Datasets:
willtheorangeguy
/
2023-Changelog-Interviews-Transcripts

Dataset card Data Studio Files
xet
 Community
text
stringlengths
0
2.46k
[5333.54 --> 5335.10] So if there's data, I don't want it to memorize.
[5335.56 --> 5338.56] Is there a way that I can know ahead of time what's going to be memorized?
[5338.64 --> 5345.46] That's the paper that we have that we actually just released on archive about forecasting what is going to be memorized before you actually train the model.
[5345.88 --> 5349.42] Is that to make it less black box, more like you deploy it and you don't know what it can do?
[5349.42 --> 5358.02] So that you can sort of understand, okay, here's the data, here's how it's trained to sort of have a more clarity of what the box actually contains versus this black box.
[5358.18 --> 5359.12] Is that why that's important?
[5359.22 --> 5361.42] That is what the field of interpretability is about in general.
[5361.58 --> 5372.10] And I would say kind of building on that, that what my research is about in particular is not just opening up that black box and looking inside and understanding what the model is actually doing,
[5372.42 --> 5378.18] but understanding where it came from and how we can build boxes that are more transparent from the ground up.
[5378.50 --> 5379.36] Predictable maybe even?
[5379.56 --> 5379.70] Yeah.
[5380.02 --> 5380.24] Yeah?
[5380.66 --> 5385.10] Because, I mean, that's one of the fears is, you know, especially with like Bing.
[5385.54 --> 5385.76] Yeah.
[5385.76 --> 5388.42] When they put that out there, I think what, it threatened the person?
[5388.54 --> 5391.56] Like there was some sort of like threat on humanity essentially.
[5391.84 --> 5396.72] And it's like you deploy this thing out into the world and you don't understand what they can actually do.
[5396.82 --> 5400.50] Is that to be more predictable, more controlled to some degree?
[5400.50 --> 5400.58] Absolutely.
[5401.58 --> 5401.74] Sorry?
[5402.10 --> 5403.26] And even designable.
[5403.46 --> 5405.58] Like say, well, forget these things, remember these things.
[5405.76 --> 5405.96] Yeah.
[5407.40 --> 5409.08] Designability is a really big component.
[5409.08 --> 5411.14] I think that's going to become huge in the future.
[5411.50 --> 5411.58] Right.
[5411.58 --> 5414.90] And really it hasn't been studied primarily because people haven't had the tools.
[5415.58 --> 5418.46] Very few model suites have intermediate checkpoints at all.
[5419.26 --> 5423.38] A lot of publicly released models weren't trained on publicly released data sets.
[5423.38 --> 5428.52] Or if they were trained on publicly released data sets, they didn't tell you what order it was trained on.
[5428.90 --> 5430.32] And it turns out that matters a lot.
[5431.12 --> 5433.04] What it saw early in training, what it saw late in training.
[5433.46 --> 5445.74] And so there's really a huge reproducibility issue in terms of under, like if you want to dig in and really understand how data by data, data point by data point, the model is learning to behave.
[5445.74 --> 5448.26] You need to be able to basically fully reproduce the training.
[5448.50 --> 5451.74] Not actually, because you're not going to spend a couple hundred thousand dollars.
[5452.24 --> 5457.50] But at least in principle, you need to be able to inspect individual data points, know when it's going to get loaded, understand kind of how it works.
[5457.50 --> 5464.02] And this is something that we've put a huge amount of resources into, both on the training side as well as kind of on the engineering side.
[5464.12 --> 5469.32] It was not easy, but you can actually reproduce our model training exactly.
[5469.60 --> 5482.88] So if you take the code base that we used to train these Pythia models and you pick a checkpoint and you load that checkpoint and you resume training from that checkpoint, you will end up with the same fully trained model that we did.
[5483.22 --> 5483.62] Exactly.
[5484.16 --> 5484.80] That's important.
[5485.10 --> 5485.82] That is really important.
[5485.82 --> 5492.08] It's important because if you want to understand how to design models, you need to understand how they're changing over the course of training.
[5492.60 --> 5499.62] And that is really persnickety and really sensitive to a lot of implementation specific details that tend to not get released.
[5500.26 --> 5508.72] How far in the future do you think, since you're at the training level, you're like the ground level of if this is the eureka moment for humanity.
[5509.00 --> 5509.18] Yeah.
[5509.26 --> 5509.44] Right.
[5509.74 --> 5514.70] How far in the future do you think and do you have fear, trepidation, hope?
