id stringlengths 66 88 | text stringlengths 46 1.7k | duration float64 2.32 30 | dnsmos float64 2.8 3.65 | sampling_rate int64 24k 24k | path stringlengths 79 101 | audio_content_token_indices listlengths 58 749 | audio_global_embedding listlengths 128 128 | audio_token_len int64 58 749 | audio_duration float64 2.32 30 |
|---|---|---|---|---|---|---|---|---|---|
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758572 | [S1] And, okay, not unrelated to them being undergrads or not, did you, like, how often does it happen that you get into, like, hot waters, like, you know, that there- [S2] [LAUGHS] [S1] ... You know, in security research, there are implicate, there are n- national defense implications, there are legal implications and so on. Like, how do you navigate that space? And how often does it happen that you're like, "Oops, I- I hope no- [S2] [LAUGHS] [S1] ... No one noticed this." [S2] [LAUGHS] | 26.12 | 3.135624 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758572.mp3 | [
5672,
5665,
11502,
9199,
11757,
12582,
8206,
8962,
4325,
10793,
8306,
9432,
11827,
7990,
8804,
7147,
11750,
2255,
8715,
10734,
7923,
7989,
1762,
3617,
9242,
8354,
11866,
8848,
8795,
6250,
8803,
8819,
9346,
12482,
9800,
9867,
5329,
9736,
123... | [
0.6153590679168701,
-5.642721176147461,
-0.02102169208228588,
-0.04485280066728592,
0.451404869556427,
-0.0002931156486738473,
0.35438841581344604,
-0.02051481232047081,
-0.07675450295209885,
-3.869802951812744,
0.17484770715236664,
0.04388783499598503,
0.07453946769237518,
0.1233540847897... | 653 | 26.12 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758576 | [S1] So for security in general, there's, there's so many, I mean, and there's, I'm sure there's a, a, two dozen YouTube channels that could probably hook you up with like incredible, um, so maybe I, I, we can send some and, and link some of those below or something. [S2] Mm-hmm. [S1] Um, I wish I could say that there was like this amazing, | 15.12 | 3.115102 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758576.mp3 | [
4651,
3619,
6802,
6122,
5520,
6108,
10539,
8418,
7885,
2766,
3560,
8184,
5620,
8069,
6634,
9160,
6116,
7317,
12129,
10144,
9051,
7969,
8010,
8257,
11611,
9474,
12204,
11286,
9995,
7106,
8607,
11638,
11412,
8053,
5963,
10539,
10419,
5764,
29... | [
0.7477782964706421,
-5.793813705444336,
-0.03255368396639824,
-0.14014369249343872,
0.5733439326286316,
-0.13353650271892548,
0.42940351366996765,
0.0052240705117583275,
0.03986523672938347,
-3.873082399368286,
0.24684679508209229,
0.01505394745618105,
-0.30064356327056885,
0.3057340085506... | 378 | 15.12 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758577 | [S1] Uh, AI, censorship, I want to say like censorship resource space where everyone can come to and, and learn how to apply AI to these techniques. [S2] Mm-hmm. [S1] Uh, something like that doesn't quite exist, but there are great, there are great resources for learning about what censorship is happening in the world. Um, so something like Ooni, uh, Ooni is O-O-N-I, it's the Open Observatory of Network Interference. [S2] Mm-hmm. [S1] It's a spin out from the Tor team that, uh, monitor censorship all over the world. Um, you can pull up their website later, but the, um, | 29.56 | 2.867807 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758577.mp3 | [
3176,
4637,
9967,
2031,
3958,
4398,
3886,
2421,
4525,
1820,
4705,
9321,
5737,
5160,
11493,
12261,
12741,
9158,
9207,
3965,
3390,
9022,
10166,
12719,
12694,
12102,
11590,
12613,
9541,
8582,
4565,
9893,
12542,
12757,
12742,
7637,
4826,
6250,
... | [
0.8952295184135437,
-5.426708698272705,
-0.10216163098812103,
-0.2053033858537674,
0.6408955454826355,
-0.2114247977733612,
0.34259992837905884,
0.040326494723558426,
0.07895620167255402,
-3.9329581260681152,
0.27965256571769714,
0.11884496361017227,
-0.29441818594932556,
0.264683723449707... | 739 | 29.56 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758579 | [S1] Excellent. Kevin, thank you so much for being here and bringing this a bit, a bit closer. I, I know more. I hope everyone else does too now. Uh, yeah, thanks. [S2] Thanks so much for having me. This has been a blast. [S1] Excellent. [S2] Super appreciate it. [S1] Bye. | 14.48 | 3.026709 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758579.mp3 | [
8784,
7449,
12387,
12726,
12653,
9321,
11620,
993,
896,
3239,
2670,
284,
4702,
9747,
9864,
4065,
4189,
9778,
5809,
8267,
8273,
9881,
12636,
7648,
5022,
11756,
11250,
3299,
6511,
2342,
4174,
6223,
4173,
5984,
3956,
12734,
12125,
12316,
8987,... | [
0.8978511095046997,
-5.755830764770508,
-0.1428508460521698,
-0.008534681051969528,
0.5513747334480286,
-0.08189789205789566,
0.4930986166000366,
-0.03071877732872963,
0.009897662326693535,
-3.852268934249878,
0.09511154145002365,
-0.03313180059194565,
-0.045663170516490936,
0.079488374292... | 362 | 14.48 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818472 | [S1] Welcome everyone. Today I have with me Armin Aghajanian and I've practiced that name 10 seconds ago and I th- I think I got it down. Uh, Armin is the first author of the CM3 paper. Welcome, uh, Armin, to the channel. [S2] Thank you for having me. [S1] So I, I saw this paper and of course you have, like, uh, some big names here. There's lots of authors, there's Facebook AI Research, um, but still, like, given all of that, it was still | 29.8 | 3.467483 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818472.mp3 | [
1081,
7261,
113,
8694,
11066,
5339,
9897,
12265,
8115,
8221,
11765,
11766,
3372,
2971,
11222,
10054,
627,
5430,
5943,
829,
4115,
3590,
3616,
11273,
12248,
4500,
7233,
3550,
12725,
12717,
12252,
12228,
11717,
6607,
9322,
11873,
8873,
10862,
... | [
1.1598042249679565,
-5.3291401863098145,
-0.11406993120908737,
-0.013326525688171387,
0.47703495621681213,
-0.15143820643424988,
0.2998678386211395,
-0.0383530929684639,
0.24586080014705658,
-3.8631105422973633,
0.37281569838523865,
0.02483963966369629,
0.15139240026474,
0.0073722735978662... | 745 | 29.8 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818473 | [S1] I mean, the goal here was kind of to have a single multimodal model that can do everything. [S2] Yeah. [S1] Um, image generation, image captioning, uh, image infilling to, uh, to even pure text tasks- [S2] Yeah. [S1] ... like summarization, but mostly focusing on this zero-shot setting specifically- [S2] Mm-hmm. [S1] ... this blocking setting. | 18.32 | 3.078185 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818473.mp3 | [
9248,
5672,
1684,
3109,
9428,
12674,
11285,
8278,
3669,
3601,
4115,
5640,
8171,
10981,
6250,
11611,
7981,
10354,
10347,
8342,
4443,
10152,
10131,
11733,
5478,
2900,
108,
2870,
3922,
2897,
4258,
12130,
3506,
12030,
11413,
9071,
3107,
8520,
4... | [
0.5557155013084412,
-5.83724308013916,
-0.13952748477458954,
-0.00207767472602427,
0.8066749572753906,
0.008885221555829048,
0.4827260673046112,
-0.0052366177551448345,
-0.12891729176044464,
-3.8470988273620605,
0.3169809877872467,
-0.4485393166542053,
-0.24942505359649658,
0.1710622906684... | 458 | 18.32 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818474 | [S1] ... in other modalities than text. [S2] Yeah. [S1] Um, this, this goes even further. This is multimodal, uh, there have been a lot of other approaches to multimodal. There is like this, this Ru, Rudolf, Rudolph even model, I don't know if you've seen that, that goes like image to text to image and so on. And they all work, let's say, with very cleaned up data. It's, it's very, you know, I want text, I want images that go with the text, which makes sense, right? How, | 29.16 | 3.347572 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818474.mp3 | [
4713,
5160,
11046,
11716,
9606,
10829,
5022,
11314,
9087,
2877,
4207,
2915,
11054,
5452,
6172,
7789,
10486,
8962,
8623,
11701,
10738,
7789,
8092,
11253,
8808,
10704,
6642,
11659,
12165,
4570,
2960,
529,
2706,
11638,
12148,
10317,
6697,
11073,... | [
1.2696338891983032,
-5.412273406982422,
-0.10436766594648361,
-0.04894949123263359,
0.29533761739730835,
-0.08554740995168686,
0.3178137540817261,
-0.059653766453266144,
0.20161062479019165,
-3.83779239654541,
0.3827197849750519,
0.16785800457000732,
0.12983262538909912,
0.0471718907356262... | 729 | 29.16 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818476 | [S1] ... to be used, um, in HTML. So after DALI came out, we thought, okay, um, there are some fundamental restrictions with DALI. So the first one being the causal, uh, approach. So they train a decoder-only left-to-right model. [S2] Mm-hmm. [S1] So in some sense, you can't do things like generate the text given the image, right? Just because of the positioning of the image. [S2] Yeah. [S1] So it's on the right side of the input, right? You can't really do image infilling either, um, which means conditioning on both the prefix and postfix of the image. | 29.24 | 3.190094 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818476.mp3 | [
5680,
3105,
8001,
8152,
6118,
5652,
5737,
4068,
10054,
12613,
7494,
4933,
4485,
4682,
4761,
7377,
12483,
9988,
9030,
3399,
719,
774,
4293,
3472,
3464,
1152,
5760,
4056,
9458,
9377,
9813,
9839,
4654,
92,
1627,
4123,
2729,
5657,
6822,
9268,... | [
0.48125365376472473,
-5.657584190368652,
-0.03706926479935646,
-0.0425427071750164,
0.9012011289596558,
-0.04911351948976517,
0.5302054286003113,
0.049074072390794754,
-0.10032369941473007,
-3.9167590141296387,
0.260729044675827,
-0.30587658286094666,
-0.27521365880966187,
0.20312340557575... | 731 | 29.24 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818477 | [S1] Exactly. Yeah. So those were kind of the, the first weaknesses, uh, that we saw there. Uh, the approach was very clever though, right? So pretty much taking continuous data, discretizing it, and just doing sequence modeling, it seems to work very, very well. [S2] Mm-hmm. [S1] So the idea went that we could kind of combine the two from the HTLM paper, which was that, uh, you know, document structure | 22.16 | 3.067929 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818477.mp3 | [
6689,
5800,
4108,
7390,
12435,
897,
4447,
4918,
3756,
9449,
674,
1117,
6184,
1539,
7299,
12742,
10717,
6134,
3885,
9377,
11996,
12597,
8877,
12589,
12532,
12581,
12525,
11445,
9963,
6233,
6048,
8144,
8043,
8049,
10477,
5270,
3738,
6765,
564... | [
0.5184067487716675,
-5.809015274047852,
-0.08292531222105026,
-0.09293520450592041,
0.8392511606216431,
-0.0328715518116951,
0.5903341770172119,
0.014271043241024017,
0.03045055642724037,
-3.881147861480713,
0.33530113101005554,
-0.27042171359062195,
-0.2628275752067566,
0.2077129036188125... | 554 | 22.16 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818479 | [S1] artificially increased the amount of images that we have available in the documents. [S2] Yeah. [S1] You actually, you look, I think we have 25 million unique images, um, | 9.8 | 3.040377 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818479.mp3 | [
6810,
9377,
5933,
7998,
2972,
3268,
5534,
5852,
8682,
8696,
948,
10086,
4298,
7081,
7648,
355,
8359,
2771,
8543,
10117,
4803,
6756,
10142,
8854,
7206,
3946,
4025,
283,
12102,
7558,
8664,
3400,
6209,
3878,
6454,
11343,
10965,
6590,
5886,
7... | [
0.67353355884552,
-5.996739864349365,
-0.16343024373054504,
0.059912871569395065,
0.8438705801963806,
0.0023237508721649647,
0.5496697425842285,
-0.04437122493982315,
-0.13104216754436493,
-3.8667595386505127,
0.16805978119373322,
-0.19599340856075287,
-0.23986268043518066,
0.1326616555452... | 245 | 9.8 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818481 | [S1] This is true, but there are pros and cons to this. So encoder, decoder-only architectures, uh, they're really good for fine-tuning, but they're not so good for prompting, is at least what we noticed. Um, and also training them is a little bit more non-trivial. So decoder-only models are quite nice 'cause you get per-token generation, uh, so you pretty much generate every token, uh, for the source. [S2] Yeah. [S1] Um, whereas for encoder-decoder, most of the time you're generating, I think, like 15% is with BERT and like BART. | 27.68 | 3.225371 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818481.mp3 | [
5672,
5145,
3202,
8117,
6115,
8144,
10668,
4050,
5512,
12241,
6632,
2964,
8025,
10955,
8353,
8054,
4198,
7880,
11198,
10525,
5589,
1109,
10602,
3570,
8094,
10421,
9494,
3456,
3294,
6166,
12579,
7152,
2877,
3389,
8742,
4363,
3553,
3657,
5968... | [
0.3670370876789093,
-5.732509136199951,
