OK today because I don't think[1797.33] [1797.33][S01]the systems are that powerful.[1798.81] [1798.81][S01]But in two, three, four years time,[1801.18] [1801.18][S01]especially when you start getting agent-like systems[1803.96] [1803.96][S01]or agentic behaviors, then, I think,[1806.96] [1806.96][S01]if something's misused by someone, or perhaps even[1810.65] [1810.65][S01]a rogue nation, state, there could be serious harm.[1814.44] [1814.44][S01]So then, I don't have a solution to that.[1817.8] [1817.8][S01]But as a community, we need to think about what does that[1820.22] [1820.22][S01]mean for open source?[1822.62] [1822.62][S01]Perhaps the frontier models need to have more checks on them,[1825.83] [1825.83][S01]and then only after they've been out for a year or two years,[1828.9] [1828.9][S01]then they can get open sourced.[1830.49] [1830.49][S01]That's the model we're following because we[1832.76] [1832.76][S01]have our own open models of Gemini called Gemma[1835.4] [1835.4][S01]because they're smaller.[1836.49] [1836.49][S01]So they're not frontier models.[1839.41] [1839.41][S01]So their capabilities are very useful still to the developer[1841.91] [1841.91][S01]because they're also easy to run on a laptop[1843.987] [1843.987][S01]because they're small numbers of parameters.[1845.82] [1845.82][S01]But the capabilities they have are well[1848.84] [1848.84][S01]understood at this point.[1850.65] [1850.65][S01]Because they're not frontier models.[1852.18] [1852.18][S01]So it's just not as powerful as the latest, say,[1854.73] [1854.73][S01]Gemini 1.5 models.[1856.92] [1856.92][S01]So I think that's probably the approach[1859.73] [1859.73][S01]that we'll end up taking is we'll have open source models,[1863.74] [1863.74][S01]but they'll be lagging maybe one year behind the most[1867.81] [1867.81][S01]cutting edge models just so that we can really assess out[1871.89] [1871.89][S01]in the open by users what those models can do-- the frontier[1875.61] [1875.61][S01]ones can do.[1876.11] [1876.11][S02] And you can really, I guess,[1877.777] [1877.777][S02]test those boundaries of the stochastic--[1879.51] [1879.51][S01] Yeah, and we'll see what those are.[1880.63] [1880.63][S01]The problem with open source is if something goes wrong,[1883.21] [1883.21][S01]you can't recall it.[1884.76] [1884.76][S01]With a proprietary model, if your bad actor[1887.25] [1887.25][S01]starts using it in a bad way, you can just close the tap off.[1892.29] [1892.29][S01]In the limit, you could switch it off.[1894.52] [1894.52][S01]But once you open source something,[1896.95] [1896.95][S01]there's no pulling it back.[1898.54] [1898.54][S01]So it's a one way door, so you should be very, very[1900.99] [1900.99][S01]sure when you do that.[1902.088] [1902.088][S02] Is it definitely possible[1903.63] [1903.63][S02]to contain an AGI, though, within the walls[1908.33] [1908.33][S02]of an organization?[1909.122] [1909.122][S01] Well, that's a whole separate question.[1911.413] [1911.413][S01]I don't think we know how to do that right now.[1913.66] [1913.66][S01]So when you start talking about AGI level powerful,[1916.84] [1916.84][S01]like human level AI--[1917.74] [1917.74][S02] Well, what about intermediary?[1919.49] [1919.49][S01] Well, intermediary, I think,[1921.323] [1921.323][S01]we have good ideas of how to do that.[1923.56] [1923.56][S01]So one would be things like secure sandboxing.[1928.543] [1928.543][S01]So you test-- that's what I'd want[1929.96] [1929.96][S01]to test the agent behaviors in is in a game environment,[1933.23] [1933.23][S01]or a version of the internet that's[1935.39] [1935.39][S01]not quite fully connected.[1937.32] [1937.32][S01]So there's a lot of security work[1939.98] [1939.98][S01]that's done and known in this space, and in fintech,[1942.86] [1942.86][S01]and other places.[1943.59] [1943.59][S01]So we'd probably borrow those ideas,[1945.41] [1945.41][S01]and then build those kinds of systems.[1947.73] [1947.73][S01]And that's how we would test the early prototype systems.[1951.39] [1951.39][S01]But we also know that's not going[1953.6] [1953.6][S01]to be good enough to contain an AGI, something that's[1956.09] [1956.09][S01]potentially smarter than us.[1957.66] [1957.66][S01]So I think we got to understand those systems better so that we[1961.16] [1961.16][S01]can design the protocols for an AGI.[1965.33] [1965.33][S01]When that time comes, we'll have better ideas[1968.0] [1968.0][S01]for how to contain that, potentially[1969.75] [1969.75][S01]also using AI systems and tools to monitor the next versions[1973.88] [1973.88][S01]of the AI system.[1974.793] [1974.793][S02] So on the subject of safety,[1976.46] [1976.46][S02]because I know that you are a very big part of the AI Safety[1979.94] [1979.94][S02]Summit at Bletchley Park in 2023, which was, of course,[1982.82] [1982.82][S02]hosted by the UK government.[1984.44] [1984.44][S02]And from the outside, I think a lot of people[1988.11] [1988.11][S02]just say the word regulation as though it's just going[1990.9] [1990.9][S02]to come in and fix everything.[1992.68] [1992.68][S02]But what is your view on how regulation should be structured?[1996.763] [1996.763][S01] Well, I think it's great[1998.43] [1998.43][S01]that governments are getting up to speed on it and involved.[2001.35] [2001.35][S01]I think that's one of the good things[2002.96] [2002.96][S01]about the recent explosion of interest is that, of course,[2005.79] [2005.79][S01]governments are paying attention.[2007.92] [2007.92][S01]And I think it's been great.[2009.212] [2009.212][S01]The UK government specifically, who[2010.67] [2010.67][S01]I've talked to a lot, and US, as well,[2012.9] [2012.9][S01]they've got very smart people in the civil service[2015.56] [2015.56][S01]staff that understand the technology now to a good degree.[2020.45] [2020.45][S01]And it's been great to see the AI safety institutes[2023.42] [2023.42][S01]being set up in the UK and US, and I[2024.98] [2024.98][S01]think many other countries are going to follow.[2026.938] [2026.938][S01]So I think these are all good precedents and protocols[2029.72] [2029.72][S01]to settle into, again, before the stakes get really high.[2033.87] [2033.87][S01]So this is a proving stage, again, as well.[2037.76] [2037.76][S01]And I do think international cooperation is going[2040.07] [2040.07][S01]to be needed, ideally around things like regulation,[2043.13] [2043.13][S01]and guardrails, and deployment norms.[2046.11] [2046.11][S01]So because AI is a digital technology, very much so,[2051.17] [2051.17][S01]it's hard to contain it within national boundaries.[2054.03] [2054.03][S01]So if the UK or Europe does something, or even the US,[2057.389] [2057.389][S01]but China doesn't, does that really help the world?[2060.17] [2060.17][S01]When we start getting closer to AGI, not really.[2062.79] [2062.79][S01]So I think my view on it is you've[2064.88] [2064.88][S01]got to be, because the technology is changing so fast,[2069.17] [2069.17][S01]we've got to be very nimble and light-footed with regulation so[2073.4] [2073.4][S01]that it's easy to adapt it to where the latest technology is[2077.09] [2077.09][S01]going.[2077.67] [2077.67][S01]If you'd regulated AI five years ago,[2079.732] [2079.732][S01]you'd have regulated something completely different to what[2082.19] [2082.19][S01]we see today, which is Gen AI.[2083.69] [2083.69][S01]But it might be different again in five years.[2085.927] [2085.927][S01]It might be these agent-based systems[2087.469] [2087.469][S01]that are the ones that carry the highest risk.[2090.12] [2090.12][S01]So right now, I would recommend to beef up existing regulations[2095.12] [2095.12][S01]in domains that already have them-- health, transport, so on.[2098.4] [2098.4][S01]I think you can update them for an AI world[2101.57] [2101.57][S01]just like they were updated for mobile and internet.[2103.83] [2103.83][S01]That's probably the first thing I'd[2105.288] [2105.288][S01]do, while doing a watching brief on making sure you understand[2108.75] [2108.75][S01]and test the frontier systems.[2110.86] [2110.86][S01]And then as things become clear and more clearly obvious,[2115.72] [2115.72][S01]then start regulating around that.[2118.59] [2118.59][S01]Maybe in a couple of years time would make sense.[2121.45] [2121.45][S01]One of the things we're missing is, again, the benchmarks--[2124.66] [2124.66][S01]the right tests for capabilities that--[2127.338] [2127.338][S01]what we'd all want to know, including[2128.88] [2128.88][S01]the industry in the field, is at what point are capabilities[2132.78] [2132.78][S01]posing some big risk?[2135.1] [2135.1][S01]And there's no answer to that at the moment beyond what I've just[2138.42] [2138.42][S01]said, which is agent-based capabilities is probably[2140.73] [2140.73][S01]our next threshold.[2141.82] [2141.82][S01]But there's no agreed-upon test for that.[2144.52] [2144.52][S01]One thing you might imagine is testing for deception,[2147.19] [2147.19][S01]for example, as a capability.[2148.84] [2148.84][S01]You really don't want that in the system[2150.75] [2150.75][S01]because then you can't rely on anything else[2152.583] [2152.583][S01]that it's reporting.[2153.79] [2153.79][S01]So that would be my number one emerging capability that I think[2158.67] [2158.67][S01]would be good to test for.[2160.06] [2160.06][S01]But there's many-- ability to achieve certain goals,[2163.67] [2163.67][S01]the ability to replicate.[2165.152] [2165.152][S01]And there's quite a lot of work going on on this now.[2167.36] [2167.36][S01]And I think the safety institutes, which are basically[2170.41] [2170.41][S01]government agencies, I think it would be great for them[2174.76] [2174.76][S01]to push on that, as well.[2176.42] [2176.42][S01]As well as the labs, of course, contributing what we know.[2179.14] [2179.14][S02] I wonder, in this picture of the world[2181.3] [2181.3][S02]that you're describing, what's the place for institutions[2185.73] [2185.73][S02]in this?[2186.23] [2186.23][S02]I mean, if we get to the stage where[2187.73] [2187.73][S02]we have AGI that's supporting all scientific research,[2191.0] [2191.0][S02]is there still a place for great institutions?[2194.09] [2194.09][S01] Yeah, I think so.[2195.6] [2195.6][S01]There's the stage up to AGI, and I[2198.1] [2198.1][S01]think that's got to be a cooperation[2199.6] [2199.6][S01]between civil society, academia, government,[2203.44] [2203.44][S01]and the industrial labs.[2205.82] [2205.82][S01]So I really believe that's the only way[2207.73] [2207.73][S01]we're going to get to the final stages of this.[2211.04] [2211.04][S01]Now, if you're asking after AGI happens, maybe[2214.18] [2214.18][S01]that is what you're asking, then AGI, of course,[2216.24] [2216.24][S01]one of the reasons I've always wanted to build it is then we[2218.74] [2218.74][S01]can use it to start answering some[2220.302] [2220.302][S01]of the biggest, most fundamental questions about the nature[2222.76] [2222.76][S01]of reality, and physics, and all of these things,[2225.37] [2225.37][S01]and consciousness, and so on.[2227.06] [2227.06][S01]It depends what form that takes, whether that[2229.84] [2229.84][S01]will be a human-expert combination with AI.[2234.04] [2234.04][S01]I think that will be the case for a while in terms[2236.56] [2236.56][S01]of discovering the next frontier.[2238.28] [2238.28][S01]So like right now, these systems can't come up[2241.03] [2241.03][S01]with their own conjectures or hypotheses.[2242.923] [2242.923][S01]They can help you prove something,[2244.34] [2244.34][S01]and I think we'll be able to prove[2246.89] [2246.89][S01]get gold medals on international maths[2248.71] [2248.71][S01]Olympiad, things like that.[2249.98] [2249.98][S01]But maybe even solve a famous conjecture.[2252.44] [2252.44][S01]I think that's within reach now.[2254.27] [2254.27][S01]But they don't have the ability to come up[2256.51] [2256.51][S01]with Riemann hypothesis in the first place,[2258.83] [2258.83][S01]or general relativity.[2260.81] [2260.81][S01]So that, really, was always my test[2263.11] [2263.11][S01]for maybe a true artificial general intelligence[2266.41] [2266.41][S01]is it will be able to do that, or invent Go.[2269.44] [2269.44][S01]And so we don't have any systems.[2271.58] [2271.58][S01]We don't really know how we would design, in theory, even,[2276.31] [2276.31][S01]a system that could do that.[2277.517] [2277.517][S02] You know the computer scientist,[2279.35] [2279.35][S02]Stuart Russell?[2279.975] [2279.975][S02]So he told me that he was a bit worried that once we get to AGI,[2283.3] [2283.3][S02]it might be that we all become like the royal princes[2287.28] [2287.28][S02]of the past--[2288.63] [2288.63][S02]the ones who never had to ascend the throne or do any work,[2291.52] [2291.52][S02]but just got to live this life of unbridled luxury[2295.41] [2295.41][S02]and have no purpose.[2297.152] [2297.152][S01] Yeah, so that is the interesting question, is it?[2299.86] [2299.86][S01]Maybe it's beyond AGI.[2300.91] [2300.91][S01]It's more like artificial superintelligence or something--[2303.52] [2303.52][S01]sometimes people call it ASI.[2305.1] [2305.1][S01]But then we should have radical abundance.[2307.56] [2307.56][S01]And assuming we make sure we distribute that fairly[2310.89] [2310.89][S01]and equitably, then we will be in this position[2313.32] [2313.32][S01]where we'll have more freedom to choose what to do.[2319.26] [2319.26][S01]And then meaning will be a big philosophical question.[2322.18] [2322.18][S01]And I think we'll need philosophers,[2325.48] [2325.48][S01]perhaps theologians even, to start[2327.84] [2327.84][S01]thinking as social scientists.[2329.28] [2329.28][S01]They should be thinking about that now.[2331.06] [2331.06][S01]What brings meaning?[2332.59] [2332.59][S01]I mean, I still think there's, of course, self-actualization,[2336.96] [2336.96][S01]and I don't think we'll all just be sitting there meditating.[2339.63] [2339.63][S01]But maybe we'll be playing computer games.[2341.38] [2341.38][S01]I don't know.[2341.922] [2341.922][S01]But is that a bad thing even, or not?[2343.99] [2343.99][S01]Who knows?[2344.83] [2344.83][S02] I don't think the princes of the past[2346.18] [2346.18][S02]came off particularly well.[2347.305] [2347.305][S01] No.[2348.097] [2348.097][S01]Traveling the stars.[2349.07] [2349.07][S01]But then there's also extreme sports people do.[2351.92] [2351.92][S01]Why do they do them?[2352.91] [2352.91][S01]I mean, climb Everest, all these.[2354.675] [2354.675][S01]I mean, there'll be-- but I think[2356.05] [2356.05][S01]it's going be very interesting.[2357.342] [2357.342][S01]And that I don't know, but that's[2359.14] [2359.14][S01]what I was saying earlier about it's[2361.48] [2361.48][S01]underappreciated what's going to happen going back to the hype[2365.41] [2365.41][S01]near-term versus far-term.[2367.045] [2367.045][S01]So if you want to call that hype,[2368.42] [2368.42][S01]even, it's definitely under hyped,[2370.22] [2370.22][S01]I think, the amount of transformation that will happen.[2373.25] [2373.25][S01]I think it will be very good in the limit.[2376.43] [2376.43][S01]We'll cure lots of diseases, and/or all diseases,[2379.18] [2379.18][S01]solve our energy problems, climate problems.[2382.03] [2382.03][S01]But then the next question comes is, is there meaning?[2385.31] [2385.31][S02] So bring us back slightly closer to AGI[2388.42] [2388.42][S02]rather than superintelligence.[2389.9] [2389.9][S02]I know that your big mission is to build artificial intelligence[2393.47] [2393.47][S02]to benefit everybody, but how do you make sure[2396.01] [2396.01][S02]that it does benefit everybody?[2397.67] [2397.67][S02]How do you include all people's preferences[2401.53] [2401.53][S02]rather than just the designers?[2403.16] [2403.16][S01] Yeah, I think what's[2404.873] [2404.873][S01]going to have to happen is--[2406.04] [2406.04][S01]I mean, it's impossible to include all preferences[2408.26] [2408.26][S01]in one system.[2409.01] [2409.01][S01]Because by definition, people don't agree.[2411.03] [2411.03][S01]We can see that in, unfortunately,[2412.62] [2412.62][S01]in the current state of the world.[2414.037] [2414.037][S01]Countries don't agree.[2415.37] [2415.37][S01]Governments don't agree.[2417.185] [2417.185][S01]We can't even get agreement on obvious things[2419.06] [2419.06][S01]like dealing with the climate situation.[2422.1] [2422.1][S01]So I think that's very hard.[2424.41] [2424.41][S01]What I imagine will happen is that we'll[2426.95] [2426.95][S01]have a set of safe architectures,[2428.54] [2428.54][S01]hopefully, that personalized AIs can be built on top of.[2433.4] [2433.4][S01]And then everyone will have, or different countries[2437.27] [2437.27][S01]will have their own preferences about what they use it for,[2440.12] [2440.12][S01]what they deploy it for, what can and can't be done with them.[2444.15] [2444.15][S01]But overall-- and that's fine.[2445.83] [2445.83][S01]That's for everyone to individually decide[2447.62] [2447.62][S01]or countries to decide themselves,[2449.04] [2449.04][S01]just like they do today.[2450.18] [2450.18][S01]But as a society, we know that there's[2453.83] [2453.83][S01]some provably safe things about those architectures.[2456.96] [2456.96][S01]And then you can let them proliferate, and so on.[2459.42] [2459.42][S01]So I think that we've got to get through the eye of a needle[2462.89] [2462.89][S01]in a way where as we get closer to AGI,[2466.57] [2466.57][S01]we've probably got to cooperate more, ideally internationally,[2470.82] [2470.82][S01]and then make sure we build AGIs in a safe architecture way.[2476.225] [2476.225][S01]Because I'm sure there are unsafe ways,[2477.85] [2477.85][S01]and I'm sure there are safe ways of building AGI.[2480.84] [2480.84][S01]And then once we get through that,[2482.56] [2482.56][S01]then we can open the funnel again,[2484.83] [2484.83][S01]and everyone can have their own personalized pocket AGIs,[2489.06] [2489.06][S01]if they want.[2489.81] [2489.81][S02] What a version of the future.[2491.53] [2491.53][S02]But then, in terms of the safe way to build it,[2494.17] [2494.17][S02]I mean, are we talking about undesirable behaviors[2496.38] [2496.38][S02]here that might emerge?[2497.62] [2497.62][S01] Yes, undesirable emergent behaviors, capabilities[2502.6] [2502.6][S01]that--[2503.1] [2503.1][S02] Deception.[2504.017] [2504.017][S01] Deception is one example that you don't want.[2507.51] [2507.51][S01]Value systems.[2508.81] [2508.81][S01]We got to understand all of these things[2510.72] [2510.72][S01]better-- what kind of guardrails work, not circumventable.[2514.828] [2514.828][S01]And there's two cases to worry about.[2516.37] [2516.37][S01]There's bad uses by bad individuals or nations, so[2521.85] [2521.85][S01]human misuse, and then there's the AI itself[2524.95] [2524.95][S01]as it gets closer to AGI going off the rails.[2527.68] [2527.68][S01]And I think you need different solutions for those two[2530.13] [2530.13][S01]problems.[2531.27] [2531.27][S01]And so, yeah, that's what we're going[2533.55] [2533.55][S01]to have to contend with as we get closer[2535.35] [2535.35][S01]to building these technologies.[2538.11] [2538.11][S01]And also, just going back to your benefiting everyone point,[2540.61] [2540.61][S01]of course, we're showing the way with things[2543.03] [2543.03][S01]like AlphaFold and isomorphic.[2544.93] [2544.93][S01]I think we could cure most diseases within the next decade[2549.18] [2549.18][S01]or two if AI drug design works.[2552.135] [2552.135][S01]And then they could be personalized medicines[2554.01] [2554.01][S01]where it minimizes the side effects on the individual[2556.95] [2556.95][S01]because it's mapped to the person's[2559.02] [2559.02][S01]individual illness, and their individual metabolism,[2561.46] [2561.46][S01]and so on.[2562.12] [2562.12][S01]So these are amazing things--[2565.56] [2565.56][S01]clean energy, renewable energy sources,[2567.63] [2567.63][S01]fusion, or better solar power, all of these types of things.[2571.335] [2571.335][S01]I think they're all within reach.[2572.71] [2572.71][S01]And then that would sort out water access[2575.13] [2575.13][S01]because you could do desalination everywhere.[2577.18] [2577.18][S01]So I just feel like an enormous good[2580.92] [2580.92][S01]is going to come from these technologies,[2584.17] [2584.17][S01]but we have to mitigate the risks, too.[2585.978] [2585.978][S02] And one way that you said[2587.52] [2587.52][S02]that you would want to mitigate the risks[2589.26] [2589.26][S02]was that there would be a moment where you would basically[2592.05] [2592.05][S02]do the scientific version of Avengers assemble.[2594.31] [2594.31][S01] Yes, sure.[2595.413] [2595.413][S02] Terence Tao, get him on the phone.[2597.33] [2597.33][S01] Exactly.[2597.87] [2597.87][S02] Bring him on down.[2599.12] [2599.12][S01] Yeah, exactly.[2600.51] [2600.51][S02] Is that still your plan?[2602.01] [2602.01][S01] Yeah, well, I think so.[2603.94] [2603.94][S01]I think if we can get the international cooperation,[2606.69] [2606.69][S01]I'd love there to be a international CERN, basically,[2610.2] [2610.2][S01]for AI.[2610.95] [2610.95][S01]Where you get the top researchers in the world,[2613.05] [2613.05][S01]and you go, look, let's focus on the final few years of this AGI[2617.28] [2617.28][S01]project, and get it really right,[2619.23] [2619.23][S01]and do it scientifically, and carefully,[2621.6] [2621.6][S01]and thoughtfully at every step--[2623.88] [2623.88][S01]the final steps.[2625.002] [2625.002][S01]I still think that would be the best way.[2626.71] [2626.71][S02] How do you know when is the time to press the button?[2629.2] [2629.2][S01] Well, that's the big question.[2630.43] [2630.43][S01]Because you can't do it too early because you[2632.46] [2632.46][S01]would never be able to get the buy-in to do that.[2634.598] [2634.598][S01]A lot of people would disagree.[2635.89] [2635.89][S01]I mean, today people disagree with the risks.[2637.765] [2637.765][S01]You see very famous people saying there's no risks,[2640.37] [2640.37][S01]and then you have people like Jeff Hinton saying[2642.37] [2642.37][S01]there's lots of risks.[2643.51] [2643.51][S01]And I'm in the middle of that.[2645.245] [2645.245][S02] I want to talk to you[2646.62] [2646.62][S02]a bit more about neuroscience.[2648.54] [2648.54][S02]How much does it still inspire what you're doing?[2652.69] [2652.69][S02]Because I noticed the other day that DeepMind had unveiled[2655.65] [2655.65][S02]this computerized rat with an artificial brain[2659.87] [2659.87][S02]that helps to change our understanding of how[2663.57] [2663.57][S02]the brain controls movement.[2665.17] [2665.17][S02]But in the first season of the podcast,[2667.03] [2667.03][S02]I remember we talked a lot about how[2669.36] [2669.36][S02]DeepMind takes direct inspiration[2671.73] [2671.73][S02]from biological systems.[2673.36] [2673.36][S02]Is that still the core of your approach?[2675.027] [2675.027][S01] No, it's evolved now[2676.527] [2676.527][S01]because I think we've got to a stage now.[2678.355] [2678.355][S01]In the last, I would say, two to three years,[2680.23] [2680.23][S01]we've gone more into an engineering phase--[2682.5] [2682.5][S01]large scale systems, massive training architectures.[2687.96] [2687.96][S01]So I would say that the influence of neuroscience[2691.59] [2691.59][S01]on that is a little bit less.[2693.39] [2693.39][S01]It may come back in.[2694.63] [2694.63][S01]So any time where you need more invention,[2697.118] [2697.118][S01]then you want to get as many sources as possible.[2699.16] [2699.16][S01]And neuroscience would be one of those sources of ideas.[2702.42] [2702.42][S01]But when it's more engineering heavy,[2704.28] [2704.28][S01]then I think that takes a little bit more of a backseat.[2707.5] [2707.5][S01]So it may be more applying AI to neuroscience now[2710.94] [2710.94][S01]like you saw with the virtual rat brain.[2712.98] [2712.98][S01]And I think we'll see that as we get closer to AGI[2715.59] [2715.59][S01]using that to understand the brain.[2717.8] [2717.8][S01]I think it would be one of the coolest use[2719.55] [2719.55][S01]cases for AGI and science.[2720.882] [2720.882][S02] I guess this stuff goes[2722.34] [2722.34][S02]through phases of the engineering challenge,[2724.2] [2724.2][S02]the intervention challenge.[2725.325] [2725.325][S01] So it's done its part for now,[2727.32] [2727.32][S01]and it's been great.[2728.65] [2728.65][S01]And we still obviously keep a close track of it[2731.04] [2731.04][S01]and take any other ideas, too.[2733.59] [2733.59][S02] OK.[2734.55] [2734.55][S02]All of the pictures of the future that you've painted[2738.24] [2738.24][S02]are still anchored quite in reality.[2740.74] [2740.74][S02]But I know that you've said that you really[2743.1] [2743.1][S02]want AGI to be able to peer into the mysteries of the universe[2746.85] [2746.85][S02]down at the Planck scale.[2748.24] [2748.24][S01] Yes![2749.073] [2749.073][S02] Like subatomic, quantum worlds.[2752.79] [2752.79][S02]Do you think that there are things that we have not even[2755.4] [2755.4][S02]yet conceived of that might end up being possible?[2759.345] [2759.345][S02]I'm talking wormholes here.[2760.47] [2760.47][S01] Completely, yes.[2761.803] [2761.803][S01]I'd love wormholes to be possible.[2763.82] [2763.82][S01]I think there is a lot of probably misunderstanding,[2767.592] [2767.592][S01]I would say, still things we don't understand about physics[2770.05] [2770.05][S01]and the nature of reality.[2771.73] [2771.73][S01]And obviously, the quantum mechanics, and unifying[2774.76] [2774.76][S01]that with gravity, and all of these things,[2776.98] [2776.98][S01]and there's all these problems with the standard model.[2779.48] [2779.48][S01]So I think there's--[2781.09] [2781.09][S01]and string theory, I mean, I just think--[2784.18] [2784.18][S02] There's giant gaping holes in physics.[2786.4] [2786.4][S01] Yes, in physics, all over the place.[2788.567] [2788.567][S01]And I talk to my physics friends about this,[2790.6] [2790.6][S01]and there's a lot of things that don't fit together.[2793.15] [2793.15][S01]I don't really like the multiverse explanation.[2795.32] [2795.32][S01]So I think that it would be great to come up[2799.3] [2799.3][S01]with new theories and then test those on massive apparatus,[2803.09] [2803.09][S01]perhaps out in space, at these tiny--[2806.847] [2806.847][S01]the reason I'm obsessed with Planck scale things--[2808.93] [2808.93][S01]Planck time, Planck space--[2810.4] [2810.4][S01]is because that seems to be the resolution of reality in a way.[2816.82] [2816.82][S01]That's the smallest quanta you can break anything into.[2820.01] [2820.01][S01]So that feels like the level you want[2822.7] [2822.7][S01]to experiment on if you had powerful apparatus perhaps[2827.83] [2827.83][S01]designed or enabled by having AGI and radical abundance.[2831.08] [2831.08][S01]You would need both to be able to afford to build[2833.47] [2833.47][S01]those types of experiments.[2834.86] [2834.86][S02] The resolution of reality.[2836.56] [2836.56][S02]What a phrase.[2837.55] [2837.55][S02]What, so as in the resolution that we're at the moment?[2840.61] [2840.61][S02]Human level is just an approximation of reality.[2843.01] [2843.01][S01] Yes, that's right.[2843.5] [2843.5][S01]And then we know there's the atomic level,[2844.87] [2844.87][S01]and below that is the Planck level, which as far as we know[2847.328] [2847.328][S01]is the smallest resolution one can even talk about things.[2850.58] [2850.58][S01]And so that, to me, would be the resolution[2853.09] [2853.09][S01]one wants to experiment on to really understand[2855.31] [2855.31][S01]what's going on here.[2856.22] [2856.22][S02] I wonder whether you're also[2857.92] [2857.92][S02]envisaging that there will be things that[2859.81] [2859.81][S02]are beyond the limits of human understanding[2862.46] [2862.46][S02]AGI will help us to uncover--[2864.37] [2864.37][S02]that actually, we're just not really capable of understanding.[2867.89] [2867.89][S02]And then I wonder if things are unexplainable or[2871.6] [2871.6][S02]ununderstandable, are they still falsifiable?[2874.383] [2874.383][S01] Yeah well, look, I mean,[2876.05] [2876.05][S01]these are great questions.[2877.133] [2877.133][S01]I think there will be a potential for an AGI system[2879.79] [2879.79][S01]to understand higher level abstractions than we can.[2882.68] [2882.68][S01]So again, going back to neuroscience,[2884.96] [2884.96][S01]we know that it's your prefrontal cortex that[2887.17] [2887.17][S01]does that.[2887.93] [2887.93][S01]And there's up to about six or seven layers of indirection one[2892.3] [2892.3][S01]could take.[2892.99] [2892.99][S01]This person's thinking this, and I'm[2894.615] [2894.615][S01]thinking this about that person thinking this, and so on.[2896.99] [2896.99][S01]And then we lose track.[2898.46] [2898.46][S01]But I think an AI system could have an arbitrarily large[2901.99] [2901.99][S01]prefrontal cortex effectively.[2903.83] [2903.83][S01]So you could imagine higher levels[2905.41] [2905.41][S01]of abstraction and patterns that it[2907.037] [2907.037][S01]will be able to see about the universe[2908.62] [2908.62][S01]that we can't really comprehend or hold in mind at once.[2912.91] [2912.91][S01]And then I think in terms of explainability point of view,[2916.52] [2916.52][S01]the way I think that is a little bit[2918.04] [2918.04][S01]different to other philosophers who've thought about this, which[2920.71] [2920.71][S01]is we'll be closer to an ant and then the AGI, in terms of IQ.[2924.57] [2924.57][S01]But I don't think that's the way to think of it.[2926.57] [2926.57][S01]I think it's we are Turing complete,[2930.89] [2930.89][S01]so we're a full general intelligence as ourselves,[2934.61] [2934.61][S01]albeit a bit slow because we run on slow machinery.[2938.0] [2938.0][S01]And we can't infinitely expand our own brains.[2941.15] [2941.15][S01]But we can, in theory, given enough time and memory,[2946.4] [2946.4][S01]understand anything that's computable.[2949.4] [2949.4][S01]And so I think it will be more like Garry Kasparov or Magnus[2954.7] [2954.7][S01]Carlsen playing an amazing chess move.[2956.54] [2956.54][S01]I couldn't have come up with it, but they can explain it to me[2959.14] [2959.14][S01]why it's a good move.[2960.59] [2960.59][S01]So I think that's what an AGI system will be able to do.[2964.25] [2964.25][S02] You said that DeepMind was a 20-year project.[2967.78] [2967.78][S02]How far through are we?[2968.87] [2968.87][S02]Are you on track?[2969.64] [2969.64][S01] I think we're on track, yeah, crazily.[2971.89] [2971.89][S01]Because usually 20-year projects stay 20 years away.[2974.56] [2974.56][S01]But yeah, we're a good way in now, and I think we're--[2977.77] [2977.77][S02] 20 years is 2030 for AGI.[2979.312] [2979.312][S01] 2030, yeah.[2980.437] [2980.437][S01]So I think the way I say is I wouldn't be surprised[2982.6] [2982.6][S01]if it comes in the next decade.[2984.37] [2984.37][S01]So I think we're on track.[2985.84] [2985.84][S02] That matches what you said last time.[2987.05] [2987.05][S02]You haven't updated your priors.[2988.383] [2988.383][LAUGHTER][2990.01] [2990.01][S01] Exactly[2991.217] [2991.217][S02] Amazing.[2992.05] [2992.05][S02]Demis, thank you so much.[2993.4] [2993.4][S02]Absolute delight.[2994.313] [2994.313][S02]Absolute delight, as always.[2995.48] [2995.48][S01] Very fun to talk, as always, as well.[2997.688] [2997.688][S01]Thank you.[2998.41] [2998.41][S02] OK, I think there are a few really important[3000.743] [3000.743][S02]things that came out of that conversation, especially[3003.15] [3003.15][S02]when you compare it to what Demis was saying last time we[3006.38] [3006.38][S02]spoke to him in 2022.[3008.36] [3008.36][S02]Because there have definitely been a few surprises[3011.39] [3011.39][S02]in the last couple of years.[3013.04] [3013.04][S02]The way that these models have demonstrated[3015.26] [3015.26][S02]a genuine conceptual understanding is one--[3018.38] [3018.38][S02]this real world grounding that came in from language[3021.8] [3021.8][S02]and human feedback alone.[3023.76] [3023.76][S02]We did not think that would be enough.[3026.72] [3026.72][S02]And then how interesting and useful, imperfect AI[3030.53] [3030.53][S02]has been to the everyday person.[3032.79] [3032.79][S02]Demis himself there admitted that he had not[3035.12] [3035.12][S02]seen that one coming.[3036.62] [3036.62][S02]And that makes me wonder about the other challenges[3039.02] [3039.02][S02]that we don't yet know how to solve[3041.305] [3041.305][S02]like long-term planning, and agency, and robust, unbreakable[3046.64] [3046.64][S02]safeguards.[3047.85] [3047.85][S02]How many of those--[3049.55] [3049.55][S02]which we're going to cover in detail in this podcast,[3051.78] [3051.78][S02]by the way--[3052.56] [3052.56][S02]are we going to come back to in a couple of years[3055.1] [3055.1][S02]and realize that they were easier than we thought?[3058.91] [3058.91][S02]And how many of them are going to be harder?[3062.0] [3062.0][S02]And then as for the big predictions[3063.87] [3063.87][S02]that Demis made, like cures for most diseases in 10 or 20 years,[3067.62] [3067.62][S02]or AGI by the end of the decade, or how[3070.32] [3070.32][S02]we're about to enter into an era of abundance,[3073.6] [3073.6][S02]I mean, they all sound like Demis is being a bit overly[3076.95] [3076.95][S02]optimistic, doesn't it?[3078.96] [3078.96][S02]But then again, he hasn't exactly been wrong so far.[3083.49] [3083.49][S02]You've been listening to \"Google DeepMind, the Podcast\" with me,[3086.38] [3086.38][S02]Professor Hannah Fry.[3087.28] [3087.28][S02]If you have enjoyed this episode, hey, why not subscribe?[3090.63] [3090.63][S02]We've got plenty more fascinating conversations[3092.97] [3092.97][S02]with the people at the cutting edge of AI[3095.76] [3095.76][S02]coming up on topics ranging from how AI is accelerating[3099.18] [3099.18][S02]the pace of scientific discoveries[3100.98] [3100.98][S02]to addressing some of the biggest[3102.72] [3102.72][S02]risks of this technology.[3104.53] [3104.53][S02]If you have any feedback, or you want to suggest a future guest,[3107.59] [3107.59][S02]then do leave us a comment on YouTube.[3109.56] [3109.56][S02]Until next time.[3110.76] [3110.76][MUSIC PLAYING][3113.81]"}