[5514.70 --> 5516.94] Like where will this take us as humanity?
[5517.40 --> 5518.36] I really don't know.
[5519.06 --> 5529.78] My kind of attitude is that the recent, like there was a really big paradigm shift in 2020 with the release of G2B3 and the aggressive focus on scaling.
[5529.78 --> 5537.00] And people really changed their attitudes towards like how to design language models and kind of how they can be used and what they can be used for.
[5537.28 --> 5539.78] In a sense, we got really lucky because it wasn't that dangerous.
[5540.18 --> 5543.48] You know, there were a lot of fears about what G2B3 could do.
[5543.48 --> 5547.16] And by and large, it turned out to be pretty safe.
[5547.60 --> 5551.66] There wasn't all that much harm done and a lot of the fears turned out to be not come to fruition.
[5552.40 --> 5559.38] And, you know, kind of looking forward, I think the really important thing to think about is we obviously can't predict the next paradigm shift.
[5559.38 --> 5569.44] But building tools that allow us to hopefully more readily adopt and adapt and respond to future paradigm shifts in large scale AI.
[5570.06 --> 5574.48] So that, you know, one day there probably will be something that gets developed that is dangerous.
[5574.66 --> 5577.16] And we want to be able to be, I guess, ready for that.
[5577.56 --> 5577.64] Yeah.
[5577.96 --> 5578.12] Yeah.
[5578.68 --> 5578.94] Cool.
[5579.02 --> 5580.24] Well, what are some touch points?
[5580.46 --> 5586.02] People who are interested in what you're up to, want to help out, want to give money, want to read more?
[5586.20 --> 5587.20] Where can people connect with you?
[5587.20 --> 5590.36] So the best place to connect with us is our Discord server.
[5590.88 --> 5596.60] We are a research institute, but we actually operate basically entirely in the public view.
[5597.06 --> 5602.70] We're distributed all over the world and we do our research in a public Discord.
[5602.92 --> 5608.44] And anyone can join, anyone can drop in, read about what we're getting up to, hang out with us, chat with us about AI.
[5608.84 --> 5612.54] So our Discord server is discord.gg slash EleutherAI.
[5613.10 --> 5616.08] There's also a link on our website, which is Eleuther.AI.
[5616.08 --> 5616.48] Nice.
[5616.48 --> 5616.88] Shockingly.
[5618.06 --> 5619.60] We'll link it up to the show notes for sure.
[5619.98 --> 5620.12] Yeah.
[5620.90 --> 5624.10] And yeah, we're always happy to take on more volunteers.
[5624.86 --> 5628.82] We have a small professional staff and a large number of volunteers that help out as well.
[5628.94 --> 5629.70] How small is small?
[5631.42 --> 5632.68] Like 10 full-time employees.
[5632.88 --> 5633.06] Okay.
[5633.06 --> 5635.94] And if they go to the Discord server, what can they do there?
[5636.00 --> 5637.42] What can they expect from the Discord server?
[5637.62 --> 5639.18] Like you're there, others are there.
[5639.54 --> 5639.76] Yeah.
[5639.84 --> 5641.86] So you can chat about AI.
[5642.10 --> 5646.84] We have a bunch of discussion channels where people talk about kind of cutting edge trends in artificial intelligence.
[5646.84 --> 5655.12] Honestly, like I don't really follow AI publication news anymore because I just follow my Discord server and everything that's important shows up for me.
[5655.14 --> 5655.46] There you go.
[5655.72 --> 5657.32] Which is a really nice place to be.
[5657.66 --> 5658.64] But you can talk with us.
[5658.66 --> 5659.60] You can talk with other researchers.
[5659.76 --> 5663.18] We have a large amount of researchers at the cutting edge of AI.
[5663.18 --> 5667.08] I can't count the number of times that someone's posted a paper and been like, hey, this is really cool.
[5667.58 --> 5669.02] Like, does anyone know anything about this?
[5669.04 --> 5671.00] And someone just like tags the guy who wrote the paper.
[5671.28 --> 5672.12] That happens all the time.
[5672.22 --> 5684.98] We have people from OpenAI, Anthropic, Meta, like all the major labs who come, DeepMind, come in and chat about language models, give advice, give perspectives on research and talk about kind of how things are going.
[5685.62 --> 5689.16] You can also get involved with ongoing research projects.
[5689.16 --> 5698.78] So we have a dozen-ish ongoing research projects ranging from learning to train, figuring out how to train better language models to training language models in other languages.