-0.020578933879733086,
-0.20617394149303436,
0.990156888961792,
-0.0943976640701294,
0.5420764684677124,
0.059657756239175797,
-0.021281380206346512,
-3.8618204593658447,
0.28670451045036316,
-0.28389695286750793,
-0.20712172985076904,
0.222584977746... | 692 | 27.68 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818482 | [S1] Or like Roberto do. It's all around that 15%. [S2] Mm-hmm. [S1] Um, so most of the times you have to go through the data multiple times. Um, for some reason they don't prompt super well. Um, | 10.32 | 2.892461 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818482.mp3 | [
4714,
8216,
3765,
2973,
5213,
2990,
7590,
9818,
6379,
174,
2899,
2838,
3108,
5482,
5423,
3020,
771,
3277,
4398,
4162,
3328,
9486,
4429,
149,
6823,
6337,
400,
8439,
7868,
9045,
3356,
2949,
12222,
12470,
8845,
4030,
7478,
6120,
6584,
5023,
... | [
0.5420716404914856,
-6.0278544425964355,
-0.10966137796640396,
0.0032979529350996017,
0.8199151158332825,
-0.06631191074848175,
0.6529113054275513,
0.014397927559912205,
-0.04051826894283295,
-3.816727638244629,
0.061538487672805786,
-0.18033035099506378,
-0.21556255221366882,
0.1689722985... | 258 | 10.32 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818483 | [S1] And the kind of the other big thing is if you want to do score-based prompting, it's kinda, it's kinda hard to do with the encoder-decoder-only architecture, right? Like if you wanna ask what's the log probability of this sequence- [S2] Yeah. [S1] ... with the masked language model, it's kinda tough to do, right? Um, so we knew that we wanted to go kind of this decoder-only route, so we introduced this new, uh, objective that we called causal masking. Um, and so the idea behind causal masking, uh, if you wanna scroll down- [S2] Yeah. [S1] ... I think there's a, uh, there's a figure there. | 29.88 | 2.947927 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818483.mp3 | [
2608,
8225,
11117,
12196,
7693,
5540,
6756,
12194,
3571,
11698,
8986,
11942,
8082,
6950,
10869,
8053,
11932,
8148,
8532,
2580,
8480,
8670,
6615,
11931,
7273,
2897,
12708,
11860,
12061,
12759,
2896,
4419,
12645,
873,
10073,
8471,
2676,
8502,
... | [
0.5756643414497375,
-5.684328079223633,
-0.023851949721574783,
-0.11806756258010864,
0.9309369921684265,
-0.109146349132061,
0.583043098449707,
0.04692626744508743,
-0.09958599507808685,
-3.8869049549102783,
0.2996024489402771,
-0.296113520860672,
-0.25113645195961,
0.21343764662742615,
... | 747 | 29.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818484 | [S1] Uh, which is actually a really, really big thing to have, right? Um, and the other thing that we noticed is that depending on the setting, prompting versus fine-tuning, the size of the mask is really important. So for fine-tuning, localized information is really important. [S2] Mm-hmm. [S1] Um, so you wanna have a lot of small masks. For prompting, we saw kind of the opposite, which is you wanna have very, very few masks, but they can be very long. [S2] Mm-hmm. [S1] Um, so the strategy that we use here is for every document, we sample from a Poisson distribution center in round one. | 29.88 | 2.951007 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818484.mp3 | [
3112,
4123,
11382,
3511,
4334,
3822,
741,
1124,
3123,
2805,
7307,
12672,
8670,
3927,
832,
6528,
7259,
11645,
5629,
9445,
12770,
10192,
2779,
9494,
9540,
6615,
3502,
2774,
3691,
347,
8150,
10924,
2707,
12108,
9477,
178,
339,
11166,
3230,
3... | [
0.4503988027572632,
-5.641013145446777,
0.015756880864501,
-0.12804768979549408,
0.9664720892906189,
-0.12438022345304489,
0.6204627156257629,
0.05036792531609535,
-0.04621398448944092,
-3.9272537231445312,
0.31301596760749817,
-0.2712109684944153,
-0.267521470785141,
0.2755862772464752,
... | 747 | 29.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818485 | [S1] Uh, so, you know, the majority of times, right, and we clip it to one, so if you get zero, it becomes one, right? [S2] Yeah. [S1] So majority of times, you're only gonna get a single mask, right? Over 50% of the time, you're only gonna get a single mask. Uh, and then you pick a, you, you uniformly, uh, sample a subset of the document of any size, um, and you can, and you kind of place that in the end. So you get these very, very long kind of infilling naturally. [S2] Mm-hmm. [S1] Uh, | 24.48 | 2.83761 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818485.mp3 | [
9425,
9829,
9340,
9974,
6766,
9843,
4640,
5136,
4645,
8821,
3959,
3830,
3894,
3895,
3830,
3766,
3614,
3110,
7258,
4177,
3873,
9096,
3035,
10487,
7983,
1238,
4707,
12548,
11973,
8020,
5495,
6367,
3584,
2989,
5659,
10476,
11566,
10945,
6603,
... | [
0.6695219874382019,
-5.740097999572754,
-0.09850458800792694,
-0.0622173435986042,
0.8231538534164429,
-0.15523479878902435,
0.6294865012168884,
0.03875702619552612,
-0.022650456055998802,
-3.9372215270996094,
0.33657002449035645,
-0.23541145026683807,
-0.36696961522102356,
0.1305101513862... | 612 | 24.48 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818486 | [S1] Uh, of sequences super, super easily. Um, so we're kind of going all in on this objective. And so we have some follow-up work looking at, um, causal masked, uh, scaling loss for text. [S2] Yeah. [S1] So this is some ongoing work that we have now. Um, so we're pushing heavily on this. Um, so the general argument that we're trying to build is that, you know, if you're doing language modeling, uh, decoding language modeling, you should be doing causal masked language modeling. So that's kind of my- [S2] It's- it's- | 26.12 | 2.891961 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818486.mp3 | [
3112,
7184,
9463,
2998,
3374,
3886,
3950,
3950,
1398,
3894,
3893,
3749,
3108,
3091,
9778,
6696,
8365,
6439,
3057,
10217,
8680,
8144,
11221,
7622,
11937,
3754,
5422,
9230,
3929,
3472,
805,
3294,
337,
9048,
6488,
9160,
8640,
10581,
6277,
88... | [
0.5980409979820251,
-5.671493053436279,
-0.0364726260304451,
-0.0983344241976738,
0.824686586856842,
-0.12409982085227966,
0.7094893455505371,
0.044949498027563095,
-0.012293742969632149,
-3.8578925132751465,
0.3267131745815277,
-0.29192236065864563,
-0.2630385756492615,
0.2082719504833221... | 653 | 26.12 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818487 | [S1] Yeah, I mean, I was, I was, it is intuitively a good trade-off. So I think here you make the case if, if I interpret this correctly, that this word nationalist right here is really important to fill in this mask. And, and if it, if it were just sort of left to right, it would be very difficult to fill this in. Yet- [S2] Yeah. [S1] ... since you move it to the end, right? And, and th- the model has to extra learn kind of to, uh, keep, | 27.56 | 3.383163 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818487.mp3 | [
9227,
3609,
6869,
9604,
10666,
11132,
1901,
6699,
12279,
4894,
12149,
6246,
11684,
12163,
6668,
4112,
5659,
10070,
7265,
8757,
8702,
11628,
2732,
10484,
10613,
3395,
5456,
2953,
3328,
11289,
9761,
8745,
12671,
7102,
7782,
5294,
10547,
11502,
... | [
1.3564056158065796,
-5.369674205780029,
-0.1406698077917099,
-0.0544280968606472,
0.3267252445220947,
-0.1608932465314865,
0.33330419659614563,
-0.03393837809562683,
0.18219465017318726,
-3.847520351409912,
0.4198440611362457,
0.09680093079805374,
0.15823857486248016,
-0.0648493841290474,
... | 689 | 27.56 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818488 | [S1] XLNet or so. [S2] Mm-hmm. [S1] They were saying, "Well, we just train on all possible paths," right, "of decoding." Like all possible sequence of masking out tokens. And it was, it was never really satisfying because I always thought, "Well, there is something to left, to right." However, sometimes, as you say, it's really important to know what's after. And, uh, and, and I think this is like a really good trade-off. | 26.88 | 3.414309 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818488.mp3 | [
6184,
6712,
8741,
12278,
12774,
9897,
9258,
11740,
8690,
6105,
5577,
4643,
12790,
11246,
3567,
3383,
2782,
3404,
8717,
6638,
7151,
7159,
7449,
10848,
11776,
1902,
3326,
386,
4499,
9201,
8600,
468,
4023,
181,
45,
2566,
1173,
8873,
9830,
73... | [
1.4088486433029175,
-5.381155967712402,
-0.18651118874549866,
-0.013152063824236393,
0.4626936912536621,
-0.012725111097097397,
0.26523661613464355,
-0.020541833713650703,
0.25794702768325806,
-3.7800791263580322,
0.3315171003341675,
0.13994541764259338,
0.1615089327096939,
-0.079078227281... | 672 | 26.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818489 | [S1] Yeah, like specifically in this example, right? Like in the zero-shot prompting case, right? Like let's say we want to tag nationalist with some entity link, right? Um, if it appears beforehand in the sequence, there's no way to prompt the language model to generate like an entity link before the entity appears, right? [S2] Yeah. [S1] Uh, | 17.64 | 2.986214 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818489.mp3 | [
7204,
8309,
12588,
12078,
7660,
6746,
12693,
10697,
10666,
7926,
5469,
9718,
7442,
9882,
12525,
6548,
8089,
3914,
10928,
8182,
6122,
8136,
8155,
8988,
5616,
8484,
11817,
3737,
7963,
12554,
12134,
6157,
3043,
10677,
3049,
8064,
926,
9820,
43... | [
0.44178247451782227,
-5.984381198883057,
-0.13057240843772888,
-0.030611790716648102,
0.9227055907249451,
-0.046715445816516876,
0.5591086149215698,
0.026132402941584587,
-0.052355699241161346,
-3.82283878326416,
0.1942618489265442,
-0.2837497591972351,
-0.2796102464199066,
0.1726559698581... | 441 | 17.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818490 | [S1] So that was kind of another reason that we had because like I said, like HTML data is very localized, right? Like in Wikipedia, this, this A tag which represents the entity link always appears before the entity. [S2] Yeah. [S1] Either, uh, we have the option of, you know, training two models, right? One left to right, one right to left. | 17.76 | 2.997638 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818490.mp3 | [
4853,
3760,
4377,
8665,
3042,
11062,
10478,
7983,
2791,
3223,
3163,
3171,
9249,
8369,
4097,
12,
3104,
5632,
8685,
11770,
7788,
7988,
6042,
1425,
12138,
5564,
12019,
8403,
12527,
2849,
4829,
12589,
8206,
8404,
8117,
10556,
6411,
8630,
3340,
... | [
0.5583714842796326,
-5.909152984619141,
-0.16515350341796875,
-0.08205653727054596,
0.8701104521751404,
-0.03970438987016678,
0.5351744890213013,
-0.008421149104833603,
-0.1854948252439499,
-3.846724033355713,
0.19781823456287384,
-0.2683768570423126,
-0.2685506343841553,
0.197734057903289... | 444 | 17.76 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818491 | [S1] Um, um, or you can kind of do this kind of clever rotation of the document. [S2] Mm-hmm. [S1] Um, you said. Uh, yeah, the XLNet approach is definitely interesting, uh, which is, you know, having different permutations of the source document. But, uh, like you said, I think there's a lot of inductive biased, um, for left to right, which is why I think left to right models are kind of de facto now. | 22.92 | 3.012765 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818491.mp3 | [
6184,
7768,
9319,
9662,
5502,
6455,
4406,
6958,
4398,
3749,
8790,
6157,
10017,
8803,
9838,
7213,
7204,
9766,
4702,
9843,
4688,
6184,
3625,
3633,
4619,
6903,
3959,
4270,
3765,
822,
3829,
3685,
3100,
3619,
4626,
7201,
10778,
8318,
7924,
809... | [
0.7058594822883606,
-5.7247514724731445,
-0.13148456811904907,
-0.07742906361818314,
0.8557871580123901,
-0.08359836786985397,
0.633318305015564,
0.0333118811249733,
0.07784249633550644,
-3.8898379802703857,
0.23404701054096222,
-0.207220658659935,
-0.23992976546287537,
0.09250593930482864... | 573 | 22.92 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818492 | [S1] Is there, just, just for my understanding, is there a reason behind these arrows? Like why do the arrows, like, are like double arrows, and there's a line, and there's like a double arrow again? Like, is, is, does that have a specific meaning? And here the arrows are only here? [S2] Yeah, so arrows pretty much, uh, was the tokens that you actually generate. [S1] Okay. [S2] So in the language model you're generating every token in, in the- [S1] Oh, so you're going- | 23.84 | 3.322117 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818492.mp3 | [
6185,
5752,
8739,
12189,
6915,
5473,
642,
2770,
12203,
11673,
11698,
8027,
8083,
5589,
4893,
9443,
8802,
10866,
11826,
6201,
11340,
7352,
7322,
5728,
5673,
3082,
4660,
2078,
2665,
6289,