{"file_name": "audio/val_000021.wav", "transcription": "[11.512][S01] The idea of creating artificial creatures[14.081] [14.081][S01]has obsessed us for millennia,[16.683] [16.683][S01]but for the purposes of this exercise before we go any further,[19.953] [19.953][S01]I’m going to ask you to purge any of the following thoughts from your head.[24.458] [24.458][S01]Checklist: Hephaestus expelled from Olympus[27.661] [27.661][S01]and then built two servant robots,[29.63] [29.63][S01]the Bhuta Vahana Yanta or spirit movement machines[32.966] [32.966][S01]of 12th Century India made to protect the relics of Buddha,[36.403] [36.403][S01]the chess playing mechanical Turk, just a bloke in a box pulling levers.[40.874] [40.874][S01]Mary Shelley’s Frankenstein - he walks, he talks, not a lot else.[45.212] [45.212][S01]Kubrick’s HAL “Don’t Call Me Dave,” C3P0, R2D2, canine, NS2,[50.684] [50.684][S01]I mean seriously these are just numbers and letters.[53.32] [53.32][S01]Robbie the Robot, robot lovers, RoboCop.[56.924] [62.429][S01]Feeling better? Okay. Let’s get going.[68.669] [71.572][S01]I’m Hannah Fry, I’m an Associate Professor in Mathematics,[74.274] [74.274][S01]and I am AI curious, and this is DeepMind,[78.846] [78.846][S01]the podcast series where we look[80.347] [80.347][S01]at the fast-moving story of artificial intelligence.[84.051] [84.051][S01]We’ve been talking to the scientists,[85.552] [85.552][S01]researchers and engineers based at DeepMind in London.[89.489] [89.489][S01]We’re looking at how they’re approaching the science of AI.[92.96] [92.96][S01]And some of the tricky decisions[94.595] [94.595][S01]the whole field is wrestling with at the moment.[98.198] [98.198][S01]So whether you just want to know more[100.2] [100.2][S01]or want to be inspired on your own AI journey,[103.303] [103.303][S01]then this is the place to be.[105.739] [108.675][S01]You see the thing is - robots sell. We’ve long lusted after the idea[113.714] [113.714][S01]of upending the natural order with human ingenuity.[117.117] [117.117][S01]We just can’t seem to leave it alone,[120.053] [120.053][S01]and in this episode we are looking at AI and robotics.[124.258] [124.258][S01]Murray Shanahan is the Senior Scientist at DeepMind -[126.527] [126.527][S01]he’s also a professor of cognitive robotics[129.696] [129.696][S01]at Imperial College London, and growing up,[132.466] [132.466][S01]Murray was utterly mesmerized by science fiction,[136.47] [136.47][S01]so you can picture his face when the Hollywood film director Alex Garland[140.674] [140.674][S01]approached him following the publication of his book[143.51] [143.51][S01]Embodiment and the Inner Life.[145.979] [145.979][S06] Alex contacted me and said: “Oh I’m writing a script for ah film[150.05] [150.05][S06]about AI and consciousness and I read your book[152.519] [152.519][S06]and it you know helped to crystallize some ideas[154.655] [154.655][S06]and would you like to chat about ah chat about it?[157.391] [157.391][S06]And ah and so of course it was you know it was a great opportunity[161.695] [161.695][S06]ah to get involved in the science fiction film and then[164.364] [164.364][S06]and then to my great good fortune it turns out to be an absolute cracker.[168.235] [169.203][S01] And that is how Murray became scientific advisor[173.04] [173.04][S01]on the Oscar-winning film Ex Machina.[176.176] [176.844][S01]I met with Murray to get a potted history of AI.[180.08] [180.08][S01]Murray, people tend to think of AI as this this this very new thing,[183.884] [183.884][S01]a very a very modern invention,[185.886] [185.886][S01]but it’s actually been around for quite a long time.[188.121] [188.121][S06] It has. The idea of artificial intelligence,[190.557] [190.557][S06]the idea of making artificial creatures dates back to Greek mythology[194.561] [194.561][S06]but the sort of modern conception of AI perhaps really dates back to Alan[199.766] [199.766][S06]Turing’s paper published in the 1950s[202.936] [202.936][S06]where he first kind of asked the question -[205.639] [205.639][S06]could a machine think?[207.14] [207.14][S06]And gave a number of kind of refutations for counter arguments[210.477] [210.477][S06]to the idea that a machine could think.[212.379] [212.379][S01] This is where the Turing test comes from.[214.615] [214.615][S06] And this was this famous paper inaugurated the so-called[217.751] [217.751][S06]Turing test because Turing didn’t call it[219.319] [219.319][S06]the Turing test [Hannah laughs][220.554] [220.554][S06]which is the idea that we should subject a machine[223.257] [223.257][S06]to a test to see whether it’s basically -[225.125] [225.125][S06]whether it’s indistinguishable from a human in dialogue.[228.228] [228.228][S06]The term “artificial intelligence” was actually coined by John McCarthy,[232.299] [232.299][S06]a Stanford professor - he was at MIT at the time,[234.801] [234.801][S06]and John McCarthy organized a conference in 1956[239.106] [239.106][S06]bringing together a lot of leading thinkers in maths[241.542] [241.542][S06]to try and scope out the idea of building a thinking machine[244.945] [244.945][S06]and he coined the term artificial intelligence.[247.247] [247.247][S01] What did they describe it as at that time -[250.217] [250.217][S01]how did they see artificial intelligence?[252.686] [252.686][S06] John McCarthy in particular - his idea of artificial intelligence[256.79] [256.79][S06]that he had in mind was a kind of system that would answer questions really[260.894] [260.894][S06]and be able to engage in dialogue with humans, so…[264.431] [264.431][S01] Something that you are actually just talking to -[265.799] [265.799][S06] So something that you’re that you’re talking to[267.501] [267.501][S06]although of course in those days it wouldn’t have been through speech,[270.304] [270.304][S06]it would have been by typing in at the keyboard,[272.439] [272.439][S06]and it was very much a disembodied notion of artificial intelligence,[275.909] [275.909][S06]so this system didn’t have a body and interact with the physical world[280.414] [280.414][S06]in the way that we do or animals do or indeed that robots do[283.217] [283.217][S06]so they weren’t really thinking about robotics at that point.[285.953] [285.953][S01] I’m sort of imagining something like um HAL in “2001 Space Odyssey,”[289.756] [289.756][S01]except typing in rather than speaking to -[291.658] [291.658][S06] Yeah, and a kind of nice version, you know.[294.027] [294.027][S06]That their approach to artificial intelligence[296.163] [296.163][S06]was to build systems that reasoned in logic[299.366] [299.366][S06]and we now think of that whole sort of approach[301.869] [301.869][S06]to artificial intelligence of using logic and reasoning[304.938] [304.938][S06]as so-called good old fashioned artificial intelligence,[307.641] [307.641][S06]or Gofi, or classical AI.[310.944] [311.512][S01] Now I know, I know I did a bad thing there.[313.347] [313.347][S01]I mentioned HAL but nobody ever said this was going to be easy.[317.05] [317.05][S01]But to recap - in classical AI,[319.453] [319.453][S01]you have to write down a complete list of rules[321.755] [321.755][S01]for how you want your agent to think.[324.558] [324.558][S01]If this happens, then do this. If that happens, then do that.[328.829] [328.829][S01]It’s a nice idea in theory, but if your agent is going to know[331.665] [331.665][S01]how to handle every possible scenario you could throw at it,[335.836] [335.836][S01]it’s going to need to be a long, long list.[339.973] [339.973][S06] There was a project called Psych[342.342] [342.342][S06]which attempted to write out all the rules of common sense[346.68] [346.68][S06]to build an enormous encyclopedic database of common sense.[350.017] [350.017][S01] Can you remember any of them?[351.018] [351.018][S06] Well I mean there would be things like -[352.419] [352.419][S06]if you’ve got a container, and you put something in that container[357.191] [357.191][S06]and then move that container somewhere else,[359.359] [359.359][S06]then the thing that was in it gets moved as well.[361.862] [361.862]You know - [S01] Common sense![362.729] [362.729][S06] Yeah, stuff like that you know,[364.364] [364.364][S06]and if you buy something and pay for it,[367.134] [367.134][S06]you’ll have less money than you had in the first place.[369.937] [369.937][S06]So do I think it’s impossible to do that?[372.239] [372.239][S06]I think it’s impossible in practice.[374.174] [374.174][S06]Because it turns out that the sheer number of rules[377.244] [377.244][S06]that you would have to write is absolutely enormous.[379.813] [379.813][S01] We might have moved away from this long list of rules by now[382.749] [382.749][S01]as a way to teach our artificial intelligence in favor of agents[386.52] [386.52][S01]that can learn the rules for themselves. But the skills that we want our AI[391.258] [391.258][S01]to have like good old fashioned common sense[395.295] [395.295][S01]are just as important now as they ever were.[398.665] [402.402][S01]Imagine one day long into the future a wealthy computer scientist[406.74] [406.74][S01]builds an AI to manage his stamp collection.[410.377] [410.377][S01]He plugs it into the internet, gives it access to his bank account[413.847] [413.847][S01]and sets up the challenge to buy as many stamps as possible.[418.051] [418.051][S01]At first the agent acts as its creator intended,[421.321] [421.321][S01]signing up to eBay, bidding on stamps. But after a while he gets another idea.[427.461] [430.597][S01]More money equals more stamps,[433.433] [433.433][S01]so why not start trading on the stock market to make more money?[437.504] [437.504][S01]And it soon realizes it can get the stamps cheaper[440.607] [440.607][S01]if it can get them at source.[442.709] [442.709][S01]So the agent buys up a factory, converts its manufacturing process[446.58] [446.58][S01]to stamp-making, and goes on with achieving its goal.[450.684] [454.121][S01]But of course the limiting factor here is paper - more paper, more stamps.[459.293] [459.293][S01]So it starts commanding forests to be felled, the wood to be processed,[463.997] [463.997][S01]all to feed its single minded ambition more stamps.[469.036] [470.404][S01]Now there’s no denying that the AI is doing what it was told,[473.841] [473.841][S01]but it’s doing so at any cost,[476.743] [476.743][S01]and any agent without some kind of commons[479.413] [479.413][S01]sense will be at risk of taking our instructions a bit too literally.[485.319] [486.086][S01]This might be a bit of an extreme example, but Victoria Krakovna,[490.057] [490.057][S01]a research scientist a DeepMind working on AI safety[494.161] [494.161][S01]is already seeing agents[495.762] [495.762][S01]that aren’t exactly behaving in the way their designers intended.[500.234] [500.234][S04] A reinforcement agent that was playing a boat racing game[503.904] [503.904][S04]and the intended behavior there was to go around the race track[507.14] [507.14][S04]and finish the race as soon as possible, and the agent was encouraged to do this[511.345] [511.345][S04]by having these little green squares along the track[514.348] [514.348][S04]that would give it rewards,[515.616] [515.616][S04]and then what the agent figured out is that[517.184] [517.184][S04]instead of actually playing the game it could be going around in circles[520.153] [520.153][S04]and heading the same green squares over and over again to rack up more points,[523.624] [523.624][S04]and then you have this this whole situation with the boat going in circles[527.761] [527.761][S04]and crashing into everything and catching on fire[529.897] [529.897][S04]and still getting more points than to otherwise[531.832] [531.832][S04]get this this kind of situations are quite common.[534.368] [534.368][S01] But because you haven’t ever stopped the AI from doing that,[537.104] [537.104][S01]or told the AI that you don’t want it to do that,[539.573] [539.573][S01]that’s a perfectly sensible solution for it to come across.[543.143] [543.143][S04] Yeah, from the perspective of the AI[544.878] [544.878][S04]it can’t really tell that the solution is a cheat.[548.115] [548.115][S04]It’s just something that gives it a lot of rewards,[550.751] [550.751][S04]so it can’t necessarily distinguish between the general solution[554.555] [554.555][S04]and an overly creative solution that humans haven’t foreseen.[559.059] [560.694][S01] There are plenty of examples like this.[563.43] [563.43][S01]One team of researchers created an agent[565.832] [565.832][S01]inside a very simple two-dimension computer game[569.703] [569.703][S01]and tasked it with building itself a body[571.972] [571.972][S01]to get itself from the start line to the finish line.[575.108] [575.108][S01]Quite quickly it worked out that it could just build itself taller[579.012] [579.012][S01]and taller and taller[580.881] [580.881][S01]until it was as high as the track was long,[583.817] [583.817][S01]and then just flopped forwards to cross the line.[587.12] [587.12][S01]And there was the agent playing the game of Tetris[589.489] [589.489][S01]which realized it could just pause the game forever and never lose.[594.261] [594.261][S01]But there is a balancing act here. We don’t want our AI misbehaving,[598.365] [598.365][S01]but we also don’t want to restrict our agents too much.[601.602] [601.602][S04] And this is part of what’s tricky about achieving safe behaviors[605.639] [605.639][S04]that we don’t want to hamper the systems’ ability[608.675] [608.675][S04]to come up with really interesting and innovative solutions[611.578] [611.578][S04]that we have not seen - so we don’t just want human imitation.[615.749] [615.749][S04]We want super human capabilities but without unsafe behavior.[619.553] [622.189][S01] There is a very fine line between a naughty algorithm[625.959] [625.959][S01]and one that’s finding innovative solutions to problems[628.629] [628.629][S01]that humans haven’t been able to solve.[631.331] [631.331][S01]The AI doesn’t know the difference between the two -[634.701] [634.701][S01]it doesn’t understand what’s really important to us,[637.771] [637.771][S01]it doesn’t have any common sense -[640.407] [640.407][S01]and that means you have to be very careful[642.876] [642.876][S01]when you’re setting up incentives and rewards for your agents.[647.314] [647.314][S04] Part of the reason that this is so difficult is this general affect[651.285] [651.285][S04]that’s called Goodhart’s law in economics -[653.453] [653.453][S04]what Goodhart’s law comes when a metric becomes a target,[656.657] [656.657][S04]it ceases to be a good metric.[658.358] [660.427][S01] A classic example of Goodhart’s Law comes from British India,[663.897] [663.897][S01]when authorities offered cash rewards for dead cobras[667.935] [667.935][S01]as a way to decrease the population of snakes.[672.105] [672.105][S01]Unbeknownst to the British, the locals started to breed cobras[676.076] [676.076][S01]in order to take advantage of the reward.[679.179] [679.179][S01]As soon as the authorities found this out,[682.015] [682.015][S01]they scrapped the scheme altogether and revoked the rewards.[686.386] [686.386][S01]But now there were these snake farms everywhere filled with worthless cobras,[691.925] [691.925][S01]so what did the locals do? Release the cobras into the country,[695.796] [695.796][S01]resulting in an increase in the cobra population.[700.467] [700.467][S01]This is what the scientists call a specification problem -[704.938] [704.938][S01]when your specified objective fails to bring about the intended behavior.[710.444] [710.444][S01]This is generally quite likely to happen because human preferences[714.114] [714.114][S01]tend to be quite complex and whenever we try to distill them[717.117] [717.117][S01]into some kind of specification or something that we say we want[719.853] [719.853][S01]it’s going to be a lot simpler than our real preferences are[724.625] [724.625][S01]and it wouldn’t necessarily capture everything that’s important to us.[728.762] [731.131][S01]Let us imagine we’re living in the future[733.333] [733.333][S01]where robot butlers are commonplace.[736.904] [736.904][S01]You clearly specify the objective for your robot -[740.04] [740.04][S01]it should serve you at all times.[743.443] [743.443][S01]Now, how is that agent going to feel about its own off switch.[749.116] [749.116][S05] Your robot always has an incentive to preserve its own function.[753.854] [753.854][S05]If it gets turned off, it can’t vacuum the floors anymore.[757.224] [757.224][S05]It can’t bring you coffee. It want to disable its off switch for example.[761.762] [761.762][S01] Jan Leike is a Senior Research Scientist at DeepMind[764.665] [764.665][S01]also working on AI safety.[766.8] [766.8][S05] If you turn it off, then it can’t vacuum the floors.[769.536] [769.536][S05]So if it understands how the off switch the off switch mechanism works,[774.474] [774.474][S05]it would want to disable it. What we want is you want our systems[778.145] [778.145][S05]to actually do something that is good for us -[782.015] [782.015][S05]ready to do something that we actually wanted, not just what we said we wanted.[787.421] [787.421][S01] But how do you get around these kinds of problems?[790.357] [790.357][S01]Well we already know that writing a long list of dos and dont’s won’t work.[795.729] [795.729][S01]However long your list gets you’re always going to forget something.[800.334] [800.334][S01]The new breed of learning agents is going to need a different approach.[804.838] [804.838][S05] So one direction that we are pursuing is learning what functions[808.509] [808.509][S05]for reinforcement learning agents. And you can kind of think of this[811.712] [811.712][S05]as learning what your systems should be doing from human feedback.[817.284] [817.284][S05]So for example in in one work that we did together with Open AI[822.389] [822.389][S05]is we trained a simulated robot to do a backflip,[826.727] [826.727][S05]and um the way this works is like the robot does some movement[831.632] [831.632][S05]and then you watch a video of that movement and you or you say two videos[836.703] [836.703][S05]and then you can compare which of those looks more like a backflip.[840.374] [840.374][S01] Jan’s experiment has a human sitting and watching a screen,[844.244] [844.244][S01]looking at an AI attempting to do a backflip.[848.182] [848.182][S01]Each time the human will feedback on whether the attempt was good enough,[852.386] [852.386][S01]slowly nudging the agent in the right direction.[856.223] [856.223][S01]Here is the key idea - with constant human feedback,[860.827] [860.827][S01]the human can communicate their preferences[864.064] [864.064][S01]without having to actually specify them,[867.701] [867.701][S01]and risk oversimplifying things in the process.[871.104] [871.104][S05] And after like a few hundred rounds of feedback,[874.141] [874.141][S05]the robot can actually perform a backflip.[876.243] [876.243][S05]It’s kind of learn what the objective should be -[878.779] [878.779][S05]that the objective should be a backflip, and what a backflip is.[882.082] [882.082][S05]This is difficult to specify precisely what a backflip would be[887.487] [887.487][S05]in so in my case like I can’t do a backflip, right?[891.792] [891.792][S05]And but I can see if the systems do a backflip,[895.629] [895.629][S05]and in some ways like this allows us to get super human capability.[901.101] [901.101][S01] The AI then is essentially being rewarded[904.738] [904.738][S01]by pleasing the human - in a way?[907.407] [907.407][S05] Exactly. And so we can we can do a little experiment.[911.111] [911.144][S01] Okay.[912.079] [912.079][S05] I’ll try to teach you to make a sound.[914.047] [914.047][S01] Okay. [S05] By giving feedback.[916.483] [916.483]So you’ll make two sounds. [S01] Mm-hmm.[918.385] [918.385][S05] And I’ll tell you which of the two sounds is closer[921.054] [921.054][S05]to what I have in mind.[922.389] [922.389][S01] Okay. [S05] Yeah, does that sound good?[924.558] [924.558][S01] Let’s do it, let’s do it! So I’m being the AI here?[926.593] [926.593][S05] Yeah you’re being the AI and I’m being the human teacher.[929.329] [929.329][S01] And you’re training me essentially with reinforcement learning,[932.533] [932.533][S01]and my reward function is getting you to be happy.[935.469] [936.069][S05] You know that’s like - so the reward function here[938.172] [938.172][S05]is like something is in my mind, right?[940.407] [940.407][S05]And I’m trying to teach you, so you can’t directly see the reward -[943.343] [943.343][S05]you can only see like my feedback.[945.612] [945.612][S01] But my objective is to get you to say you like the sound that I’m making.[949.85] [949.85][S05] Exactly. [S01] Okay, let’s think of two sounds.[951.785] [951.785][S01]Let’s go for: Meep! And: Meeeeep![956.056] [956.056][S05] The first one.[958.292] [958.292][Hannah laughs] [S01] Okay.[960.594] [960.594][S01]Have you actually got something in your mind or are you just making -[962.763] [962.763][S01] This went on for a while. Beep Beep and [more noises here].[969.002] [969.002][S01]But with Jan’s feedback, we got Beep and Beep -[973.34] [973.34][S05] I’d go with the first one - [S01] Absolutely nowhere![976.677] [976.677][Jan laughs][977.778] [977.778][S05] This is really hard because of the expression problem actually.[981.748] [981.782][S01] I mean ultimately - but if I was an actual AI[984.351] [984.351][S01]I’d have I’d have gone through 10,000 iterations by now.[987.855] [987.855][S05] well there’s you have the, have a human loop[990.257] [990.257][S05]that reviews all of that so it is it is kind of slow um[994.394] [994.394][S05]but this is actually a problem that we have in our systems is that,[998.298] [998.298][S05]like in order to give useful feedback, you have to have useful examples, right?[1003.604] [1003.604][S05]In this case you produced sounds that are very similar,[1006.84] [1006.84][S05]and like the sound I had in mind[1008.008] [1008.008]was like very different [S01] Totally different[1009.476] [1009.476][S05] So I, I don't really like have the opportunity[1012.412] [1012.412][S05]to give you a very useful feedback, right?[1014.248] [1014.248][S01] Damn My optimization strategy was terrible here, okay [laughs][1017.251] [1017.251][S05] But this is like basically this is the same thing would come up[1021.088] [1021.088][S05]with a backflip right like if your robot just like lies on the floor[1024.291] [1024.291][S05]in like two different ways, like what are you going to do?[1026.36] [1026.36][S01] But you would hope though after I’d have maybe a hundred go’s at it,[1030.297] [1030.297][S01]I’d end up with something that began to approach what you had in mind.[1034.468] [1034.468][S05] yeah, definitely, definitely.[1036.703] [1036.703][S01] Is it like - oh no, I’m not allowed to ask any questions[1038.639] [1038.639][S01]am I, because I’m an AI.[1039.74] [1039.74][S05] I mean ideally this is at some point[1041.341] [1041.341][S05]this is what we want our assistants to be able to do, right?[1044.378] [1044.378][S05]Or just like describe the sound to you in actual language[1047.347] [1047.347][S05]and then you could just do it, that would be really cool, right?[1049.149] [1049.183][S01] Yeah.[1050.117] [1050.117][S05] So this is the kind of research that we want to do in the future,[1052.553] [1052.553][S05]that’s like the kind of systems that we want to figure out how to build.[1055.789] [1055.789][S01] Oh, okay, actually. Everything that I have done so far has been like voices.[1059.993] [1059.993][S01]You didn’t say that it had to be a vocal thing, did you?[1063.564] [1063.564][S05] Yeah - [S01] Okay, hang on.[1065.165] [1065.165][S01]Okay, alright, how about this, how about this.[1067.067] [1067.067][S01][clicks tongue]. And [whistles].[1070.637] [1070.637][S05] The first one. [S01] Ooh.[1071.972] [1071.972][S05] The first one is actually what I had in mind.[1074.107] [1074.107][S05][Laughs]. This is great.[1075.008] [1075.008][S01] I was so far away.[1077.744] [1080.147][S01]The only problem with setting up reinforcement[1082.382] [1082.382][S01]learning with a human standing over it offering feedback at every stage[1087.321] [1087.321][S01]is that it is monumentally slow.[1090.524] [1090.524][S01]Now it would take an agent months if not years to master a game like Atari.[1095.963] [1095.963][S01]And even then, you’d need to hire a pretty large group of poor students[1100.1] [1100.1][S01]to do the boring job of supplying feedback in real-time.[1104.438] [1104.438][S01]But there is an alternative.[1106.24] [1106.24][S01]You can rustle up a slightly more sophisticated learning partnership.[1111.612] [1112.212][S05] So what we do when we actually built these systems is[1115.749] [1115.749][S05]that we don’t literally do the experiment that we do now.[1118.852] [1118.852][S05]But instead we have - we train a neural network -[1122.389] [1122.389][S05]a second neural network that learns how I would give feedback as a human[1128.362] [1128.362][S05]and then the neural network can teach you[1130.531] [1130.531][S05]because it can just oversee all of the things that we are doing.[1133.333] [1133.333][S01] By now neural networks have become really good at spotting patterns.[1137.738] [1137.738][S01]Dog, not dog. Backflip, not backflip.[1141.341] [1141.341][S01]Perhaps you don’t always need a human in the loop laboriously giving feedback.[1146.513] [1146.513][S01]Why not have two agents - one attempting the task,[1150.284] [1150.284][S01]and the other deciding if it succeeded.[1153.62] [1153.62][S05] The reason why this works very well[1155.355] [1155.355][S05]is because evaluating the objective[1158.091] [1158.091][S05]is an easier task than producing the behavior that achieves it.[1163.197] [1163.197][S05]So you can have something like the for example[1165.432] [1165.432][S05]the and the back-flipping robot all the neural network has to do[1168.869] [1168.869][S05]is like look at what the robot is doing and see whether it’s a backflip.[1173.44] [1174.308][S01] You’re listening to DeepMind the podcast, a window into AI research.[1179.446] [1180.147][S01]But of course where this stuff really comes alive[1182.749] [1182.749][S01]is where you take the ideas - take them out of a computer simulation[1187.554] [1187.554][S01]and allow your imagination to roam into the world of embodied AI.[1193.16] [1193.16][S01]Robots that learn how to cook, robots that learn how to pack fruit in crates,[1198.465] [1198.465][S01]tuck you into bed and perform backflips.[1201.869] [1201.869][S01]It’s time to visit the DeepMind robotics laboratory.[1205.973] [1207.274][S03] we are standing right outside of the DeepMind robotics lab.[1210.978] [1210.978][S01] This is your lab? This is where you spend your days?[1213.447] [1213.447][S01]Well it’s not just my lab, but yes, this is our lab.[1215.482] [1215.482][S03] Well it’s not just my lab, but yes. It is our lab.[1216.917] [1216.917][S01] Can we go inside? [S03] Yes. We’ll get our badge . . .[1218.752] [1218.752][S01] This is Jackie Kay, a software engineer.[1222.055] [1222.055][S01]Packed to the rafters with robots.[1224.124] [1224.124][S03] Yes. [S01] It’s very cool![1226.56] [1226.56][S03] We’re basically running out of space.[1228.495] [1228.495][S01] It’s quite noisy in here.[1230.497] [1230.497][S03] Yeah, we have a lot going on today.[1232.332] [1232.332][S01] This place isn’t quite the high security basement laboratory[1235.869] [1235.869][S01]you’d imagine.[1237.204] [1237.204][S01]There are half assembled robot arms and machine parts[1240.841] [1240.841][S01]scattered all around the place. Some look like mechanical hands,[1244.745] [1244.745][S01]others quite a lot like red kitchen aide style food mixers.[1249.116] [1249.116][S01]And curiously, there is a SpongeBob Squarepants[1252.119] [1252.119][S01]mascot hanging from the ceiling.[1254.454] [1254.454][S07] Hey kids, check it out! [Jackie laughs][1256.657] [1256.657][S03] I think we use it as kind of a punishment of somebody[1259.96] [1259.96][S03]like leaves a tool lying around when they are supposed to put it back,[1264.198] [1264.198][S03]we’ll like put a SpongeBob in his space or something[1266.5] [1266.5][Jackie laughs] you know. [S01] That seems like a fair punishment.[1268.268] [1268.268][S01]Much of the work in this room focusses on getting robotics arm[1271.872] [1271.872][S01]to learn how to do simple tasks.[1275.042] [1275.042][S01]Each robot is bolted to the floor in its own cordoned off area,[1278.879] [1278.879][S01]and as we come in the door, there’s one that’s caught my eye.[1282.382] [1282.382][S01]You know that game that you play when you’re a kid um[1285.285] [1285.285][S01]and you’re you have a tennis bat and a ball attached to it[1289.356] [1289.356][S01]and then you kind of play tennis on your own,[1291.491] [1291.491][S01]it sort of looks like a robotic version of that.[1293.927] [1293.927][S03] It’s called the ball in a cup. [S01] Oh, so it actually is![1296.864] [1296.864][S03] Yes. It is exactly that.[1298.532] [1298.532]It is trying to swing the ball into that little basket. [S01] Oh![S03][1302.469] [1302.469][S03]So you can see it’s swinging the ball around a little,[1305.572] [1305.572][S03]and it’s actually checking position of the yellow ball um[1308.909] [1308.909][S03]and it’s trying to minimize the distance[1311.712] [1311.712][S03]from that ball to the area inside the basket, um -[1315.582] [1315.582][S01] Quick! it did it! [S03] Oh - it got it -Yes![1317.451] [1318.418][S01] Not exactly a smooth delivery of the ball into the cup -[1322.489] [1322.489][S01]more a slightly clumsy, lucky flip.[1325.759] [1325.759][S01]Of course if you’re ultimate goal was to make a perfect cup[1329.696] [1329.696][S01]and ball playing robot,[1331.431] [1331.431][S01]you could directly program a machine to do that without fail.[1335.669] [1335.669][S01]You can build robots that perform simple tasks in much more elegant ways.[1340.941] [1340.941][S01]But that’s not really the point.[1343.11] [1343.11][S01]Here’s Raia Hadsell, Senior Research Scientist at DeepMind.[1347.114] [1347.114][S02] A robot is a machine that can take over a task.[1351.885] [1351.885][S02]One can say that in the broadest sense[1353.253] [1353.253][S02]that your dishwasher at home and your vacuum cleaner are both robots[1356.557] [1356.557][S02]because they do something complex, they run on their own,[1359.193] [1359.193][S02]they have some autonomy to them of course[1361.295] [1361.295][S02]we also like to think about robots about having some intelligence[1364.898] [1364.898][S02]and then you get more into the realm of a robot with AI.[1368.502] [1368.502][S02]And I’m particularly interested in saying[1371.071] [1371.071][S02]how can we take this AI technology[1373.574] [1373.574][S02]and make it work on robots so that we have embodied AI.[1378.078] [1379.546][S01] And that is an important distinction.[1381.882] [1381.882][S01]The focus of the research in this robotics lab[1385.352] [1385.352][S01]isn’t to create robots that are told what to do,[1389.122] [1389.122][S01]but have them learn their own skills[1391.525] [1391.525][S01]much in the same way that other agents do here.[1394.494] [1394.494][S01]The robots here are what’s known as embodied AI.[1398.932] [1398.932][S03] So let’s say yeah, the task is a robot that’s trying to lift a box um[1405.372] [1405.372][S03]I as the programmer in the kind of traditional setting[1409.576] [1409.576][S03]would say go to box, open hand,[1413.881] [1413.881][S03]move hand um you know 30 centimeters towards the box, close hand.[1420.954] [1420.954][S01] But this is different.[1422.322] [1422.322][S03] Yeah, this is sort of taking what we want[1426.527] [1426.527][S03]and then figuring out how to accomplish what we want.[1429.363] [1429.363][S01] Every few seconds this particular robot[1431.164] [1431.164][S01]will have an attempt of getting at ball into the cup[1434.535] [1434.535][S01]before pausing, resetting, and having another go.[1438.605] [1438.605][S01]And every now and then, by chance, the ball lands in the cup[1443.01] [1443.01][S01]and the robot is rewarded with a positive score.[1445.879] [1445.879][S01]Just like if it was playing a computer game.[1448.882] [1448.882][S03] The reset in between the training episodes[1452.119] [1452.119][S03]where it untangles itself or flips the ball out of the cup,[1455.522] [1455.522][S03]those are scripted but then when it actually tries to accomplish the task,[1461.195] [1461.195][S03]that is a policy which it has taught itself through experience.[1464.431] [1464.431][S01] Over time from everything it learns from the schools it receives,[1468.635] [1468.635][S01]the robot builds a picture of how the ball moves,[1472.706] [1472.706][S01]and how this relates to the robot’s own movement.[1476.31] [1476.31][S01] So had we come in right at the very beginning bit, what would we see?[1480.714] [1480.714][S03] Just completely random noise,[1483.65] [1483.65][S03]probably very chaotic swinging around, ah - oh![1487.688] [1487.688][S01] Oh it did it in one! It did it in one![1489.489] [1489.489][S03] That was quite - that was wow! That was like in three seconds.[1493.627] [1493.627][S01] So will it know now that that movement gave it a successful result?[1499.266] [1499.299][S03] Yes.[1500.3] [1500.3][S01] And so the next time that we see this do we expect it[1502.936] [1502.936][S01]to be better than when we walked in?[1504.905] [1504.905][S03] Yeah, umm it will try similar actions[1508.175] [1508.175][S03]ah that gave it a positive reward or a success.[1512.145] [1512.145][S01] It’s just sort of sitting there quite smugly right now.[1514.047] [1514.047][S03] Yes.[1515.749] [1523.023][S01] There is a big advantage to getting the AI to figure out tasks for itself.[1528.095] [1528.095][S01]You’ll end up with a robot that is much more flexible out of the box.[1532.566] [1532.566][S01]It doesn’t matter what you want it to do - tie a knot,[1535.502] [1535.502][S01]stack some bricks, peel a banana.[1538.305] [1538.305][S01]Just as long as you can clearly communicate your objective,[1541.675] [1541.675][S01]you don’t need to specify a long list of instructions for these robots.[1546.346] [1546.346][S01]And perhaps they’ll come up with a new way of banana peeling[1549.149] [1549.149][S01]that you haven’t thought of.[1551.518] [1551.518][S01]But there is also a big drawback in training AI[1555.122] [1555.122][S01]that has a physical body - they are much slower to learn[1559.193] [1559.193][S01]than all of the disembodied agents you’ll find in this building.[1562.93] [1562.93][S01]The ones that only exist inside a computer.[1566.5] [1567.067][S03] Other researchers are able to take advantage of parallelism -[1570.904] [1570.904][S03]that is they run their environments in simulation, on computers,[1575.742] [1575.742][S03]and they can run them in parallel on hundreds of compute -[1579.479] [1579.479][S03]different computers.[1580.781] [1580.781][S03]We're all gathering data about this environment[1583.483] [1583.483][S03]they’re trying to learn something about.[1585.285] [1585.285][S03]We only have let’s see there’s 4 robots here[1589.122] [1589.122][S03]and they are frequently not running the same experiment,[1591.959] [1591.959][S03]so it might just be one robot collecting data.[1595.095] [1595.095][S03]And that means our training will be orders of magnitudes slower[1599.032] [1599.032][S03]for comparable tasks,[1600.501] [1600.501][S03]compared to ones that are running in simulation on computers.[1603.537] [1604.404][S01] Progress is slower in this room, and you can tell.[1608.809] [1608.809][S01]The agents here are a little less accomplished.[1613.247] [1613.247][S01]In another corner another robot is trying to pick up a Lego brick[1617.751] [1617.751][S01]with a hooked gripping device, kind of like a claw,[1621.021] [1621.021][S01]and there’s a rather ominous box of mangled Lego bricks next to it.[1625.692] [1625.692][S03] In the exploration phases of training[1627.895] [1627.895][S03]it will sort of randomly open and close its gripper,[1630.23] [1630.23][S03]and I think this one has some shaping to ah close its gripper[1634.535] [1634.535][S03]when it detects it’s close to the brick.[1636.57] [1636.57][S01] Ah - close![1638.639] [1638.639][S03] And then it also has a fixed training time [S01] ah![S03][1641.375] [1641.375][S03]so after some number of seconds, it just will give up[1644.845] [1644.845][S03]and go back to the start and then try again.[1648.482] [1648.482][S03]Lego is sort of this building block to general purpose manipulation.[1655.989] [1655.989][S03]If you know if we can stack you know two bricks together,[1658.959] [1658.959][S03]we can then do kind of arbitrarily[1660.794] [1660.794][S01] It’s going to break it![1662.329] [1662.329][Jackie laughs] [S03] Oh it’s fine.[1664.398] [1664.398][S01] sorry I got distracted by the robot smashing up the Lego.[1667.267] [1668.268][S01] There is real potential here, and so Jackie and the team[1672.539] [1672.539][S01]are constantly trying to find ways of speeding up that learning process.[1677.211] [1677.211][S03] One technique ah that we are looking into in order to[1681.248] [1681.248][S03]in that some of the researchers here[1683.183] [1683.183][S03]have done some really cool work on in the past is something we call[1687.02] [1687.02][S03]Sim to reel - or simulation to reality transfer[1691.859] [1691.859][S03]which is where you take a simulation on a computer that models your robot -[1698.365] [1698.365][S03]and we can learn kind of in broad strokes what the robot is like,[1702.336] [1702.336][S03]how its actions affects its environment[1704.404] [1704.404][S03]and how it can do something similar to a task it’s trying to learn in real life.[1710.577] [1710.577][S03]So once we figure out all of that in simulation,[1713.38] [1713.38][S03]without even touching the real robot,[1715.315] [1715.315][S03]we can transfer the data it’s collected onto real hardware.[1719.586] [1719.586][S01] So you’re - you can cheat basically.[1721.522] [1721.555][S03] Yeah.[1722.456] [1722.456][S01] You can cheat by imagining the real robot within a computer[1727.327] [1727.327][S01]calculating all the physics that would happen in real life[1730.13] [1730.13][S01]and then use the same techniques.[1731.765] [1731.765][S01]You used that army of computers to give you a bit of a head start[1735.002] [1735.002][S01]before you even apply it to the real physical robot.[1738.105] [1738.105][S03] Exactly.[1739.039] [1739.039][S01] So the real robots end up acting the same way as the simulated robots do.[1743.911] [1743.911][S03] Well they started out acting the same way[1746.079] [1746.079][S03]as the simulated robots do and then as they train more,[1750.384] [1750.384][S03]they might start behaving slightly differently or better[1753.153] [1753.153][S03]when they go into reality.[1754.988] [1757.324][S01] Using simulation might give you a head start on reality[1760.394] [1760.394][S01]but it’s never going to match precisely.[1762.729] [1762.729][S01]The real robot has to contend with grip, friction, gravity, wear and tear.[1768.602] [1768.602][S01]All of which play important roles in the real world,[1771.772] [1771.772][S01]but none of which will be perfectly represented inside the computer.[1776.476] [1776.476][S01]And all of that means, well -[1778.679] [1778.679][S01]I think science fiction may have set some false expectations.[1783.383] [1784.284][S01] One thing um that I am a little bit surprised about being in here[1789.022] [1789.022][S01]um don’t take this the wrong way, but these robots are a bit rubbish[1793.227] [1793.227][S03] Yes, it’s true, I mean, we’ve got a lot of work to do.[1797.931] [1797.931][S01] But as with so much of what happens here at[1800.1] [1800.1][S01]DeepMind, it’s not so much about these exact agents,[1803.604] [1803.604][S01]it’s not about cup and ball[1805.639] [1805.639][S01]or Lego stacking, it’s about the type of intelligence being acquired,[1810.677] [1810.677][S01]and how that fits into the bigger picture.[1813.38] [1814.114][S03] we want to demonstrate general purpose physical intelligence.[1818.018] [1818.018][S01] Physical intelligence? [S03] Right.[1819.786] [1819.786][S03]So contrasting that with kind of an intelligence that’s not embodied,[1825.225] [1825.225][S03]ah which maybe it can learn to play games,[1827.261] [1827.261][S03]or maybe even understand language, physical intelligence is looking[1832.399] [1832.399][S03]at how the physical actions of your body affect the real world[1836.937] [1836.937][S03]ah so we want to take a wide variety of tasks,[1840.04] [1840.04][S03]playing with objects, using tools,[1843.677] [1843.677][S03]ah maybe walking around or running in the future[1847.481] [1847.481][S03]and we want to show that robots can teach themselves[1851.485] [1851.485][S03]those tasks- how to do those tasks.[1853.72] [1853.72][S01] In the physical world?[1855.122] [1855.122][S03] In the physical world, yes.[1856.823] [1859.226][S01] But physical intelligence is of course[1861.762] [1861.762][S01]only one type of intelligence - one string to the robot’s bow,[1866.567] [1866.567][S01]and DeepMind as we have seen dares to dream big.[1870.938] [1872.206][S01]Here’s Murray Shanahan with the big finish.[1874.942] [1874.942][S06] The holy grail of AI research[1876.944] [1876.944][S06]is to build artificial general intelligence[1879.513] [1879.513][S06]so to build AI that is as good at doing an enormous variety of tasks[1885.752] [1885.752][S06]as we humans are,[1887.521] [1888.155][S06]so we’re not specialists in that kind of way,[1890.624] [1891.491][S06]you know a young adult human can learn to do a huge number of things.[1895.195] [1895.195][S06]You can learn to make food, you can learn to um make a company,[1900.033] [1900.033][S06]you can learn to build things, to fix things, you can do[1902.87] [1902.87][S06]so many things to have conversations, to rear children, so all of those things[1907.307] [1907.307][S06]- and we really want to be able to build AI[1909.71] [1909.71][S06]that has the same level of generality as that.[1912.679] [1915.249][S01] If you want to know more about robotics and technical AI safety,[1919.686] [1919.686][S01]then head over to the show notes where you can also explore the world of[1923.09] [1923.09][S01]AI research beyond DeepMind, and we’d welcome your feedback[1926.693] [1926.693][S01]or your questions on any aspects of artificial intelligence[1929.997] [1929.997][S01]that we’re covering in this series.[1931.999] [1931.999][S01]So if you want to join in the discussion or point us to stories or resources[1936.003] [1936.003][S01]that you think other listeners would find helpful, then please let us know.[1939.339] [1939.339][S01]You can message us on Twitter or you can email us -[1942.242] [1942.242][S01]podcasts@deepmind.com[1944.378]"}
{"file_name": "audio/val_000022.wav", "transcription": "[0.0][S02] I think there are just[1.625] [1.625][S02]a huge number of enormously interesting[3.74] [3.74][S02]philosophical questions that AI gives rise to.[7.33] [7.33][S02]What is the nature of the human mind?[10.2] [10.2][S02]What is the nature of mind?[11.76] [11.76][S01] What about consciousness?[13.53] [13.53][S02] I do think that is the wrong question,[15.96] [15.96][S02]and I think it's wrong in many ways.[17.58] [17.58][S01] How good do you think that AI is at reasoning?[20.04] [20.04][S02] Well, that's a very interesting and open[22.67] [22.67][S02]question, and somewhat controversial.[25.35] [25.35][S02]It really is astonishing to think[26.9] [26.9][S02]that every single child born today,[29.94] [29.94][S02]they're going to grow up in a world where they've never known,[32.61] [32.61][S02]they've never known a world in which machines[35.27] [35.27][S02]can't talk to them.[37.14] [37.14][MUSIC PLAYING][39.4] [41.66][S01] Welcome back to \"Google DeepMind,\" the podcast.[44.52] [44.52][S01]My guest on this episode is Murray Shanahan,[47.31] [47.31][S01]Professor of Cognitive Robotics at Imperial College London[50.48] [50.48][S01]and Principal Research Scientist at Google DeepMind.[53.73] [53.73][S01]Now, we have all heard the stories[55.37] [55.37][S01]about people falling in love with their chat bots,[57.86] [57.86][S01]about people pushing large language models to contemplate[61.59] [61.59][S01]their own existence, or questioning[63.99] [63.99][S01]the limits of their conceptual understanding of reality.[67.0] [67.0][S01]But these kinds of questions about self-identity and thinking[71.85] [71.85][S01]and metacognition have been puzzling philosophers[75.21] [75.21][S01]for millennia already.[77.71] [77.71][S01]And so it makes sense that they should[79.53] [79.53][S01]be turning to AI to interrogate the most profound questions[83.82] [83.82][S01]about the nature of AI's intelligence,[86.38] [86.38][S01]of its current capabilities, even its consciousness[89.31] [89.31][S01]or other ones.[90.73] [90.73][S01]Murray Shanahan has been working in the field of AI since[93.96] [93.96][S01]the 1990s.[95.14] [95.14][S01]And if you've been following this podcast for a while,[97.99] [97.99][S01]you will remember him as the man that consulted on the 2014[102.18] [102.18][S01]science fiction film \"Ex Machina,\"[104.68] [104.68][S01]about a computer programmer who gets the chance to test[107.67] [107.67][S01]the intelligence of a female robot, Ava,[111.22] [111.22][S01]and ultimately questions whether she is conscious.[114.16] [114.16][S01]Welcome back to the podcast, Murray.[116.1] [116.1][S01]Just thinking back, because I know that you played a key role[119.18] [119.18][S01]in \"Ex machina,\" shall we say, the Alex Garland film.[122.07] [122.07][S01]What do you think you got right in that film[125.31] [125.31][S01]and in other science fiction films that were around[127.58] [127.58][S01]at the time?[128.25] [128.25][S01]I mean, thinking back to 10, 15 years ago,[130.979] [130.979][S01]were we on the right track?[132.75] [132.75][S02] So one respect in which \"Ex Machina\"[135.65] [135.65][S02]really did a great service was that it does raise a whole load[139.1] [139.1][S02]of very interesting and provocative questions about[141.59] [141.59][S02]consciousness and about AI and consciousness, and therefore,[145.13] [145.13][S02]about consciousness itself.[147.12] [147.12][S02]So that's one huge success.[150.0] [150.0][S02]But it's interesting that just very shortly before \"Ex Machina\"[152.72] [152.72][S02]came out, \"Her\" came out.[154.65] [154.65][S02]So Spike Jones's movie \"Her\" came out.[156.84] [156.84][S02]And at the time, I really wasn't all that keen on \"Her\"[160.13] [160.13][S02]as a movie, because I just thought it was so implausible[163.46] [163.46][S02]that a person could fall in love with this kind of disembodied[166.55] [166.55][S02]voice, even if it's Scarlett Johansson's.[168.392] [168.392][S01] Yeah.[169.1] [169.1][S02] I mean, how wrong was that?[170.933] [170.933][S02]As a bit of prediction, I think \"Her\"[173.42] [173.42][S02]really did amazingly well at predicting the world we've got[176.82] [176.82][S02]now.[177.46] [177.46][S02]Now, we don't know quite how things[179.04] [179.04][S02]are going to unfold in the next few years[180.748] [180.748][S02]because maybe robotics will progress rapidly as[185.31] [185.31][S02]well in the way that language has in AI.[189.22] [189.22][S02]But at the moment, it's all about disembodied language.[192.85] [192.85][S02]And also, \"Her\" showed how people can, in fact,[196.5] [196.5][S02]very much form relationships, whatever, in the broadest sense,[201.78] [201.78][S02]with disembodied AI systems, which is an extraordinary thing,[205.84] [205.84][S02]really.[206.34] [206.34][S01] OK.[206.74] [206.74][S01]We're talking 10, 15 years ago, but your involvement in AI[209.31] [209.31][S01]goes back much further than this.[210.685] [210.685][S01]You knew John McCarthy.[211.71] [211.71][S02] I did know John McCarthy.[213.46] [213.46][S02]I knew him very well.[214.75] [214.75][S02]John McCarthy was a professor of computer science[218.28] [218.28][S02]and artificial intelligence.[219.93] [219.93][S02]Back in the day, he actually coined the phrase artificial[223.14] [223.14][S02]intelligence and was one of the authors[226.47] [226.47][S02]of the proposal for the very famous Dartmouth Conference that[230.94] [230.94][S02]took place in 1956, which was the first AI[233.73] [233.73][S02]conference in the world.[235.27] [235.27][S02]And that conference really mapped out the whole field.[237.52] [237.52][S02]People just weren't thinking about this kind of thing[239.728] [239.728][S02]seriously at all.[240.59] [240.59][S02]There was just a handful.[241.64] [241.64][S02]So I think he was a real radical thinker and always was.[244.695] [244.695][S01] OK.[245.32] [245.32][S01]That choice of words, artificial intelligence, back in 1955,[248.33] [248.33][S01]was it a good choice of words?[249.833] [249.833][S02] Yeah.[250.75] [250.75][S02]I mean, I still think it was.[252.35] [252.35][S02]I mean, I know that some people don't[254.11] [254.11][S02]think that perhaps it wasn't a good choice of words, but I--[257.746] [257.746][S01] Give us some of their arguments.[259.579] [259.579][S02] So first of all, there[261.49] [261.49][S02]is the word intelligence.[263.24] [263.24][S02]So intelligence itself is, in some ways, a very contentious[267.37] [267.37][S02]concept, especially if people think[271.33] [271.33][S02]about IQ tests and that kind of thing, and the idea[274.792] [274.792][S02]that intelligence is something that[276.25] [276.25][S02]can be quantified on a straightforward, simple scale,[281.095] [281.095][S02]and then some people are more intelligent than others.[283.7] [283.7][S02]And I think in psychology, it's well-recognized today[287.413] [287.413][S02]that there are many different kinds of intelligence.[289.58] [289.58][S02]And this is a really important point, right?[291.68] [291.68][S02]There is that concern about that word there.[293.873] [293.873][S02]So what would you have used differently?[295.54] [295.54][S02]Well, maybe artificial cognition or something.[298.18] [298.18][S02]I often use the word cognition to mean[301.68] [301.68][S02]kind of thinking and processing information and so on.[305.73] [305.73][S02]But it doesn't have the same ring to it, does it?[307.833] [307.833][S02]Let's be honest.[308.5] [308.5][S01] No.[309.125] [309.125][S01]Especially not now.[310.09] [310.09][S01]I think we're too far down this road, aren't we?[312.13] [312.13][S02] Yeah.[312.42] [312.42][S02]The word artificial, I don't really[314.04] [314.04][S02]have a problem with the word artificial.[315.893] [315.893][S02]That seems like the right kind of thing.[317.56] [317.56][S02]It's alluding to the fact that it's something that we've built[320.143] [320.143][S02]and it hasn't evolved in nature, and so that[322.44] [322.44][S02]seems the right sort of word.[323.648] [323.648][S01] The objection to that word, I guess,[325.648] [325.648][S01]is that ultimately, everything that artificial intelligence is[328.59] [328.59][S01]built on is, at some level, constructed by humans.[331.543] [331.543][S02] Sure.[332.46] [332.46][S02]Yes.[333.61] [333.61][S02]But it is.[334.62] [334.62][S02]So what's wrong with the word in that case?[337.72] [337.72][S02]I mean, I think that's true.[339.545] [339.545][S01] You were working on symbolic AI, right?[341.67] [341.67][S01]Just talk to us about the difference[343.17] [343.17][S01]between that and the other types and where we're at now with--[346.12] [346.12][S02] Yeah, absolutely.[347.11] [347.11][S02]Yeah.[347.61] [347.61][S02]Yeah, so the so-called symbolic paradigm[349.74] [349.74][S02]of artificial intelligence was very much preeminent,[352.76] [352.76][S02]very much dominant for decades, for many decades.[356.3] [356.3][S02]So the idea there is that it's all about the manipulation[360.76] [360.76][S02]of symbols and of language-like sentences and symbols[366.79] [366.79][S02]and using reasoning processes with those symbols.[371.9] [371.9][S02]So the classic example would be an expert system.[374.81] [374.81][S02]So where back in the 1980s, people were building these[379.3] [379.3][S02]expert systems, and the idea there was that you would try[381.82] [381.82][S02]to encode medical knowledge, say, in a set of rules,[386.33] [386.33][S02]and the rules would be something like, oh,[388.78] [388.78][S02]if the patient has temperature of 104 and their skin is purple[396.34] [396.34][S02]and then there's a 0.75% probability that they've got[401.5] [401.5][S02]skinny-itis or something.[403.63] [403.63][S02]You can tell that I'm not a medical doctor.[406.362] [406.362][S01] Just about.[407.32] [407.32][S02] Yeah.[408.237] [408.237][S02]And then so you'd have thousands and thousands of these sorts[410.99] [410.99][S02]of rules would be put into a kind of big knowledge base,[415.11] [415.11][S02]and then you'd have what was called an inference engine which[419.03] [419.03][S02]would carry out logical reasoning over all[421.49] [421.49][S02]of these rules and come to some conclusion about what[424.4] [424.4][S02]the likely disease was in.[425.783] [425.783][S01] But it was a lot of, if this, then that.[427.95] [427.95][S02] It was a lot of, yeah, if, then type rules,[430.37] [430.37][S02]largely.[430.92] [430.92][S02]And one of the big problems with that[432.26] [432.26][S02]is that, where do the rules come from?[433.98] [433.98][S02]Well, somebody has to write them all out, basically.[437.22] [437.22][S02]And so there was a whole field of knowledge elicitation[440.09] [440.09][S02]where you go around to experts and you try and extract[442.64] [442.64][S02]from them their understanding in their domain, which[445.94] [445.94][S02]could be medical diagnosis, it could be fixing photocopiers,[450.12] [450.12][S02]it could be the law.[451.65] [451.65][S02]And you try and codify all of this[453.59] [453.59][S02]into a computer comprehensible, very precise rule.[457.95] [457.95][S02]That is a very cumbersome process.[459.43] [459.43][S02]And also what you ended up with at the end[461.18] [461.18][S02]was very, very brittle.[462.3] [462.3][S02]It would go wrong in all kinds of ways.[464.82] [464.82][S02]Another big area of research was common sense,[467.66] [467.66][S02]because often it was realized that we implicitly[472.475] [472.475][S02]have an enormous amount of common sense[474.1] [474.1][S02]knowledge about the everyday world to do[476.38] [476.38][S02]with just everyday objects.[478.19] [478.19][S02]The fact that they're solid.[479.9] [479.9][S02]The fact that they move in certain ways,[482.57] [482.57][S02]they fit into each other in certain ways,[484.84] [484.84][S02]liquids and gases and gravity and all kinds[488.08] [488.08][S02]of things like that.[489.078] [489.078][S02]And we actually bring all of that knowledge[490.87] [490.87][S02]to bear all the time in what we're doing,[492.89] [492.89][S02]but it's sort of unconscious.[494.12] [494.12][S02]So then there was a big project, there[495.55] [495.55][S02]were various big projects to try and codify all[497.44] [497.44][S02]of that common sense knowledge.[498.8] [498.8][S02]And then trying to turn that into axioms and logic[501.46] [501.46][S02]and rules and everything was a nightmare.[503.36] [503.36][S02]So eventually, I think by about the early 2000s,[507.88] [507.88][S02]I really thought that this research paradigm was kind[510.82] [510.82][S02]of doomed, to be honest.[512.72] [512.72][S02]I sort of started moving away from it.[514.917] [514.917][S01] But then, of course, along came things[517.0] [517.0][S01]like neural networks and so on.[518.72] [518.72][S02] Yes.[519.595] [519.595][S01] Which was much less about if, then rules[522.46] [522.46][S01]and much more about extracting information[525.04] [525.04][S01]from a large amount of data.[526.85] [526.85][S02] Yeah.[527.08] [527.08][S01] But then I sort of wonder now about,[529.08] [529.08][S01]now that language is effectively cracked,[531.67] [531.67][S01]have we reached a higher level of abstraction[536.04] [536.04][S01]where we can go back to some more[538.5] [538.5][S01]of those symbolic techniques, some of those more[540.81] [540.81][S01]symbolic ideas?[541.435] [541.435][S02] Yeah.[542.352] [542.352][S02]Well, we certainly have, because nowadays[544.41] [544.41][S02]one of the hot topics at the moment with large language[548.04] [548.04][S02]models is reasoning.[550.29] [550.29][S02]So you have these so-called chain of thought models that[553.47] [553.47][S02]actually carry out a whole--[555.63] [555.63][S02]rather than simply generating an answer to a question,[559.03] [559.03][S02]they generate a whole chain of reasoning[561.63] [561.63][S02]before they issue the answer.[563.49] [563.49][S02]And that can be very, very effective.[565.51] [565.51][S02]So it's interesting how that harks[567.42] [567.42][S02]back in many ways to the kind of thing[570.42] [570.42][S02]that people were looking at back in the days of symbolic AI.[574.127] [574.127][S02]But the underlying substrate for doing all of that[576.21] [576.21][S02]is very, very different indeed, because it's not[578.43] [578.43][S02]hardcoded rules.[580.17] [580.17][S02]As you mentioned, it's neural networks that have learned.[584.015] [584.015][S01] Let me pick up on that point about reasoning.[586.39] [586.39][S01]As a philosopher, background in logic,[588.61] [588.61][S01]how good do you think the AI is at reasoning?[591.25] [591.25][S02] Well, that's a very interesting and open[593.91] [593.91][S02]question, and somewhat controversial.[595.96] [595.96][S02]So computer scientists and AI people,[598.54] [598.54][S02]they have a particular notion of reasoning, a particular concept[602.01] [602.01][S02]of reasoning, which very much harks back to formal logic[605.25] [605.25][S02]and theorem proving.[606.99] [606.99][S02]So in the days of symbolic AI, for example, then you[610.29] [610.29][S02]had systems that were really very good at doing theorem[613.2] [613.2][S02]proving with formal logic.[614.865] [614.865][S02]And so people think, well, that's proper reasoning.[616.99] [616.99][S02]That's your hardcore kind of reasoning.[620.22] [620.22][S02]And today's large language models,[622.69] [622.69][S02]they can't match the performance of a hand-coded theorem prover[629.01] [629.01][S02]or logic engine of the sort that's been around for decades.[632.62] [632.62][S01] Give me an example of a type of theorem[635.01] [635.01][S01]that might be able to be proved by a hardcoded system.[637.96] [637.96][S02] So it would be where you've got[639.96] [639.96][S02]maybe 20 or 30 axioms of logic.[643.435] [643.435][S01] So it might be something like the number that[645.81] [645.81][S01]follows one is to.[646.56] [646.56][S02] Well, I mean, it could be something like that.[649.185] [649.185][S02]It could be in the domain of number theory or something[651.72] [651.72][S02]very mathematical.[652.63] [652.63][S02]But it could be something much more everyday.[655.15] [655.15][S02]So for example, suppose that you've[656.79] [656.79][S02]got some very difficult logistical planning[659.34] [659.34][S02]problem where maybe you have hundreds of lorries and depots[664.14] [664.14][S02]and goods and all kinds of things like that,[667.06] [667.06][S02]and you need to plan the routes and the deployment[670.41] [670.41][S02]of the lorries and where they're going to go.[672.31] [672.31][S02]So that's a very difficult problem, computationally,[676.66] [676.66][S02]and it can be expressed very precisely in formal rules.[680.707] [680.707][S02]And that's the kind of situation where[682.29] [682.29][S02]you might want to use a good old-fashioned, straightforward[685.2] [685.2][S02]algorithm, planning algorithm of that's[687.81] [687.81][S02]been around for a long time.[688.98] [688.98][S02]Now, contemporary large language models[691.083] [691.083][S02]are getting better and better at this kind of thing,[693.25] [693.25][S02]but they're still--[694.2] [694.2][S02]you don't have those kinds of mathematical guarantees[696.75] [696.75][S02]that they're always going to come up[698.43] [698.43][S02]with exactly the right answer.[700.09] [700.09][S02]And it's very easy to make examples[702.57] [702.57][S02]where you have more and more axioms and so on,[705.3] [705.3][S02]where they're going to slip up.[706.867] [706.867][S02]There's a whole separate research direction,[708.7] [708.7][S02]which is to try and build more hand-coded things that[712.05] [712.05][S02]combine today's AI techniques with more[714.57] [714.57][S02]old-fashioned symbolic techniques[717.54] [717.54][S02]specifically for mathematical theorem proving.[719.813] [719.813][S02]And DeepMind has done some amazing work along those lines.[722.23] [722.23][S02]But that's different from large language models.[723.79] [723.79][S02]So with large language models, we're[725.29] [725.29][S02]thinking of these chatbots that can talk[728.55] [728.55][S02]about anything under the sun.[729.88] [729.88][S02]And one of the things they happen to be able to do[731.963] [731.963][S02]is a kind of reasoning.[733.68] [733.68][S02]That's not going to be, at the moment, quite as good[736.32] [736.32][S02]as you could do by hand building something for that.[739.18] [739.18][S01] It's kind of interesting[740.68] [740.68][S01]because hand building something is,[742.77] [742.77][S01]I mean, you end up with something that's very rigid.[745.42] [745.42][S02] That's the problem.[746.17] [746.17][S02]Yeah.[746.67] [746.67][S01] And brittle.[747.91] [747.91][S02] Yes, absolutely.[748.93] [748.93][S01] But then at the same time,[750.513] [750.513][S01]the flexibility that you get from the generative AI[753.39] [753.39][S01]approach, it's too floppy, as it were.[756.61] [756.61][S01]You want the rigidity in there.[758.16] [758.16][S02] Well, maybe or maybe not.[759.91] [759.91][S02]I mean, I think many examples of human affairs[762.41] [762.41][S02]are just not as black and white as that.[766.1] [766.1][S02]You do maybe want things to be a bit more blurry, even[769.52] [769.52][S02]in simple, everyday things.[771.05] [771.05][S02]Like, what would be good flowers to put over[773.72] [773.72][S02]in this corner of the garden?[775.26] [775.26][S02]Well, we've already got some roses in that corner there,[778.83] [778.83][S02]and those roses are yellow, so we'd-- but we can't have too[781.673] [781.673][S02]much yellow, so maybe we'd need to move them to the other corner[784.34] [784.34][S02]of the garden.[785.13] [785.13][S01] But then at the same time,[786.713] [786.713][S01]though, is this real reasoning or is this just the AI mimicking[791.9] [791.9][S01]well-structured arguments that have existed[794.96] [794.96][S01]in the training data, but just in a sort of novel environment?[799.353] [799.353][S02] Yeah.[800.27] [800.27][S02]Well, of course, that begs the question,[801.937] [801.937][S02]what is real reasoning?[803.48] [803.48][S02]I don't think-- it's not written in the sky what[806.66] [806.66][S02]real reasoning is.[808.02] [808.02][S02]It's up to us to define the concept of real reasoning[811.79] [811.79][S02]or of reasoning.[813.03] [813.03][S02]And so we have that--[814.91] [814.91][S02]we were talking earlier on about mathematical reasoning[817.25] [817.25][S02]of the sort that logicians do and that[820.525] [820.525][S02]is done by theorem provers in the past and so on and today.[826.71] [826.71][S02]When people were first using the terms like reasoning,[829.627] [829.627][S02]they weren't thinking of that kind of thing.[831.46] [831.46][S02]And when we used the word reasoning in everyday life,[833.68] [833.68][S02]we're not thinking about that sort of thing.[835.513] [835.513][S02]So if you're chatting away to a large language[837.6] [837.6][S02]model about your garden and you say,[840.215] [840.215][S02]oh, I'm thinking about what plants are right,[842.09] [842.09][S02]and it says, well, maybe you should consider[844.23] [844.23][S02]this kind of plant in that kind of location[846.06] [846.06][S02]because that's best for the soil and given[847.81] [847.81][S02]you said that it's windy there.[850.59] [850.59][S02]We would just say that there is supplying reasons.[853.09] [853.09][S02]I mean, it is supplying reasons.[854.7] [854.7][S02]Now, where they come from is another matter.[856.653] [856.653][S02]So people might say, well, it's just mimicking[858.57] [858.57][S02]what's in the training set.[860.35] [860.35][S02]But it's probably never seen exactly[862.02] [862.02][S02]that kind of scenario exactly before.[864.7] [864.7][S02]So it's moving beyond the training[866.76] [866.76][S02]set to a certain extent.[868.05] [868.05][S02]I think it's just using the everyday concept of reasoning[871.47] [871.47][S02]in an everyday way to call that reasoning.[873.22] [873.22][S01] I'm just thinking back[874.637] [874.637][S01]to some of the different characteristics[876.66] [876.66][S01]that the earlier philosophers wanted artificial intelligence[879.61] [879.61][S01]to have, and reasoning being one of them.[882.52] [882.52][S01]But then also the Turing test, which, of course, gets[885.07] [885.07][S01]brought up all the time, about a way[887.56] [887.56][S01]to test for the capability of an artificial intelligence.[890.27] [890.27][S01]I mean, it's kind of controversial,[892.46] [892.46][S01]I suppose, in terms of how good it ever[894.91] [894.91][S01]would have been as a test for the capability of AI.[897.83] [897.83][S01]What was your take on it?[899.45] [899.45][S01]Do you think it was ever a good test?[901.43] [901.43][S02] No.[902.62] [902.62][S02]I've always thought it was a terrible test, but a really[905.59] [905.59][S02]great spur to philosophical discussion about things.[909.83] [909.83][S02]And again, with a bit of hindsight,[911.63] [911.63][S02]maybe I might backtrack a little bit on a few of my views,[915.23] [915.23][S02]because I was certainly very, very[916.9] [916.9][S02]much of the opinion that embodiment[919.3] [919.3][S02]was a critical facet of intelligence,[923.08] [923.08][S02]was critical for achieving intelligence.[925.17] [925.17][S01] Which doesn't come anywhere near the Turing test[927.67] [927.67][S01]at all, right?[928.325] [928.325][S02] No.[929.158] [929.158][S02]The Turing test is absolutely, explicitly[930.88] [930.88][S02]nothing to do with embodiment because in the Turing test--[933.68] [933.68][S02]so just to remind people what it is.[936.22] [936.22][S02]So in the Turing test, you have two subjects, as it were.[938.835] [938.835][S02]One is a human and the other is the computer.[940.71] [940.71][S02]And then you have a judge.[941.793] [941.793][S02]The human judge can't see which is the computer[945.952] [945.952][S02]and which is the humans, and they're only[947.66] [947.66][S02]talking to these subjects through a kind[950.69] [950.69][S02]of chat-like interface.[952.35] [952.35][S02]They can't see whether they're embodied or not,[954.84] [954.84][S02]so we can easily suppose that the computer might[958.82] [958.82][S02]be one of today's large language models.[960.72] [960.72][S02]In which case, I have to say that today they[963.05] [963.05][S02]pretty much would pass the Turing test.[964.74] [964.74][S02]I mean, we've got to that point, which is amazing, really.[967.84] [967.84][S02]So I used to think that it was a bad test because it didn't test[970.88] [970.88][S02]any of these embodied skills.[972.66] [972.66][S02]So you'd need a robot, really, to test whether something[975.74] [975.74][S02]was capable of the kind of everyday cognition[978.11] [978.11][S02]that we all put to use when we're, for example, making[981.26] [981.26][S02]a cup of tea or something.