8346,
3106,
3617,
8780,
12482,
7169,
4108,
10113,
1071... | [
1.1547608375549316,
-5.523612022399902,
-0.0885607898235321,
-0.05276701599359512,
0.5159457325935364,
-0.18585604429244995,
0.38299450278282166,
-0.05341155454516411,
0.09163365513086319,
-3.880824327468872,
0.2844753563404083,
-0.06540333479642868,
-0.02808551676571369,
0.065143458545207... | 596 | 23.84 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818493 | [S1] ... go like this. Okay, I see, I see. [S2] Yeah. [S1] 'Cause I was, I was like, okay, is, is there some meaning? But yes, there is. And this shows that in the mask language model object, if you only actually generate very small number of tokens, and you, you wouldn't even get like a loss for the other tokens. | 17.16 | 3.334201 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818493.mp3 | [
9131,
10549,
8429,
7854,
10398,
3758,
1187,
6178,
12411,
10920,
8240,
8248,
7217,
7707,
3221,
5469,
11774,
7436,
11482,
1755,
8916,
3931,
11181,
7007,
4462,
4957,
9201,
3048,
2897,
2889,
3336,
3904,
4633,
9454,
10605,
3228,
11547,
4594,
972... | [
1.5161118507385254,
-5.391561985015869,
-0.2617987394332886,
-0.02983153611421585,
0.3638658821582794,
-0.22876127064228058,
0.2940629720687866,
-0.051951996982097626,
0.19411225616931915,
-3.871201276779175,
0.39528387784957886,
0.08201995491981506,
0.12567339837551117,
-0.124344803392887... | 429 | 17.16 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818494 | [S1] Uh, yeah. So there's a couple of ways to approach this. So the, the very first thing is that modeling, and I think we mentioned this quickly in the paper, but modeling image tokens versus text tokens, it's quite different actually. So for like text usually follows, like textual tokens follow like a Zipfian distribution, right? Whereas I think in appendix we have a figure, it's pretty much uniform for images. [S2] Yeah. [S1] Uh, so there's different, | 25.6 | 3.142817 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818494.mp3 | [
3104,
4641,
8949,
6455,
6382,
8438,
7990,
7983,
6391,
6236,
5724,
6699,
7323,
9891,
12442,
8426,
11931,
9882,
9897,
5674,
6153,
6185,
5153,
9298,
6192,
6868,
5680,
7388,
5216,
3617,
3609,
2793,
5641,
7372,
12610,
11212,
11683,
11125,
3957,
... | [
0.5301077961921692,
-5.709949016571045,
-0.14261822402477264,
-0.07400034368038177,
0.9070348739624023,
-0.08009347319602966,
0.5856414437294006,
-0.013717480935156345,
-0.04881018027663231,
-3.858583450317383,
0.30508291721343994,
-0.3479403555393219,
-0.3015248775482178,
0.19020749628543... | 640 | 25.6 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818495 | [S1] ... you know, images would be being optimized for, but text kind of stayed flat. So we don't really have explanations for why this is happening. Um, I think there needs to be future, like, scaling laws looking at multimodal sequence modeling. [S2] Mm-hmm. [S1] And when I say multimodal, I'm not just talking about, like, images and, like, natural language text. I meant, like, you can even include code as a different modality, right? [S2] Yeah. | 23.24 | 3.348522 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818495.mp3 | [
6185,
5193,
1180,
6805,
12547,
7303,
8541,
3830,
3165,
6754,
9825,
8361,
7904,
9248,
5169,
9755,
11741,
11602,
7714,
10467,
11740,
8770,
10177,
8663,
6559,
1353,
3912,
1809,
3180,
5293,
8991,
8200,
6240,
7074,
10054,
4357,
5665,
4068,
12678... | [
0.6812966465950012,
-5.687626838684082,
-0.14964663982391357,
-0.008826239965856075,
0.8995286822319031,
-0.06994454562664032,
0.5125292539596558,
0.03266770765185356,
-0.16639961302280426,
-3.8617911338806152,
0.2819024324417114,
-0.24768505990505219,
-0.21201950311660767,
0.1805310547351... | 581 | 23.24 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818496 | [S1] Uh, yeah, it seems appropriate that an image could be expressed in something like a sequence of tokens. It's- [S2] Mm-hmm. [S1] It's just a bit... I'm not too big of a fan of how this is currently done because the tokens, they also... Like, they al- already, they seem to be a bit localized in the image and so on. Like, I don't- [S2] Yeah. [S1] I- I think there's like a better... I don't think there's a better way... If- if you're a human, you're not... That's not really what you do with an image. You- you see more like... | 29.88 | 3.412359 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818496.mp3 | [
3688,
2083,
11887,
9663,
957,
4470,
4470,
3958,
3958,
3518,
6006,
3164,
3169,
8801,
11827,
8353,
3624,
9749,
7365,
7299,
9857,
9929,
9606,
11243,
11689,
11637,
6645,
4772,
12637,
7435,
10256,
3554,
3024,
12118,
12674,
11908,
7708,
10909,
34... | [
1.5987828969955444,
-5.187533855438232,
-0.066188283264637,
-0.027952272444963455,
0.43298572301864624,
-0.19841349124908447,
0.2675727903842926,
0.0006297763902693987,
0.3052566647529602,
-3.872225284576416,
0.5745472311973572,
0.07498157769441605,
0.14353755116462708,
-0.1158081516623497... | 747 | 29.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818497 | [S1] Um, it took a little bit of hand-holding to work, especially the, the 13 billion parameter model, took a little bit of hand-holding to work. So a lot of the times the pathologies we see is, are things like gradient underflow or overflow. [S2] Mm-hmm. [S1] Uh, gradient explosions happen, although they're more, they usually happen in much bigger models like the 100 billion scale. Um, but the surprising thing was that we almost used exactly the same hyper-parameters, um, as this paper that came out from Vesto and his group. Um, | 29.88 | 3.120978 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818497.mp3 | [
11508,
10676,
8043,
11051,
10540,
10611,
11484,
8294,
8798,
12571,
12572,
9948,
12508,
12508,
12500,
12435,
9736,
10889,
5738,
544,
3097,
12436,
11628,
10004,
8392,
10704,
6122,
10540,
11875,
6562,
10546,
7780,
2716,
11125,
10987,
10461,
3091,
... | [
0.6296629905700684,
-5.5373101234436035,
-0.06742647290229797,
-0.0707489475607872,
0.954764187335968,
-0.10964055359363556,
0.5392454862594604,
0.09172938019037247,
-0.020251629874110222,
-3.8815107345581055,
0.3165019154548645,
-0.35131698846817017,
-0.20431803166866302,
0.17886973917484... | 747 | 29.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818498 | [S1] So the surprising thing is it kind of just worked out of the box. Apart from having to tune, I think we tuned, tuned like learning rate, um, we had to tune weight t-k- [S2] Mm-hmm. [S1] Uh, and batch size. Apart from tuning those things, it just worked almost straight out of the box. Um, and what you said is actually correct, which is if you look at the large model, it's actually not done training. Um, so | 22.12 | 3.085381 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818498.mp3 | [
6176,
2337,
8481,
8648,
3036,
11061,
7982,
7918,
7902,
5271,
5254,
3220,
8866,
12441,
9881,
11891,
7265,
5666,
5289,
3609,
3624,
5145,
8082,
10686,
6121,
8144,
8999,
10538,
4081,
5612,
10681,
9985,
3538,
11986,
9245,
5768,
11755,
11457,
117... | [
0.7639814019203186,
-5.65359354019165,
-0.13365598022937775,
-0.08160878717899323,
0.8842290639877319,
-0.10396628826856613,
0.5884767770767212,
0.07535403966903687,
-0.045985300093889236,
-3.886378526687622,
0.2744047939777374,
-0.41015079617500305,
-0.21050845086574554,
0.083352133631706... | 553 | 22.12 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818499 | [S1] The good news is once, once CM3 is released, we're going to release the, the checkpoint that we used for this model. Um, uh, I think the model that we have now is continuing training, so we'll release that one too. So, uh, people will be able to play around with both. [S2] Excellent. [S1] Uh, but one thing I'd like to point out is that the multimodal scaling laws are a little bit different than, than text scaling laws. Uh, one thing, | 22.12 | 3.108213 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818499.mp3 | [
5672,
5153,
8073,
8622,
9307,
9627,
10659,
11475,
10891,
12099,
11078,
6096,
5587,
10659,
5956,
3016,
3923,
2731,
5365,
10677,
11564,
10756,
3482,
8153,
8154,
1939,
11885,
3692,
2675,
8053,
11500,
11467,
8169,
6104,
6040,
5568,
12172,
11652,
... | [
0.7206982374191284,
-5.744968891143799,
-0.15620113909244537,
-0.013273310847580433,
0.7883843183517456,
-0.10311498492956161,
0.6463605761528015,
0.07336322218179703,
-0.0660715252161026,
-3.8545680046081543,
0.3284039795398712,
-0.311431884765625,
-0.2196815460920334,
0.1771397739648819,... | 553 | 22.12 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818500 | seems to be that scale plays a slightly larger role, um, in multi-modal than it does in text. [S1] Mm-hmm. [S2] Um, so I, I think the qua- the quantitative thing that we saw is that if you looked at the data efficiency jumps between like, uh, I'm forgetting the exact numbers, but like, like let's make them up, like the 1.3 billion model and the 13 billion model from, from Vesta's paper. [S1] Mm-hmm. [S2] Um, | 26.24 | 2.9587 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818500.mp3 | [
5680,
6675,
3561,
10720,
5520,
4048,
12612,
12483,
7773,
3622,
8426,
5089,
8080,
5504,
8216,
5520,
6617,
8966,
5155,
8800,
6621,
12164,
12549,
12677,
7237,
9,
2980,
6143,
4716,
7872,
5617,
5600,
8152,
2448,
11866,
5019,
11668,
11740,
7804,
... | [
0.7904265522956848,
-5.64358377456665,
-0.13902850449085236,
-0.13046309351921082,
0.8706899285316467,
-0.1420954018831253,
0.5741438865661621,
0.050903528928756714,
-0.14007706940174103,
-3.8730216026306152,
0.35854363441467285,
-0.31360289454460144,
-0.24984857439994812,
0.11221479624509... | 656 | 26.24 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818501 | [S1] And the data efficiency there, well, let's say it was like the larger model was five times more efficient in terms of data. So, uh, in order to reach the same perplexity, it would need five times less data. Uh, using these same exact models, we saw that in the multimodal case, it was 10x. So there was all, almost a two times difference for some reason. [S2] Mm-hmm. [S1] Uh, and that's why I think it's really important to kind of chase these multimodal scaling laws and fundamentally understand what's going on here. Uh, there's a lot of unknowns here. | 29.36 | 3.141938 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818501.mp3 | [
2600,
5152,
10981,
11758,
9861,
8781,
12749,
4822,
3161,
9771,
5787,
6874,
4635,
524,
18,
5312,
10741,
9430,
8256,
8667,
11217,
11665,
8465,
8545,
11108,
3562,
6640,
8629,
11605,
12697,
4562,
9302,
8971,
6040,
5003,
7558,
3075,
2963,
11767,... | [
0.747488260269165,
-5.61313009262085,
-0.0506150908768177,
-0.0882011130452156,
0.8534215688705444,
-0.14760227501392365,
0.5148462653160095,
0.0481768399477005,
-0.1430002599954605,
-3.898984432220459,
0.36811134219169617,
-0.3566192388534546,
-0.27860477566719055,
0.14393950998783112,
... | 734 | 29.36 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818502 | [S1] It's just really, really tedious. So one of the main things is, um, you know, whenever you have a ton of nodes that you're running, there's infrastructure issues that pop up, right? [S2] Yeah. [S1] So like if one GPU goes down, right, then, then all of training is paused, right? So infrastructure issues are kind of a big thing, and we have some automated systems in place to take care of that. [S2] Mm-hmm. [S1] Um, the other things are like, | 22.48 | 2.972791 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818502.mp3 | [
6184,
4646,
10069,
11728,
10184,
12160,
8155,
8484,
6040,
3904,
3652,
619,
347,
10653,
10412,
2837,
12118,
6340,
171,
2900,
10982,
2644,
3486,
7556,
11856,
8080,
12256,
4891,
10054,
4357,
3852,
4292,
4997,
1391,
1379,
6304,
2888,
4242,
6705... | [
0.6559805274009705,
-5.723817348480225,
-0.16170595586299896,
-0.049537088721990585,
0.8207542896270752,
-0.042625848203897476,
0.5271332263946533,
0.021287858486175537,
-0.1011388748884201,
-3.851170301437378,
0.281141996383667,
-0.3661023676395416,
-0.2043776959180832,
0.1429226100444793... | 562 | 22.48 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818503 | [S1] Yeah, because of the computers, one model. So it, it really comes down to intuition. Um, so both Mike Lewis and Namung Goyo, who are on the paper, have trained these really, really big models before. Um, so they had a ton of great intuition about how to get things to work, um, in terms of these very large models. [S2] Mm-hmm. [S1] Uh, yeah. | 21.36 | 3.064457 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818503.mp3 | [