[982.86] [982.86][S01] Because otherwise it's[984.277] [984.277][S01]a very, very narrow form of intelligence.[986.887] [986.887][S02] Yes, it's all to do with language[988.97] [988.97][S02]and reasoning, and not to do with the kinds of things[992.57] [992.57][S02]that evolution developed in us and in other animals[996.5] [996.5][S02]before language, which is the ability to manipulate and move[1000.73] [1000.73][S02]around with and navigate and exploit,[1004.645] [1004.645][S02]in the best sense of the word, the everyday physical world.[1007.705] [1007.705][S01] So actually, that's really interesting.[1009.83] [1009.83][S01]That's so interesting, because I often think about how--[1013.27] [1013.27][S01]fine, maybe the large language models we have at the moment[1016.353] [1016.353][S01]can pass the Turing test, but they[1017.77] [1017.77][S01]don't flinch if you throw a ball at your computer.[1019.94] [1019.94][S02] Oh, no, indeed.[1020.98] [1020.98][S01] And in a sense, there[1022.355] [1022.355][S01]are these, as you say, these much deeper forms.[1025.7] [1025.7][S01]Maybe we wouldn't class them as intelligence in the way[1029.043] [1029.043][S01]that we talk about it.[1029.96] [1029.96][S01]But ultimately, it sort of is a form of intelligence too.[1032.756] [1032.756][S02] It very much is a form of intelligence.[1035.089] [1035.089][S02]And moreover, I think that in the biological case--[1037.243] [1037.243][S02]and now I have to caveat all these things by saying,[1039.41] [1039.41][S02]in the biological case.[1041.079] [1041.079][S02]Our ability to think and to reason and to talk[1044.44] [1044.44][S02]is very much grounded in our interaction[1046.869] [1046.869][S02]with the everyday world.[1048.02] [1048.02][S02]If you think about almost all of your everyday speech[1053.17] [1053.17][S02]is using spatial metaphors.[1055.55] [1055.55][S02]I mean, they completely permeate our everyday speech.[1058.7] [1058.7][S02]Even the word permeate.[1059.84] [1059.84][S01] Yeah.[1060.07] [1060.07][S01]Absolutely.[1060.73] [1060.73][S02] Grounded.[1061.813] [1061.813][S02]I used the word grounded.[1063.13] [1063.13][S02]So we just use those kinds of things all the time.[1066.357] [1066.357][S01] Because we're fundamentally physical beings.[1068.69] [1068.69][S02] Because we're fundamentally physical beings,[1070.16] [1070.16][S02]and because our brains have evolved[1071.89] [1071.89][S02]to help us to survive and reproduce[1075.22] [1075.22][S02]in this physical world.[1078.02] [1078.02][S01] Yeah.[1078.728] [1078.728][S02] While interacting[1080.145] [1080.145][S02]with all these other beings that are doing the same thing.[1082.84] [1082.84][S01] Because there are some alternatives when[1084.64] [1084.64][S01]you are trying to test for the capability[1086.348] [1086.348][S01]of an artificial intelligence.[1088.15] [1088.15][S01]Just talk me through some of the potential alternatives[1091.03] [1091.03][S01]that we have.[1091.572] [1091.572][S02] Well, I think perhaps you've[1093.447] [1093.447][S02]got in mind the Garland test, what I call the Garland[1095.92] [1095.92][S02]test, which is--[1097.96] [1097.96][S02]so that goes back to the film \"Ex Machina,\"[1100.91] [1100.91][S02]which was directed by Alex Garland, of course.[1103.49] [1103.49][S02]And there's a bit in the script where Nathan, the billionaire[1106.81] [1106.81][S02]guy, is talking to Caleb, and Caleb,[1110.02] [1110.02][S02]who's the guy who's been brought in to interact with Ava,[1113.63] [1113.63][S02]the robot, and Caleb says, oh, I'm[1115.59] [1115.59][S02]here to conduct a Turing test on Ava.[1118.87] [1118.87][S02]And Nathan says, oh, no, we're way past that.[1121.87] [1121.87][S02]Ava could pass the Turing test easily.[1124.27] [1124.27][S02]The point is to show you she's a robot[1126.66] [1126.66][S02]and see if you still think she's conscious.[1129.13] [1129.13][S01] Wow.[1129.87] [1129.87][S02] And that's what I call the Garland test.[1131.59] [1131.59][S02]And it's different from the Turing test in two respects.[1134.02] [1134.02][S02]So first of all, the judge, as it were--[1136.62] [1136.62][S02]who, in that case, is Caleb--[1139.8] [1139.8][S02]can see that she's a robot.[1141.76] [1141.76][S02]So in the Turing test, the judge can't see which is which.[1145.3] [1145.3][S02]But here, the idea is that Caleb knows that she's a robot,[1149.83] [1149.83][S02]knows that her brain is an AI brain.[1153.52] [1153.52][S01] And yet, still attributes these characteristics[1156.72] [1156.72][S01]to her.[1157.24] [1157.24][S02] And the characteristic in question[1158.79] [1158.79][S02]also is different, because it's not intelligence.[1160.66] [1160.66][S02]It's not, can she think?[1161.66] [1161.66][S02]But, is she conscious?[1162.78] [1162.78][S02]Or, is it conscious?[1164.25] [1164.25][S02]Which is an entirely different test.[1166.06] [1166.06][S02]And I think intelligence and consciousness[1168.217] [1168.217][S02]are different things that we can disentangle those two things,[1170.8] [1170.8][S02]dissociate them.[1172.11] [1172.11][S02]So when I first read the script of the film,[1175.6] [1175.6][S02]those particular lines were in there for Caleb and Nathan,[1179.2] [1179.2][S02]and I wrote next to it in my version,[1181.43] [1181.43][S02]spot on with an exclamation mark,[1183.05] [1183.05][S02]because I just thought Alex had totally nailed[1185.5] [1185.5][S02]a really important idea there.[1187.37] [1187.37][S02]And so in my writing, I call this the Garland test.[1189.535] [1189.535][S02]And quite a few people have picked up on that[1191.41] [1191.41][S02]and called it the Garland test as well.[1193.04] [1193.04][S01] Is there a test that would really[1195.79] [1195.79][S01]impress you if an AI were able of passing it?[1198.482] [1198.482][S02] So I always was very impressed[1200.44] [1200.44][S02]with Francois Chollet's ARC test.[1204.35] [1204.35][S02]And that's A-R-C, which stands for abstract reasoning corpus.[1207.92] [1207.92][S02]So these are little sequences of images of the sort that you[1212.32] [1212.32][S02]get in IQ tests and things.[1214.78] [1214.78][S02]And the images are arranged in pairs.[1217.07] [1217.07][S02]So you have the first image.[1218.96] [1218.96][S02]It's kind of pixelated image.[1220.85] [1220.85][S02]It's got little cells with little kind of things[1224.59] [1224.59][S02]that you can interpret as objects or lines[1226.57] [1226.57][S02]and so on in the images.[1228.02] [1228.02][S02]And you're interested in--[1229.333] [1229.333][S02]the challenge is to work out a rule that[1231.0] [1231.0][S02]takes you from one image to the second one.[1233.14] [1233.14][S02]Then you've got to apply that rule to a third image.[1236.95] [1236.95][S02]First of all, he held out, made completely secret[1240.54] [1240.54][S02]all of the test ones, so you couldn't game[1242.94] [1242.94][S02]it by knowing what the actual test versions were.[1247.01] [1247.01][S01] Or using it in a training set.[1248.76] [1248.76][S02] Or using it in a training set.[1249.79] [1249.79][S02]That's what I mean by gaming it.[1251.71] [1251.71][S02]And also, he very carefully designed them[1253.98] [1253.98][S02]so that it was very different rules.[1256.03] [1256.03][S02]Each rule was completely different to the other rules.[1260.658] [1260.658][S02]And you usually had to find some kind[1262.2] [1262.2][S02]of intuitive application of often our everyday common sense[1266.13] [1266.13][S02]knowledge is seeing this as a liquid that's[1268.47] [1268.47][S02]moving in this direction or imagining this thing moving,[1272.25] [1272.25][S02]growing or something.[1273.182] [1273.182][S01] So it required grounding, in a way.[1275.14] [1275.14][S02] Well, it seemed to.[1276.64] [1276.64][S02]But recently, people have been able to make[1279.12] [1279.12][S02]significant progress on these in a more brute force kind of way.[1284.67] [1284.67][S02]So I feel that the solutions are not really[1291.134] [1291.134][S02]getting at the spirit of the original test quite so much.[1293.622] [1293.622][S01] Well, that's it, I guess, in a way,[1295.58] [1295.58][S01]is that as soon as you set a metric, as soon as you[1298.39] [1298.39][S01]set a bar for once we've crossed this threshold,[1301.76] [1301.76][S01]then we will have capability intelligence and consciousness,[1305.21] [1305.21][S01]whatever it might be.[1306.585] [1306.585][S01]It sort of changes the whole nature of the test in itself.[1310.143] [1310.143][S02] Yeah.[1311.06] [1311.06][S02]Well, people are going to start--[1312.3] [1312.3][S01] Optimizing for it.[1313.12] [1313.12][S02] --for the test, right?[1314.745] [1314.745][S02]It's Goodhart's law.[1316.0] [1316.0][S01] Absolutely.[1317.05] [1317.05][S01]A lot of people who've come on this podcast[1318.842] [1318.842][S01]have expressed real need for caution[1322.51] [1322.51][S01]about anthropomorphizing these things.[1324.34] [1324.34][S01]Are you one of those people who thinks that we shouldn't?[1327.07] [1327.07][S02] Well, I think there are different ways[1328.12] [1328.12][S02]of looking at this.[1328.94] [1328.94][S02]And I think there are good and bad forms[1331.93] [1331.93][S02]of anthropomorphization.[1333.86] [1333.86][S02]So on the one hand, people can start to form relationships[1338.98] [1338.98][S02]as they see it with AI systems, friendships and companionships[1343.54] [1343.54][S02]and mentorships.[1344.96] [1344.96][S02]And that can potentially be a bad thing[1348.75] [1348.75][S02]if they are misled into thinking that they can trust[1352.565] [1352.565][S02]the thing that they're talking to[1353.94] [1353.94][S02]or that they're really in love with it,[1355.66] [1355.66][S02]or that it really cares about them.[1357.285] [1357.285][S02]On the other end of the spectrum,[1358.66] [1358.66][S02]if an AI system is just using the word I,[1362.32] [1362.32][S02]then I think that that's a pretty harmless sort of form[1366.18] [1366.18][S02]of self-anthropomorphization.[1368.01] [1368.01][S02]We even see buses that say things like, on the side,[1370.74] [1370.74][S02]I am out of service.[1371.583] [1371.583][S02]And we don't have a problem with that kind of thing,[1373.75] [1373.75][S02]so I don't see why we should have[1374.67] [1374.67][S02]a problem with that with large language models either.[1377.23] [1377.23][S02]But I think we do tend to anthropomorphize things.[1380.02] [1380.02][S02]When we had satnavs in cars that weren't just in our phones,[1382.69] [1382.69][S02]I used to anthropomorphize the satnav all the time.[1384.82] [1384.82][S02]I used to think, oh, stupid thing.[1386.41] [1386.41][S02]It thinks we're doing this.[1388.11] [1388.11][S02]It's a natural human tendency, I think.[1390.135] [1390.135][S01] What about the other words that we use?[1392.26] [1392.26][S01]I mean, the example that you gave of the satnav saying,[1394.752] [1394.752][S01]oh, it thinks we're in the car park[1396.21] [1396.21][S01]or, oh, it believes that this is--[1399.78] [1399.78][S01]it got this wrong.[1400.96] [1400.96][S01]It misunderstood this.[1402.36] [1402.36][S01]Those are all very human-centric words, aren't they?[1405.16] [1405.16][S02] Yeah, yeah.[1406.327] [1406.327][S02]Absolutely.[1407.47] [1407.47][S02]There are examples of what philosophers often[1409.65] [1409.65][S02]call folk psychology.[1411.04] [1411.04][S02]So we have this folk psychology where we use words like belief.[1414.06] [1414.06][S02]We have concepts like belief, desire, and intention, which[1416.88] [1416.88][S02]we can apply not just to other humans and other animals,[1421.72] [1421.72][S02]but we can apply to objects as well.[1424.63] [1424.63][S02]It's what the philosopher Dan Dennett called[1427.59] [1427.59][S02]taking the intentional stance.[1429.55] [1429.55][S02]So we adopt the intentional stance towards something[1432.45] [1432.45][S02]if we talk about it and think about it[1434.61] [1434.61][S02]as if it acted on the basis of having beliefs and goals[1439.53] [1439.53][S02]and carrying out rational decisions for what it does[1442.47] [1442.47][S02]on the basis of those things.[1444.25] [1444.25][S02]And that's a very useful way of thinking[1447.03] [1447.03][S02]about many, many things, such as even our satnav or a chess[1449.94] [1449.94][S02]computer.[1450.97] [1450.97][S02]So for Dan Dennett, that was one of the examples that he used.[1453.93] [1453.93][S02]A chess computer, that, oh, it wants to get the queen forward[1457.47] [1457.47][S02]because it thinks I'm going to use my rook to defend this rank[1461.91] [1461.91][S02]or something.[1462.7] [1462.7][S02]And that's full of this kind of intentional folk psychological[1466.71] [1466.71][S02]language about beliefs and goals and things.[1469.027] [1469.027][S01] Is that problematic, then,[1470.61] [1470.61][S01]if we start using that idea of beliefs and intentions[1473.46] [1473.46][S01]and desires about the AI?[1475.442] [1475.442][S02] So it's only problematic[1477.15] [1477.15][S02]if we start to use these things in ways that mislead us[1481.53] [1481.53][S02]into thinking that things have capabilities[1485.67] [1485.67][S02]that they don't really have.[1486.97] [1486.97][S02]So I think that's where it becomes problematic.[1489.76] [1489.76][S02]So the Encyclopedia Britannica, the physical volume[1493.98] [1493.98][S02]of the Encyclopedia Britannica doesn't[1495.93] [1495.93][S02]know that Argentina won the World Cup in this[1498.81] [1498.81][S02]because it's too old.[1500.86] [1500.86][S02]So if you made that remark, it would make perfect sense.[1504.905] [1504.905][S02]You might say that and it's fine.[1506.57] [1506.57][S02]But if somebody said to you, why don't you[1508.32] [1508.32][S02]have a conversation with it about England's football prowess[1512.22] [1512.22][S02]or lack thereof, that would be ridiculous.[1515.02] [1515.02][S02]Right?[1515.52] [1515.52][S02]Now, the interesting thing is that now[1517.103] [1517.103][S02]we've got these large language models,[1518.77] [1518.77][S02]you can have a conversation with them.[1520.75] [1520.75][S02]You can tell it things so that it kind of pushes[1524.37] [1524.37][S02]the boundary of where we might start to say, well,[1527.56] [1527.56][S02]it doesn't really XYZ.[1529.33] [1529.33][S02]It pushes that a little bit further out.[1531.1] [1531.1][S01] I wonder if there's something even deeper here[1533.517] [1533.517][S01]about this human need, or maybe it's just a desire,[1537.9] [1537.9][S01]to really want AI to have these characteristics,[1543.06] [1543.06][S01]to be anthropomorphized.[1544.51] [1544.51][S02] Yeah.[1544.86] [1544.86][S02]Yeah.[1545.36] [1545.36][S02]Well, that's a really interesting question, isn't it?[1547.75] [1547.75][S02]So I don't think it kind of comes back to that.[1549.708] [1549.708][S02]It comes back to language.[1551.01] [1551.01][S02]In this case, we're inclined to anthropomorphize things[1554.7] [1554.7][S02]because they're really good at using language.[1556.772] [1556.772][S02]And for us, the only things that are good at using language[1559.23] [1559.23][S02]are other humans.[1560.26] [1560.26][S02]And so it's very strange, in a way, to be suddenly in a world[1563.46] [1563.46][S02]where we have language using things.[1567.375] [1567.375][S02]It's not just humans that can talk.[1569.32] [1569.32][S02]That's astonishing.[1570.4] [1570.4][S01] Yeah.[1571.38] [1571.38][S01]I mean, it is astonishing.[1572.71] [1572.71][S02] It is astonishing.[1573.39] [1573.39][S02]So it really is astonishing to think[1575.25] [1575.25][S02]that every single child born today,[1578.082] [1578.082][S02]they're going to grow up in a world[1579.54] [1579.54][S02]where they've never known--[1580.96] [1580.96][S02]they've never known a world in which[1582.97] [1582.97][S02]machines can't talk to them.[1585.79] [1585.79][S02]Isn't that an extraordinary thing?[1588.262] [1588.262][S01] Yeah.[1588.97] [1588.97][S02] I mean, it really is.[1590.553] [1590.553][S02]And so what the implications are of that for us[1594.1] [1594.1][S02]all is really hard to say.[1595.773] [1595.773][S01] I'm just thinking back[1597.19] [1597.19][S01]to what you were saying about how grounded humans are[1602.23] [1602.23][S01]in the physical world.[1604.13] [1604.13][S02] Yes.[1605.57] [1605.57][S01] It does feel like the kind[1607.27] [1607.27][S01]of embodied aspect of AI has lagged behind this language[1612.37] [1612.37][S01]aspect quite a bit.[1614.12] [1614.12][S01]Do you think that we're going to see[1616.06] [1616.06][S01]a big up-step in intelligence, however you want to define it,[1620.93] [1620.93][S01]or broader capabilities once we get good and effective embodied[1625.45] [1625.45][S01]AI?[1626.268] [1626.268][S02] Well, I think it might make a big difference[1628.81] [1628.81][S02]because the large language models we have at the moment,[1631.46] [1631.46][S02]it's really difficult to discern, actually,[1633.59] [1633.59][S02]to be honest, right now, where the limits are[1636.25] [1636.25][S02]for how good they're going to get,[1638.12] [1638.12][S02]whether we really are on the road to producing[1641.64] [1641.64][S02]general intelligence that is comparable[1643.5] [1643.5][S02]to human general intelligence.[1646.12] [1646.12][S02]And often when you get to the boundaries of the capabilities[1651.21] [1651.21][S02]of these kinds of things, sometimes you[1653.16] [1653.16][S02]get the impression that the AI system doesn't really quite[1656.13] [1656.13][S02]grok something.[1657.99] [1657.99][S02]It doesn't really deeply understand something.[1659.95] [1659.95][S02]You reach some kind of limit and you[1662.22] [1662.22][S02]realize that it's been faking it a little bit.[1665.17] [1665.17][S02]But it may be that that sort of general ability[1668.13] [1668.13][S02]to really get things on a deep level,[1670.87] [1670.87][S02]on a deep, common sense level, maybe, that that does still[1675.09] [1675.09][S02]require a bit of embodiment.[1677.53] [1677.53][S02]It does still basically require training data[1680.31] [1680.31][S02]that involves interacting with a real world of physical objects[1684.99] [1684.99][S02]with their spatial organization.[1687.01] [1687.01][S02]And there's something fundamental about that.[1689.02] [1689.02][S01] OK.[1689.645] [1689.645][S01]If understanding, then, however we define it,[1692.61] [1692.61][S01]is something that can emerge as just a consequence of more[1697.86] [1697.86][S01]and more data, what about consciousness?[1700.79] [1700.79][S01]I mean, I'm sure you've been asked a thousand times about AI[1704.23] [1704.23][S01]consciousness and whether it's something[1706.09] [1706.09][S01]that we can expect to happen or has already happened.[1709.373] [1709.373][S02] Yeah, yeah.[1710.54] [1710.54][S02]The very first thing to point out[1712.48] [1712.48][S02]is that I do think we can dissociate intelligence[1716.41] [1716.41][S02]or cognition and cognitive capabilities,[1719.42] [1719.42][S02]we can dissociate that from consciousness.[1721.67] [1721.67][S02]So I think we can imagine things that are very capable[1726.13] [1726.13][S02]and that we want to say are very intelligent because of the way[1730.78] [1730.78][S02]they can achieve their goals and so on,[1732.56] [1732.56][S02]but that we don't want to ascribe consciousness to.[1735.47] [1735.47][S02]But actually, what does that even mean,[1737.95] [1737.95][S02]to ascribe consciousness to something at all?[1740.59] [1740.59][S02]I think the concept of consciousness[1742.09] [1742.09][S02]itself can be broken down into many parts.[1745.64] [1745.64][S02]It's a multifaceted concept.[1748.37] [1748.37][S02]So, for example, we might talk about awareness of the world.[1751.99] [1751.99][S02]In the scientific study of consciousness,[1754.4] [1754.4][S02]there are all of these experimental protocols[1756.82] [1756.82][S02]and paradigms, and many of them are to do with perception.[1760.58] [1760.58][S02]You're looking at whether a person is aware of something,[1764.16] [1764.16][S02]is consciously perceiving something in the world.[1766.89] [1766.89][S02]Large language models are not aware of the world[1769.31] [1769.31][S02]at all in that respect.[1771.177] [1771.177][S02]But there are other facets of consciousness.[1773.01] [1773.01][S02]We also have self-awareness.[1774.89] [1774.89][S02]Now our self-awareness, part of that[1776.42] [1776.42][S02]is awareness of our own body and where it is in space.[1780.08] [1780.08][S02]But another aspect of self-awareness[1781.58] [1781.58][S02]is a kind of awareness of our own inner machinations[1784.94] [1784.94][S02]of our stream of consciousness, as William James called it.[1788.52] [1788.52][S02]So we have that kind of self-awareness as well.[1791.34] [1791.34][S02]And we have what some people call metacognition as well.[1794.91] [1794.91][S02]We have the ability to think about what we know.[1797.7] [1797.7][S02]And then additionally, there's the emotional side[1800.39] [1800.39][S02]or the feeling side of consciousness or sentience.[1803.61] [1803.61][S02]So the capacity to feel, the capacity to suffer.[1808.64] [1808.64][S02]And that's another aspect of consciousness.[1810.66] [1810.66][S02]Now, I think we can dissociate all of these things.[1812.84] [1812.84][S02]Now in humans, they all come as a big package, a big bundle.[1815.47] [1815.47][S02]We only actually have to think about non-human animals[1817.72] [1817.72][S02]to realize that we can start to separate these things[1821.23] [1821.23][S02]a little bit, because I think that much as I love cats,[1825.11] [1825.11][S02]I think there's a limited self-awareness going on in cats.[1828.918] [1828.918][S01] How dare you.[1829.96] [1829.96][S02] Well, I'm a big cat person, I have to say,[1832.418] [1832.418][S02]so I do say that with some hesitation.[1834.925] [1834.925][S01] There's little metacognition, shall we say.[1837.35] [1837.35][S02] Well, yeah.[1838.16] [1838.16][S02]Certainly they don't have an awareness[1839.743] [1839.743][S02]of their own ongoing stream of verbal consciousness[1842.2] [1842.2][S02]because they don't have it.[1844.15] [1844.15][S02]They're not thinking about what they did yesterday[1846.79] [1846.79][S02]in verbal terms or what they want to do with their lives.[1850.04] [1850.04][S02]So if we think about robots, you may[1851.95] [1851.95][S02]have a very sophisticated robot, even your robot vacuum cleaner.[1855.65] [1855.65][S02]And you may say that, well, it does actually[1857.74] [1857.74][S02]have a kind of awareness of the world.[1859.76] [1859.76][S02]And that's not an inappropriate use[1861.94] [1861.94][S02]of that phrase, awareness of the world.[1863.582] [1863.582][S02]Do I want to call it consciousness?[1865.04] [1865.04][S02]Well, then I seem to be bringing on board all of this other stuff[1867.87] [1867.87][S02]as well.[1868.37] [1868.37][S02]But you don't have to.[1869.66] [1869.66][S02]You can break down the concept of consciousness[1871.81] [1871.81][S02]into these different aspects.[1873.018] [1873.018][S01] Because your robot vacuum[1874.56] [1874.56][S01]can know exactly where it is in a space and how--[1876.83] [1876.83][S02] Yeah, and respond[1878.247] [1878.247][S02]in an intelligent and sensitive way to where it is[1880.517] [1880.517][S02]and the objects around it.[1881.6] [1881.6][S01] Achieve its ends.[1882.14] [1882.14][S02] And achieve its ends and so on.[1884.14] [1884.14][S02]So there's a kind of awareness of the world there.[1886.43] [1886.43][S02]There's no self-awareness.[1887.757] [1887.757][S02]There's certainly no capacity for suffering.[1889.59] [1889.59][S02]And so in a large language model,[1891.03] [1891.03][S02]there might not be awareness of the world[1893.33] [1893.33][S02]in that perceptual sense, but maybe there's[1896.21] [1896.21][S02]some kind of self awareness or reflexive capabilities,[1901.32] [1901.32][S02]reflexive cognitive capabilities.[1903.3] [1903.3][S02]They can talk about the things that they've[1905.96] [1905.96][S02]talked about earlier in the conversation, for example,[1908.21] [1908.21][S02]and can do so in a reflective manner, which kind of feels[1913.25] [1913.25][S02]a little bit like some aspects of self-awareness[1916.16] [1916.16][S02]that we have a little bit.[1918.09] [1918.09][S02]I don't think that it's appropriate to think of them[1920.45] [1920.45][S02]in terms of having feelings.[1923.13] [1923.13][S02]They can't experience pain because they don't have a body.[1925.68] [1925.68][S02]I think we can take the concept apart, basically.[1928.59] [1928.59][S01] So then is the question,[1930.15] [1930.15][S01]can AI be conscious or not, as though it's a binary thing?[1933.82] [1933.82][S01]It's the wrong question from the off?[1935.68] [1935.68][S02] I do think that is the wrong question,[1938.02] [1938.02][S02]and I think it's wrong in many ways.[1939.61] [1939.61][S02]So just then we were talking about the fact[1942.03] [1942.03][S02]that it's actually a sort of multifaceted concept.[1945.1] [1945.1][S02]But also I think that we tend to have[1946.68] [1946.68][S02]these very deep metaphysical commitments[1950.37] [1950.37][S02]to the idea of consciousness as some sort of magical thing[1955.41] [1955.41][S02]that is a metaphysical thing.[1958.87] [1958.87][S02]So the question of whether something[1960.63] [1960.63][S02]is conscious or is not a matter of consensus[1964.53] [1964.53][S02]or a matter of just our language,[1966.07] [1966.07][S02]but it's something that is out there[1968.19] [1968.19][S02]in the metaphysical reality or in the mind of God[1970.77] [1970.77][S02]or in the platonic heaven or something like that.[1972.91] [1972.91][S02]But ultimately, I do think that that's the wrong way of thinking[1975.577] [1975.577][S02]about consciousness.[1976.66] [1976.66][S01] Let's take one aspect of consciousness, then,[1979.035] [1979.035][S01]that you described about an emotional side,[1981.4] [1981.4][S01]an ability to suffer, but not necessarily physical pain.[1984.46] [1984.46][S01]Emotional pain, too, and a sense of self in the emotional way.[1989.17] [1989.17][S01]Do you think this is something that[1991.89] [1991.89][S01]will just emerge as a natural consequence of intelligence.[1995.417] [1995.417][S01]If you build something that is intelligent enough,[1997.5] [1997.5][S01]at some point, this is going to happen.[1999.45] [1999.45][S01]Or is there something unique about biological creatures.[2002.81] [2002.81][S01]And I guess the process of evolution[2004.42] [2004.42][S01]that we've been through that has resulted in that that can't[2006.94] [2006.94][S01]be replicated in a machine.[2008.27] [2008.27][S02] I don't think there is a right or wrong answer[2010.96] [2010.96][S02]to your question there.[2012.52] [2012.52][S02]I think we just have to wait and see[2014.14] [2014.14][S02]what things we bring into the world[2016.6] [2016.6][S02]and how we end up treating them and talking about them[2019.39] [2019.39][S02]and thinking about them.[2020.45] [2020.45][S02]And I don't think we really know until they're[2023.77] [2023.77][S02]among us, as it were, these things that we're building.[2026.08] [2026.08][S02]Then we will just be led to think about them[2029.14] [2029.14][S02]and talk about them and treat them in a particular way.[2033.26] [2033.26][S02]So an example I like to think in this regard is the octopus.[2038.0] [2038.0][S02]So octopuses have recently been brought[2041.47] [2041.47][S02]into, UK legislation, brought into the category of things[2045.4] [2045.4][S02]that we have to care about the welfare of.[2047.84] [2047.84][S02]That's as a result of lots of things, I think, happening.[2051.78] [2051.78][S02]So the public has been exposed to being with octopuses[2056.51] [2056.51][S02]a lot more.[2057.21] [2057.21][S02]Now, you don't have to literally be under the water[2059.63] [2059.63][S02]and poking around with octopuses to know what it's[2062.887] [2062.887][S02]like to be with them, because there's[2064.429] [2064.429][S02]all kinds of wonderful documentaries[2066.05] [2066.05][S02]and wonderful books by--[2067.672] [2067.672][S02]Peter Godfrey-Smith has these great books[2069.38] [2069.38][S02]about interacting with octopuses and so on.[2072.679] [2072.679][S02]So those sort of narratives and documentaries,[2074.875] [2074.875][S02]they give us a feel for what it's[2076.25] [2076.25][S02]like to be with an octopus, what it's like to have[2078.56] [2078.56][S02]an encounter with an octopus.[2080.969] [2080.969][S02]And then you can't help yourself but to see it[2084.05] [2084.05][S02]as a fellow conscious creature.[2086.48] [2086.48][S02]But complementing that is the scientific progress as well.[2089.46] [2089.46][S02]So at the same time, scientists study the nervous systems[2093.199] [2093.199][S02]of octopuses and realize the extent[2097.16] [2097.16][S02]to which their nervous systems are similar to ours in the way[2100.34] [2100.34][S02]that we experience pain.[2103.08] [2103.08][S02]You can find analogous aspects of their nervous systems[2106.88] [2106.88][S02]to ours.[2108.03] [2108.03][S02]So taking all these things together,[2109.53] [2109.53][S02]I think that tends to affect the way we think about them[2112.48] [2112.48][S02]and the way we talk about them and the way we treat them.[2115.46] [2115.46][S02]So I think the same kind of thing[2117.73] [2117.73][S02]is going to happen with AI systems.[2120.29] [2120.29][S02]Do I think there's a right or wrong answer to,[2123.67] [2123.67][S02]could we be misled there?[2126.08] [2126.08][S02]I think that's a really, really deep and difficult[2128.68] [2128.68][S02]metaphysical philosophical question.[2130.945] [2130.945][S01] I do wonder, though--[2132.32] [2132.32][S01]I mean, that point about suffering to me[2134.44] [2134.44][S01]seems different to the others, because metacognition,[2138.71] [2138.71][S01]the sense of the world, et cetera,[2140.71] [2140.71][S01]there's not these ethical implications necessarily[2143.98] [2143.98][S01]about those.[2144.74] [2144.74][S01]But I think with suffering, you wouldn't want your shoes[2147.67] [2147.67][S01]to be conscious.[2148.45] [2148.45][S01]You know?[2149.17] [2149.17][S01]You wouldn't want a forklift truck to be sort of conscious.[2152.091] [2152.091][S02] No, no.[2153.091] [2153.091][S02]Unless they happen to really like being a forklift truck.[2155.54] [2155.54][S01] Sure, sure.[2156.53] [2156.53][S01]But then do we have to be a tiny bit more[2158.238] [2158.238][S01]careful about that particular aspect of it?[2160.34] [2160.34][S02] I think we do.[2161.632] [2161.632][S02]If there were the prospect of bringing into being something[2164.92] [2164.92][S02]that is genuinely capable of suffering,[2166.97] [2166.97][S02]then we should think very hard about whether we should do it[2168.95] [2168.95][S02]or not.[2169.53] [2169.53][S02]I tend to think that that's not the case with anything[2172.34] [2172.34][S02]that we've got at the moment.[2173.85] [2173.85][S02]But some people will push back against that.[2178.32] [2178.32][S02]We take the example of large language models.[2180.27] [2180.27][S02]Well, OK.[2180.96] [2180.96][S02]So there's one level in which what they do[2183.53] [2183.53][S02]is next token prediction, next word prediction.[2186.63] [2186.63][S02]But in order to be able to do that really,[2190.62] [2190.62][S02]really, really well in the way that they can at the moment,[2194.13] [2194.13][S02]then they've had to learn and acquire[2198.11] [2198.11][S02]all kinds of emergent mechanisms.[2200.01] [2200.01][S02]So who knows whether or not there's[2202.07] [2202.07][S02]some kind of emergent mechanism has[2203.99] [2203.99][S02]been learned in the weights of this enormous, staggeringly huge[2207.38] [2207.38][S02]number, hundreds of billions of weights in a language[2209.93] [2209.93][S02]model, whether some mechanism hasn't been learned there that[2215.195] [2215.195][S02]has, for example, genuine understanding in it,[2217.83] [2217.83][S02]whatever that means, or even consciousness.[2220.71] [2220.71][S02]Coming back to embodiment again, I have always been of the view[2225.05] [2225.05][S02]that it's only really legitimate to talk about consciousness[2227.79] [2227.79][S02]in the context of something we can share a world with and have[2233.1] [2233.1][S02]that kind of encounter with that we have[2234.81] [2234.81][S02]with an octopus or a dog or a horse,[2237.24] [2237.24][S02]whatever, and being together in the world with that animal[2240.45] [2240.45][S02]and responding to things together.[2242.56] [2242.56][S02]Then I'm in no doubt that they are conscious.[2245.377] [2245.377][S02]That's the kind of primal case for me.[2246.96] [2246.96][S02]Now with a large language model, you[2248.46] [2248.46][S02]can't be in the same world as them in that kind of way,[2251.44] [2251.44][S02]and you can't hang out with them and interact[2254.31] [2254.31][S02]with physical objects with today's large language models.[2257.61] [2257.61][S02]So to my mind, using the language of consciousness[2260.34] [2260.34][S02]in that context is what Wittgenstein would say,[2263.44] [2263.44][S02]it's taking language on holiday.[2265.15] [2265.15][S02]It's using it so far outside of its normal use,[2268.8] [2268.8][S02]maybe it's inappropriate.[2270.01] [2270.01][S02]But that can change.[2271.05] [2271.05][S02]And the more I interact with large language models,[2273.725] [2273.725][S02]the more I have these sophisticated and interesting[2275.85] [2275.85][S02]conversations with them, the more I'm inclined to think,[2279.58] [2279.58][S02]well, maybe I want to extend the language of consciousness,[2282.64] [2282.64][S02]bend it, change it, distort it, make up some new words,[2286.43] [2286.43][S02]break it apart in ways that are going[2288.37] [2288.37][S02]to fit these new things that I'm interacting with all the time.[2292.76] [2292.76][S01] I know you've spent a lot of time interacting[2295.24] [2295.24][S01]with these large language models.[2296.615] [2296.615][S01]I've actually seen you described as a renowned prompt whisperer.[2299.78] [2299.78][S01]What's your secret?[2300.675] [2300.675][S02] Well, one secret[2302.05] [2302.05][S02]is to talk to the large language model as if it were human.