9911,
6825,
4108,
12613,
11211,
11252,
8183,
1462,
6437,
4379,
9377,
9317,
5558,
8702,
2324,
2595,
5416,
11724,
9932,
12641,
6643,
10611,
2944,
5586,
11428,
2769,
11694,
6242,
12634,
4587,
11295,
8866,
7592,
9156,
8015,
6404,
3502,
6358,
34... | [
0.7409189939498901,
-5.765048503875732,
-0.18782426416873932,
-0.0693642646074295,
0.829537034034729,
-0.09320662915706635,
0.6646050214767456,
-0.0038552256301045418,
-0.013655648566782475,
-3.8209526538848877,
0.2910911738872528,
-0.23902243375778198,
-0.16891388595104218,
0.132843479514... | 534 | 21.36 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818504 | [S1] Yeah. Um, yeah, that's a great question. So I, I think at the beginning of the project the, the push was really to have a single model, uh, that can do any image task in the zero shot case. [S2] Mm-hmm. [S1] Um, and so kind of the story that we built around it is, can we describe all the tasks that we're interested in, um, | 20.24 | 3.050452 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818504.mp3 | [
10840,
9936,
7381,
12101,
11175,
6007,
6438,
6370,
9370,
12580,
12596,
8429,
12565,
9972,
12509,
9910,
12014,
12398,
4694,
9842,
6769,
8729,
6454,
10621,
10549,
10613,
11054,
11062,
11054,
8438,
8798,
5661,
3109,
6746,
9313,
12384,
11818,
623... | [
0.5921516418457031,
-5.811474323272705,
-0.15524977445602417,
-0.029538581147789955,
0.8115456104278564,
-0.07711499184370041,
0.7324839234352112,
-0.0057883006520569324,
-0.009338914416730404,
-3.8098347187042236,
0.20814239978790283,
-0.2833087742328644,
-0.20876343548297882,
0.216498211... | 506 | 20.24 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818506 | [S1] So some of them didn't work. Some of them only worked at scale, and we can kind of go, go through this. Um, specifically, like one thing is that, like, the captioning only worked at scale. [S2] Okay. [S1] So the 13 billion model was the only model that could caption well. [S2] And the captioning, you go mainly with the alt text of the image. [S1] Alt or the title. Either one. [S2] Yeah. [S1] Yeah. | 19.56 | 3.024697 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818506.mp3 | [
7274,
5744,
4754,
6032,
5595,
7995,
10470,
3221,
3307,
7666,
8152,
5603,
8121,
11370,
7851,
8364,
5656,
3019,
11185,
8732,
6677,
5184,
8660,
11164,
8259,
11117,
8726,
8807,
2733,
5309,
5541,
5572,
11036,
9313,
12450,
4458,
2355,
11625,
8144... | [
0.9237847924232483,
-5.627383232116699,
-0.18362149596214294,
-0.04513449966907501,
0.6892415881156921,
-0.08378496766090393,
0.602705180644989,
-0.023922458291053772,
-0.018701517954468727,
-3.8874757289886475,
0.28114771842956543,
-0.050910621881484985,
-0.18081773817539215,
0.0927539393... | 489 | 19.56 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818507 | [S1] No. But like the figure that you're on now, I think is kind of interesting. So we can kind of get unconditional image generation by just asking the model to generate a sequence of tokens after the image tag. [S2] Yeah. [S1] So we saw one interesting behavior is that the model, for some reason, almost always wanted to first generate the alt text before generating the image. Um, for it was actually easier to condition on the text before generating the image than doing this type of free-form generation. Um- [S2] When you | 29.52 | 3.1289 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818507.mp3 | [
8280,
6875,
9354,
9939,
6421,
6454,
5999,
3183,
6747,
9372,
10022,
12532,
12654,
10020,
7201,
10904,
2870,
5339,
10749,
10004,
9322,
8082,
9638,
8680,
3569,
10664,
9482,
8994,
8091,
6210,
8100,
11707,
4364,
9548,
5469,
4253,
8446,
10870,
86... | [
0.4780593812465668,
-5.656517028808594,
0.007854422554373741,
-0.11560782790184021,
0.8826935887336731,
-0.11308593302965164,
0.5160194039344788,
0.037720683962106705,
-0.09160512685775757,
-3.8814265727996826,
0.2963036298751831,
-0.29134756326675415,
-0.27484840154647827,
0.2774980366230... | 738 | 29.52 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818508 | [S1] When you say it wanted to, it, that's just what it did. [S2] Yeah. [S1] Like when you, when you sampled, did you, like, I mean, this, when you say it wanted to, it could also be that in the internet, humans most of the time write alt first and then the source. [S2] Yeah, so we actually looked into this. So, uh, | 18.56 | 3.31566 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818508.mp3 | [
9454,
4406,
8933,
3893,
6453,
3308,
11252,
9365,
11604,
9099,
6428,
6049,
2952,
12212,
11689,
10116,
4821,
12701,
4429,
3624,
3620,
2669,
2667,
2667,
2862,
3319,
10430,
8500,
11412,
10614,
11940,
10848,
6705,
5377,
7080,
11629,
10395,
5254,
... | [
1.2677958011627197,
-5.62701940536499,
-0.0937219187617302,
-0.15268321335315704,
0.4958823621273041,
-0.1831662654876709,
0.38475310802459717,
-0.04599102586507797,
0.23716376721858978,
-3.786436080932617,
0.2782163619995117,
-0.003484078450128436,
0.0454070009291172,
0.0128241628408432,
... | 464 | 18.56 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818509 | [S1] A lot of text does have alt, um, but it's around like, I want to say like 70% to 80% mark, if I recall correctly. So it wouldn't explain why the model almost always wants to generate alt text. Now the, the theory that we kind of have is that without alt text, you have much higher perplexities for images. [S2] Mm-hmm. [S1] So the model, you know, even, 'cause, because we're doing like sampling, right? So it's gonna pick out high probability, low perplexity tokens. [S2] Yeah. [S1] Which most of the case means picking out the alt. | 29.24 | 3.290174 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818509.mp3 | [
2600,
11801,
11126,
8116,
5213,
3290,
5404,
7990,
10545,
12522,
5450,
11131,
3290,
4905,
8384,
11168,
7156,
11770,
12124,
11355,
913,
1354,
524,
5377,
8631,
8053,
4577,
5584,
3936,
9532,
11261,
8951,
2914,
4329,
9257,
8759,
2870,
2870,
7862... | [
0.5933179259300232,
-5.6310906410217285,
-0.0033600558526813984,
-0.17749454081058502,
0.8434569835662842,
-0.12494909018278122,
0.5699974298477173,
0.08194992691278458,
-0.15041260421276093,
-3.8770105838775635,
0.37523433566093445,
-0.3175373673439026,
-0.26160475611686707,
0.28865310549... | 731 | 29.24 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818510 | [S1] Yeah, that's true. There, there isn't, for VQ-V again, there isn't something explicit, but the, I think the way that the layers are constructed, you do still get some implicit dependencies across the- [S2] Yeah. [S1] ... tokens. And so I think this is what the Transformer's kind of pulling apart here. [S2] Yeah. [S1] Um, yeah. And to be honest, I think there's still a lot of work to be done on the discretizing images front. [S2] Mm-hmm. [S1] So, uh, | 25.84 | 2.928408 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818510.mp3 | [
6184,
9360,
9995,
10178,
10713,
10664,
11252,
2998,
6498,
10217,
6962,
9387,
9378,
3658,
7210,
1026,
3586,
8181,
11645,
6938,
465,
4553,
7632,
12753,
3035,
2901,
404,
798,
3790,
3097,
2962,
11189,
10724,
5581,
4308,
9887,
5184,
6125,
8684,
... | [
0.6676140427589417,
-5.6946611404418945,
-0.13837943971157074,
-0.0901891365647316,
0.8938692212104797,
-0.10362544655799866,
0.5781000256538391,
0.024841634556651115,
-0.09163253009319305,
-3.890094518661499,
0.2598116099834442,
-0.2627161145210266,
-0.24230487644672394,
0.138041406869888... | 646 | 25.84 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818511 | [S1] Um, so being able to generate trivially, um, being able to compute log probabilities, I think tokens are probably the easiest way to go. [S2] Yeah. [S1] Uh, and one thing is you can naturally increase the resolution of tokens images just by increasing how many tokens you use per image. [S2] Mm-hmm. [S1] So in some sense, if you have enough compute, you can scale up to arbitrary resolutions, right? | 19.84 | 2.818317 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818511.mp3 | [
6697,
5680,
4099,
6903,
3375,
8558,
3446,
3894,
4341,
3766,
6758,
3109,
1571,
4122,
4106,
3945,
10688,
5548,
10911,
8225,
6628,
11715,
12171,
9293,
12117,
12773,
6236,
5366,
3155,
8072,
5620,
8790,
8328,
12673,
10734,
12718,
10316,
11623,
3... | [
0.8316604495048523,
-5.761127471923828,
-0.15706178545951843,
-0.0709485113620758,
0.8844632506370544,
-0.11606218665838242,
0.38777437806129456,
0.010337515734136105,
-0.08643754571676254,
-3.854665517807007,
0.36988556385040283,
-0.34129390120506287,
-0.2918381989002228,
0.11233513057231... | 496 | 19.84 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818512 | [S1] Yeah, I mean, yeah, down to probably, probably you could at some point get more, more tokens than pixels. I wouldn't know what, what that would mean, but, um, I guess the resolution isn't even limited by the resolution of the image itself. [S2] Yeah. [S1] Uh, so there's, there's this, this interesting thing you can do, as you said, infilling, uh, by letting the model generate sort of middle tokens. Now, | 27.56 | 3.370621 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818512.mp3 | [
3104,
4163,
10060,
11653,
9207,
1958,
7470,
4029,
3622,
12125,
12549,
7692,
5641,
3610,
3104,
2052,
12677,
11239,
4542,
1397,
4909,
9449,
9811,
12410,
6176,
3632,
9216,
3566,
7101,
8958,
8366,
11852,
8288,
6536,
7643,
5895,
1164,
8297,
7922... | [
1.434423804283142,
-5.34345817565918,
-0.11280728876590729,
-0.04982763156294823,
0.378972589969635,
-0.23315776884555817,
0.38042157888412476,
-0.06517625600099564,
0.22759827971458435,
-3.8107495307922363,
0.49150213599205017,
0.1226712167263031,
0.16243745386600494,
-0.09622466564178467... | 689 | 27.56 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818513 | [S1] Yeah, so actually because of our objective, 'cause we sampled the number of masks, right? [S2] Yeah. [S1] Um, you can actually mask out like five, six, seven masks. [S2] Yeah. [S1] And it should still work. Um, I don't think there was any specific reason that we stuck to masking out a single thing. [S2] Mm-hmm. [S1] Um, I'm sure it would work with multiple as well. | 18.08 | 2.813773 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818513.mp3 | [
12597,
8950,
7595,
6819,
11667,
10648,
11573,
1444,
6616,
5595,
7910,
9966,
10747,
9525,
3507,
12761,
2851,
11161,
7237,
35,
10656,
11726,
7314,
12641,
4084,
10546,
5898,
448,
10612,
8859,
874,
5871,
10997,
8053,
5525,
5518,
3029,
3348,
937... | [
0.8247446417808533,
-5.6932573318481445,
-0.16568860411643982,
0.010489452630281448,
0.8772872686386108,
-0.060554202646017075,
0.6474851965904236,
0.01446821354329586,
0.04034986346960068,
-3.8629162311553955,
0.2014649659395218,
-0.26030421257019043,
-0.1918824166059494,
0.10335259139537... | 452 | 18.08 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818514 | [S1] Um, then you also give the ground truth, which is here on the right, and then there's one model that does infilling unconditional, so just looking at the image, and then there's one model that does it conditionally, and the conditional is, uh, conditioned with this thing right here as the, the alt text. So the understand- [S2] Mm-hmm. [S1] Okay. So understand it correctly. Um, | 22.64 | 3.350773 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818514.mp3 | [
4258,
4176,
4645,
7343,
4470,
1526,
1389,
1973,
950,
1469,
4398,
3100,
558,
5670,
4643,
8737,
8376,
6176,
10376,
7280,
7254,
2132,
3336,
12727,
12790,
11756,
11763,
11859,
10892,
10830,
10582,
7462,
9313,
11826,
8368,
7189,
7236,
12483,
583... | [
1.2821205854415894,
-5.421082496643066,
-0.15236715972423553,
0.03619888797402382,
0.3798403739929199,
-0.20739401876926422,
0.39242252707481384,
-0.058382220566272736,
0.30163493752479553,
-3.798346519470215,
0.4814143478870392,
0.01854604482650757,
0.10742906481027603,
-0.085916101932525... | 566 | 22.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818515 | [S1] I, I'm not, I'm not sure the unconditionality has something much to do with it because there is no, this doesn't look like natural. You know, you know what I mean, a little bit? Like- [S2] Yeah, a little bit. [S1] This, this, this shouldn't be like, just because it's not conditioned on it. If it's not conditioned on text, I would expect it to be maybe a red bench, right? Or, or something, you know, something, uh, that is conceivable in nature but is not, | 29.16 | 3.223518 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818515.mp3 | [
9387,
5680,
10862,
8695,
7573,
6752,
11819,
8702,
7271,
7789,
11340,
8050,
10417,
4260,
6688,
9768,
11126,
7094,
7717,
8916,
10492,
10994,
9832,
11010,
12698,
12754,
7632,
8039,
8044,
5451,
5451,
2883,
5517,
3022,
3022,
6356,
7849,
11361,
1... | [
1.32594895362854,
-5.324089050292969,
-0.1118880957365036,
-0.04358567297458649,
0.4568369388580322,
-0.16779015958309174,
0.34057265520095825,
-0.0011326689273118973,
0.23863840103149414,
-3.8218929767608643,
0.356128066778183,
0.16570040583610535,
0.14922602474689484,
-0.0048095826059579... | 729 | 29.16 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818516 | [S1] Yeah. So, so one theory that we kind of have here, um, is that the, the model needs to understand the continue, continuation of the, the horizontal lines, right? [S2] Mm-hmm. [S1] And that re- that requires some semantic understanding that this is, for example, uh, a bench, right? And actually, if you look at the, the mass-style input, the horizontal n- lines are not completely horizontal, so the bottom of the bench is at a different angle than- [S2] Yeah. | 23.88 | 2.909264 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818516.mp3 | [
6184,
5217,
4819,
9989,
3447,
3830,
5794,
11945,
8866,
12389,
12683,
10220,
8437,
12077,
12716,
12661,
12526,
6753,
7074,
11224,
5611,
7978,
7909,
5270,
3234,
6689,
6311,
6641,
11208,
2988,
10413,
2652,
5235,
7995,
12019,
8720,
3616,
8138,
... | [
0.6072186231613159,
-5.725311756134033,
-0.11779740452766418,
-0.049552321434020996,
0.8766210079193115,
-0.0634957104921341,
0.5453144907951355,
-0.001316215144470334,
-0.06752132624387741,
-3.814415693283081,
0.35298117995262146,
-0.377679705619812,
-0.3015633225440979,
0.174203738570213... | 597 | 23.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818517 | [S1] the top of the bench. So I think the model has a tough time understanding the, the high level semantic content of the image, which is fixed by feeding in text. [S2] Yeah. [S1] Uh, now I think of course if you have, I think if you have a larger model that's trained for longer with a higher resolution, uh, this probably should not be an issue. Um, but VQ-VE again, it, it blurs out a lot of things, number one. [S2] Yeah. [S1] Uh, and number two, it's, um, | 26.8 | 3.185842 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818517.mp3 | [
6169,
6810,
4133,
3778,
4512,
8958,
2941,
6692,
5865,
108,
3609,
5457,
8807,
3112,
10602,
11638,
9579,
9300,
8913,
12681,
7449,
9321,
12013,
9397,
9331,
9453,
7845,
9902,
9830,
8349,
9820,
7338,
5650,
4009,
6096,
3501,
10542,
7918,
5847,
... | [
0.6328297853469849,
-5.682229518890381,
-0.07112960517406464,
-0.10037156939506531,
0.8876149654388428,
-0.1163826733827591,
0.5560495853424072,
0.023947985842823982,
-0.12102312594652176,
-3.8426430225372314,
0.26819223165512085,
-0.33862364292144775,
-0.2632578909397125,
0.24062845110893... | 670 | 26.8 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818518 | [S1] Just if you change the tokens even a little bit, the, the blurring aspect happens very, very quickly with VQ-VAE again. Compared to, for example, the VQ-VAE from DALI, um, which requires more tokens, so 1024 tokens versus the 256 we use here. Um, but, uh, it's more direct in some sense. [S2] Yeah. [S1] Um, so, yeah, I think the main thing here is just that you need to get some, like, high level, | 26 | 3.097037 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818518.mp3 | [
9384,
3170,
11649,
10707,
10731,
8673,
5512,
4754,
12701,
3418,
9696,
9093,
8454,
11465,
12746,
9177,
11762,
12626,
8907,
10192,
5761,
8174,
6676,
8392,
12712,
9150,
7973,
3149,
12449,
9054,
8215,
6411,
2896,
6691,
12685,
12044,
5418,
11931,
... | [
0.7546611428260803,
-5.674359321594238,
-0.11130039393901825,
-0.07110528647899628,
0.8754115700721741,
-0.1574537456035614,
0.4583186209201813,
-0.010050252079963684,
-0.12293554097414017,
-3.8821451663970947,
0.33835771679878235,
-0.3290744721889496,
-0.28044310212135315,
0.1480865478515... | 650 | 26 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818519 | [S1] The, in the case that it doesn't generate a car at all, it just generates mountains, right? Just because of what landscapes are easier to generate. Um, the other thing that we saw kind of tough compared to Dali is, you know, the data that we used only came from Wikipedia or Common Crawl News. So none of it was fictional in some sense, right? Like we don't have any, like, art. [S2] Yeah. [S1] Um, so, like, our images always tried to be as non-fictional as possible, which is, it acts weird if you try to give it, | 27.8 | 2.929623 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818519.mp3 | [
6177,
5296,
3589,
531,
7808,
3446,
3375,
1327,
1254,
1766,
3886,
1246,
9322,
6688,
12134,
11277,
8340,
9719,
11938,
12768,
7660,
12254,
7629,
3408,
5485,
4838,
3486,
3083,
7872,
10726,
5558,
6485,
3520,
8926,
8198,
12681,
9173,
6134,
7261,
... | [
0.5893973708152771,
-5.6403326988220215,
0.03560522943735123,
-0.11963874846696854,
0.9066162705421448,
-0.0947156623005867,
0.43427667021751404,
0.057472553104162216,
-0.16748052835464478,
-3.899660348892212,
0.2802180051803589,
-0.34124958515167236,
-0.24267061054706573,
0.24077361822128... | 695 | 27.8 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818520 | [S1] Um, like really fantasy based prompts. [S2] Yeah. [S1] Uh, so that's kind of one downside. And actually this is one criticism I have, uh, of the evaluation that we did for the FID matrix, which is a way to measure, uh, you know, the, the quality of images, which is, uh, we actually took the table from Glide, um, for the FID numbers on the conditional generation. Uh, one thing was, is that MS Coco is all like | 24.64 | 3.159062 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818520.mp3 | [
4129,
1555,
6838,
4030,
3446,
1405,
4861,
3613,
5669,
102,
358,
4821,
9825,
4322,
164,
3469,
11045,
2598,
350,
7430,
7557,
1234,
9834,
2650,
8680,
9136,
6584,
12766,
10222,
9902,
8914,
3024,
7493,
2378,
833,
12614,
7493,
2593,
10665,
1008... | [
0.5355765223503113,
-5.76883602142334,
-0.14722995460033417,
-0.07933824509382248,
0.8800201416015625,
-0.027360307052731514,
0.47509321570396423,
0.022226683795452118,
-0.042190223932266235,
-3.8326406478881836,
0.32727083563804626,
-0.4629775285720825,
-0.21829617023468018,
0.20872546732... | 616 | 24.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818521 | [S1] almost all non-fiction. [S2] Yeah. [S1] Uh, like non-fantasy images. So this is really sh- like, it's under-representing Dali. So I think if you casted a wider net here and had something that included a wider array, uh, a, a bigger distribution of images, I think Dali's results here would be, uh, much, much stronger. | 21.52 | 3.071398 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818521.mp3 | [
5680,
5680,
9261,
7933,
10844,
10421,
8438,
6560,
6464,
1187,
11838,
5367,
7990,
5887,
5878,
8375,
5149,
5269,
203,
3722,
9313,
6739,
8263,
7685,
8718,
11742,
8447,
5885,
7278,
9356,
4520,
9201,
6640,
4079,
2390,
12451,
5025,
12744,
5016,
... | [
0.8584367632865906,
-5.722309589385986,
-0.18048404157161713,
-0.09692195057868958,
0.9004741311073303,
-0.0847947970032692,
0.47499316930770874,
0.06416263431310654,
-0.04871782660484314,
-3.8794891834259033,
0.2819502353668213,
-0.38593828678131104,
-0.24250298738479614,
0.12945629656314... | 538 | 21.52 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818522 | [S1] Mm-hmm. [S2] Yeah. [S1] You, you did, you did discuss a little bit. You also said you ha- you, you saw subsampled, uh, web data in, and, and you cited some concerns as well. Um, but there is also quality issue with sort of the, the wider you cast the net. | 19.16 | 3.33269 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818522.mp3 | [
8787,
8854,
9427,
11478,
2067,
7186,
7849,
5650,
3749,
6445,
6380,
3829,
3308,
9451,
58,
5152,
9832,
9265,
7192,
6176,
7192,
7904,
8736,
7249,
6160,
9770,
6232,
7201,
7193,
7866,
7256,
6168,
5288,
4641,
7784,
12675,
12653,
6892,
7323,
699... | [
1.3496577739715576,
-5.5114216804504395,
-0.2365690916776657,
-0.03574926406145096,
0.3793889880180359,
-0.20334675908088684,
0.40454018115997314,
-0.04094036668539047,
0.307661771774292,
-3.8297762870788574,
0.3788904547691345,
-0.006323389708995819,
0.18617935478687286,
-0.07418725639581... | 479 | 19.16 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818523 | [S1] I think at the beginning we had some ethical concerns of like, like I said, we have very weak alignment, so you can prompt with anything, right? [S2] Yeah. [S1] We had some ethical concerns about images that you can generate if you were just trained on all of Common Chrome. [S2] Mm-hmm. [S1] Um, so we tried to think about what are like large scale data sets that we can get that are somewhat filtered. Uh, Wikipedia is definitely one of them. Um, but even, and actually Wikipedia itself has a gender bias, and I think this is a new, I think other, | 26 | 2.836699 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818523.mp3 | [
8714,
9258,
11260,
12627,
5704,
8161,
11601,
11857,
9059,
11644,
3104,
8081,
11565,
2602,
8682,
12038,
11993,
11677,
11682,
8851,
12051,
11285,
628,
3869,
10002,
7466,
11700,
11565,
8256,
3434,
2960,
5806,
8231,
7261,
9251,
12278,
10743,
7324... | [
0.769744873046875,
-5.640860080718994,
-0.09483033418655396,
-0.07872729748487473,
0.8501927852630615,
-0.0811985433101654,
0.47605210542678833,
0.05546186864376068,
0.013260915875434875,
-3.9038448333740234,
0.36896494030952454,
-0.37507152557373047,
-0.3572378158569336,
0.114001579582691... | 650 | 26 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818525 | [S1] ... from Convo. So you could probably do something like this here. [S2] Yeah. [S1] Um, I, I questioned the efficacy just because very large models, they only need to see a data point a couple times in order to pick it up. Um, so, uh, I think there's like some very fundamental engineering work that, that's being done in, for scaling up these datasets to like, um, like, uh, trillions of tokens, essentially. [S2] Um- [S1] Um, and images. [S2] Yeah, I mean, | 29.28 | 2.864815 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818525.mp3 | [
3748,
9314,
6734,
548,
3797,
1858,
4009,
164,
5661,
7340,
9514,
3829,
3765,
3742,
3686,
8802,
8939,
562,
6307,
8868,
6228,
8926,
12540,
12077,
12589,
11509,
12644,
10083,
9435,
6616,
3357,
9026,
8517,
12129,
5486,
8426,
3050,
8041,
7866,
... | [
0.7790442705154419,
-5.628880023956299,
-0.059953946620225906,
-0.10613711178302765,
0.7764776945114136,
-0.11056230962276459,
0.5010915398597717,
0.028427235782146454,
-0.008226552978157997,
-3.926746368408203,
0.2952783405780792,
-0.28693583607673645,
-0.3286925256252289,
0.1271604150533... | 732 | 29.28 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818526 | [S1] Sort of the, let's say diversity makes, is, is probably the best, so you can always choose which one you wanna, you wanna use. I don't know. I'm sorry, this is just a rant by now. [S2] [LAUGHS] [S1] Um, you, you do have some- [S2] Yeah. [S1] Sorry, go ahead. | 14.56 | 3.371556 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818526.mp3 | [
4716,
6712,
4362,
3408,
8502,
8043,
3397,
5516,
11062,
5404,
3449,
4842,
5312,
7078,
12037,
9477,
10054,
9670,
6812,
12451,
12377,
9779,
5849,
9753,
7728,
9776,
3112,
10859,
10121,
9956,
5804,
7272,
4713,
7754,
1243,
557,
660,
5999,
12719,
... | [
1.5207031965255737,
-5.4955010414123535,
-0.155094712972641,
0.0015473091043531895,
0.5042995810508728,
-0.3227607309818268,
0.33524778485298157,
-0.0525294654071331,
0.3158743977546692,
-3.867234706878662,
0.2412707358598709,
0.09879428148269653,
0.08441774547100067,
-0.1803397834300995,
... | 364 | 14.56 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818527 | [S1] I was going to say, uh, with respect to what you're saying, there's, the solution doesn't necessarily have to lie on the language model side. [S2] Yeah. [S1] Um, so one thing is you can think of language modeling as just pure density estimation over tokens, right? So if you're doing that, like, of course you're going to model like 4chan, for example, right? But it's up to your generative sampling strategy to remove that part of the density and only sample from, you know, | 26.56 | 2.814624 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818527.mp3 | [
9249,
11940,
11764,
6859,
4307,
10029,
7605,
7404,
9469,
3893,
9845,
10685,
5813,
5801,
9378,
11517,
11259,
5355,
11574,
2952,
6363,
11428,
11477,
5601,
3024,
11758,
5062,
9315,
10074,
9835,
10558,
6134,
3365,
2726,
3227,
3171,
2661,