[2305.195] [2305.195][S02]So if you think that what they're doing[2306.82] [2306.82][S02]is role playing a human character, such[2309.61] [2309.61][S02]as, say, a very smart and helpful intern,[2313.223] [2313.223][S02]then you should treat them like a smart and helpful intern[2315.64] [2315.64][S02]and talk to them as if they were a smart and helpful intern.[2318.95] [2318.95][S02]For example, just being polite and saying, is that clear[2322.67] [2322.67][S02]and please and thank you.[2323.78] [2323.78][S02]And in my experience, you get better responses out of things[2327.59] [2327.59][S02]if you do things that way.[2329.237] [2329.237][S01] Do you say please and thank you?[2331.07] [2331.07][S02] You can say please and thank you.[2332.69] [2332.69][S02]Yeah.[2333.19] [2333.19][S02]Now there's a good reason, good scientific reason[2335.74] [2335.74][S02]why that might get--[2337.96] [2337.96][S02]again, it just depends.[2339.79] [2339.79][S02]Models are changing all the time.[2341.38] [2341.38][S02]Why that might get better performance out of it.[2344.71] [2344.71][S02]Because if it's role playing, say it's role[2347.73] [2347.73][S02]playing a very smart intern.[2349.74] [2352.54][S02]Then it's going to just role play, maybe[2354.48] [2354.48][S02]being a bit more stroppy if they're not[2356.58] [2356.58][S02]being treated politely.[2358.83] [2358.83][S02]It's just mimicking what humans would do in that scenario.[2364.3] [2364.3][S02]So the mimicry might extend to being a bit more--[2368.43] [2368.43][S02]not being as responsive if their boss is a bit of a stroppy--[2375.103] [2375.103][S01] So-and-so.[2376.02] [2376.02][S02] Bossy boss.[2377.187] [2377.187][S01] I absolutely love that.[2378.73] [2378.73][S01]I think I want to return to where we started, which[2381.57] [2381.57][S01]is about how we think of that AI and the language[2386.04] [2386.04][S01]we use to describe it, and how we frame it in our minds.[2390.12] [2390.12][S01]Do you think that we need a new way of talking about AI?[2393.673] [2393.673][S02] I do.[2394.59] [2394.59][S01] Both acknowledges its potential[2397.08] [2397.08][S01]without overestimating it, but then similarly,[2400.32] [2400.32][S01]isn't dismissive of the things that it can do.[2402.752] [2402.752][S02] I think that's exactly what we need.[2404.96] [2404.96][S02]In one of my papers, I used the phrase[2406.543] [2406.543][S02]exotic mind-like entities to describe large language models.[2411.47] [2411.47][S02]So I think that they are, to a degree,[2413.92] [2413.92][S02]exotic mind-like entities.[2416.06] [2416.06][S01] Lovely.[2416.972] [2416.972][S02] So they are kind of mind-like,[2418.93] [2418.93][S02]and they're increasingly mind-like.[2420.98] [2420.98][S02]Now there's a very important reason[2422.62] [2422.62][S02]for using the little hyphen like there,[2424.97] [2424.97][S02]which is because I want to hedge my bets as to whether they[2428.47] [2428.47][S02]really qualify as minds.[2430.1] [2430.1][S02]And so I can wriggle out of that problem by just using mind-like.[2433.24] [2433.24][S02]They're exotic because they're not like us, language use,[2436.258] [2436.258][S02]but in other respects, they're disembodied.[2438.05] [2438.05][S02]For a start, there's really weird conceptions[2440.65] [2440.65][S02]of self-hood that are applicable to them, maybe.[2443.21] [2443.21][S02]But so they are quite exotic entities as well.[2446.09] [2446.09][S02]So I think of them as exotic, mind-like entities,[2449.06] [2449.06][S02]and we just don't have the right kind of conceptual framework[2453.85] [2453.85][S02]and vocabulary for talking about these exotic, mind-like entities[2457.13] [2457.13][S02]yet.[2457.63] [2457.63][S02]We're working on it.[2459.26] [2459.26][S02]And the more they are around us, the more[2463.43] [2463.43][S02]we'll develop new kinds of ways of talking and thinking[2466.85] [2466.85][S02]about them.[2467.67] [2467.67][S01] It is interesting, though, that you are still[2470.045] [2470.045][S01]going for the Turing-like approach of a creature,[2473.61] [2473.61][S01]almost, rather than the [INAUDIBLE].[2476.302] [2476.302][S02] Well, entity is a pretty neutral term, isn't it?[2479.01] [2479.01][S02]I suppose you could just say thing.[2480.98] [2480.98][S02]Exotic, mind-like thing, if you prefer.[2483.468] [2483.468][S01] Yeah, let's go with that.[2485.01] [2485.01][S01]I think let's push for that for the new name.[2488.367] [2488.367][S02] OK.[2489.2] [2489.2][S02]OK.[2489.7] [2489.7][S02]But I mean, I can't, Hannah, because I've[2492.29] [2492.29][S02]used the word entity in that context in many publications[2495.08] [2495.08][S02]now, so.[2496.065] [2496.065][S01] Exotic, mind-like entities.[2497.69] [2497.69][S01]I like it.[2498.45] [2498.45][S01]I like it a lot.[2499.44] [2499.44][S01]Murray, thank you so much for joining us.[2501.45] [2501.45][S02] It's been a pleasure, Hannah.[2502.65] [2502.65][S02]Thank you.[2503.37] [2503.37][S01] One of the nice things[2504.787] [2504.787][S01]about having done this podcast for a number of years[2507.02] [2507.02][S01]is that you really get to see how[2509.96] [2509.96][S01]the people at the frontier of AI, how their opinions change[2513.92] [2513.92][S01]and shift over time.[2516.08] [2516.08][S01]The last few years have been a real game changer[2518.99] [2518.99][S01]in all sorts of ways about the extent to which intelligence[2523.54] [2523.54][S01]requires a physical body, about how much[2527.2] [2527.2][S01]we need to expand our definition of consciousness[2530.54] [2530.54][S01]to account for the subtly different ways[2532.99] [2532.99][S01]that these mind-like entities can operate.[2536.45] [2536.45][S01]And the next few years, well, who knows?[2538.91] [2538.91][S01]But if past predictions are any indication,[2542.06] [2542.06][S01]the only thing we know about tomorrow's science[2544.45] [2544.45][S01]and technology is that it will be radically different to what[2547.9] [2547.9][S01]we imagine today.[2549.59] [2549.59][S01]You have been listening to \"Google DeepMind,\"[2551.93] [2551.93][S01]the podcast with me, Professor Hannah Fry.[2554.35] [2554.35][S01]If you enjoyed this episode, then[2556.12] [2556.12][S01]do subscribe to our YouTube channel.[2558.08] [2558.08][S01]You can also find us on your favorite podcast platform.[2561.38] [2561.38][S01]And of course, we have plenty more episodes[2564.16] [2564.16][S01]on a whole range of topics to come, so do check those out.[2568.04] [2568.04][S01]See you next time.[2570.06]"}
{"file_name": "audio/val_000023.wav", "transcription": "[0.0][S02] It sort of seems a bit bonkers to me[2.0] [2.0][S02]that we're in 2025 and this hasn't been done already.[5.97] [5.97][S02]I mean, do we not know where the forests are?[8.01] [8.01][S01] A lot of the information[8.78] [8.78][S01]that we want about the natural world has never been recorded[11.28] [11.28][S01]like that on paper necessarily-- where[13.61] [13.61][S01]all the different species of trees are[15.36] [15.36][S01]or where the habitats are or where[16.777] [16.777][S01]the boundaries between the rough grassland and wetland are.[20.423] [20.423][S01]So we need to actually create that information[22.34] [22.34][S01]for the first time.[23.06] [23.06][S02] Thing is, I can see the potential for this in terms[25.685] [25.685][S02]of scientific interest, but I do wonder whether,[28.49] [28.49][S02]let's say, we get to a point where you can understand what[31.55] [31.55][S02]dolphins are saying, communicate with the higher[33.62] [33.62][S02]animals on the planet.[34.62] [34.62][S02]Does it change how we view ourselves[36.32] [36.32][S02]and our place in the world?[37.59] [37.59][S01] I think yeah.[38.07] [38.07][S01]I think it absolutely has that potential.[39.87] [39.87][MUSIC PLAYING][42.323] [42.323][S02] Welcome back to \"Google DeepMind--[44.24] [44.24][S02]The Podcast.\"[44.91] [44.91][S02]I'm Professor Hannah Fry.[46.19] [46.19][S02]Now, we focus a lot in this podcast on AI[48.98] [48.98][S02]and its interactions with humans.[50.94] [50.94][S02]But there is another story that's unfolding, one in which[54.38] [54.38][S02]AI could help protect our planet.[57.0] [57.0][S02]Think oceans and forests and deserts and fragile ecosystems[61.08] [61.08][S02]and the millions of animal species[63.21] [63.21][S02]that are crawling, swimming, and flying across this Earth.[66.88] [66.88][S02]AI has the potential to tackle one[69.66] [69.66][S02]of the biggest challenges of our time, the damage[72.63] [72.63][S02]to nature and ecosystems.[74.59] [74.59][S02]But with a problem that's this vast, where do you even begin?[78.58] [78.58][S02]Well, Drew Purves, Nature Lead at Google DeepMind,[81.25] [81.25][S02]is my guide through the rich terrain of AI for nature.[85.3] [85.3][S02]He has got two decades of experience[87.15] [87.15][S02]in ecological research and has been at Google DeepMind[89.61] [89.61][S02]for almost 10 years now.[91.36] [91.36][S02]Drew, thank you so much for joining me.[93.07] [93.07][S01] Oh, thanks.[93.52] [93.52][S01]It's great to be here.[94.45] [94.45][S02] I think most people probably[96.117] [96.117][S02]agree by now that the environment is[98.58] [98.58][S02]an important aspect of our future, something that deserves[102.03] [102.03][S02]preserving and looking after.[103.95] [103.95][S02]What's holding us back in this area?[106.48] [106.48][S02]What's making it a difficult problem?[108.37] [108.37][S01] The short answer to that[109.912] [109.912][S01]often is lack of information.[111.19] [111.19][S01]So you're right.[111.94] [111.94][S01]There's this groundswell now of agreement[113.97] [113.97][S01]about the importance of biodiversity, ecosystems,[117.03] [117.03][S01]and nature.[118.32] [118.32][S01]And then we have signs of action in different sectors,[122.02] [122.02][S01]in the private sector, in the government sector.[124.02] [124.02][S01]If you look at some of the numbers,[125.478] [125.478][S01]for example, there are 189 countries[127.05] [127.05][S01]around the world that have signed up to the 30[128.967] [128.967][S01]by 30 plan, which is to protect 30% of ecosystems on land[133.74] [133.74][S01]and in the oceans by the year 2030, things[136.343] [136.343][S01]that there's so much to feel good about.[138.01] [138.01][S01]But often, when it comes down to actually taking action[140.43] [140.43][S01]on the ground to either protect or restore biodiversity[143.76] [143.76][S01]and ecosystems, there's just a lack of basic information.[146.135] [146.135][S02] So what are the big questions that AI could[148.427] [148.427][S02]help answer in this space then?[150.002] [150.002][S01] I mean, there are a number of them, obviously.[152.46] [152.46][S01]But for example, if you're thinking about a protection[155.4] [155.4][S01]scenario, protecting biodiversity,[157.218] [157.218][S01]you need to know where the biodiversity is.[159.01] [159.01][S01]It might be focal species, or it might be biodiversity hotspots,[164.07] [164.07][S01]and so on.[164.77] [164.77][S01]It might be a particular endemic species.[166.72] [166.72][S01]If it's a restoration scenario, you[169.02] [169.02][S01]want to find particular places around the world that[172.83] [172.83][S01]have the highest potential for ecological restoration.[176.17] [176.17][S01]But for example, you might need to know which species of trees[179.48] [179.48][S01]to plant or which species of animals to reintroduce.[182.88] [182.88][S01]So down at this local level-- it's[184.58] [184.58][S01]because biodiversity is so place-specific.[186.93] [186.93][S01]You need that locally relevant information for communities[191.03] [191.03][S01]or whatever local action is happening[192.688] [192.688][S01]to guide the action down there.[193.98] [193.98][S01]And often, it's just missing.[195.188] [195.188][S02] Is that the big goal, then, of Google DeepMind,[198.02] [198.02][S02]using AI in nature to really fill[200.84] [200.84][S02]in the information gaps in the same way[203.09] [203.09][S02]that we have done for the human world?[205.05] [205.05][S01] Well, here at Google DeepMind,[207.09] [207.09][S01]we're growing a portfolio of work around AI for nature.[211.08] [211.08][S01]And there are at least three key categories of AI[214.4] [214.4][S01]for nature to explore.[215.73] [215.73][S01]So the first of those is AI for data.[218.028] [218.028][S01]And that can mean bringing in data from the field[220.07] [220.07][S01]from things like cameras, microphones,[221.75] [221.75][S01]and so on, or identifying and bringing[223.678] [223.678][S01]in data from the literature because there's[225.47] [225.47][S01]a huge amount of data there.[227.258] [227.258][S01]The second category, though, is taking all of that data[229.55] [229.55][S01]and combining it with lots of other sources of data,[231.717] [231.717][S01]like satellite data and so on, to create[233.57] [233.57][S01]that derived information that decision-makers need[236.6] [236.6][S01]to protect or enhance nature.[238.7] [238.7][S01]And then, the third category, which is easy to miss,[240.87] [240.87][S01]is that all the information in the world is fine,[243.0] [243.0][S01]but human beings can get overwhelmed by that amount[245.51] [245.51][S01]of data.[246.06] [246.06][S01]So then there's this key role for active deployment of AI[248.93] [248.93][S01]to help decision-making, to help people[251.42] [251.42][S01]make sense of all that data.[252.65] [252.65][S02] And I guess, in part because this[254.81] [254.81][S02]is quite a new area, right?[256.92] [256.92][S02]I mean, this hasn't been around for as long as some[259.07] [259.07][S02]of the other applications of AI.[260.73] [260.73][S02]It's not as obvious of how to solve biodiversity[264.56] [264.56][S02]as it might be to solve disease, for instance.[267.425] [267.425][S01] I think that's true.[268.8] [268.8][S01]It does feel like a growth area in AI.[270.57] [270.57][S01]It's kind of interesting, actually.[272.028] [272.028][S01]That ecology obviously has a long history.[274.17] [274.17][S01]But interestingly, mathematical ecology and statistical ecology[277.58] [277.58][S01]and even machine learning ecology has a long history, too.[280.52] [280.52][S01]Fisher, for instance, the famous Fisherian statistics,[283.092] [283.092][S01]a lot of that was developed in an ecological context.[285.3] [285.3][S01]I think that's because ecology is inherently quite challenging.[288.0] [288.0][S01]There's no obvious way to record and frame ecology.[291.0] [291.0][S01]You've got a vast variety of species.[292.56] [292.56][S01]You've got processes operating at different scales.[294.69] [294.69][S01]The temporal signals are always really noisy.[297.12] [297.12][S01]And so it's interesting.[298.27] [298.27][S01]It's always been a little bit of a challenging domain that[301.26] [301.26][S01]has led to innovation in statistics and machine learning.[304.24] [304.24][S01]And so we very much expect that, in the work[307.02] [307.02][S01]we're doing now, trying to apply this deep learning,[309.94] [309.94][S01]for instance, the recent AI revolution to ecology, that[313.05] [313.05][S01]also will be a way to innovate.[314.598] [314.598][S02] OK, well, let's dig into some of this stuff then.[317.14] [317.14][S02]So in particular, in that middle layer of taking in data[320.88] [320.88][S02]and deriving something from it-- because I mean, a lot of this[324.93] [324.93][S02]involves mapping, that fundamental stage.[329.23] [329.23][S02]Tell us a little bit about mapping the biosphere.[331.74] [331.74][S01] So many of the decisions in environment[334.65] [334.65][S01]and ecology come down to place.[336.58] [336.58][S01]And that's because, in ecology and biodiversity,[339.062] [339.062][S01]each place is so different-- different species,[341.02] [341.02][S01]different habitats, and often different issues[342.75] [342.75][S01]that you need to deal with, whether it's agriculture[344.917] [344.917][S01]or wildfires, et cetera.[347.02] [347.02][S01]So all of that \"missing information\" I mentioned before,[351.13] [351.13][S01]most of the time, it's missing geo information.[354.0] [354.0][S01]And so, very simply, you can think of that as maps.[356.16] [356.16][S01]And so we need to map habitats.[358.15] [358.15][S01]That's the substrate over which ecology plays because substrate[361.2] [361.2][S01]is largely defined by plants.[362.74] [362.74][S01]So habitats and plants are closely related,[364.96] [364.96][S01]so whether that's forest, grasslands, et cetera.[367.57] [367.57][S01]But of course, then we want to map all the species ideally.[370.35] [370.35][S01]That's a very challenging problem[371.37] [371.37][S01]just because there are millions of them, and a lot of them[373.02] [373.02][S01]are very small and can't be seen, and so on.[374.8] [374.8][S01]Of course, there's maps for now, but also, we[376.29] [376.29][S01]might want to understand historical change.[378.37] [378.37][S01]Those baselines of change are very important.[380.68] [380.68][S01]And of course, being able to project maps[382.92] [382.92][S01]into the future, which is a whole different challenge--[385.44] [385.44][S01]all roads lead to maps in ecology.[387.94] [387.94][S02] OK, that's interesting.[389.44] [389.44][S02]I'm sort of surprised that this doesn't exist already, though.[391.75] [391.75][S02]I mean, can you not just get this from Google Earth?[393.44] [393.44][S01] [CHUCKLES] It's a really interesting question[395.19] [395.19][S01]because we got so used to-- and, of course, it's been amazing,[397.773] [397.773][S01]the geographic information we have[399.48] [399.48][S01]through things like Google Earth and Google Maps,[401.63] [401.63][S01]a real human achievement to bring all that together and make[404.13] [404.13][S01]it available.[404.8] [404.8][S01]But it's very human-centric so far, that data.[406.87] [406.87][S01]So of course, we've got roads, and you've[408.578] [408.578][S01]got every shopping center and every post office, everything[411.5] [411.5][S01]like that.[412.0] [412.0][S01]But actually, it's much less developed[413.79] [413.79][S01]in terms of mapping the natural world.[415.798] [415.798][S02] It's focused more on people than the environment.[418.34] [418.34][S01] Yeah, that's right.[419.06] [419.06][S01]A lot of the information that we want about the natural world[421.602] [421.602][S01]has never been recorded like that on paper,[423.46] [423.46][S01]necessarily, where all the different species of trees are[426.243] [426.243][S01]or where the habitats are or where[427.66] [427.66][S01]the boundaries between the rough grassland and wetland are.[431.623] [431.623][S01]So we need to actually create that information[433.54] [433.54][S01]for the first time.[434.48] [434.48][S02] It sort of seems a bit bonkers to me[436.48] [436.48][S02]that we're in, what, 2025, and this hasn't been done already.[440.72] [440.72][S02]There are attempts at this.[443.6] [443.6][S02]They exist.[444.32] [444.32][S02]Landsat was developed in the 1970s of taking satellite images[449.26] [449.26][S02]and categorizing different regions.[450.77] [450.77][S02]I think there was some stuff before that[452.437] [452.437][S02]for military purposes using satellite images.[454.88] [454.88][S02]But is this just a different scale to those things?[457.86] [457.86][S01] It amazes me too.[459.11] [459.11][S01]One thing that amazes me is that the satellites even exist.[461.87] [461.87][S01]I mean, that's incredible.[463.15] [463.15][S01]It just blows my mind that actually, for many decades,[465.82] [465.82][S01]we've basically had these floating high-res digital[468.34] [468.34][S01]cameras looking down at the Earth for decades.[470.652] [470.652][S01]That's absolutely incredible.[471.86] [471.86][S01]It's also incredible the amount of geographic information we[474.52] [474.52][S01]actually do have.[475.34] [475.34][S01]There are amazing data sets or amazing maps[477.5] [477.5][S01]of all kinds of things that are relevant to the environment.[480.0] [480.0][S01]And yet, at the same time, sometimes you[481.73] [481.73][S01]come in with the most basic questions like,[483.522] [483.522][S01]where are the forests?[484.53] [484.53][S02] Do we not know where the forests are?[486.27] [486.27][S01] That's right.[486.75] [486.75][S01]And honestly, it surprises me too.[488.37] [488.37][S01]But we don't have gold-standard accepted global maps[491.723] [491.723][S01]for most of the habitats that we need to understand,[493.89] [493.89][S01]including forests.[494.64] [494.64][S01]So there is no absolutely universally accepted gold[497.93] [497.93][S01]standard for forest/non-forest.[499.49] [499.49][S01]Now, having said that, there are some pretty good ones[501.74] [501.74][S01]nowadays appearing, but there's still room for improvement.[504.3] [504.3][S01]On the other hand, what we're looking[505.842] [505.842][S01]at is even if you had that to guide decision making,[508.43] [508.43][S01]you often need to distinguish different kinds of forest,[511.28] [511.28][S01]for example, the most basic split-- natural forest[514.039] [514.039][S01]from planted forest.[515.19] [515.19][S01]And there definitely isn't a gold-standard accepted map[517.52] [517.52][S01]for that.[517.95] [517.95][S01]And some of the work we're doing at the moment[519.867] [519.867][S01]is trying to do our best to provide the best yet[522.02] [522.02][S01]available map of that.[523.14] [523.14][S01]We call it the Natural Forests of the World Project.[525.03] [525.03][S02] Why are forest maps useful?[526.655] [526.655][S02]What can you use them for?[527.843] [527.843][S01] Well, forests are incredibly important,[530.01] [530.01][S01]harboring biodiversity and carbon.[531.6] [531.6][S01]And for that reason, there's a lot of government policy[535.11] [535.11][S01]and a lot of international regulation around forests.[537.38] [537.38][S01]And forests are a major focus of concern for conservation groups.[540.69] [540.69][S01]And in all cases, they need the best maps[542.417] [542.417][S01]so they can know where to protect if they're[544.25] [544.25][S01]looking to protect forests, where to potentially restore[546.89] [546.89][S01]forests, where to monitor for problems like forest diseases[550.98] [550.98][S01]and on and on.[551.67] [551.67][S02] Can I see it?[552.48] [552.48][S02]Have you got it?[552.99] [552.99][S01] Yes, I have.[554.032] [554.032][S01]So firstly, what we produced is a map for the whole globe[557.03] [557.03][S01]at 10-meter resolution.[558.63] [558.63][S01]So each 10-meter pixel is classified.[560.69] [560.69][S01]And we give it a probability that it's natural forest.[563.812] [563.812][S02] I should probably do a little audio description[566.27] [566.27][S02]for the people that are listening rather than watching.[568.83] [568.83][S02]We have essentially in a browser what looks superficially like--[573.86] [573.86][S02]well, exactly like Google Earth.[575.76] [575.76][S02]But overlaid on top are these teal colored pixels[579.11] [579.11][S02]that align but not exactly with what[581.9] [581.9][S02]you can see to be trees from the satellite images.[584.177] [584.177][S01] That's right.[585.26] [585.26][S01]So you can illustrate that, for instance, in--[587.177] [587.177][S01]and this is why it's so important but also[589.748] [589.748][S01]difficult to tell the difference between natural forest[592.04] [592.04][S01]and planted forest because the natural forest will typically[594.54] [594.54][S01]be much higher in biodiversity and often carbon[596.55] [596.55][S01]if it's somewhat old growth.[597.64] [597.64][S01]So this will be an area that you would[598.71] [598.71][S01]tend to want to protect more.[600.43] [600.43][S01]And then equally, if you detect forest loss,[602.703] [602.703][S01]then the loss of natural forest will typically[604.62] [604.62][S01]be much more concerning than, for instance,[606.13] [606.13][S01]the loss of planted forest, many of which[607.838] [607.838][S01]are planted in order to cut them down, like timber.[610.33] [610.33][S01]So they can actually be sustainable.[611.83] [611.83][S01]But in this case, in the Southeast US, for example,[613.955] [613.955][S01]there are lots of loblolly pine plantations, for example.[616.548] [616.548][S01]So these are a type of planted forest[618.09] [618.09][S01]which is planted to give timber very quickly, after about[621.0] [621.0][S01]five years.[621.76] [621.76][S01]And so you can see there we've got a mix of areas[624.84] [624.84][S01]that we classify as natural and areas[626.97] [626.97][S01]that we classify as planted forest.[629.14] [629.14][S01]And if you zoom in and out, most of the time-- of course,[632.342] [632.342][S01]there are inaccuracies.[633.3] [633.3][S01]But we're over 90% accurate when we test it.[635.8] [635.8][S01]So you can see here, as we fade up our predictions there,[640.337] [640.337][S01]then if we look at those remaining parts, and you go in,[642.67] [642.67][S01]you'll find these neat rows.[644.0] [644.0][S02] Look at that.[644.52] [644.52][S01] And that's telling us it's planted forest.[645.81] [645.81][S02] That's extraordinary.[646.02] [646.02][S01] So to be able to do this at scale-- yeah.[647.61] [647.61][S02] So, OK, I notice in this example[649.443] [649.443][S02]that it's sort of natural forest/not natural forest.[653.247] [653.247][S02]But I also notice that you've got a confidence threshold[655.58] [655.58][S02]there.[656.1] [656.1][S01] Yes.[656.24] [656.24][S02] So are you classifying[658.1] [658.1][S02]this with a probability rather than just saying,[660.36] [660.36][S02]yes, it's natural forest, not?[661.34] [661.34][S01] Yes, we are.[661.83] [661.83][S01]And it's something we're quite passionate about, really,[664.163] [664.163][S01]because there is a tendency, if you[665.75] [665.75][S01]see a map, that is classified in a black and white way,[668.88] [668.88][S01]to believe that.[670.49] [670.49][S01]And even if it's relatively accurate most of the time,[673.23] [673.23][S01]it's really important to be aware of the inaccuracies.[676.237] [676.237][S01]And so that's why we much prefer to present things[678.32] [678.32][S01]as this uncertainty map.[681.11] [681.11][S01]Now, of course, one thing you'd often do[682.85] [682.85][S01]is to think, OK, I'll choose some threshold then[685.31] [685.31][S01]which I will use, for my purposes,[687.967] [687.967][S01]to distinguish forests and non-forest.[689.55] [689.55][S01]But depending on what you're after,[690.65] [690.65][S01]you may choose different thresholds.[692.01] [692.01][S01]So if you were really looking to make[693.65] [693.65][S01]sure you're protecting all of the remaining[695.63] [695.63][S01]natural forest in an area, you would[697.64] [697.64][S01]set your confidence threshold-- it's actually setting it low so[700.4] [700.4][S01]that you pick all of that up.[701.683] [701.683][S01]But of course, you'll also pick up[703.1] [703.1][S01]a whole load of planted forest, but you[704.725] [704.725][S01]don't mind along the way.[706.02] [706.02][S01]Whereas on the other hand, if you had very limited resources,[708.562] [708.562][S01]let's say, and you wanted to verify[710.09] [710.09][S01]that the loss of natural forest had occurred,[712.35] [712.35][S01]and you can only afford to visit certain places,[714.615] [714.615][S01]you may put a very high threshold on it[716.24] [716.24][S01]to make sure that you're really visiting the places where you're[718.907] [718.907][S01]most confident about and so on.[720.21] [720.21][S01]So we're providing that to downstream users[723.74] [723.74][S01]in this uncertainty form.[725.04] [725.04][S01]We're open sourcing all the maps.[726.51] [726.51][S01]We're open sourcing actually the data[728.06] [728.06][S01]and also the new models we've developed,[730.11] [730.11][S01]open sourcing out to the community.[732.21] [732.21][S01]And what it means for downstream users[734.09] [734.09][S01]is that they can get to their own high-quality,[737.81] [737.81][S01]remote-sense maps with much less compute than before,[742.85] [742.85][S01]lower data requirements, and, all importantly,[745.41] [745.41][S01]lower skills requirements as well.[746.88] [746.88][S01]So if we do it right, something like that[748.588] [748.588][S01]can really help to democratize remote sensing on the outside.[751.55] [751.55][S02] In terms of what's going on behind the scenes,[754.11] [754.11][S02]you're using a vision transformer, right?[756.905] [756.905][S02]How does that work?[758.16] [758.16][S01] So what, overall, we're doing[759.98] [759.98][S01]is bringing in this massive satellite information.[762.9] [762.9][S01]So these are enormous images for a start,[765.15] [765.15][S01]the sheer number of pixels.[766.38] [766.38][S01]And then for each pixel, it's not just RGB[768.38] [768.38][S01]like it would be with an image, but it's often[769.85] [769.85][S01]many different bands--[770.91] [770.91][S01]infrared, et cetera, all these different bands.[773.21] [773.21][S01]And those bands are going up through time.[774.96] [774.96][S01]Then you've got multiple satellites.[776.46] [776.46][S01]Each satellite is giving you records every few days.[778.98] [778.98][S01]You have a lot of missing data from things like clouds.[781.88] [781.88][S01]So you need to take this enormous amount of data[784.28] [784.28][S01]and somehow crunch it all down to pull out[787.0] [787.0][S01]the thing that you actually want to know, like, is it a forest?[789.625] [789.625][CHUCKLING][790.64] [790.64][S01]And the model in the middle that we're using[792.83] [792.83][S01]is this vision transformer model.[794.31] [794.31][S02] Transformers, people usually[796.01] [796.01][S02]associate them with large language models,[797.99] [797.99][S02]like paying attention to different bits of sentences[800.595] [800.595][S02]more important than others.[801.72] [801.72][S02]How does it feed in when it comes to maps?[803.58] [803.58][S01] That's right.[804.663] [804.663][S01]So transformers were mostly developed to work on language,[807.25] [807.25][S01]this idea of these attention heads and so on,[809.24] [809.24][S01]that were then adapted into vision transformers[811.82] [811.82][S01]to do similar things on images, attending[813.89] [813.89][S01]to different parts of images.[815.13] [815.13][S01]What we've done here is then expand that out[817.13] [817.13][S01]to a special vision transformer that's[819.47] [819.47][S01]set up to deal with the challenges of satellite data.[822.54] [822.54][S01]So it's actually now a multimodal, temporal, spatial[825.5] [825.5][S01]vision transformer.[826.527] [826.527][S01]But it's still a form of vision transformer.[828.36] [828.36][S01]It's a great story about how you can get this exchange of methods[831.09] [831.09][S01]between different areas, from language modeling[833.048] [833.048][S01]into vision and, in this case, into remote sensing[835.63] [835.63][S01]and out there into environmental policy.[837.85] [837.85][S02] So when it's working out whether it's forest or not[840.6] [840.6][S02]forest, it's paying attention, as it[843.09] [843.09][S02]were, to sometimes the infrared data, sometimes the image,[848.44] [848.44][S02]sometimes the topology of the area, that kind of thing.[852.847] [852.847][S01] That's right.[853.93] [853.93][S01]The transformer architecture is extremely expressive[857.43] [857.43][S01]architecture that gives the model a huge amount of freedom[860.19] [860.19][S01]to choose what it attends to and how it then[862.8] [862.8][S01]combines that information downstream into its prediction.[865.93] [865.93][S01]And that's particularly valuable in an area like remote sensing[868.87] [868.87][S01]where you've got such a wide variety of modalities coming in[871.86] [871.86][S01]and all the data sets are so huge.[874.3] [874.3][S01]And so it's an area that really benefits[876.3] [876.3][S01]from this kind of extra expressivity and flexibility[879.69] [879.69][S01]of the model.[880.36] [880.36][S02] But that expressivity, in the end,[882.277] [882.277][S02]comes down to forest/not forest.[883.643] [883.643][S01] That's what's funny, right?[885.31] [885.31][S01]All of that gets crunched down into just,[887.41] [887.41][S01]in this case, that single map.[888.97] [888.97][S02] Well, OK, I'm just picking up on something[891.22] [891.22][S02]that you said in your answer there--[892.992] [892.992][S02]because you were talking about how you have these satellite[895.45] [895.45][S02]images over time.[896.84] [896.84][S02]So then rather than just mapping where the forests are or aren't,[900.65] [900.65][S02]does that mean that you can look at how the forests are changing?[903.71] [903.71][S01] Yeah, that's a very important point.[905.15] [905.15][S01]Yeah, indeed.[905.75] [905.75][S01]So the interesting thing with remote sensing and satellite[909.04] [909.04][S01]data is, because the satellites have been up for years,[911.39] [911.39][S01]if you can do this mapping, say you[913.6] [913.6][S01]take one year's worth of satellite data,[915.77] [915.77][S01]say for this year, and we can see[917.26] [917.26][S01]where the forests are this year, we can automatically[919.468] [919.468][S01]redo that from the past to estimate the picture of change.[922.82] [922.82][S01]And this is really important.[924.31] [924.31][S01]So there's a project called Global Forest[927.16] [927.16][S01]Watch, which Google has supported before,[928.87] [928.87][S01]that does this.[929.54] [929.54][S01]And so each year, the latest map comes out,[931.6] [931.6][S01]and you can look at the changes in forest cover through time.[934.58] [934.58][S01]More recently, we've created this stack of maps[938.02] [938.02][S01]that we call deforestation drivers.[940.36] [940.36][S01]And what this is actually looking at[941.86] [941.86][S01]is it's taking each unit of forest loss[944.56] [944.56][S01]and categorizing the cause of that loss[946.72] [946.72][S01]in terms of logging, agricultural expansion,[949.58] [949.58][S01]and so on and so on.[950.64] [950.64][S01]But we've applied that for the last 20 years.[952.74] [952.74][S01]So we have 20 years of not just deforestation but 20 years[956.12] [956.12][S01]of the causes of deforestation.[958.02] [958.02][S01]So in collaboration with our colleagues at the World[960.32] [960.32][S01]Resources Institute, at WRI, we've[962.51] [962.51][S01]worked up this global map of the causes of deforestation[966.02] [966.02][S01]for each year from the year 2000 to today.[968.807] [968.807][S02] And what do you find when you do that?[970.89] [970.89][S01] It's really interesting[972.39] [972.39][S01]to look at these patterns.[973.79] [973.79][S02] Go on, show me, show me.[975.29] [975.29][S02]I know you've got it.[975.87] [975.87][S02]I know you've got it.[976.53] [976.53][S02]I want to see it.[977.1] [977.1][S01] So the first thing you[977.75] [977.75][S01]can see when you look across the world[978.86] [978.86][S01]is that there's forest loss everywhere.[980.37] [980.37][S02] Yeah.[980.93] [980.93][S02]Quite a lot in Northern Europe.[982.32] [982.32][S01] Yes, for example.[983.22] [983.22][S01]So it's so easy to think that all of these problems[985.46] [985.46][S01]are just occurring in the tropics or in the Global South.[987.998] [987.998][S02] Brazil.[988.79] [988.79][S01] Right, exactly.[989.28] [989.28][S01]And the second thing, though, is the causes[991.28] [991.28][S01]of the loss does change between different regions.[993.65] [993.65][S01]So if you have a look at-- zoom in on Brazil here in South[997.25] [997.25][S01]America, you can see--[998.56] [998.56][S01]again, you can go right in and see these--[1000.31] [1000.31][S02] But lots of permanent agriculture.[1001.88] [1001.88][S01] That's right.[1002.42] [1002.42][S01]Yeah, so this sort of darker yellow color.[1004.252] [1004.252][S01]And you'll find when you fade in and out,[1005.96] [1005.96][S01]you'll often find that this is associated with clearings[1008.5] [1008.5][S01]and so on.[1009.08] [1009.08][S01]So you can see how that's working.[1010.497] [1010.497][S02] You can literally see[1011.872] [1011.872][S02]how it picks out the edges of fields effectively.[1014.28] [1014.28][S01] That's right.[1015.363] [1015.363][S01]Because all that fine-grain data is actually in the satellite[1017.97] [1017.97][S01]signals.[1018.47] [1018.47][S01]There have been previous maps, but this one[1020.262] [1020.262][S01]is 10 times finer resolution.[1021.703] [1021.703][S01]And that's what really enables you to pull out[1023.62] [1023.62][S01]these local patterns.[1025.21] [1025.21][S01]And that's where a lot of the decision-making happens,[1027.46] [1027.46][S01]is at local scales.[1028.295] [1028.295][S01]And so understanding things at local scale is really important.[1030.92] [1030.92][S02] Putting all of this together, long into the future,[1036.349] [1036.349][S02]is there a real-time aspect to this?[1038.69] [1038.69][S02]Could you prevent illegal logging[1040.359] [1040.359][S02]from happening before it did?[1042.143] [1042.143][S01] It's a really good question[1043.81] [1043.81][S01]because the work that we've been talking about up to now[1046.96] [1046.96][S01]tends to be taking satellite data from an extended period[1051.67] [1051.67][S01]to estimate how things were either right now[1054.618] [1054.618][S01]or looking into the past.[1055.66] [1055.66][S02] Retrospectively.[1056.63] [1056.63][S01] Right.[1057.422] [1057.422][S01]Whereas this idea of rapid, real-time alerts[1059.86] [1059.86][S01]and then also longer-term future forecasting[1062.98] [1062.98][S01]is some of the real growth areas for us.[1064.81] [1064.81][S01]There are already actually existing organizations[1068.96] [1068.96][S01]that are using satellites to try to pick up[1071.36] [1071.36][S01]deforestation alerts live.[1072.96] [1072.96][S01]So that actually exists.[1074.49] [1074.49][S01]One of the challenges, though, is[1075.89] [1075.89][S01]that in order to pick up most of the real deforestation,[1079.178] [1079.178][S01]they've had to use a methodology that[1080.72] [1080.72][S01]picks up a lot of false positives at the same time.[1082.86] [1082.86][S02] I see.[1083.19] [1083.19][S01] So it's actually really challenging.[1084.27] [1084.27][S01]So there's the so-called deforestation alerts.[1085.83] [1085.83][S01]It's an incredible achievement.[1087.3] [1087.3][S01]But at the same time, it feels like that could be improved.[1089.798] [1089.798][S01]So obviously, if we could get it down[1091.34] [1091.34][S01]to the point where the alerts were much more accurate,[1094.74] [1094.74][S01]then it would be much more easy for people to act on them.[1097.167] [1097.167][S02] I guess this goes beyond just satellite data then.[1099.75] [1099.75][S02]You have to start working in other types of data,[1101.97] [1101.97][S02]wouldn't you?[1102.89] [1102.89][S01] Yes.[1103.85] [1103.85][S01]Well, I think so.[1104.61] [1104.61][S01]I mean, the interesting thing is that, you[1106.43] [1106.43][S01]can produce a model of change.[1108.53] [1108.53][S01]And you can train that on past data,[1110.61] [1110.61][S01]and you can do that very scientifically,[1112.29] [1112.29][S01]and you do all of your held out data and your validation and all[1115.032] [1115.032][S01]that clever machine learning stuff.[1116.49] [1116.49][S01]The problem is, on the other hand,[1117.907] [1117.907][S01]that the whole thing around climate change[1120.543] [1120.543][S01]and environmental change is the future is different to the past.[1123.21] [1123.21][S01]So how can we be confident that we can reapply[1125.84] [1125.84][S01]that model into the future?[1127.4] [1127.4][S02] This isn't like predicting the weather a week[1130.2] [1130.2][S02]in advance.[1131.43] [1131.43][S02]The whole underlying model is different.[1133.267] [1133.267][S01] That's right.[1134.35] [1134.35][S01]And when you look at, say, this process of deforestation[1137.16] [1137.16][S01]that we've been talking about, that is a human-driven process.[1139.93] [1139.93][S01]And humans can change behavior very, very quickly in responses[1143.28] [1143.28][S01]to things like changes in the market.[1145.39] [1145.39][S01]Perhaps the price of soybeans goes up,[1148.41] [1148.41][S01]and then that encourages more rapid deforestation.[1150.642] [1150.642][S01]On the other hand, government regulation and policies[1152.85] [1152.85][S01]or local regulation and policies and so on-- so[1156.36] [1156.36][S01]traditional machine learning approaches[1158.012] [1158.012][S01]are going to really struggle to deal with that because they[1160.47] [1160.47][S01]can only deal with numbers and they can only[1162.303] [1162.303][S01]deal with the past.[1163.21] [1163.21][S01]And I think that's one of the reasons why we're[1164.31] [1164.31][S01]really excited about hybridizing,[1166.03] [1166.03][S01]in a sense, the best of the scientific tradition[1168.03] [1168.03][S01]of simulation modeling, with the machine[1169.92] [1169.92][S01]learning, data-driven approach to forecasting with actually[1173.565] [1173.565][S01]things like the large language models, like Gemini,[1175.69] [1175.69][S01]which potentially could pick up on changes in government policy,[1179.59] [1179.59][S01]news, media, perhaps integrate together things[1182.13] [1182.13][S01]like agricultural prices and so on to help[1185.25] [1185.25][S01]us at least to slightly adjust and identify[1188.43] [1188.43][S01]some of the uncertainty scenarios[1190.32] [1190.32][S01]around those future predictions.[1192.255] [1192.255][S02] If that's vegetation,[1193.63] [1193.63][S02]though, if that's trees and hedgerows and shrubs,[1196.15] [1196.15][S02]what about--[1196.71] [1196.71][S02]I mean, we also care quite a lot about the creatures[1198.54] [1198.54][S02]that are living inside them.[1199.81] [1199.81][S02]How do you bring that in?[1201.22] [1201.22][S01] So habitats is one thing.[1202.81] [1202.81][S01]But then most species are not visible from space,[1205.71] [1205.71][S01]and they don't form a habitat in that sense.[1207.88] [1207.88][S01]Instead, it's all the insects, the fungi, the birds, et cetera.[1211.12] [1211.12][S01]And so much of our concern around biodiversity and so much[1214.32] [1214.32][S01]of the action around conservation[1216.03] [1216.03][S01]is actually around those species.[1217.69] [1217.69][S01]And so you look surprised earlier[1219.42] [1219.42][S01]when you said there isn't an agreed map of forest,[1222.333] [1222.333][S01]and that's true.[1223.0] [1223.0][S01]But that's also true for most of the species in the world.[1225.94] [1225.94][S01]Now, again, it's really important to realize everything[1228.54] [1228.54][S01]we're doing in ecology and biodiversity builds on[1231.72] [1231.72][S01]this huge foundation of organizations and hard work[1234.447] [1234.447][S01]from the past.[1235.03] [1235.03][S01]So there's an amazing organization called the IUCN[1237.24] [1237.24][S01]that does provide maps for 140,000 species worldwide.[1240.85] [1240.85][S01]It's amazing.[1242.16] [1242.16][S01]On the other hand, those maps are very coarse.[1245.1] [1245.1][S01]They're driven by expert opinion.[1246.77] [1246.77][S01]Those experts, they really know what they're talking about.[1249.228] [1249.228][S01]But necessarily, they're quite coarse maps, so 50 kilometer[1252.7] [1252.7][S01]or so cells.[1254.28] [1254.28][S01]And they're only refreshed about every 10 years.[1256.28] [1256.28][S02] Oh, wow.[1256.88] [1256.88][S01] But there's an obvious potential role there[1259.213] [1259.213][S01]for producing, in theory, much better maps[1262.18] [1262.18][S01]using the latest in AI and remote sensing, et cetera.[1265.18] [1265.18][S01]So we're exploring that as well.[1266.78] [1266.78][S02] OK, but then how do you do it, though?[1268.97] [1268.97][S02]Because you haven't got the equivalent of satellites[1272.92] [1272.92][S02]in space photographing insects.[1275.147] [1275.147][S01] That's right.[1276.23] [1276.23][S01]You could view it as a sort of prediction problem.[1277.79] [1277.79][S01]What you're saying is the signals[1279.165] [1279.165][S01]I do have, like the satellite signals and other input data,[1283.73] [1283.73][S01]can help me produce a sort of probabilistic estimate of how[1287.08] [1287.08][S01]likely something would be then.[1288.735] [1288.735][S01]And there's probably always quite[1290.11] [1290.11][S01]a large residual uncertainty on that.[1292.78] [1292.78][S01]So how do you go about doing that?[1294.23] [1294.23][S01]Interestingly, at the absolutely coarsest level,[1296.9] [1296.9][S01]it's the same old, in a sense, machine learning approach.[1300.607] [1300.607][S01]In other words, you take your training data.[1302.44] [1302.44][S01]And you take your input data, like the satellites.[1304.58] [1304.58][S01]And you bring in a special model, and you fit the model,[1306.44] [1306.44][S01]and you get the output, and you evaluate it.[1308.273] [1308.273][S01]It's the classic stuff.[1309.24] [1309.24][S01]But in this case, it's really challenging[1311.18] [1311.18][S01]because the data we're bringing in mostly[1313.01] [1313.01][S01]is citizen science data.[1314.6] [1314.6][S01]So this is people out in the field noticing--[1316.475] [1316.475][S02] Who've spotted something, yeah.[1317.4] [1317.4][S01] That's right.[1317.88] [1317.88][S01]It's an incredible platform called iNaturalist.[1320.04] [1320.04][S01]It's on your phone now.[1321.33] [1321.33][S01]And if you see something, you can take a snapshot of it,[1324.33] [1324.33][S01]and then you can upload that to iNaturalist.[1326.562] [1326.562][S01]And iNaturalist actually have their own machine[1328.52] [1328.52][S01]learning models so they can classify that image[1330.478] [1330.478][S01]to species, which is fantastic.[1332.43] [1332.43][S01]And this is now the main source of[1334.91] [1334.91][S01]these on-the-ground observations of species[1337.468] [1337.468][S01]globally, is coming from this amazing phone[1339.26] [1339.26][S01]app called iNaturalist.[1340.29] [1340.29][S01]So that's fantastic.[1341.34] [1341.34][S01]But actually, that data is still very biased and patchy.[1343.903] [1343.903][S02] Why?[1344.57] [1344.57][S01] So, for example, people typically[1346.94] [1346.94][S01]are close to their home when they take a picture.[1349.73] [1349.73][S01]And so where people go, when they go,[1352.01] [1352.01][S01]they'll typically go on a nice day.[1353.85] [1353.85][S01]And also what they choose to look at and take pictures[1356.12] [1356.12][S01]of-- so what you'll find is you'll[1357.62] [1357.62][S01]get lots of pictures of brightly colored birds close[1359.96] [1359.96][S01]to cities on sunny days.[1361.918] [1361.918][S02] Right, I see.[1362.96] [1362.96][S01] And what you don't get[1364.418] [1364.418][S01]is many pictures of small brown mushrooms in remote places[1368.24] [1368.24][S01]when it's raining.[1370.04] [1370.04][S01]And although I'm joking about that,[1371.64] [1371.64][S01]there's actually a much more serious element to this,[1373.52] [1373.52][S01]too, which is when you look globally[1375.08] [1375.08][S01]at the distribution of citizen science data-- there are[1376.88] [1376.88][S01]major areas of the world with very little-- sub-Saharan[1378.83] [1378.83][S01]Africa, for example.[1379.98] [1379.98][S01]Because actually, iNaturalist depends[1381.68] [1381.68][S01]on having a certain amount of leisure time,[1383.472] [1383.472][S01]really, to do these sorts of activities.[1385.14] [1385.14][S02] Yeah, and also speed of internet connection[1387.8] [1387.8][S02]and quality of your smartphone camera.[1390.0] [1390.0][S01] That's right.[1390.33] [1390.33][S01]And this is a real irony, then, for two reasons, in a way.[1392.747] [1392.747][S01]One is from the fundamental ecology.[1394.743] [1394.743][S01]If you weren't careful, if you naive about it,[1396.66] [1396.66][S01]you would conclude that all species like[1398.69] [1398.69][S01]to live near to middle-class neighborhoods[1400.85] [1400.85][S01]on the edge of cities, right?[1402.5] [1402.5][S01]Because that's where most of the data is.[1404.63] [1404.63][S01]And then, globally, there's a real irony there,[1407.04] [1407.04][S01]which is we're lacking data in many of the places that[1410.09] [1410.09][S01]have the most biodiversity, in many of the places where[1413.372] [1413.372][S01]the current threats to biodiversity are the greatest,[1415.58] [1415.58][S01]and where people's livelihoods depend most on biodiversity,[1418.14] [1418.14][S01]which is just kind of awful.[1419.43] [1419.43][S01]And so the magic, in a sense, that we're[1421.46] [1421.46][S01]looking for from deep learning in this space[1423.96] [1423.96][S01]is to see if we can somehow generalize[1425.91] [1425.91][S01]on this very patchy, biased data that we[1427.83] [1427.83][S01]do have to ecologically plausible distributions for all[1432.66] [1432.66][S01]of the species worldwide.[1433.87] [1433.87][S01]And that's a big challenge.[1435.01] [1435.01][S01]And at the moment, I think you would say the jury is out as[1437.468] [1437.468][S01]to whether it's possible.[1438.55] [1438.55][S01]But we are seeing some evidence of it.[1440.08] [1440.08][S02] But does the things that you've learned and that you[1442.747] [1442.747][S02]now \"know\" in inverted commas about the vegetation then feed[1445.47] [1445.47][S02]into where you expect the species to be?[1448.2] [1448.2][S02]Does one help the other?[1449.2] [1449.2][S01] Firstly, that's a really good point,[1451.242] [1451.242][S01]that we know enough about ecology to know[1453.6] [1453.6][S01]that, actually, things like climate and elevation[1455.97] [1455.97][S01]and habitat are very strong predictors of species.[1458.785] [1458.785][S01]So you're absolutely right.[1459.91] [1459.91][S01]We've already talked about the fact[1461.368] [1461.368][S01]that we can classify habitats from space down to a fine grain.[1464.52] [1464.52][S01]So that itself we would expect to be highly predictive[1466.77] [1466.77][S01]of which species are there.[1468.51] [1468.51][S01]And so then bringing in all this observation data,[1471.163] [1471.163][S01]then we can look at the overlap between these in the areas[1473.58] [1473.58][S01]where we do have data and hopefully pick up[1476.01] [1476.01][S01]a general enough understanding to then be[1478.08] [1478.08][S01]able to reapply that globally.[1480.497] [1480.497][S01]Because there is another challenge here, which is,[1482.58] [1482.58][S01]you have these long-tail distributions in ecology[1485.228] [1485.228][S01]all over the place.[1486.02] [1486.02][S01]You have a few common species and then lots of very rare ones.[1489.16] [1489.16][S01]So we need to do the same thing across the taxonomic tree[1492.1] [1492.1][S01]of life.[1493.112] [1493.112][S01]Can we learn the key relationships for the more[1495.07] [1495.07][S01]common species that are data-rich,[1497.15] [1497.15][S01]enabling us to make the best predictions we can for species[1500.24] [1500.24][S01]actually fairly poor, data-poor?[1502.25] [1502.25][S01]So if we can learn about--[1504.85] [1504.85][S01]so, for example, if we bring in species trait data,[1507.29] [1507.29][S01]we might find out that, in this part of the tree of life,[1510.23] [1510.23][S01]let's say, it's the larger-bodied species that[1513.76] [1513.76][S01]tend to live in the colder environments.[1516.02] [1516.02][S01]And we might be able to discover laws like that in places[1519.762] [1519.762][S01]where we do have lots of data like Europe[1521.47] [1521.47][S01]and the US in a way one day where[1523.488] [1523.488][S01]we can apply that with a certain level of confidence in areas[1526.03] [1526.03][S01]where we have much less data, like the Global South.[1529.702] [1529.702][S01]Not that we'd ever have a blind guess.[1532.34] [1532.34][S01]But what it would really do is lower the data requirements[1534.91] [1534.91][S01]in those areas.[1535.54] [1535.54][S01]Once we've learned some of the more general rules,[1537.59] [1537.59][S01]it would mean that actually a relatively small amount[1539.798] [1539.798][S01]of observational data may be enough to anchor[1541.9] [1541.9][S01]the distribution of that species.[1543.45] [1543.45][S01]Such an exciting space.[1544.62] [1544.62][S02] I do wonder about other types of data here.[1546.96] [1546.96][S02]You're talking about data about the landscape, about elevation,[1550.08] [1550.08][S02]about the vegetation.[1551.94] [1551.94][S02]But are there other elements of data like photographs,[1555.63] [1555.63][S02]for instance, that you can also bring in here?[1557.703] [1557.703][S01] Yes, there are.[1558.87] [1558.87][S01]And in general, you could think of that[1560.495] [1560.495][S01]as this trend towards multimodal models and multimodal AI.[1564.21] [1564.21][S01]And that particularly lights up, I[1565.9] [1565.9][S01]think, in the world of ecology and biodiversity science[1569.66] [1569.66][S01]because you've got such a wide variety of organisms[1572.0] [1572.0][S01]at different scales, and so different modalities suit[1575.78] [1575.78][S01]monitoring different kinds of species.[1578.39] [1578.39][S01]So talking there, you've got images,[1581.137] [1581.137][S01]let's say, humans take deliberately with a camera.[1583.22] [1583.22][S01]They're the ones that, say, iNaturalist specializes in.[1586.107] [1586.107][S01]Often, though, say in the Serengeti Projects and others,[1588.44] [1588.44][S01]we're talking about so-called camera traps.[1590.37] [1590.37][S01]These are cameras that are motion-activated cameras.[1592.802] [1592.802][S01]And a lot of the time, they've got infrared lights on them.[1595.26] [1595.26][S01]So they're actually taking infrared images at night.[1597.51] [1597.51][S01]So there are different kinds of images you can take.[1600.75] [1600.75][S01]You've then got aerial imagery, let's say,[1602.6] [1602.6][S01]from drones or from planes.[1604.245] [1604.245][S01]You've got the remote-sensing satellite[1605.87] [1605.87][S01]imagery we're talking about.[1607.17] [1607.17][S01]And then also, you've got bioacoustics,[1609.21] [1609.21][S01]so-called bioacoustics, is where you're deploying microphones[1611.87] [1611.87][S01]in the field.[1613.31] [1613.31][S01]That might be familiar to a lot of people watching this,[1615.745] [1615.745][S01]for example, around the apps that[1617.12] [1617.12][S01]identify birdsong, Merlin and eBird and similar.[1620.113] [1620.113][S02] Because those already exist.[1621.78] [1621.78][S02]I've played with them a little bit.[1622.88] [1622.88][S02]You're outside.[1623.7] [1623.7][S02]You turn it on.[1624.42] [1624.42][S02]It tells you what birds you're listening to.[1626.253] [1626.253][S01] That's right.[1627.337] [1627.337][S02] But that's supervised learning, is it,[1629.6] [1629.6][S02]that manages to do that?[1631.037] [1631.037][S01] Yeah, that's right.[1632.37] [1632.37][S01]So in all of these areas, you need[1635.94] [1635.94][S01]what we call a corpus of labeled data to start with.[1638.848] [1638.848][S01]So you need a picture with a label along with it.[1640.89] [1640.89][S01]This is a monarch butterfly, or this is a--[1644.47] [1644.47][S02] A crow, crow squawking.[1645.945] [1645.945][S01] Yeah, exactly.[1647.07] [1647.07][S01]All of this sort of stuff.[1647.97] [1647.97][S01]And with enough of that, you can then[1648.92] [1648.92][S01]train a model that generalizes.[1650.217] [1650.217][S01]So that's what's called supervised learning.[1652.05] [1652.05][S01]That's right.[1652.55] [1652.55][S01]And that's the power of most of these developments in each[1654.29] [1654.29][S01]of the individual modalities, whether that's[1656.42] [1656.42][S01]images, remote sensing, et cetera.[1658.8] [1658.8][S01]Yeah.[1659.3] [1659.3][S02] You've got a project, though, called Perch,[1661.8] [1661.8][S02]right?[1662.3] [1662.3][S02]Tell me about Perch.[1663.77] [1663.77][S01] So Perch is in this space of bioacoustics.[1666.23] [1666.23][S01]And the idea of deploying microphones in the field[1668.02] [1668.02][S01]is really attractive because there's almost always[1670.103] [1670.103][S01]something making a sound.[1671.152] [1671.152][S02] Yeah.[1671.86] [1671.86][S01] And it's incredibly rich,[1673.1] [1673.1][S01]the information we can extract from sound, actually.[1675.41] [1675.41][S01]And you can leave the microphones out for days on end[1678.7] [1678.7][S01]to record all that data.[1680.282] [1680.282][S01]And of course, sound is not necessarily just species,[1682.49] [1682.49][S01]but it can be behavior.[1684.02] [1684.02][S01]We might be able to distinguish juveniles from adults.[1686.27] [1686.27][S01]You might be able to pick up warning signals.[1688.33] [1688.33][S01]Insects make sounds.[1689.51] [1689.51][S01]Reptiles make sounds.[1690.77] [1690.77][S01]Birds make sounds.[1691.79] [1691.79][S01]Mammals make sounds.[1692.78] [1692.78][S01]And you pick all of that up.[1693.947] [1693.947][S01]So you don't have to have line of sight to the organism either.[1697.73] [1697.73][S01]It can work at night.[1698.75] [1698.75][S01]So it's an incredibly promising technology for pulling out data[1702.82] [1702.82][S01]from the field.[1703.71] [1703.71][S02] But if it's already been done, if you do already[1706.21] [1706.21][S02]have these apps that can tell if you're listening to a crow,[1708.71] [1708.71][S02]then what else is there to do?[1710.81] [1710.81][S01] Perch is taking a slightly different approach.[1713.268] [1713.268][S01]It's a foundational model for natural sound.[1715.78] [1715.78][S01]And what that does is, like other foundational models,[1718.25] [1718.25][S01]it's not out of the box, it's not just[1719.92] [1719.92][S01]providing you with the ability to identify[1722.738] [1722.738][S01]lots of different birds, for example.[1724.28] [1724.28][S01]It does do that.[1725.45] [1725.45][S01]But what it's really designed to do[1727.75] [1727.75][S01]is to allow people to rapidly create their own detectors[1730.48] [1730.48][S01]for things that haven't been in the detector already.[1733.173] [1733.173][S01]So for instance, if you're working[1734.59] [1734.59][S01]on a particularly rare species, it won't be in one of these apps[1737.35] [1737.35][S01]already.[1737.69] [1737.69][S01]It's too rare.[1738.273] [1738.273][S01]They haven't put it in there.[1739.55] [1739.55][S01]But you might also be working on a more familiar species,[1742.07] [1742.07][S01]but you'd like to divide the juveniles[1743.653] [1743.653][S01]from the adults or a local accent, for instance,[1746.24] [1746.24][S01]of a birdsong.[1747.745] [1747.745][S02] Because birds do have accents, do they?[1749.87] [1749.87][S01] Oh, [LAUGHS] well, yeah, yeah.[1751.45] [1751.45][S01]I think that's right.[1752.03] [1752.03][S01]They certainly have local variants on their songs.[1754.26] [1754.26][S01]So you'd be able to track these sorts of things.[1756.26] [1756.26][S02] So how does bioacoustic modeling actually[1757.78] [1757.78][S02]work then?[1758.692] [1758.692][S01] So the way Perch works[1760.15] [1760.15][S01]is it brings in these audio data files,[1762.91] [1762.91][S01]and they're huge, so lots and lots of data.[1765.23] [1765.23][S01]And it crunches all that down into a much smaller amount[1767.89] [1767.89][S01]of data, which is this so-called embedding.[1770.93] [1770.93][S01]And then that embedding, then, can[1772.78] [1772.78][S01]have all these downstream uses, the most obvious of which[1775.63] [1775.63][S01]is developing new detectors.[1777.243] [1777.243][S01]So you can think of it as a big kind of process that[1779.41] [1779.41][S01]takes huge amounts of data in, and it crunches it down[1781.66] [1781.66][S01]to a small amount of data that retains[1784.09] [1784.09][S01]all of the salient signal that you'd[1786.04] [1786.04][S01]want for understanding ecology and identifying species[1790.382] [1790.382][S01]and has thrown away as much of the noise as possible.[1792.59] [1792.59][S01]However, there's also, for example,[1794.12] [1794.12][S01]at least one other important use case, which is called Search.[1796.703] [1796.703][S01]So if you take different audio clips and crunch them down here,[1799.64] [1799.64][S01]you can also then easily go and find similar audio clips.[1802.015] [1802.015][S01]Because often, one of the main barriers[1803.64] [1803.64][S01]is, let's say, you're looking to monitor some rare species,[1806.2] [1806.2][S01]and you put the microphone out for hours.[1808.64] [1808.64][S01]Well, where's the chirp?[1810.56] [1810.56][S01]You've got to listen to the thing.[1813.02] [1813.02][S01]But if you can just find one or two here, then you can say,[1816.2] [1816.2][S01]right now, I can find the similar ones[1818.5] [1818.5][S01]because I'm looking for the ones that have got a similar Perch[1821.11] [1821.11][S01]embedding to these.[1822.0] [1822.0][S01]And even a third one, which I don't mind speculating around,[1824.5] [1824.5][S01]an idea I've been talking with a lot of colleagues about[1826.833] [1826.833][S01]is, if you're looking to do something verify[1829.57] [1829.57][S01]ecological restoration in the field, you can imagine,[1832.33] [1832.33][S01]if I put microphones in a good place, like a nature reserve[1835.083] [1835.083][S01]and in some other place like my garden--[1836.75] [1836.75][S01]I've been rewilding my garden--[1838.73] [1838.73][S01]then through time, you can see whether the Perch embeddings,[1842.0] [1842.0][S01]say, in my garden become more similar to the local nature[1844.4] [1844.4][S01]reserve.[1844.98] [1844.98][S02] OK, so wait, wait, let me understand that then.[1846.69] [1846.69][S02]So let's say you're doing a project where[1848.398] [1848.398][S02]you want to increase the occupancy of a particular type[1851.66] [1851.66][S02]of species.[1852.66] [1852.66][S02]You can put a microphone in there[1854.3] [1854.3][S02]and work out whether your intervention is[1856.28] [1856.28][S02]increasing the population of whatever it might be.[1858.363] [1858.363][S01] That's absolutely right.[1859.905] [1859.905][S01]So there's at least two scenarios to think about.[1862.118] [1862.118][S01]One is, yes, if you're focused on one[1863.66] [1863.66][S01]or more particular species, then Perch[1865.513] [1865.513][S01]would be massively helpful because you could develop[1867.68] [1867.68][S01]detectors for those species.[1869.04] [1869.04][S01]And we also know it's not just the species, yes/no,[1871.165] [1871.165][S01]but the frequency of the calls and even some[1872.998] [1872.998][S01]of the variation in the calls.[1874.26] [1874.26][S01]So in some cases, you can tell individuals apart, by the way,[1876.802] [1876.802][S01]here.[1877.35] [1877.35][S01]So we can say, is it one bird singing all day long,[1880.372] [1880.372][S01]or do we have four or five birds singing?[1882.08] [1882.08][S01]And Perch can even help with that.[1884.27] [1884.27][S02] That specific bird--[1885.845] [1885.845][S01] Specific bird.[1886.97] [1886.97][S02] Wow.[1887.285] [1887.285][S02]OK.[1887.6] [1887.6][S01] Exactly, yeah.[1888.44] [1888.44][S01]It's really important to know I've got lots of blackbird song,[1891.09] [1891.09][S01]but is it 1 blackbird, or is it 10 blackbirds?[1893.04] [1893.04][S01]So in this way, you can use bioacoustics in combination[1896.54] [1896.54][S01]with Perch to increase your ability[1898.23] [1898.23][S01]to measure not just the presence or absence of species[1900.882] [1900.882][S01]but the abundance and then the change in that through time.[1903.34] [1903.34][S01]It's getting more abundant.[1904.465] [1904.465][S01]So that's one scenario, very species-focused view.[1907.0] [1907.0][S01]And this is how most people currently approach biodiversity[1910.35] [1910.35][S01]and ecosystems, is through this species[1912.0] [1912.0][S01]lens, which is very important.[1913.45] [1913.45][S01]But it's also interesting to me--[1914.92] [1914.92][S01]and I'm sharing an idea here that I've[1916.59] [1916.59][S01]shared with lots of colleagues, which[1917.73] [1917.73][S01]is in some areas of the world where you've got a really high[1920.37] [1920.37][S01]biodiversity and much less data, much less understanding, what[1923.91] [1923.91][S01]you could do instead is almost view the Perch[1927.0] [1927.0][S01]distillation as a distillation of the audio ecology of a place,[1930.1] [1930.1][S01]even if you can't identify all the species.[1932.08] [1932.08][S01]And then if I pick a good place, like a local nature reserve,[1935.74] [1935.74][S01]and I pick a place that I'm looking to restore,[1938.092] [1938.092][S01]then I could just say on aggregate,[1939.55] [1939.55][S01]are the Perch embeddings becoming more similar?[1941.34] [1941.34][S01]And I may not know which species are involved.[1943.257] [1943.257][S01]I've got all these things going, pops and clicks and whistles[1947.745] [1947.745][S01]and et cetera.[1950.55] [1950.55][S01]Because what Perch, in a sense, is[1953.25] [1953.25][S01]a kind of deep learning numerical answer[1955.89] [1955.89][S01]to the question, what does the nature sound like in this place?[1959.84] [1959.84][S01]And so if, through time, I could say[1961.54] [1961.54][S01]my place is becoming, in terms of that sound,[1965.157] [1965.157][S01]more similar to some reference place--[1966.74] [1966.74][S02] Well, let me ask though, because--[1968.657] [1968.657][S02]OK, so Perch, given the name, sounds quite a lot like it's[1971.29] [1971.29][S02]mainly focused on birds.[1972.305] [1972.305][S02]But can you put this into other environments?[1974.18] [1974.18][S02]What about underwater?[1975.218] [1975.218][S02]Does it work there?[1976.01] [1976.01][S01] So you're right.[1977.218] [1977.218][S01]Perch started by just classifying very large numbers[1979.96] [1979.96][S01]of bird species.[1981.798] [1981.798][S01]But amazingly, it turned out that then when[1983.59] [1983.59][S01]they tried it underwater on these hydrophones,[1985.782] [1985.782][S01]it seems to work really well.[1986.99] [1986.99][S02] What, the same model?[1987.95] [1987.95][S01] Exactly the same model worked really well.[1989.8] [1989.8][S02] What?[1990.11] [1990.11][S01] And I think what this[1990.85] [1990.85][S01]teaches us is that there's something[1992.44] [1992.44][S01]about natural sound, the way that that's evolved,[1995.24] [1995.24][S01]that actually has--[1996.41] [1996.41][S01]now, of course, it won't be perfect,[1997.91] [1997.91][S01]but it has a surprisingly high degree of transfer[2000.27] [2000.27][S01]into underwater.[2002.05] [2002.05][S01]There's this whole area of acoustic ecology,[2004.145] [2004.145][S01]and we know that-- well, we think[2005.52] [2005.52][S01]there's evidence for the fact that species have evolved away[2008.4] [2008.4][S01]from each other in terms of their use of the audio spectrum[2011.43] [2011.43][S01]because it's easy to forget that most of the sound you're picking[2014.25] [2014.25][S01]up in bioacoustics is deliberately[2017.19] [2017.19][S01]produced by organisms.[2018.298] [2018.298][S01]It's communication.[2019.09] [2019.09][S01]That's what's amazing.[2019.87] [2019.87][S01]They're communicating with each other.[2021.453] [2021.453][S01]And just like us, they need to think about the spectral bands.[2024.12] [2024.12][S02] They want to be heard.[2024.91] [2024.91][S01] Yeah, audio spectral bands.[2025.87] [2025.87][S01]So they're differentiated.[2027.18] [2027.18][S01]So it looks like--[2027.93] [2027.93][S02] So some use low sounds,[2029.14] [2029.14][S02]some use high sounds because you want[2030.81] [2030.81][S02]to make sure that you're communication is successful.[2033.337] [2033.337][S01] Yes, exactly.[2034.42] [2034.42][S01]So yeah, overall high and low.[2035.8] [2035.8][S01]And then you think about the temporal thing.[2037.71] [2037.71][S01]Is it like da, da, da, da, da, or is it ou, ou, ou thing?[2041.17] [2041.17][S01]So you can imagine that this is somehow[2043.26] [2043.26][S01]evolved as an interaction not just within species,[2045.82] [2045.82][S01]communication, but between all the species.[2047.98] [2047.98][S01]They've all responded to each other somehow[2049.29] [2049.29][S01]and settled down on these different patterns.[2051.19] [2051.19][S02] And what do you find when[2052.139] [2052.139][S02]you put microphones underwater?[2053.469] [2053.469][S02]What do you get from these hydrophones?