603,
3... | [
0.5496330261230469,
-5.772242546081543,
-0.08685548603534698,
-0.20934058725833893,
0.9633209109306335,
-0.03765704482793808,
0.5725804567337036,
0.04217405244708061,
-0.08399484306573868,
-3.846281051635742,
0.22265057265758514,
-0.3437137305736542,
-0.20998835563659668,
0.199017092585563... | 664 | 26.56 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818528 | [S1] Uh, select in some sense the mode of the slice of the density. [S2] Yeah. [S1] Right? Um, and so something probably similarly can be done here. Like a great example is like, take Codex, for example, right? I think in the Codex paper, what they do is they generate a ton of samples and then they re-rank the samples | 19.52 | 3.07124 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818528.mp3 | [
3617,
10841,
2278,
8623,
3439,
3950,
3895,
3886,
3950,
3886,
3886,
3813,
3747,
6817,
12377,
12377,
12305,
5289,
5169,
4178,
6505,
8160,
11224,
3035,
7863,
2658,
5478,
10748,
12573,
9313,
3857,
11883,
7454,
10821,
6121,
5537,
5384,
3044,
105... | [
0.5525416135787964,
-5.885440826416016,
-0.09965677559375763,
-0.11601842939853668,
0.8742133378982544,
-0.05397270619869232,
0.6408370733261108,
0.03453966975212097,
-0.12505307793617249,
-3.8659379482269287,
0.2903550863265991,
-0.26794177293777466,
-0.22942236065864563,
0.20956324040889... | 488 | 19.52 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818529 | [S1] Uh, in terms of perplexity, so average log probability. And then they take the mode, so essentially the exact mode of that, uh, density estimation, right? [S2] Yeah. [S1] So one thing to argue is that, you know, you could, you could train language models that do pure density estimation over all the text that we have, and then have | 17.56 | 2.871121 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818529.mp3 | [
6697,
5144,
4191,
6838,
10678,
5494,
6007,
2991,
3447,
5998,
3878,
6817,
8873,
8804,
12631,
10254,
10888,
11216,
6642,
8045,
8468,
8230,
6430,
3009,
333,
8559,
5487,
5798,
2715,
2865,
4257,
6705,
8170,
3562,
3221,
10601,
7088,
2788,
8126,
... | [
0.5594155192375183,
-5.907697677612305,
-0.11178291589021683,
-0.06063157692551613,
0.8960236310958862,
-0.06530411541461945,
0.5430735945701599,
-0.0005859180819243193,
-0.10187441855669022,
-3.829085111618042,
0.26088595390319824,
-0.3093873858451843,
-0.2026548683643341,
0.1402161568403... | 439 | 17.56 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818531 | [S1] Entity linking, right? So in some sense, the genre objective was a subset of our much more general objective. [S2] Yeah. [S1] Right? Uh, and it's not too surprising we beat out genre just because our models are bigger, um, in the, in our fine-tuned case. But the really, really cool thing, I think, was that we can do this, the zero shot, which is exactly what I've shown in the first figure. Um, you know, if you mask out the entity, if you know that you want this entity | 24.64 | 3.213307 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818531.mp3 | [
6184,
8225,
12021,
7150,
7260,
6993,
4049,
4642,
8528,
10010,
10117,
3759,
2595,
2708,
4901,
4693,
9762,
10129,
2325,
7242,
3612,
3610,
4625,
1124,
2844,
2948,
9007,
10063,
8265,
8857,
9308,
10038,
10485,
12076,
12470,
7340,
6297,
8144,
542... | [
0.573757529258728,
-5.763712406158447,
-0.09786508977413177,
-0.10292373597621918,
0.963760495185852,
-0.017107943072915077,
0.5913746356964111,
0.052982136607170105,
-0.03138650581240654,
-3.9105801582336426,
0.32107722759246826,
-0.2860727310180664,
-0.30318620800971985,
0.17916932702064... | 616 | 24.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818532 | [S1] You want to disambiguate this entity. You can place a mask there with this A tag, right? Uh, and then our model will fill in what it thinks the disambiguation is. [S2] Yeah. [S1] So that's kind of cool. Uh, I couldn't find any, like, zero-shot baselines like this, so I think this is kind of the first paper to do this, um, type of zero-shot entity linking disambiguation. | 20.68 | 3.153219 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818532.mp3 | [
6184,
5169,
4314,
9564,
7901,
8061,
8740,
11341,
2752,
5598,
4066,
8144,
9187,
11294,
8490,
9678,
7233,
8977,
5838,
5486,
12758,
8272,
1362,
7478,
5520,
4766,
12219,
10772,
6536,
1491,
5712,
6993,
12677,
6866,
12644,
12661,
12012,
8301,
939... | [
0.6914229393005371,
-5.776357173919678,
-0.15154962241649628,
-0.007957025431096554,
0.7744473814964294,
-0.0881090834736824,
0.549540102481842,
0.0006279288791120052,
-0.1134653314948082,
-3.9005374908447266,
0.27232155203819275,
-0.2064083069562912,
-0.20306271314620972,
0.12092515081167... | 517 | 20.68 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818534 | [S1] Good, actually, at least from a semantic level. [S2] Mm-hmm. [S1] So one problem is that we don't actually generate in the style of, uh, I think MS Coco here. Um, so we didn't report like blue four, uh, numbers or like the standard numbers. But if you look at the semantic, uh, uh, similarity using BERT score, the, the CM3 captioning with clip as a re-ranker is actually a very, very strong baseline. | 27.16 | 3.252657 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818534.mp3 | [
5665,
3601,
3093,
2580,
11811,
11730,
8102,
423,
1502,
4293,
5705,
3723,
7185,
8362,
6696,
12455,
6590,
3516,
4845,
11939,
7640,
12680,
1875,
2643,
2845,
2379,
4690,
10166,
9205,
5814,
5395,
12683,
12164,
2960,
4945,
6137,
5534,
10790,
8054... | [
0.5501031279563904,
-5.7240095138549805,
-0.037700723856687546,
-0.16829009354114532,
0.8874644041061401,
-0.02261306159198284,
0.6495147347450256,
0.022568441927433014,
-0.014927552081644535,
-3.909754991531372,
0.24190576374530792,
-0.31484103202819824,
-0.22760269045829773,
0.2434035837... | 679 | 27.16 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818535 | [S1] Uh, and so you can kind of see the style here is weird. It tries to explicitly state what type of airplane it is. [S2] Yeah. [S1] Uh, uh, but that's kind of a interesting behavior. Uh, so I think definitely at scale, uh, you know, you could get a single model that I think could be competitive with MS Coco. | 18.08 | 2.973956 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818535.mp3 | [
4267,
5648,
7351,
3966,
6454,
6007,
4278,
3382,
1334,
3838,
6765,
3164,
9377,
12385,
12393,
8744,
5673,
3099,
7867,
7209,
6178,
7721,
7187,
7423,
8205,
3930,
11168,
3536,
10613,
7980,
7918,
2775,
1178,
4707,
12675,
10644,
8859,
12712,
9373,... | [
0.6614694595336914,
-5.795372009277344,
-0.11141392588615417,
-0.10258540511131287,
0.8820825219154358,
-0.04488777741789818,
0.6135650277137756,
0.04458494111895561,
-0.11017721891403198,
-3.840183734893799,
0.20424476265907288,
-0.28355857729911804,
-0.2268991619348526,
0.215213492512702... | 452 | 18.08 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818536 | [S1] Um, from, you know, our, I think our 100 million parameter model to our 13 billion parameter model, around the 60 billion mark is where we'll see grounding in this setup. [S2] Okay. [S1] It's kind of a linear plot. So our expectation is that if you scale this up to 60 billion, that you should be able to achieve, I think, language image grounding, which is kind of a cool result that I think a lot of people have been chasing here. Um, and that's 60- [S2] It's, it's, | 25 | 2.903306 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818536.mp3 | [
10009,
5665,
9460,
6582,
3446,
5999,
6447,
3894,
3446,
3894,
3830,
4206,
5662,
6175,
3092,
3602,
3611,
4618,
6658,
3105,
8785,
5658,
3561,
8682,
9201,
4089,
5526,
5550,
6006,
6006,
5943,
6455,
6766,
8222,
8286,
4694,
9313,
5715,
9861,
121... | [
0.5338654518127441,
-5.645522594451904,
-0.03622150793671608,
-0.04927439987659454,
0.8621363639831543,
-0.13120698928833008,
0.4821726083755493,
-0.006160935387015343,
-0.16158968210220337,
-3.9336278438568115,
0.2852034568786621,
-0.27370041608810425,
-0.25192826986312866,
0.193814679980... | 625 | 25 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818537 | [S1] You can actually make these predictions, which is, which is, you know, it's, it's cool. Like, I'm amazed by this. [S2] Yeah. I definitely don't think we're going to be like an order of magnitude off, right? [S1] Yeah. [S2] So, so I think with the 100 billion parameter, 100 billion and 175 billion like 2-3 size, we can get some very, very non-trivial behavior, um, to the point of being competitive across all tasks. Um, | 25.68 | 3.164018 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818537.mp3 | [
3090,
9801,
7380,
9931,
8966,
1568,
5666,
12485,
5446,
4201,
12761,
9709,
8774,
4959,
12093,
9213,
9900,
10155,
10184,
3218,
12117,
7237,
3109,
3118,
6182,
9702,
12781,
11729,
11659,
9827,
10146,
2273,
17,
467,
12566,
12677,
4947,
8073,
419... | [
0.8953903913497925,
-5.533348560333252,
-0.04351796954870224,
-0.09412535279989243,
0.7386606335639954,
-0.1772143840789795,
0.47738316655158997,
-0.02903137356042862,
0.011160250753164291,
-3.9532644748687744,
0.32654038071632385,
-0.24817147850990295,
-0.19892679154872894,
0.073106169700... | 642 | 25.68 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818538 | Um, so, yeah, it's interesting to see what the next, you know, step-wise changes in behavior will be if you scale this up. Um- [S1] With respect to the HTML, right- [S2] Mm-hmm. [S1] ... uh, that you use, which is, I- I thought it was, it was pretty cool because it is data that is, you know, so available, and your argument is a little bit that | 25.04 | 3.011238 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818538.mp3 | [
7413,
3680,
4718,
8574,
5494,
6007,
3895,
6455,
8806,
6307,
9378,
9299,
7334,
9902,
4252,
4144,
3793,
8608,
4048,
8046,
8046,
2862,
5343,
5775,
3212,
6298,
11930,
7270,
4827,
9809,
6657,
3616,
2720,
1569,
3681,
4384,
7653,
11231,
10671,
8... | [
1.1455217599868774,
-5.5781660079956055,
-0.0843309536576271,
-0.13460680842399597,
0.5460092425346375,
-0.08362966775894165,
0.35195329785346985,
-0.055667776614427567,
0.062145158648490906,
-3.85332989692688,
0.4076891243457794,
-0.029649117961525917,
0.038054950535297394,
0.043059155344... | 626 | 25.04 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818539 | [S1] Yeah. So, so in some sense, do, do we want to model every single token? So in, in the case that you have infinite compute? Sure. [S2] Mm-hmm. [S1] Right. Um, but here, there's kind of a min-max problem that you have to solve, right? Which is you want to kind of, you want to maximize the amount of semantic information that is available while minimizing the amount of tokens that you have, right? Um, and, and, and this is, | 24.32 | 3.260757 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818539.mp3 | [
6184,
5673,
6667,
4814,
9605,
11245,
8183,
1462,
3811,
9378,
9371,
9901,
8301,
9453,
6821,
9906,
6696,
4306,
8168,
8673,
6608,
10662,
7990,
10542,
10534,
7902,
5342,
662,
3226,
8801,
9322,
8219,
4262,
6640,
8584,
3547,
10550,
7918,
3222,
... | [
0.5741584300994873,
-5.736083984375,
-0.1057298481464386,
-0.030532846227288246,
0.8990451097488403,
0.020806441083550453,
0.5845118165016174,
0.04744845628738403,
-0.18413224816322327,
-3.867203950881958,
0.2711511254310608,
-0.29120686650276184,
-0.20789115130901337,
0.19763624668121338,... | 608 | 24.32 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818541 | [S1] ... the document structure is described in the minimal set of tokens. [S2] Mm-hmm. [S1] Um, so maybe that's, you know, th- that's a pure engineering project as well. Um- | 8.8 | 3.09608 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818541.mp3 | [
6162,
4619,
3597,
5697,
8172,
12141,
5184,
5588,
7925,
8051,
9389,
8984,
4324,
10790,
11574,
2458,
7040,
3993,
2901,
8053,
5860,
10008,
10129,
5534,
8837,
3907,
1300,
6168,
3010,
5934,
8664,
4953,
9451,
405,
11254,
7527,
3164,
5852,
9311,
... | [
0.8165577054023743,
-5.8433003425598145,
-0.20970144867897034,
-0.035951171070337296,
0.6994810104370117,
-0.00976146012544632,
0.7343937158584595,
-0.05591447651386261,
-0.10539602488279343,
-3.9584991931915283,
0.03257292881608009,
-0.1876157820224762,
-0.24116572737693787,
0.09688514471... | 220 | 8.8 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818542 | [S1] When you, when you think of HTML and the DOM, it is a tree, right? [S2] Mm-hmm. [S1] Which is different from a linear sequence. Uh, do you, do you think there is, do you think there's value in treating the tree as a tree? Do you think it's mainly a limitation of the models we have? They go, let's say, like, | 22.36 | 3.392074 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818542.mp3 | [