[2055.179] [2055.179][S01] So we've cued up actually a couple[2057.36] [2057.36][S01]of sound recordings from coral reefs[2059.245] [2059.245][S01]that I'd like to listen to.[2060.37] [2060.37][S02] OK, go for it.[2061.75] [2061.75][CRACKLING, GURGLING][2062.969] [2062.969][S02]I mean, oh, some grumbles going on, some squeaks.[2069.0] [2069.0][S02]Quite crackly as well.[2070.302] [2070.302][S01] Yeah, and then here's another one.[2072.26] [2072.26][CRACKLING][2073.71] [2073.71][S02] Much quieter.[2075.67] [2075.67][S02]No grumbles.[2078.4] [2078.4][S02]I mean, what does that say about this coral reef, though?[2081.032] [2081.032][S01] So maybe now it won't[2082.449] [2082.449][S01]surprise you to hear that the first one of those[2084.449] [2084.449][S01]was a much healthier coral reef than the second.[2086.872] [2086.872][S02] Oh, really?[2087.83] [2087.83][S01] So that diversity and richness[2089.889] [2089.889][S01]of ecological sound, animal communication,[2092.889] [2092.889][S01]is a strong indication of a healthy ecosystem, you see?[2095.872] [2095.872][S01]So even as humans, we can hear that difference.[2097.83] [2097.83][S02] Wow.[2098.497] [2098.497][S02]So you really have that richness of understanding[2101.77] [2101.77][S02]the environment itself just from the audio signal.[2103.853] [2103.853][S01] That's right.[2104.937] [2104.937][S01]And that's an illustration of how even just that overall--[2107.39] [2107.39][S01]if we can embed that overall signal in a way,[2110.04] [2110.04][S01]that can just give us a general indication[2111.79] [2111.79][S01]of the health of ecosystems.[2113.48] [2113.48][S01]Of course, with more specific species data,[2115.58] [2115.58][S01]we can actually start to tease out individual species,[2118.16] [2118.16][S01]so if we're particularly concerned[2119.89] [2119.89][S01]with an endangered species, maybe even invasive species.[2123.34] [2123.34][S01]And we can pull out seasonal and daily temporal patterns[2127.24] [2127.24][S01]in animal behavior.[2128.99] [2128.99][S01]And we can start to understand geographic differences[2132.22] [2132.22][S01]in where species live, all just by dropping these microphones,[2135.2] [2135.2][S01]in this case, into the ocean.[2136.62] [2136.62][S02] But this is an explosion in possibilities.[2138.87] [2138.87][S02]Because if, previously, you stuck a microphone underwater,[2141.703] [2141.703][S02]you'd have to listen through hours[2143.12] [2143.12][S02]and hours and hours of tapes and have no real way of--[2146.067] [2146.067][S02]I mean, it's basically the human brain[2147.65] [2147.65][S02]trying to pick up on patterns.[2149.09] [2149.09][S02]But now, if you can single out not just individual species[2153.93] [2153.93][S02]but individual voices, as it were,[2155.88] [2155.88][S02]from particular individuals, that's gigantic, right?[2159.32] [2159.32][S02]Gigantic potential.[2160.112] [2160.112][S01] Yeah, it's absolutely huge.[2161.778] [2161.778][S01]That's why so many of us are so excited by the power[2164.21] [2164.21][S01]of bioacoustics combined with deep learning in this way[2166.913] [2166.913][S01]and especially with this foundational approach[2168.83] [2168.83][S01]to bioacoustics that Perch is bringing.[2170.84] [2170.84][S01]And it has at least two advantages[2172.263] [2172.263][S01]that you've hinted at there.[2173.43] [2173.43][S01]One is that it can just massively accelerate[2175.765] [2175.765][S01]what humans could do.[2176.64] [2176.64][S01]So you could take a problem that humans could do.[2178.682] [2178.682][S01]It's just that you'd have to have[2180.2] [2180.2][S01]hundreds of people listening to thousands of hours of audio.[2183.21] [2183.21][S01]So it was very, very inefficient.[2184.932] [2184.932][S01]But you're right.[2185.64] [2185.64][S01]There are also areas where you're going[2187.265] [2187.265][S01]go beyond human capabilities.[2189.24] [2189.24][S01]Humans are not necessarily all that[2191.06] [2191.06][S01]good at parsing out natural sounds[2193.04] [2193.04][S01]or identifying species from sound.[2194.49] [2194.49][S01]It's quite hard skill to learn.[2195.66] [2195.66][S01]And there may be cases like pulling out[2197.285] [2197.285][S01]different individuals might eventually be even too[2199.4] [2199.4][S01]subtle for humans to do.[2200.637] [2200.637][S02] Staying underwater for a moment,[2202.47] [2202.47][S02]there are some species--[2203.47] [2203.47][S02]I'm thinking about whales and dolphins[2205.25] [2205.25][S02]here-- where there is some evidence[2207.44] [2207.44][S02]that their communication is sophisticated, right?[2210.39] [2210.39][S02]Can you use these ideas for that?[2212.76] [2212.76][S02]Can you-- I don't know-- learn to speak dolphin?[2216.51] [2216.51][S01] [CHUCKLES] I like to think[2218.27] [2218.27][S01]that we will be able to talk to animals using AI at some point[2221.01] [2221.01][S01]and that these embedding foundational modeling approaches[2223.88] [2223.88][S01]like Perch will be really important in that.[2225.95] [2225.95][S01]But, of course, you'd need some other elements there.[2228.207] [2228.207][S01]Interestingly, some colleagues of mine[2229.79] [2229.79][S01]have been involved in producing this thing called DolphinGemma,[2233.58] [2233.58][S01]which is a large language model that's[2236.87] [2236.87][S01]been adapted to be suitable for beginning to decode dolphin[2240.8] [2240.8][S01]communication.[2242.04] [2242.04][S01]It takes the sounds, separates them out, tokenizes them,[2245.76] [2245.76][S01]and basically brings it into the world of large language[2248.17] [2248.17][S01]modeling.[2248.67] [2248.67][S01]So that's an example of it's early days,[2250.54] [2250.54][S01]but that's an example of AI actively being[2252.29] [2252.29][S01]used to study animal communication at a level[2254.91] [2254.91][S01]that we really couldn't do before.[2256.45] [2256.45][S02] The thing is, I can see[2257.04] [2257.04][S02]the potential for this in terms of scientific interest.[2259.87] [2259.87][S02]But I do wonder whether--[2261.93] [2261.93][S02]let's say we get to a point where[2263.91] [2263.91][S02]you can understand what dolphins are saying, communicate[2266.76] [2266.76][S02]with the higher animals on the planet.[2269.07] [2269.07][S02]Does it change how we view ourselves[2271.08] [2271.08][S02]and our place in the world?[2272.35] [2272.35][S01] I think yeah.[2272.83] [2272.83][S01]I think it absolutely has that potential.[2274.782] [2274.782][S01]Most of the work we're doing at the moment, as I mentioned,[2277.24] [2277.24][S01]is filling known information gaps[2279.54] [2279.54][S01]in known processes, et cetera.[2281.46] [2281.46][S01]But sometimes you think that the real change[2284.07] [2284.07][S01]can come, in the long run, from almost these moments[2287.82] [2287.82][S01]of awakening where people almost overnight can change[2291.3] [2291.3][S01]their relationship with nature.[2292.81] [2292.81][S01]And I think we've seen at least two examples in the past anyway.[2295.477] [2295.477][S01]One was the first picture of the Earth on the moon.[2297.91] [2297.91][S01]And I guess historians can debate this,[2300.22] [2300.22][S01]but they often trace some of the more recent increases[2302.82] [2302.82][S01]in the modern conservation movement to that picture[2306.077] [2306.077][S01]when people looked at the Earth for the first time[2308.16] [2308.16][S01]and realized there was just one in this dark universe[2310.48] [2310.48][S01]and that all of us shared it together, et cetera.[2313.57] [2313.57][S01]Another one is actually whale song,[2315.13] [2315.13][S01]and this is just listening to whales, even if we[2317.688] [2317.688][S01]can't tell what they're saying.[2318.98] [2318.98][S01]That really changed the way people thought about whales.[2321.313] [2321.313][S01]So the idea of being able to do that for more species[2323.608] [2323.608][S01]with things like Perch but then being[2325.15] [2325.15][S01]able to decode that to say what they're actually saying[2327.2] [2327.2][S01]and maybe one day even have some kind of conversation.[2329.45] [2329.45][S01]And if AI can help to empower that,[2330.925] [2330.925][S01]that might, in the long run, be the most powerful role of AI.[2333.467] [2333.467][S02] Well, OK, we've covered a lot of ground[2335.592] [2335.592][S02]in this episode.[2336.8] [2336.8][S02]So maybe I'll finish by asking about the future.[2340.01] [2340.01][S02]How will AI change the kinds of questions[2342.76] [2342.76][S02]that ecologists can ask?[2344.473] [2344.473][S01] So I think up to now,[2345.89] [2345.89][S01]as we've discussed, amazing progress in our ability[2348.43] [2348.43][S01]to monitor and extract data from the field[2350.62] [2350.62][S01]and from the literature.[2351.77] [2351.77][S01]Amazing progress, I think, in mapping and geo.[2354.7] [2354.7][S01]It's only going to get better.[2356.27] [2356.27][S01]I think the couple of the future directions[2358.24] [2358.24][S01]to explore-- one is, what if we really could confidently predict[2362.32] [2362.32][S01]the future of ecosystems?[2364.55] [2364.55][S01]All the plants, all the insects, everything, the soil,[2368.42] [2368.42][S01]the soil organisms, the fungi under a range[2370.57] [2370.57][S01]of different scenarios.[2371.87] [2371.87][S01]If I do this, this is what the ecosystem[2373.61] [2373.61][S01]will look like in 10 years.[2374.735] [2374.735][S01]If I do this, this will happen.[2376.14] [2376.14][S01]And also in a way that took into account future climate change,[2379.36] [2379.36][S01]et cetera.[2379.86] [2379.86][S01]So these future-proof, robust, conditional predictions[2383.72] [2383.72][S01]of the future of ecosystems down at the fine grain.[2386.14] [2386.14][S01]If we could do that, of course, that[2387.64] [2387.64][S01]could unlock an entirely different sort of relationship[2390.35] [2390.35][S01]between us and nature.[2391.91] [2391.91][S01]We could use that simulation ability[2393.502] [2393.502][S01]to then be able to find all the optimal ways that we should[2395.96] [2395.96][S01]behave, identify the key trade-offs-- which species[2398.48] [2398.48][S01]to plan, build new kinds of regenerative agriculture systems[2402.17] [2402.17][S01]and agroforestry with mixed species.[2404.313] [2404.313][S01]We've also got rewilding, and we're also[2405.98] [2405.98][S01]bringing the soil carbon back, and it's all[2407.45] [2407.45][S01]robust to climate change.[2408.51] [2408.51][S01]So if we could do that, then every person[2410.42] [2410.42][S01]that's affecting nature or taking a decision that they[2412.67] [2412.67][S01]think affects nature would be fully[2414.128] [2414.128][S01]informed as to the future consequences of those actions.[2417.5] [2417.5][S01]And that feels, to me, like an amazing potential.[2421.04] [2421.04][S01]Of course, it would depend on values as well,[2423.06] [2423.06][S01]but I would hope that, on average, most people[2424.977] [2424.977][S01]would use that ability to do good for nature.[2427.647] [2427.647][S02] Amazing.[2428.48] [2428.48][S02]Drew, thank you so much for joining me.[2430.105] [2430.105][S01] Oh, thanks for the invite.[2431.73] [2431.73][S01]It's been great, really.[2432.83] [2432.83][S02] I think it can be tempting to imagine that ecology[2435.47] [2435.47][S02]is a bit of a niche application for AI,[2437.64] [2437.64][S02]a nice, worthy project that sits behind all[2440.39] [2440.39][S02]of the big, flashy advances like video generation and drug[2443.57] [2443.57][S02]discovery.[2444.54] [2444.54][S02]But in reality, this is just a few steps behind the others.[2448.995] [2448.995][S02]They're still at the data gathering and data assimilation[2451.37] [2451.37][S02]phase of the process, but the roadmap here[2454.49] [2454.49][S02]is just as ambitious.[2456.63] [2456.63][S02]And in the long run, I think there is real potential[2459.23] [2459.23][S02]here not to just use AI to conserve what we have already[2463.38] [2463.38][S02]but to completely change our relationship[2465.71] [2465.71][S02]with the natural world itself.[2467.705] [2467.705][S02]You have been listening to \"Google DeepMind--[2469.58] [2469.58][S02]The Podcast\" with me, Professor Hannah Fry.[2471.87] [2471.87][S02]If you enjoyed this episode, then[2473.63] [2473.63][S02]do subscribe to our YouTube channel[2475.49] [2475.49][S02]or leave a review on your favorite podcast platform.[2478.14] [2478.14][S02]And of course, we have plenty more episodes[2480.98] [2480.98][S02]on a whole range of topics to come, so do check those out.[2484.38] [2484.38][S02]See you next time.[2485.3] [2485.3][MUSIC PLAYING][2488.35]"}
{"file_name": "audio/val_000024.wav", "transcription": "[4.037][S02] So here we are, the last episode in this series of the DeepMind podcast.[9.576] [9.576][S02]My name is Hannah Fry. I am a mathematician,[12.045] [12.045][S02]and someone who is deeply intrigued by artificial intelligence.[16.049] [16.049][S02]Much like you, I imagine, since you made it this far.[19.219] [19.219][S02]Now we’ve been toying with the big questions in this series.[22.456] [22.456][S02]What is intelligence? How does an algorithm learn?[25.526] [25.526][S02]And what to do with the AI future once we get there?[29.696] [29.696][S02]And I have been asking the team of scientists and engineers[32.432] [32.432][S02]here at DeepMind to give us their take of where things are at[36.403] [36.403][S02]and where they’re going. But now for this final episode,[40.24] [40.24][S02]we have a chance to catch up with Demis Hassabis,[43.31] [43.31][S02]the co-founder and CEO of DeepMind[46.113] [46.113][S02]to hear what he has to say on these questions and more.[50.05] [52.252][S02]Demis Hassabis grew up in North London in the 1970s.[55.822] [55.822][S02]By the age of 13, he was ranked 2nd in the world at Chess in his age group.[61.762] [61.762][S02]At 16 he worked as a games designer-[64.131] [64.131][S02]remember Theme Park? That was him.[66.567] [66.567][S02]Then he went on to study computer science[68.802] [68.802][S02]and then neuroscience before setting up[71.004] [71.004][S02]DeepMind with his two co-founders Shane Legg and Mustafa Suleyman.[75.442] [75.442][S02]His accomplishments are ferociously intimidating.[79.513] [79.513][S02]But as an article in the Times put it,[81.582] [81.582][S02]Demis doesn’t even have the good grace to be socially deficient.[85.886] [85.886][S02]But if all of that makes it sound like Demis is a man with a plan,[89.857] [89.857][S02]then you’d be right.[91.625] [92.226][S01] I had in mind something like creating a company like DeepMind[95.162] [95.162][S01]to research AI from a long time ago,[97.698] [97.698][S01]so I was sort of working back from the end state[100.367] [100.367][S01]which is what would I need, what skills would I need,[102.936] [102.936][S01]what experiences would I need to even stand a chance[105.472] [105.472][S01]of even building something like that.[106.907] [106.907][S02] Because these different aspects of your life -[108.475] [108.475][S02]the chess, the neuroscience, the games design,[110.277] [110.277][S02]they’re not disconnected, I mean they do build to a bigger picture.[113.313] [113.313][S01] They do. And it’s hard to say which way round it is -[115.749] [115.749][S01]so I picked those subjects and those things to study -[118.685] [118.685][S01]say computer science at Cambridge, and then cognitive neuroscience at UCL[121.989] [121.989][S01]because I wanted this component of computer science[125.025] [125.025][S01]and neuroscience to come together[127.16] [127.16][S01]and obviously that’s what we do at DeepMind,[129.563] [129.563][S01]but even the game stuff, that taught me about creative thinking,[133.967] [133.967][S01]also a massive engineering projects,[136.637] [136.637][S01]and then of course it ended up that we used games as our main vehicle[140.941] [140.941][S01]for proving out our AI algorithms. One other thing I’ve learnt from games[145.979] [145.979][S01]is to use every scrap of asset that you have - like in games,[150.017] [150.017][S01]you always have a limited resource pool like you know, if it’s a chess game,[153.52] [153.52][S01]it’s the chess pieces you have left on the board,[155.789] [155.789][S01]and one way to think about games is maximising[158.759] [158.759][S01]the use of the of the assets you have left.[161.361] [161.361][S01]Perhaps that’s why I was biassed towards using games,[163.564] [163.564][S01]but I also felt it was the logical way to go about building AI.[167.634] [167.634][S02] What was your Phd in?[169.736] [169.736][S01] My PhD was in cognitive neuroscience[172.573] [172.573][S01]and I actually decided to study[176.543] [176.543][S01]how memory and imagination works in the brain,[179.847] [179.847][S01]and the reason I chose cognitive neuroscience[181.782] [181.782][S01]is I wanted to better understand[183.25] [183.25][S01]how the brain does certain cognitive functions[186.186] [186.186][S01]so that perhaps we could be inspired on new types of algorithms[189.756] [189.756][S01]based on how the brain works,[191.692] [191.692][S01]and so it’s a good idea to pick functions[194.828] [194.828][S01]that we don’t know how to do in AI[197.097] [197.097][S01]and ah I went to study with Eleanor Maguire at UCL[201.134] [201.134][S01]and she’s one of the world’s leading experts in the hippocampus,[204.471] [204.471][S01]which is critical for memory.[206.173] [206.173][S01]But I told her that what I really wanted to look at was imagination[209.643] [209.643][S01]which is um you can also think of it[211.445] [211.445][S01]as simulating things in the future, in your mind.[215.115] [215.115][S01]Ah for obviously it’s useful for planning but also for creativity.[218.452] [218.452][S01]And the reason I was interested in that is of course[219.953] [219.953][S01]that’s an incredibly important part of human intelligence[223.19] [223.19][S01]and it’s also something I used a lot in my games design career.[227.227] [227.227][S01]So I used a lot of visualisation techniques[229.796] [229.796][S01]and imagining how would a player viscerally like play this game,[233.934] [233.934][S01]like Theme Park and then you try and change something about it -[236.904] [236.904][S01]all in your mind before or in sketches -[239.673] [239.673][S01]before you went to the trouble of programming it all.[242.075] [242.075][S01]And it felt to me that we were using a similar type of process[246.113] [246.113][S01]to the way when we’ve lucidly remember things that have happened to us,[249.917] [249.917][S01]so I thought there maybe there would be a connection[252.352] [252.352][S01]with this kind of you could imagine like a simulation engine of the mind[255.956] [255.956][S01]that was being used both for imagination and memory,[258.759] [258.759][S01]and that’s what I wanted to work on during my PhD.[262.095] [262.095][S01]And we ended up discovering something quite important that that[264.364] [264.364][S01]in fact the hippocampus was at the core of both of those two types of function.[269.903] [269.903][S01]It’s critical for memory, which we already know,[272.372] [272.372][S01]but it’s also critical for imagination.[274.741] [274.741][S01]And you can’t really imagine vividly without your hippocampus[279.079] [279.079][S01]so we ended up discovering this important thing[282.115] [282.115][S01]and then subsequently that’s been at the heart of a lot[284.551] [284.551][S01]of what we try to do in AI[286.854] [286.854][S01]is build memory and imagination abilities into our AI systems,[291.325] [291.325][S01]and we’re still doing that now.[292.826] [292.826][S02] When it comes to bringing those ideas across[294.695] [294.695][S02]and trying to implement them in AI,[296.463] [296.463][S02]where do you find that balance between just directly copying[298.932] [298.932][S02]what the brain’s doing and and and using it for inspiration?[302.269] [302.269][S01] So that’s very important sign post[304.338] [304.338][S01]when you’re um scrabbling around in the dark in the unknown of science,[308.308] [308.308][S01]any signpost is really valuable,[310.344] [310.344][S01]and the brain is the only existence proof[312.312] [312.312][S01]we have in the universe that intelligence is possible.[315.449] [315.449][S01]So it always felt to me that it would be crazy to ignore that[319.119] [319.119][S01]as a source of information of how to build AI.[322.055] [322.055][S01]So we use neuroscience for two things -[324.825] [324.825][S01]one is inspiration for new ideas about algorithms or architectures[329.997] [329.997][S01]to representations that the brain uses.[331.965] [331.965][S01]And then we can get some inspiration for that for new types of algorithms.[335.536] [335.536][S01]The second way we use neuroscience is what I call for validation.[339.406] [339.406][S01]So we may already have some idea from coming from engineering or mathematics[344.344] [344.344][S01]about how to build a learning system that say reinforcement learning -[347.915] [347.915][S01]that came from engineering disciplines and operational research first,[352.152] [352.152][S01]but then in the 90s we discovered that the brain[355.022] [355.022][S01]also implements a form of reinforcement learning[357.224] [357.224][S01]and what that means is that from an AI perspective[360.16] [360.16][S01]you can be sure that reinforcement learning[362.963] [362.963][S01]could plausibly be a component part of an AI system[366.466] [366.466][S01]because it’s in the brain and we know the brain is a general intelligence.[370.003] [370.003][S01]So that means - that’s really important if you’re thinking about[372.806] [372.806][S01]where to put your engineering resources and effort -[375.375] [375.375][S01]you know that if it doesn’t work right now,[376.977] [376.977][S01]and things never work first time in research or engineering,[379.546] [379.546][S01]it’s worthwhile pushing that harder[382.182] [382.182][S01]because you know eventually this must work[385.352] [385.352][S01]because the proof of concept is the brain.[387.554] [387.554][S01]Having said that though, there is another school of thought from[390.557] [390.557][S01]AI practitioners and neuroscientists that we need to slavishly copy the brain[395.596] [395.596][S01]completely from the bottom up like on a neuronal level[399.6] [399.6][S01]and I think that’s also the wrong approach.[401.235] [401.235][S01]What we’re after is what I call a systems neuroscience approach[404.771] [404.771][S01]which is that you’re interested in the algorithms[407.174] [407.174][S01]and the architectures that the brain is using.[410.077] [410.077][S01]Not necessarily the exact implementation details,[413.046] [413.046][S01]because I think that’s likely to be different for in silicon systems -[417.384] [417.384][S01]like in computers compared to carbon based systems like our minds.[421.221] [421.221][S01]There's no reason to think that we would implement[423.757] [423.757][S01]exactly the same implementation details in a silicon[426.86] [426.86][S01]based system that’s going to have different strengths and weaknesses[429.93] [429.93][S01]than a carbon based system like our minds.[432.299] [435.335][S02] Normally tech startups have customers, they have products,[438.772] [438.772][S02]but this is sort of more like a start-up research facility.[442.342] [442.342][S01] Yes.[443.544] [443.544][S02] How do you get something like that off the ground?[445.479] [445.479][S01] It’s pretty hard, I mean -[447.114] [447.114][S01]you’re right in that it’s a very unusual company.[450.717] [450.717][S01]What I try to do is basically take the best from start-up world,[453.654] [453.654][S01]the kind of focus and energy and um pace[457.324] [457.324][S01]that you get in the best start-ups um say in Silicon Valley.[460.794] [460.794][S01]And I wanted to combine that with the best from academia,[463.73] [463.73][S01]which is blue sky thinking, incredibly bright people,[467.634] [467.634][S01]working on you know long-term, big research questions[471.505] [471.505][S01]and stepping into the unknown all the time,[474.141] [474.141][S01]and you know obviously I spend some time in academia myself,[477.644] [477.644][S01]and there’s very great aspects about academia,[480.28] [480.28][S01]but there’s also some things that are frustrating.[482.749] [482.749][S01]Um mostly around um the organisational aspects[486.22] [486.22][S01]and the pace of it can sometimes be slower than you would like.[489.556] [489.556][S01]It’s difficult to get momentum behind things in the way you can a start-up[493.36] [493.36][S01]if it if it’s you know if it things are going well,[496.263] [496.263][S01]but I when I was in academia and I’d already obviously started[499.466] [499.466][S01]and been involved with a few start-ups before going back to do my PhD,[502.436] [502.436][S01]so I’d experience both sides, and I didn’t feel like there[505.506] [505.506][S01]was any reason why these should be mutually exclusive environments.[509.209] [509.209][S01]Although they have generally been treated as very different,[511.812] [511.812][S01]almost opposite environments.[513.547] [513.547][S01]And there’s a lot of things that do seem opposite,[516.116] [516.116][S01]but I felt if you if you were kind of smart about it,[519.052] [519.052][S01]you could extract the best of both those worlds[521.722] [521.722][S01]and combine them into some kind of hybrid organisation,[524.892] [524.892][S01]and I feel that’s what DeepMind is.[526.96] [526.96][S01]And I don’t think many people have ever done that and and so[529.563] [529.563][S01]that’s why it seems quite strange.[531.565] [531.565][S01]Probably as an organisation I think we’ve shown with our scientific output,[536.336] [536.336][S01]even if you measure it by normal measures, Nature, Science papers,[538.972] [538.972][S01]this kind of thing that normal academic labs[540.607] [540.607][S01]would measure themselves by, we’ve been very successful,[543.577] [543.577][S01]and I think you know, we’ve been also on the other side things[545.946] [545.946][S01]we’ve been able to produce really big breakthroughs[549.049] [549.049][S01]that took a lot of engineering effort as well like AlphaGo[552.085] [552.085][S01]which would have been very difficult I think to do in academia,[554.421] [554.421][S01]in a in a normal, you know, small academic lab.[557.324] [557.324][S02] One of the thing that makes DeepMind I guess a bit unusual[560.594] [560.594][S02]is that you publish your work -[563.096] [563.096][S02]I mean other companies don’t really do this - is there some way[565.666] [565.666][S02]that you’re kind of giving away your competitive advantage, really.[568.635] [568.635][S01] Yeah, it’s an interesting issue actually,[570.27] [570.27][S01]so we’ve always published everything we’ve done,[572.639] [572.639][S01]and lots of companies do wonder you know why we’re doing that[576.577] [576.577][S01]because a lot of other companies don’t necessarily do that.[579.947] [579.947][S01]We feel it’s part of the scientific discourse,[582.449] [582.449][S01]like that’s the right way to do science.[584.585] [584.585][S01]We really believe in sort of peer review journals[587.254] [587.254][S01]so that your work is scrutinised at the highest level by your peers,[590.691] [590.691][S01]which is the you know, the gold standard in science.[593.193] [593.193][S01]That’s why we publish in the top journals like Nature and Science.[596.096] [596.096][S01]Also you do get more exposure for your ideas like that,[599.9] [599.9][S01]so some of our top papers had been cited like more[602.669] [602.669][S01]than 5,000 times now in the last 2-3 years,[605.072] [605.072][S01]some of the most cited papers in the world, so that’s great.[608.175] [608.175][S01]You know, and I think that if a community shares ideas like[610.244] [610.244][S01]that the whole field can advance much more quickly[613.28] [613.28][S01]than if everyone was to keep their ideas secret,[615.883] [615.883][S01]but you know, there are some interesting aspects to that -[618.519] [618.519][S01]one is the competitive side,[620.12] [620.12][S01]I mean I feel there that you just need to carry on innovating[623.924] [623.924][S01]than a faster pace than anyone else. That’s the most important thing,[627.094] [627.094][S01]not trying to keep hold of the ideas you already innovated.[630.364] [630.364][S01]You should just be - by the time you’ve you know, you’ve published it,[632.833] [632.833][S01]you should be one or two ideas even further down the line[635.736] [635.736][S01]if you’re continuing to work at the same pace[638.639] [638.639][S01]and the same level of innovation,[640.44] [640.44][S01]so I think that’s really the biggest sort of protection against competitors[644.778] [644.778][S01]- is your pace of innovation.[646.813] [646.813][S02] In the early days, I mean back in sort of you know 2009 -[649.216] [649.216][S02]2010, when AI wasn’t the hot topic that perhaps it is today,[653.187] [653.187][S02]did you find it difficult to get attention from the people,[658.025] [658.025][S02]that I mean, I think I remember reading[659.56] [659.56][S02]that you decided to go straight for billionaires[661.862] [661.862][S02]rather than millionaires when it came to investment.[664.064] [664.064][S01] Yes. Well, that was incredibly tough,[666.433] [666.433][S01]so it’s really hard to remember now, even for me, like 10 years ago,[670.704] [670.704][S01]you know no one was investing in AI.[672.339] [672.339][S01]It was an impossible thing to get money for actually.[675.142] [675.142][S01]You know no one would invest, and it’s still very difficult today[677.711] [677.711][S01]I think on what I would call a deep technology[680.514] [680.514][S01]or a science based start-up, right?[682.983] [682.983][S01]And with no clear product in mind. I mean what we were basically saying[686.32] [686.32][S01]is we were going to build this incredible general purpose technology[690.524] [690.524][S01]that as it got more powerful[692.426] [692.426][S01]there should be a myriad of things you could just apply it to.[695.262] [695.262][S01]But you know, that sounds probably pretty far-fetched[699.366] [699.366][S01]to a normal kind of investor.[701.602] [701.602][S01]It was almost like well that’s what academia is for, isn’t it?[705.038] [705.038][S01]You should be just - if you don’t know - you know it’s blue sky research,[707.975] [707.975][S01]you don’t know when it’s going to work, it’s pure research,[710.844] [710.844][S01]then um go and do that for another 10 years in academia,[713.981] [713.981][S01]and then come and talk to them when it’s working.[716.083] [716.083][S01]But that would have been too slow. And I could see that within academia,[719.753] [719.753][S01]so then that’s why I decided not to go after normal venture capitalists[723.924] [723.924][S01]who certainly in Europe or in the UK they would want to make 10x return,[728.529] [728.529][S01]maybe within 3 - 5 years, right? That’s the kind of time horizon.[732.032] [732.032][S01]Of course that’s no good for a research based company,[735.169] [735.169][S01]you know you haven’t barely got going after 3 years right?[737.804] [737.804][S01]So what you need is a profile investor who is more interested in 1,000x return[742.776] [742.776][S01]but they’re willing to wait 10 years, maybe even 20 years,[746.647] [746.647][S01]and that kind of profile of an investor basically just does not exist in Europe.[751.251] [751.251][S01]Certainly didn’t back in 2010.[753.12] [753.12][S01]Really you’re talking about Silicon Valley, self-made billionaires[757.457] [757.457][S01]I guess, who both have deep enough pockets[759.726] [759.726][S01]to take that kind of bet, and if it doesn’t work, it’s okay.[763.063] [763.063][S01]But they’re also personally interested in these types of topics[766.066] [766.066][S01]and have seen incredibly ambitious things work[769.269] [769.269][S01]because usually that’s how they’ve made their money.[772.339] [772.339][S02] Just going back to the idea of how everything[774.474] [774.474][S02]sort of slots into place - all the different aspects,[776.61] [776.61][S02]all the different passions that you have slots into place.[778.512] [778.512][S02]Didn’t chess play a role in you getting your first funding?[781.281] [781.281][S01] Yes, it did. So um chess has been key you know core part of my personality[785.919] [785.919][S01]I guess because I’ve been playing it for so long,[788.188] [788.188][S01]and I think a lot of my thought processes developed[790.858] [790.858][S01]because of that - so planning,[793.227] [793.227][S01]and thinking about problem-solving, all of these aspects[796.196] [796.196][S01]which I think are useful for for anything that you do in your life.[800.133] [800.133][S01]But it also turned out to be useful directly.[802.202] [802.202][S01]Because one of the first investors we talked[804.605] [804.605][S01]to was a chess player themselves[806.974] [806.974][S01]and a pretty strong junior chess player in the US[809.977] [809.977][S01]and when we were doing sort of our background reading on this,[813.046] [813.046][S01]we spent about a year preparing for this meeting almost[815.849] [815.849][S01]and there was an important meeting[817.084] [817.084][S01]because we knew that not very many people would get what we were doing,[820.22] [820.22][S01]and this was one of the people who felt that we would.[822.289] [822.289][S01]So it was important and we couldn’t get a meeting cause[825.192] [825.192][S01]we had no contacts in Silicon Valley,[827.294] [827.294][S01]you know I didn’t know anyone over there in California[829.73] [829.73][S01]-nor did Shane and Mustafa -[831.064] [831.064][S01]you know it’s sort of um how do you break into that world?[834.434] [834.434][S01]And so we finally managed to get asked to a conference[837.938] [837.938][S01]where this billionaire was sponsoring, and then we knew we would meet him[841.508] [841.508][S01]at some kind of after-conference party, but the problem is,[844.311] [844.311][S01]is that there are hundreds of people all trying to pitch him their ideas.