3112,
7704,
621,
11253,
9812,
11533,
12035,
8462,
1617,
6697,
1133,
10740,
9732,
11533,
11076,
3933,
10144,
3880,
11697,
11473,
9762,
11417,
8565,
10934,
936,
2334,
12605,
11750,
11725,
7235,
11585,
10185,
4865,
11088,
10133,
8670,
7399,
8558... | [
1.365696907043457,
-5.461874008178711,
-0.18791699409484863,
-0.07458525151014328,
0.48125478625297546,
-0.11823632568120956,
0.3086365759372711,
-0.053443703800439835,
0.1630043387413025,
-3.8053536415100098,
0.4696930944919586,
0.08169235289096832,
0.06827466934919357,
-0.089421145617961... | 559 | 22.36 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818543 | [S1] Yeah. So one thing about transformers is it seems that they can learn the inductive bias of the data fairly well. [S2] Mm-hmm. [S1] And it's not necessarily encoded. Um, so my argument to this is that usually for these large scale runs, the best thing is just to keep it as simple as possible. [S2] Mm-hmm. [S1] Mostly just because they're risky, right? You get one chance. Uh, but the other reason is that transformers are actually highly capable of picking up, um, this type of, | 26.08 | 3.02013 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818543.mp3 | [
6184,
3121,
4610,
4245,
10058,
10626,
10657,
11242,
6071,
6518,
7012,
4458,
4458,
5106,
6122,
10713,
8144,
8171,
10546,
7915,
7900,
10453,
3286,
3737,
9306,
9370,
9370,
12377,
12377,
8728,
1048,
7169,
9770,
4648,
6161,
3616,
6673,
2720,
125... | [
0.6643928289413452,
-5.644629001617432,
-0.04753477871417999,
-0.12482832372188568,
0.8305262327194214,
-0.016311343759298325,
0.4736465513706207,
0.0510065071284771,
0.008662700653076172,
-3.8618321418762207,
0.34627875685691833,
-0.39495471119880676,
-0.24365273118019104,
0.2045432478189... | 652 | 26.08 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818544 | [S1] Uh, structure. [S2] Yeah. [S1] So this isn't in the paper, but we looked at attention scores and, and then you can, you can see very clearly that the model knows what are like boundaries between HTML elements, um, for example. But I, but again, there's also a ton of work to be done as well. So like some exciting work is, I think you also interviewed like Ofer, uh, for the Alibi work, right? Like, | 23.12 | 3.153943 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818544.mp3 | [
6177,
5744,
6162,
4406,
3886,
4342,
3318,
886,
821,
3765,
3100,
3620,
3164,
6689,
5729,
3105,
10019,
8177,
9680,
10057,
9441,
354,
5550,
2934,
9387,
9442,
9680,
10200,
1515,
5590,
3533,
8987,
9954,
9889,
8802,
4126,
5676,
3620,
4299,
1224... | [
0.6702606081962585,
-5.6556243896484375,
-0.09331805258989334,
-0.14130598306655884,
0.797846794128418,
-0.11307184398174286,
0.5256128311157227,
0.024330051615834236,
-0.135187029838562,
-3.9208548069000244,
0.2621646821498871,
-0.33346688747406006,
-0.19316408038139343,
0.120858274400234... | 578 | 23.12 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818545 | [S1] That work is really clever, right? 'Cause it introduces an explicit inductive bias that the further away a token is, the probably less likely that you are to look at it. And it gets rid of the need for, you know, positional, uh, representations. [S2] Yeah. [S1] So you can imagine, like, an extension of Alibi here that would directly encode a tree-like | 17.64 | 3.203776 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818545.mp3 | [
6681,
3090,
1555,
8529,
11771,
8122,
5733,
3379,
8027,
5741,
11481,
4068,
3908,
1472,
2794,
8084,
10925,
5468,
5078,
11931,
12778,
3297,
8043,
10672,
9370,
2995,
2918,
275,
282,
3495,
11765,
12107,
9376,
12065,
6006,
6405,
897,
9127,
7363,
... | [
0.6349491477012634,
-5.812437057495117,
-0.16584734618663788,
-0.058094970881938934,
0.8347009420394897,
-0.017427615821361542,
0.4903106987476349,
-0.048117753118276596,
-0.07567955553531647,
-3.8663461208343506,
0.3523997366428375,
-0.24229192733764648,
-0.226718008518219,
0.147258892655... | 441 | 17.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818547 | [S1] So, this is all stuff that needs to be done in the future as well. Uh, but that being said, I think if you have enough compute, these models can learn anything. It mostly becomes an efficiency angle- [S2] Yeah. [S1] ... uh, going forward. [S2] So, a- about this paper, so what, what I have a bit of a trouble with is, you know, too many things in one paper, which in this case is, it's this idea of using HTML and so on, although there was a previous paper, uh, | 28.6 | 2.888895 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818547.mp3 | [
2600,
5665,
4891,
6552,
3536,
6007,
5423,
5342,
3214,
3732,
6810,
12572,
9444,
4706,
3601,
6340,
1090,
2972,
8556,
5592,
3026,
10461,
2944,
3532,
6755,
10805,
5438,
7926,
10998,
5879,
5879,
2606,
5286,
2838,
723,
6298,
9754,
3346,
8168,
8... | [
1.065171241760254,
-5.512720584869385,
-0.11135951429605484,
-0.015655873343348503,
0.5396607518196106,
-0.07767954468727112,
0.3935372233390808,
-0.08045855164527893,
0.039674706757068634,
-3.854276657104492,
0.25036972761154175,
-0.04074300825595856,
0.02725587598979473,
0.10765967518091... | 715 | 28.6 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818549 | [S1] I think both Google and Codex have similar sizes. [S2] Mm-hmm. [S1] While being able to have a bidirectional, uh, bidirectional option. [S2] Yeah. [S1] So, um, there are a couple teams within Facebook that are trying out this objective with some success. Um, so, uh, there will be future work about this. [S2] Excellent. [S1] Absolutely. | 19.76 | 3.01203 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818549.mp3 | [
6217,
4130,
7415,
3073,
7060,
12366,
9765,
2634,
19,
3097,
6193,
10528,
2871,
7982,
9796,
1298,
8289,
9323,
6555,
7917,
10397,
3214,
11931,
3359,
10412,
10342,
7710,
7774,
7782,
5326,
147,
7267,
7406,
10094,
12702,
12711,
10159,
7534,
8789,... | [
0.8227458000183105,
-5.72928524017334,
-0.17533573508262634,
-0.06289814412593842,
0.7253779768943787,
0.015758901834487915,
0.6482430100440979,
-0.024681303650140762,
-0.12339013069868088,
-3.8842244148254395,
0.3485275208950043,
-0.3002658486366272,
-0.16704711318016052,
0.06641860306262... | 494 | 19.76 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818550 | [S1] ... Speech tokens, right? [S2] Mm-hmm. [S1] Um, very simple. I think they use K-means. I might be wrong, though. Uh, uh, just to find discrete tokens for speech. So imagine how you have a single model that has video, images, you know, text. [S2] Yeah. [S1] Uh, right. Speech, everything kind of put into one, right? Like, what level of grounding and what level of zero-shot prompting can you get here? Um, and I think a lot of people are kind of chasing this at, at the bigger companies. | 25.68 | 3.154738 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818550.mp3 | [
8088,
4249,
9195,
11650,
6915,
11585,
12747,
4808,
11600,
4596,
2862,
661,
12458,
5020,
4198,
6157,
4301,
1984,
4241,
620,
389,
11630,
12687,
6800,
3212,
8907,
6492,
3733,
1163,
9762,
5730,
7180,
8371,
4676,
1045,
1172,
6760,
5665,
8745,
... | [
0.953518271446228,
-5.608548164367676,
-0.13336127996444702,
0.036005470901727676,
0.7650227546691895,
-0.13604699075222015,
0.4859637916088104,
0.009496111422777176,
-0.03198367357254028,
-3.9043350219726562,
0.3956448435783386,
-0.3507716953754425,
-0.27314308285713196,
0.075426630675792... | 642 | 25.68 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818552 | [S1] Is it just 'cause we don't have, like, bidirectional, like, masks? Is that one? Is it because we only mask for, like, causal models in upper triangular matrix? Is there something more fundamental there? I think kind of peeling that apart and figuring out what's going on there is kind of important too. Um, but I think we're very early on. Uh, I think, I think this year is gonna be the year of multimodal. [S2] Yeah. [S1] I think, you know, they kind of kick stuff off. So I'm kind of excited to see what other groups | 28.68 | 3.068643 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818552.mp3 | [
5672,
8804,
10006,
5504,
10716,
12565,
11520,
11713,
8549,
5584,
9379,
9194,
3340,
898,
5363,
12676,
8907,
7997,
8350,
7474,
12213,
11134,
3172,
3361,
2653,
3446,
7581,
9931,
11875,
5083,
11873,
9818,
5345,
4641,
3617,
7233,
5352,
4128,
102... | [
0.7815670967102051,
-5.566966533660889,
-0.017502833157777786,
-0.05448583886027336,
0.7984830141067505,
-0.052847735583782196,
0.5760988593101501,
0.021310575306415558,
-0.04271530732512474,
-3.8599984645843506,
0.34944602847099304,
-0.3337513208389282,
-0.2743844985961914,
0.211369454860... | 717 | 28.68 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818553 | [S1] ... So working on. [S2] It seems like it, yeah. Um, is there anything else about the paper or the research direction you want to shout out? You want people to know that we haven't mentioned so far? [S1] Yeah, I mean, we'll be releasing all this code really, really soon. Um, we're just waiting on some internal approvals so people will get to play around with it. Um, I think we'll release 3 billion model, but the 13 billion model is the one that really shines. [S2] Yeah. [S1] So if people can get that running, I think it's really cool. I've spent hours just playing around with it. It's | 28.32 | 3.331201 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818553.mp3 | [
3172,
5080,
6592,
6391,
1189,
44,
3125,
3421,
3413,
11874,
6754,
515,
6158,
6357,
3814,
3830,
9390,
3748,
9386,
9825,
6680,
6152,
4620,
2652,
7190,
6678,
9772,
7190,
4173,
2659,
6801,
8284,
4890,
5201,
336,
1812,
12612,
8853,
8957,
3914,
... | [
1.1760562658309937,
-5.422362804412842,
-0.038247838616371155,
-0.07319270074367523,
0.5774392485618591,
-0.20787201821804047,
0.4421811103820801,
-0.016926752403378487,
0.193706214427948,
-3.9227957725524902,
0.3589385151863098,
-0.14553046226501465,
-0.15620839595794678,
0.05418905988335... | 708 | 28.32 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818554 | [S1] Nice. What does it, what does it take to, like, just to forward propagate? What's, like, the minimal configuration? [S2] Um, so with the recent Deep Speed stuff that was released for inference, I'm not really sure, 'cause I think they said that you can use one GPU for, like, a 6.7 billion model. [S1] Yeah. [S2] So if you do model parallelism, I think you need two GPUs. | 22.24 | 3.295453 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818554.mp3 | [
9504,
3661,
8405,
3958,
9140,
4397,
10134,
10139,
3425,
3812,
3418,
3282,
5713,
6305,
8296,
3104,
3803,
6216,
12645,
12603,
8957,
4852,
9434,
6884,
174,
10740,
9825,
8128,
10546,
2826,
2962,
12180,
7169,
3291,
3830,
6956,
3308,
10612,
9874,... | [
0.9448638558387756,
-5.707254886627197,
-0.07434029877185822,
-0.1463700234889984,
0.5685898661613464,
-0.11460467427968979,
0.45863375067710876,
0.0007776094134896994,
0.07413630187511444,
-3.9117443561553955,
0.37723323702812195,
-0.1061607375741005,
-0.18074746429920197,
0.1126175224781... | 556 | 22.24 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818555 | [S1] But if, like, without that, just give us a ballpark, you know, what, what, what would it, what would it be like forward propping through this model? [S2] Yeah. So, so one thing is you could do it on a CPU if you have a strong enough CPU. [S1] Yeah. [S2] Uh, but, but for, for inference, I think what I used was four V100s. [S1] Yeah. [S2] Model parallel. So, uh, less than a node. | 22.4 | 3.256148 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818555.mp3 | [
4715,
1562,
10474,
10623,
1252,
4642,
6430,
9436,
12565,
8053,
6005,
3501,
6903,
9436,
5980,
10747,
11757,
9810,
10211,
3292,
610,
3675,
3291,
3163,
5217,
3156,
2858,
3688,
2578,
6213,
6691,
3604,
5214,
8212,
6185,
6155,
5150,
6162,
3097,
... | [
1.0906426906585693,
-5.604650497436523,
-0.12375261634588242,
-0.12659530341625214,
0.5787709951400757,
-0.11491426825523376,
0.43499815464019775,
-0.022936176508665085,
0.0061752405017614365,
-3.8497467041015625,
0.25343137979507446,
-0.08501987904310226,
-0.06201673299074173,
0.108740597... | 560 | 22.4 |
bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818556 | [S1] Cool. Excellent. Well, Armin, thank you so much for being here. This was really cool. Um, really value the, like, also the kind of behind the scenes in insights we got here. And- [S2] Yeah. [S1] ... I hope to see you again very soon with even, like, CM4. [S2] [LAUGHS] Yeah, thank you for having me. [S1] Excellent. | 20.44 | 3.392068 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p171_BV17h4y1j7aZ_p171_m4-dialogue_0818556.mp3 | [