[847.781] [847.781][S01]So if you’re just if you’re just another one pitching your another crazy idea,[851.685] [851.685][S01]it’s very unlikely you’re going to get noticed, right?[853.787] [853.787][S01]So I thought instead of that, I’d take this calculated risk[856.723] [856.723][S01]and talk to him about chess instead,[858.592] [858.592][S01]but then you have to have something interesting to say[860.794] [860.794][S01]about chess that may be you hadn’t thought of -[862.796] [862.796][S01]so I used my number one fact on chess that even surprises grand masters,[867.367] [867.367][S01]which is that thinking about it from a game designer point of view,[870.704] [870.704][S01]you know why is chess such a great game?[872.573] [872.573][S01]How did it evolve into such a great game?[874.708] [874.708][S01]And what is it that makes it so great? And my belief is that it’s actually[878.478] [878.478][S01]because of the creative tension of the bishop and knight.[882.182] [882.182][S01]The bishop and knight are basically worth the same -[884.618] [884.618][S01]they are three points each, but they have completely different powers,[889.556] [889.556][S01]and that asymmetry - that sort of creative asymmetry[892.526] [892.526][S01]that happens with the bishop and knights being swapped into various positions[896.163] [896.163][S01]I think is what makes chess a fascinating game.[898.966] [898.966][S01]And so I basically pretty much led out with that line -[901.635] [901.635][S01]I don’t know how I managed to crow bar that into a drinks party -[905.272] [905.272][S01]but I did, but it made him sort of stop and think,[907.908] [907.908][S01]which is exactly what I was hoping. And then he invited us back the next day[911.879] [911.879][S01]to do a proper pitch of our business idea,[914.147] [914.147][S01]so we actually got half an hour with him,[916.183] [916.183][S01]rather than one minute over some drinks,[918.485] [918.485][S01]so ah that actually you know worked out.[920.854] [920.854][S01]So you could say chess worked on two levels then -[922.99] [922.99][S01]the meta level of planning for that[924.925] [924.925][S01]and also actually chess is an intriguing subject in itself.[928.228] [928.228][S02] So actually getting funding from billionaires is really simple -[930.531] [930.531][S02]spend a year studying their interests, come up with a genius idea[934.201] [934.201][S02]that will catch their attention and then and then away you go.[936.537] [936.537][S01] And you have to do a good pitch after that, easy.[939.273] [939.273][S02] Simple! Very straightforward.[941.375] [943.31][S02]One thing that you hear time and time again in this building[946.547] [946.547][S02]and actually throughout this podcast series is how[949.082] [949.082][S02]DeepMind wants to use artificial intelligence to solve everything.[953.82] [953.82][S02]So here is the chance to ask - what do they actually mean by that?[958.091] [958.091][S02]Do they want to be the ones addressing every one of the world’s problems[961.728] [961.728][S02]once intelligence is cracked?[964.231] [965.832][S01] So I’ve you know I have listed -[967.501] [967.501][S01]a working list of a dozen to a couple of dozen scientific problems[971.872] [971.872][S01]that I feel are these kinds of root node problems,[975.976] [975.976][S01]and if we could crack all of those,[978.278] [978.278][S01]then I feel like that would transform society for the better,[981.582] [981.582][S01]and open up all sorts of areas in you know medicine and science[985.385] [985.385][S01]for us to make breakthroughs in.[986.787] [986.787][S02] Go on - give me the list [S01] I can’t -[988.622] [988.622][S02] Some of the list[990.524] [990.524][S01] Well some of the list of things like[992.226] [992.226][S01]I think a key thing we need to crack is cheap, abundant energy[995.596] [995.596][S01]that is renewable and clean.[997.097] [997.097][S01]And so, if that’s fusion, or just way better solar panels[1001.201] [1001.201][S01]with way better batteries with room temperature superconductors,[1004.838] [1004.838][S01]that would also solve that problem.[1006.607] [1006.607][S01]So you know I think there’s a number of solutions to that -[1009.576] [1009.576][S01]some are material science solutions, some are physics solutions,[1012.513] [1012.513][S01]and we should have a go at cracking all of those,[1014.414] [1014.414][S01]but if you crack that, then that would open up all sorts of new issues,[1018.685] [1018.685][S01]so I’ll give you an example - water access - access to clean water.[1022.422] [1022.422][S01]It’s going to become increasingly more important[1024.825] [1024.825][S01]as the population in the world grows, right?[1027.528] [1027.528][S01]And it’s already becoming in some countries more valuable than oil[1030.597] [1030.597][S01]because you know there’s just so little, fresh clean water around, right?[1034.234] [1034.234][S01]for a lot of communities, especially poor communities,[1036.703] [1036.703][S01]it’s an incredible problem.[1038.105] [1038.105][S01]But we have a solution already - it’s desalination, right?[1041.742] [1041.742][S01]70% of the earth is water but it’s salt water,[1045.045] [1045.045][S01]so how do we deal with that? Well desalination technologies exist,[1048.882] [1048.882][S01]the problem is they cost too much energy.[1050.651] [1050.651][S01]They’re too costly. So some rich countries can do it -[1053.287] [1053.287][S01]so I think Israel gets a lot of their water like this[1055.522] [1055.522][S01]and some other countries like that,[1057.09] [1057.09][S01]but of the poor countries it’s too expensive.[1059.593] [1059.593][S01]So if you’ve solved the renewable, cheap, clean energy problem,[1064.798] [1064.798][S01]then you would automatically solve the water access problem.[1069.77] [1069.77][S01]Almost straight away because um that’s actually the issue.[1073.173] [1073.173][S02] So where do you want to be[1074.575] [1074.575][S02]when the water salivation problem is solved?[1077.644] [1077.644][S02]Where do you fit into that story?[1080.08] [1080.08][S01] I hope that we will have been integral[1083.283] [1083.283][S01]in coming up with those solutions[1084.651] [1084.651][S01]by doing something say in fusion or in material science[1089.423] [1089.423][S01]more likely where we’ve come up with using an AlphaZero like system,[1093.794] [1093.794][S01]you know a battery that is 50% more efficient[1097.698] [1097.698][S01]and costs 1/10 of the price of current batteries[1100.567] [1100.567][S01]and lasts you know, 10 times longer.[1103.203] [1103.203][S01]You know or we would come up with a solar panel -[1105.339] [1105.339][S01]photovoltaic material that is twice as efficient at converting heat energy[1110.377] [1110.377][S01]into electrical energy.[1112.145] [1112.145][S01]So it would be something like that would then unlock the possibility[1116.65] [1116.65][S01]of making desalination within reach of every community.[1120.587] [1120.587][S01]Perhaps that has to be some improvement[1122.055] [1122.055][S01]with the desalination technology itself as well[1125.025] [1125.025][S01]and maybe we can be involved in that.[1126.86] [1126.86][S01]But you know we’re a relatively small company,[1128.896] [1128.896][S01]so and we’re going to stay relatively small,[1130.697] [1130.697][S01]so we have to be efficient with the solutions that we work on.[1134.835] [1135.969][S01]This all just comes, at least from my perspective,[1138.472] [1138.472][S01]from rationally, logically thinking out what’s the best thing you can do[1142.409] [1142.409][S01]and what’s happened so far looking at civilisation in this[1145.479] [1145.479][S01]I mean maybe you could say as a slightly strange way of looking at it[1148.849] [1148.849][S01]but I think it’s the correct way to look at things,[1150.817] [1150.817][S01]and I think most people just don’t think about questions in the right way,[1154.588] [1154.588][S01]and maybe that’s what I’ve done in my whole life[1156.19] [1156.19][S01]is try to ask the right questions[1159.526] [1159.526][S01]and I feel like this is the obvious answer.[1161.995] [1163.063][S02] This is DeepMind, the podcast. An introduction to AI -[1167.267] [1167.267][S02]one of the most fascinating fields in science today.[1170.537] [1172.306][S01] Like I always say to people -[1174.208] [1174.208][S01]whatever your question, the answer is AI.[1176.944] [1176.944][S01]Because it sort of is in the limit, right?[1179.079] [1179.079][S01]I mean that’s a little bit flippantly, but I mean in the limit, it must be,[1182.749] [1182.749][S01]because the answer so far, you know why we’re here,[1186.086] [1186.086][S01]why we’re talking, why we’re using these amazing computers and devices[1189.59] [1189.59][S01]is because of intelligence, human intelligence,[1191.391] [1191.391][S01]and I think it’s miraculous, and the scientific method.[1194.361] [1194.361][S01]Another miraculous thing, I think the greatest discovery of all[1198.098] [1198.098][S01]is that the scientific method works,[1199.766] [1199.766][S01]you know the enlightenment and why should it be?[1202.135] [1202.135][S01]I mean also you must have to question things like that -[1204.071] [1204.071][S01]why should the universe work like that - that the scientific method works,[1207.541] [1207.541][S01]you know it could be a little bit more random,[1208.842] [1208.842][S01]then it would be really confusing, right?[1210.31] [1210.31][S01]If sometimes the sun rose, and sometimes it didn’t -[1212.579] [1212.579][S01]it would be quite hard to do science, then. Right?[1214.715] [1214.715][S01]Sometimes you repeated the experiment with exactly the same conditions,[1217.017] [1217.017][S01]something else, something different happened.[1219.119] [1219.119][S01]But it doesn’t this world doesn't’ seem to work like that,[1221.021] [1221.021][S01]it seems to be repeatable, it seems to be consistent,[1223.991] [1223.991][S01]so therefore knowledge is possible.[1225.926] [1225.926][S01]And incredibly strangely our brains even though they’re evolved for hunting[1230.464] [1230.464][S01]gathering can somehow deal with it - which is kind of miraculous in itself.[1235.002] [1235.002][S01]So how could you not want to a.) work on those questions, and b.)[1238.906] [1238.906][S01]why would there be limits to what that is capable of doing?[1241.909] [1241.909][S02] But the ultimate goal in all of this[1243.977] [1243.977][S02]is to create artificial general intelligence,[1247.247] [1247.247][S02]exactly what is meant by that?[1250.35] [1251.285][S01] Yeah, artificial - I mean there’s not an agreed definition[1254.021] [1254.021][S01]of artificial general intelligence,[1255.989] [1255.989][S01]but the way that Shane and I think about artificial general intelligence[1260.594] [1260.594][S01]is a system that is capable of a wide range of tasks,[1265.299] [1265.299][S01]and if we think about human level artificial general intelligence,[1268.235] [1268.235][S01]then we’re talking about system that can pretty much[1270.47] [1270.47][S01]do the full spectrum of cognitive tasks[1274.374] [1274.374][S01]that humans can at least as good as humans are able to do,[1278.345] [1278.345][S01]that’s you know one reasonable definition[1280.113] [1280.113][S01]of artificial general intelligence.[1281.782] [1281.782][S02] What’s the threshold for AGI then,[1283.517] [1283.517][S02]how will you know when you’re done?[1285.385] [1285.886][S01] That’s a philosophical issue of like[1287.621] [1287.621][S01]how do we know we’re done with building AGI?[1290.624] [1290.624][S01]Certainly for me I’m waiting to see a lot of key moments,[1294.661] [1294.661][S01]for example I think a really big moment will be when an AI system[1299.032] [1299.032][S01]comes up with a new scientific discovery that’s of Nobel prize-winning level,[1305.038] [1305.038][S01]that to me would be a big watershed moment[1308.075] [1308.075][S01]and I think ah an important step in the capabilities of these systems,[1312.346] [1312.346][S01]so you know, capable of some kind of true creativity in some sense.[1316.383] [1316.383][S01]I think other big points will be when it can use language[1320.254] [1320.254][S01]and converse with us in a naturalistic way.[1323.323] [1323.323][S01]It’s capable of learning abstract concepts -[1326.293] [1326.293][S01]these are all things that I think are high-level cognitive abilities[1330.564] [1330.564][S01]that we’re nowhere near yet -[1332.099] [1332.099][S01]and I think will be big signposts on the way.[1334.835] [1334.835][S02] When were you convinced that that all of this was possible?[1338.305] [1338.305][S01] Well I had this in mind since my early teens,[1342.643] [1342.643][S01]I probably read way too much sci-fi I’m guessing.[1345.245] [1345.245][S01]Some of the really formative things on me were Asimov’s Foundation series,[1349.116] [1349.116][S01]so interestingly, not the robotics books,[1351.451] [1351.451][S01]I haven’t really read any of his robot books,[1353.02] [1353.02][S01]but the Foundation series was this really amazing series of sci-fi novels,[1357.691] [1357.691][S01]and then Ian Banks’ Culture Series which is his sort of Space Opera[1362.262] [1362.262][S01]about how the universe would look after humanity[1364.798] [1364.798][S01]has built AI and co-exists with it.[1367.134] [1367.134][S01]And then a really big scientific book for me[1369.169] [1369.169][S01]was when I was writing Theme Park,[1372.105] [1372.105][S01]obviously I was working on AI and building AI for the game,[1376.41] [1376.41][S01]but I was also reading books like Gödel, Escher, Bach, by Hofstadter[1380.013] [1380.013][S01]which I suppose is more of a philosophy book,[1382.716] [1382.716][S01]but it’s an incredible piece of work,[1385.085] [1385.085][S01]tying together Gödel’s incompleteness theorem about mathematics,[1389.256] [1389.256][S01]with Escher’s drawings and Bach’s fugues[1392.292] [1392.292][S01]and showing that they’re all related in some way,[1394.661] [1394.661][S01]this repeating cycle of patterns,[1396.53] [1396.53][S01]this infinite patterns that they all exhibit,[1399.333] [1399.333][S01]and then he tied it to consciousness, and intelligence,[1402.536] [1402.536][S01]and it was just really inspiring for me[1405.339] [1405.339][S01]and made me think about these deep questions,[1407.474] [1407.474][S01]and I was discussing this with a lot of my friends[1409.843] [1409.843][S01]who were you know we were writing games together,[1411.979] [1411.979][S01]and we were doing that 24/7[1413.814] [1413.814][S01]and we would discuss these things about what the limits of AI could be[1417.885] [1417.885][S01]if we could not just use it for what we were doing in games,[1420.988] [1420.988][S01]but actually advance it to the level[1422.489] [1422.489][S01]where it would become the same level as human,[1424.992] [1424.992][S01]and they just felt like the sky was the limit.[1428.395] [1428.395][S01]I mean maybe another way I can put it is if you look around us today,[1432.099] [1432.099][S01]you know you look at modern civilisation and its incredible,[1434.968] [1434.968][S01]well what built modern civilisation, intelligence did, right?[1439.106] [1439.106][S01]That’s what built it - human intelligence -[1441.141] [1441.141][S01]and if you were to take us back to our hunter gatherer days -[1445.179] [1445.179][S01]10, 20,000 years ago, 30,000 years ago,[1448.549] [1448.549][S01]and you were to say one day we’re going to build Manhattan,[1452.019] [1452.019][S01]and then fly over from London to Manhattan[1454.221] [1454.221][S01]on a 747 regularly, above the clouds,[1458.125] [1458.125][S01]I mean, what would you have said, that it would be mind boggling right?[1461.895] [1461.895][S01]And yet humanity’s done that incredibly,[1464.598] [1464.598][S01]and I don’t think we stop to think how amazing that is enough[1468.001] [1468.001][S01]because the other thing about the human brain[1469.603] [1469.603][S01]is that it’s incredibly adaptable - right?[1471.238] [1471.238][S01]As soon as something, you know we do something[1473.373] [1473.373][S01]and then it becomes kind of boring and mundane and it’s trivial, right?[1476.743] [1476.743][S01]But I always think about that when I’m taking a transatlantic flight,[1479.913] [1479.913][S01]about how have we with our monkey brains[1483.45] [1483.45][S01]managed to come up with these types of technologies, it’s unbelievable.[1488.856] [1488.856][S01]A 100 tonne of metal flying through the sky above the clouds,[1493.026] [1493.026][S01]so reliably, and so, if you think about that,[1495.562] [1495.562][S01]then, if we now build something like AGI[1498.165] [1498.165][S01]and enhance our own capabilities with this amazing tool,[1502.002] [1502.002][S01]then I feel like almost anything might be possible within the laws of physics[1506.74] [1506.74][S01]and perhaps even beyond the laws of physics[1508.475] [1508.475][S01]because with AI we might discover more about the laws of physics[1513.614] [1513.614][S01]or some holes or flaws in our understanding of the laws of physics,[1516.316] [1516.316][S01]so if you think about that and you extrapolate it a few hundred more years[1520.521] [1520.521][S01]with these kinds of technologies like AGI around,[1523.59] [1523.59][S01]and then what we might be able to build with AGI,[1526.426] [1526.426][S01]I think it could be truly incredible where we’ll be[1528.562] [1528.562][S01]- and I feel it will be something like[1530.163] [1530.163][S01]with the realisation of the true potential of humanity.[1533.901] [1536.904][S02] But however exciting the grand ambition of AGI[1540.073] [1540.073][S02]is, there is also a need to proceed with caution.[1544.545] [1544.545][S01] We’re cognisant of some of the technical questions around AGI -[1548.415] [1548.415][S01]making sure that they do exactly what we want,[1550.684] [1550.684][S01]how do we programme in our values, how do we specify our goals?[1555.856] [1555.856][S01]So these are all the theoretical and technical questions around AGI[1559.126] [1559.126][S01]that people like Nick Bostrom are worried about,[1561.161] [1561.161][S01]and so we and Shane leads our safety team[1565.499] [1565.499][S01]that works on a lot of these questions[1567.901] [1567.901][S01]from a research and technical perspective,[1569.87] [1569.87][S01]and I think there’s a lot more work that needs to be done there.[1572.306] [1572.306][S01]And I think that’s what we’re going to see over the next decade or two.[1575.876] [1578.879][S02] I mean even if you proceed with caution,[1581.615] [1581.615][S02]even if you act as safely as you possibly can based on the information[1585.152] [1585.152][S02]that you have in front of you at that moment,[1587.588] [1587.588][S02]you can’t really mitigate against bad actors,[1592.192] [1592.192][S02]you can’t stop someone else coming in and mucking it up for everyone.[1596.964] [1596.964][S02]Or can you?[1598.232] [1598.232][S01] I think there are ways of minimising that,[1600.667] [1600.667][S01]I think we’ll have to think carefully about that, like at the moment[1603.904] [1603.904][S01]we are at the stage where these systems are still quite nascent.[1607.14] [1607.14][S01]They can do impressive things like play Go,[1609.443] [1609.443][S01]but they are not yet properly general purpose[1612.446] [1612.446][S01]or you know you couldn’t use them for anything very dangerous,[1615.516] [1615.516][S01]or in the real world, so we’ve committed that we won’t do certain applications,[1620.454] [1620.454][S01]you know obviously like things like military and surveillance,[1623.09] [1623.09][S01]so which we think would be bad for society,[1625.392] [1625.392][S01]we don’t think AI should be applied to those things,[1627.361] [1627.361][S01]and we certainly wouldn’t do it ourselves,[1629.129] [1629.129][S01]but if you publish some great algorithms,[1631.698] [1631.698][S01]you have to think about the indirect impact of that if other people -[1635.102] [1635.102][S01]bad actors and so on around the world -[1636.87] [1636.87][S01]potentially use your algorithms for things[1639.64] [1639.64][S01]that you would not have agreed with, we have to use this time[1643.277] [1643.277][S01]now to think about what principles need to be put in place,[1646.713] [1646.713][S01]whether that’s carefully thought out regulation,[1649.917] [1649.917][S01]whether that is technical solutions to these problems,[1653.42] [1653.42][S01]mathematical proofs, more engineering solutions,[1657.024] [1657.024][S01]so like one of the projects we have here we call it[1659.793] [1659.793][S01]the virtual brain analytics project[1662.062] [1662.062][S01]and it’s inspired by what we do in neuroscience with FMRI machines[1666.366] [1666.366][S01]to brain scan people[1668.202] [1668.202][S01]while they’re doing tasks to see what parts of the brain light up[1671.205] [1671.205][S01]so we can understand what the brain is doing better,[1673.974] [1673.974][S01]and we should be doing the equivalent of that with our virtual brains,[1676.944] [1676.944][S01]our artificial neural networks,[1679.546] [1679.546][S01]so I think there’s both sort of behavioural,[1681.648] [1681.648][S01]there’s experimental understanding[1683.317] [1683.317][S01]and then there’s mathematical understanding of these systems,[1685.686] [1685.686][S01]and we should do all of those,[1686.954] [1686.954][S01]and what I’m hoping is that we will eventually in the next few years[1689.723] [1689.723][S01]have a much better understanding of these systems than we have today[1693.46] [1693.46][S01]and we may even have some mathematical proofs about[1696.997] [1696.997][S01]if you want to limit a system in a certain way,[1699.233] [1699.233][S01]what do you need to do? What components do you require?[1701.935] [1701.935][S01]Also you know we have to think about publications[1704.238] [1704.238][S01]and other stuff that we talked about earlier -[1706.173] [1706.173][S01]whether um this free exchange of information and knowledge[1710.043] [1710.043][S01]is okay even under situations[1712.145] [1712.145][S01]where there are potentially dangerous applications.[1715.716] [1715.716][S01]And where I would take our lead from this -[1717.15] [1717.15][S01]this is not the first time again this has happened -[1719.286] [1719.286][S01]you know this has happened a lot in biology[1721.255] [1721.255][S01]with designing synthetic biology and viruses, embryology, now with CRISPR[1727.06] [1727.06][S01]so biology actually has a long-standing multi-decade experience[1732.466] [1732.466][S01]of coming together with wider society in the scientists and regulatory bodies[1737.171] [1737.171][S01]and figuring out what are the rules of the road that is safe for everybody,[1741.041] [1741.041][S01]and I think that’s the kind of coordination[1743.41] [1743.41][S01]we’re going to have to have over the next decade.[1746.013] [1746.013][S02] Where do you see DeepMind as sitting in terms[1748.048] [1748.048][S02]of the brand tech industry in terms of ethics and safety,[1752.452] [1752.452][S02]are people following your lead?[1754.254] [1754.254][S01] We’re only one company but we are the biggest[1757.324] [1757.324][S01]probably group anywhere in the world,[1758.825] [1758.825][S01]and we you know are definitely one of the world leaders and acknowledge as so.[1762.796] [1762.796][S01]So that gives a powerful platform to set an example.[1767.467] [1767.467][S01]We can create initiatives, like we helped create the Partnership on AI,[1771.071] [1771.071][S01]which is a cross-industry initiative[1772.539] [1772.539][S01]to talk about some of these issues in products,[1775.576] [1775.576][S01]and we’ve also sponsored a lot of academic groups,[1779.646] [1779.646][S01]we’ve given them money for post-docs[1781.248] [1781.248][S01]and other things at arm’s length to study this,[1783.584] [1783.584][S01]we have close contacts with a lot of the institutes[1786.186] [1786.186][S01]that work on these things, like the Future of Humanity Institute[1788.322] [1788.322][S01]which is just down the road in Oxford,[1789.957] [1789.957][S01]we talk with them all the time, and we ourselves as leaders at DeepMind,[1794.161] [1794.161][S01]we’ve always talked about ethics, and I’ve always had that in mind,[1797.297] [1797.297][S01]and the reason we have from the beginning of DeepMind[1799.066] [1799.066][S01]is that we plan for success,[1801.134] [1801.134][S01]so if we’re planning all these ambitious things for AI to do,[1803.937] [1803.937][S01]then we also need to think carefully about what would that mean,[1807.207] [1807.207][S01]and so we’re doing a lot of things behind the scenes,[1810.077] [1810.077][S01]and I think by sort of being an example,[1812.746] [1812.746][S01]that will influence everybody else I’m hoping,[1815.282] [1815.282][S01]given that we’re in the lead technologically, so, for the moment.[1819.319] [1819.319][S01]That’s why I think also it’s very important we have a technical lead,[1822.222] [1822.222][S01]because why would anyone listen to what you have to say on the ethics front[1826.26] [1826.26][S01]if you’re not one of the leaders technologically right then[1828.695] [1828.695][S01]you could just be any group saying that.[1830.764] [1830.764][S01]So I think to have a seat at the table, whether that’s a country,[1834.201] [1834.201][S01]or a company, or even individuals, you need for credibility purposes,[1838.138] [1838.138][S01]you have to be close to the forefront of the technical side itself.[1841.975] [1841.975][S02] Where does the public fit into all of this,[1843.81] [1843.81][S02]do you need to have the trust in you as a technology company,[1848.282] [1848.282][S02]or do you need to be on board at all - can you kind of do it without them?[1851.118] [1851.118][S01] No, it’s really critical that this is discussed at large with society[1856.557] [1856.557][S01]and everybody and the general public, and I think they need to engage,[1859.96] [1859.96][S01]I mean the problem is that a lot of these technical things are very complex,[1863.13] [1863.13][S01]and you need PhDs to understand them and so on.[1865.732] [1865.732][S01]But some of the fundamentals are quite easy to understand,[1868.368] [1868.368][S01]and what you really need to understand is the consequences of it[1871.839] [1871.839][S01]and then society has to decide as I said before about how these things get used,[1876.91] [1876.91][S01]and how the benefits accrue to different people in society[1880.247] [1880.247][S01]and make sure that is fair,[1881.849] [1881.849][S01]and I think that’s the key thing is to engage with the public now[1885.319] [1885.319][S01]and try and educate them about some of the complexities of the technology,[1889.256] [1889.256][S01]but also the implications,[1890.557] [1890.557][S01]and this is partly what things like this podcast series is about,[1893.126] [1893.126][S01]but also we’ve done a lot of public engagement[1895.395] [1895.395][S01]with the Royal Society, the ‘You and AI’ series that we did last year[1900.534] [1900.534][S01]and I think a lot more of that has to be done.[1903.036] [1903.036][S02] Are you optimistic about the future?[1905.606] [1905.606][S01] Yeah I’m very optimistic about the future,[1907.241] [1907.241][S01]but the reason I’m optimistic is[1908.675] [1908.675][S01]because I think AI is coming down the road[1912.112] [1912.112][S01]and I feel like if we build it in the right way[1914.481] [1914.481][S01]and we deploy it in the right way for the benefit of everyone,[1917.451] [1917.451][S01]I think it’s going to be the most amazing transformative technology[1921.388] [1921.388][S01]that humanity’s ever invented.[1923.257] [1923.257][S01]And I would be quite pessimistic[1925.092] [1925.092][S01]about some of the problems we are facing as a society,[1928.562] [1928.562][S01]like climate change and sustainability or inequality.[1932.199] [1932.199][S01]I think these are going to get exacerbated in the next few years.[1935.269] [1935.269][S01]And I’d be pessimistic about our ability to solve[1938.372] [1938.372][S01]that if there wasn’t something like AI on the way,[1941.074] [1941.074][S01]in the near future.[1942.376] [1942.376][S01]I think in order to solve some of these big problems,[1945.112] [1945.112][S01]challenges society has,[1947.214] [1947.214][S01]we either need an exponential improvement in human behaviour,[1950.317] [1950.317][S01]more cooperation, less selfishness, more collaboration,[1954.254] [1954.254][S01]or we need an exponential improvement in technology,[1957.624] [1957.624][S01]and unfortunately the way politics is going right now,[1960.093] [1960.093][S01]I don't really see much evidence of the former[1962.963] [1962.963][S01]and um you know we don’t seem to be able to get our act together[1966.6] [1966.6][S01]globally to do something about climate[1968.635] [1968.635][S01]anywhere near fast enough to deal with the problem[1971.505] [1971.505][S01]and so I think we have to double down on a technical solution.[1975.075] [1975.075][S01]At least um we should try both,[1977.11] [1977.11][S01]but I think we need some technology bets.[1980.18] [1983.417][S02] It’s hard not to be inspired[1984.885] [1984.885][S02]by Demis’s insatiable thirst for knowledge,[1987.855] [1987.855][S02]and find yourself drawn in by his positive view of the future,[1991.592] [1991.592][S02]because if I’m honest, I’m normally quite skeptical.[1995.596] [1995.596][S02]I don’t have a lot of time for overly optimistic marketing talk,[1999.266] [1999.266][S02]and one of my favourite hobbies[2000.934] [2000.934][S02]is to roll my eyes at self-titled futurists at tech conferences,[2006.573] [2006.573][S02]but you know over the last 12 months that I’ve spent hanging around here,[2009.71] [2009.71][S02]I’ve come to the conclusion that there really[2011.645] [2011.645][S02]is something quite special going on at the cutting edge of AI.[2016.817] [2016.817][S02]After 50 years of quite slow progress,[2019.82] [2019.82][S02]it really feels as though the field is finally beginning to deliver.[2024.157] [2024.157][S02]The problems that everyone thought were completely out of reach[2026.894] [2026.894][S02]only a few short years ago are tumbling one by one.[2031.765] [2031.765][S02]And the science is moving forward at a blistering pace,[2036.203] [2036.203][S02]both here and at research labs all around the world.[2040.174] [2040.174][S02]And the questions that people are working on are profound and important.[2045.779] [2045.779][S02]Including some of the potential pitfalls[2048.248] [2048.248][S02]and ethical concerns of this kind of technology.[2052.019] [2052.019][S02]So with all of that in mind,[2053.82] [2053.82][S02]I think I’m going to join Demis in being optimistic about the future,[2058.425] [2058.425][S02]and the potential of AI to be a real positive force for good.[2063.73] [2063.73][S02]But don’t just take my word for it, as we’ve said throughout this podcast,[2067.067] [2067.067][S02]we hope to inspire you on your own AI journey,[2070.304] [2070.304][S02]maybe even by finding the answers[2072.339] [2072.339][S02]to some of the biggest questions there are.[2075.209] [2077.344][S01] I’m an entrepreneur second, I’m a scientist first.[2079.88] [2079.88][S01]It’s just that this was the right vehicle to make this happen,[2082.616] [2082.616][S01]and it seems to have born out,[2084.284] [2084.284][S01]but if I could have made it happen in academia,[2086.987] [2086.987][S01]I would have just done it in academia,[2088.322] [2088.322][S01]it just wasn’t possible under the constraints academia has.[2091.391] [2091.391][S01]And that’s why we also sold the company to Google[2093.46] [2093.46][S01]because it was about what accelerates the mission and the science,[2096.964] [2096.964][S01]and ultimately that’s what I want to do with my life[2099.099] [2099.099][S01]is I want to understand what’s going on here in the universe.[2102.636] [2102.636][S01]Both inside here in the brain[2104.838] [2104.838][S01]and externally out there in the universe.[2107.14] [2107.14][S01]And I guess that’s what’s always driven me[2109.176] [2109.176][S01]is this deep desire to understand what seems to me[2113.08] [2113.08][S01]to be an incredibly interesting and fascinating mysteries[2119.219] [2119.219][S01]that are going all around us and I don’t really understand[2121.889] [2121.889][S01]why more people don’t think about that all the time -[2124.691] [2124.691][S01]I can barely sleep because I’m just fascinated[2127.694] [2127.694][S01]and also troubled by the things around us[2131.999] [2131.999][S01]that we seemingly don’t understand -[2133.534] [2133.534][S01]all the big questions, you know the meaning of life -[2135.802] [2135.802][S01]how did the universe start, what is consciousness,[2137.538] [2137.538][S01]all these questions, which I feel like a blaring klaxon in my mind[2141.408] [2141.408][S01]that I would like to understand and my attempt at doing[2144.578] [2144.578][S01]that is to build AI first.[2147.114] [2150.784][S02] If you would like to find out more about some of the things[2153.52] [2153.52][S02]that we’ve talked about in this episode,[2155.522] [2155.522][S02]or explore the world of AI research beyond DeepMind,[2159.226] [2159.226][S02]you’ll find plenty of useful links in the show notes for each episode,[2162.896] [2162.896][S02]and if there are stories or resources[2164.531] [2164.531][S02]that you think other listeners would find helpful,[2166.9] [2166.9][S02]than let us know -you can message us on Twitter,[2169.436] [2169.436][S02]or email the team at podcasts@deepmind.com.[2173.14] [2173.14][S02]You can also use that address[2174.608] [2174.608][S02]to send us your questions or feedback on the series.[2178.045] [2178.645][S02]DeepMind the podcast has been a Whistledown production,[2181.815] [2181.815][S02]binaural sound recordings were by Lucinda Mason-Brown.[2185.319] [2185.319][S02]The music from this series has been especially composed by Eleni Shaw.[2189.189] [2189.189][S02]Producers were Amy Racs and Dan Hardoon.[2192.125] [2192.125][S02]The Senior Producer was Louisa Field and the series editor was David Prest.[2196.763] [2199.9][S02]I’m Hannah Fry, thank you for listening.[2202.236]"}