5665,
12699,
2931,
7833,
10392,
2653,
3746,
12385,
8865,
9306,
7218,
9992,
6768,
11803,
11765,
11619,
11370,
9707,
8664,
5568,
10484,
5341,
886,
6174,
12298,
8272,
4050,
4124,
1132,
3310,
12150,
8375,
3230,
6038,
4971,
9385,
9313,
7849,
720... | [
1.5059810876846313,
-5.360167026519775,
-0.18910932540893555,
-0.07109256833791733,
0.44633978605270386,
-0.20187582075595856,
0.33668822050094604,
-0.049946822226047516,
0.21716631948947906,
-3.8334686756134033,
0.4099733829498291,
0.03670360520482063,
-0.012910324148833752,
-0.0679633989... | 511 | 20.44 |
bilibili_data_1642432739_BV1cL4y1g74Y_p30_BV1cL4y1g74Y_p30_m4-dialogue_0894696 | [S1] It's where things happen. [S2] Time. It's when things happen. [S1] We can measure where things are and when things take place, but in modern physics, we realize when and where are actually part of the same question. [S2] Because when it comes to understanding the universe, we need to replace three-dimensional space plus time with a single concept, four-dimensional space-time. | 26.56 | 3.089146 | 24,000 | audio/en/bilibili_data_1642432739_BV1cL4y1g74Y_p30_BV1cL4y1g74Y_p30_m4-dialogue_0894696.mp3 | [
3729,
5745,
8171,
4606,
4867,
9859,
7435,
7650,
5984,
3041,
416,
979,
3026,
9524,
12459,
9827,
12572,
12403,
9363,
6688,
9435,
9482,
6705,
11494,
11165,
11666,
4112,
10825,
833,
3017,
3552,
3739,
2660,
243,
2675,
3814,
7517,
10085,
12709,
... | [
0.31994614005088806,
-5.6269426345825195,
-0.0986611396074295,
0.15097784996032715,
0.7636340260505676,
-0.03395744040608406,
0.5438377857208252,
-0.05487122759222984,
0.31222692131996155,
-3.9342737197875977,
0.436442106962204,
0.057387422770261765,
-0.19877976179122925,
0.028675794601440... | 664 | 26.56 |
bilibili_data_1642432739_BV1cL4y1g74Y_p30_BV1cL4y1g74Y_p30_m4-dialogue_0894697 | [S1] We'll explore and explain space-time in this series of animations. [S2] Animations? [S1] Yeah. [S2] Well, we're not very animated, are we? [S1] Sure we are. Look, I can go from here to here. [S2] Oh, how'd you get from here to there? How fast did you go? Did you run, walk? Did you even go in a straight line? [S1] Ah, to answer that, you'll need to make our cartoon physics look more like physics physics. You'll need more panels. [S2] More panels, please. | 29.44 | 2.901307 | 24,000 | audio/en/bilibili_data_1642432739_BV1cL4y1g74Y_p30_BV1cL4y1g74Y_p30_m4-dialogue_0894697.mp3 | [
9425,
12701,
11692,
10030,
7452,
11994,
3764,
7979,
11721,
7891,
12247,
12379,
7068,
2984,
4049,
3931,
8937,
1828,
5477,
10871,
10853,
10853,
10862,
7974,
4774,
5053,
4646,
8898,
2015,
7184,
11476,
4073,
4432,
9561,
6194,
10592,
5853,
4023,
... | [
0.32622992992401123,
-5.519548416137695,
0.05476725473999977,
0.12968939542770386,
0.8538985252380371,
-0.019434822723269463,
0.5071538090705872,
-0.03255452960729599,
0.31163880228996277,
-4.021285057067871,
0.30704620480537415,
0.03025745414197445,
-0.21562090516090393,
0.075058020651340... | 736 | 29.44 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481719 | [S1] Yeah. The news is so negative. I mean, it makes me sad to read the news, frankly. [S2] I, I, well, let me ask you a question. I don't watch, I don't watch network news and I don't read any newspapers. They couldn't pay me enough money- [S1] [LAUGHS] Yeah. [S2] ... to do that. [S1] Every time I, I, I'll accidentally read the news and I'll just be sad. [S2] [LAUGHS] It's insane. [S1] It's, | 22.4 | 3.119184 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481719.mp3 | [
4322,
8865,
9717,
12188,
6947,
6817,
9558,
9987,
3928,
3738,
12758,
11735,
9647,
1959,
1398,
4388,
9442,
9901,
8356,
9844,
7325,
6821,
6820,
9827,
3729,
9135,
5186,
12757,
10578,
11035,
6410,
858,
11574,
3418,
2912,
411,
11069,
8028,
6668,
... | [
0.5739399194717407,
-5.681005001068115,
-0.0761839896440506,
-0.04056963324546814,
0.5425477027893066,
-0.07011830061674118,
0.655407190322876,
-0.07866238802671432,
-0.0963224470615387,
-3.8916563987731934,
0.12403877079486847,
-0.07530731707811356,
0.024703821167349815,
0.051118575036525... | 560 | 22.4 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481721 | [S1] Yeah, we've got more access to food, energy, water, healthcare, education on the planet than ever before. I mean, people would just start watching horror movies, I think, instead. [S2] Yeah. [S1] Yeah, th- the challenge is, it's, uh, it's our neural nets, the wiring of our brains, you know, evolved in a world of constant danger. And so we're sort of just wired for fear and scarcity constantly. [S2] Yeah. Uh, well, | 26.6 | 2.803035 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481721.mp3 | [
10148,
3955,
6250,
10156,
9708,
12771,
10090,
7402,
1106,
9550,
9117,
9196,
7658,
9206,
4900,
6243,
10031,
12653,
10037,
6817,
5988,
12172,
162,
10075,
11191,
10749,
4825,
4099,
5733,
10486,
10604,
5973,
9135,
6526,
11709,
7341,
9578,
3400,
... | [
0.35159265995025635,
-5.749480724334717,
0.09324388951063156,
-0.1633920669555664,
0.37361639738082886,
-0.05935313180088997,
0.6562464237213135,
-0.025054533034563065,
-0.014584065414965153,
-3.9583280086517334,
0.15014851093292236,
-0.04778297618031502,
-0.10422588139772415,
0.1985448598... | 665 | 26.6 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481723 | [S1] And, and, you know, the reality is, you know, the news media has one job, to deliver your eyeballs to their advertisers. And when we pay ten times more attention to negative news and positive news, that's all we get 24/7. So, I mean, listen, I, I- [S2] It's inevitable. [S1] Yeah, it, it is. I mean, I, I do get my, my news | 21.24 | 2.943186 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481723.mp3 | [
7524,
7324,
12398,
12213,
11253,
11181,
11932,
10765,
3266,
4244,
9875,
6812,
9965,
11253,
11988,
10325,
3282,
10677,
6364,
4257,
9378,
9492,
6745,
12117,
11989,
8029,
5431,
11030,
3860,
6493,
6795,
8850,
6178,
3596,
4145,
9997,
9491,
6290,
... | [
0.3477233350276947,
-5.854247570037842,
0.06358630955219269,
-0.06272687017917633,
0.33390212059020996,
-0.13172604143619537,
0.5000746250152588,
0.003876785282045603,
-0.06764929741621017,
-3.9292845726013184,
0.09794729202985764,
-0.08940735459327698,
-0.05490098148584366,
0.192083612084... | 531 | 21.24 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481724 | [S1] ... invariably, you know, on my, on my feed on X, but I also get all the great things happening in the world 'cause I can selectively choose to watch that. But when you're watching TV or, you know, in the newspaper, uh, some editor someplace or some producers deciding what gets fed into your mind and it can really screw with your mindset. [S2] Um, yes, exactly. [S1] Yeah. | 24.84 | 3.196662 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481724.mp3 | [
5673,
5170,
9830,
11614,
8966,
3846,
3355,
5859,
2858,
3058,
12767,
5110,
6581,
8084,
5509,
11660,
8101,
5669,
8042,
5277,
10661,
11146,
11651,
11588,
11077,
11651,
11597,
4240,
8234,
6193,
11925,
9862,
8452,
1463,
372,
294,
2783,
5327,
315... | [
0.5105673670768738,
-5.699851989746094,
0.22163674235343933,
-0.1561499685049057,
0.31128454208374023,
-0.22489403188228607,
0.5451126098632812,
0.04674771800637245,
0.011972049251198769,
-3.9932234287261963,
0.2390805333852768,
-0.0852101594209671,
-0.06530223786830902,
0.1719691157341003... | 621 | 24.84 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481725 | [S1] Like, I mean, I mean, obviously we go back a ways, um, but the, I remember when we were at Adeo's party in Brazil. [S2] Yes. Yeah. [S1] How long ago was that? [S2] I don't know. He was, it was his 40th or was it his 30th? I don't know. [S1] No, it was, it was his 30th. [S2] Wow. [S1] This, this was just when SpaceX was- [S2] Yeah, it was- [S1] ... and we're just when I was filming SpaceX. [S2] Yeah. | 25.4 | 2.874678 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481725.mp3 | [
9371,
2153,
6757,
10619,
12115,
12385,
3352,
714,
8922,
5848,
3739,
10147,
9515,
9499,
9955,
6307,
4451,
4764,
8997,
9694,
7587,
7387,
7251,
6185,
11985,
11938,
5290,
5154,
4772,
8877,
9260,
12029,
7038,
7406,
12605,
12534,
7907,
12691,
115... | [
0.6521169543266296,
-5.571157455444336,
0.003599427407607436,
-0.15074856579303741,
0.3821430206298828,
-0.03410143405199051,
0.6216983795166016,
-0.05405736342072487,
-0.09610505402088165,
-3.901283025741577,
0.2064809650182724,
0.0761306881904602,
-0.1500876098871231,
0.05048080161213875... | 635 | 25.4 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481727 | [S1] Right? I mean, what's more, what's, what's a more important metric? You know, here are the numbers. 90% of the world was in global extreme poverty in, in 1800s. In 1981, it was 42%. Today, it's under 10% of the world, right? [S2] Yeah. [S1] Uh- [S2] Hun- hunger, hunger is actually rare, and it used to be common. [S1] It- | 21.28 | 3.227569 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481727.mp3 | [
6193,
3136,
2086,
6446,
2964,
5943,
4086,
7647,
6606,
6614,
4163,
4170,
6820,
8427,
9315,
9908,
3511,
7399,
10909,
12741,
12421,
8396,
3685,
2675,
10673,
11931,
6545,
3408,
8797,
10854,
10780,
2798,
7853,
8438,
4193,
4198,
2611,
10683,
1168... | [
0.4738005995750427,
-5.7817769050598145,
-0.0033695935271680355,
-0.030040891841053963,
0.38578540086746216,
-0.0642073005437851,
0.39298054575920105,
-0.04959545657038689,
-0.1602928191423416,
-3.9549779891967773,
0.23464620113372803,
-0.004106639884412289,
0.006437910720705986,
0.1213140... | 532 | 21.28 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481728 | [S1] Yes, I, I mean, uh, I mean, at, at Tesla, we've, we've made a couple of presentations, one sort of simplistic, and then one in, in extreme detail, um, on how to make Earth completely, uh, self-sustaining from an energy standpoint. [S2] Sure. [S1] Um, and, and demonstrating that there, that there is no, that if you break down all of the raw materials for a lithium-ion battery and for solar, um, you can easily make Earth | 27.92 | 3.213239 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481728.mp3 | [
8226,
3812,
3684,
8298,
3114,
6177,
9941,
9539,
11684,
4597,
2325,
4512,
4002,
2205,
10933,
10740,
9581,
6752,
8886,
10014,
8508,
6758,
11691,
9539,
7759,
6738,
8354,
9316,
9828,
8811,
4708,
3732,
12131,
9330,
12049,
12514,
12514,
12506,
11... | [
0.5126827955245972,
-5.512977123260498,
0.03257971256971359,
-0.183034285902977,
0.3789333999156952,
-0.06620814651250839,
0.4763217568397522,
-0.050994232296943665,
-0.23033420741558075,
-3.8681862354278564,
0.18289238214492798,
0.01384265348315239,
0.06800240278244019,
0.0962862148880958... | 698 | 27.92 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481729 | [S1] and, and platinum. And even though the Earth's crust is, you know, 8% bauxite, you know, basically aluminum, it was just so energetically difficult. And it wasn't that it was scarce, it just wasn't in usable form yet. And that's what technology does. It takes something which is scarce and not usable and makes it usable, right? So- [S2] Yeah. [S1] And- | 21.04 | 3.261149 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481729.mp3 | [
7265,
6172,
12021,
9726,
5047,
1967,
4951,
6150,
9359,
7302,
1537,
8449,
7332,
11956,
2479,
5647,
2059,
10521,
9201,
2851,
5558,
4542,
12221,
4892,
6793,
8846,
2942,
3764,
551,
4710,
9321,
8298,
7270,
4709,
3244,
4701,
4709,
4252,
6877,
6... | [
0.2846561074256897,
-5.850759506225586,
0.04041415452957153,
-0.1348801553249359,
0.396831214427948,
-0.16852040588855743,
0.46577492356300354,
0.008405067026615143,
0.02384212426841259,
-3.924745798110962,
0.10341702401638031,
-0.012569678947329521,
0.006129756104201078,
0.133826434612274... | 526 | 21.04 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481730 | [S1] Uh, yeah, you do need a lot of energy to, um, turn, uh, aluminum oxide into aluminum. Um, but, but, but yeah, it, it, it, in World War II, there was a massive scarcity of aluminum for aircraft. [S2] Sure. Sure. [S1] Um, and, uh, the, in fact, in Britain, the, the mosquito, uh, sort of, uh, fighter bomber was, uh, made of, mostly of wood. Um, and, but it was, it was done with, uh, | 28.04 | 3.263307 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481730.mp3 | [
9308,
12150,
6126,
3951,
6503,
9135,
6567,
7071,
3943,
6559,
4876,
4242,
7323,
11298,
8226,
9932,
11651,
11181,
1533,
1460,
2193,
9389,
12574,
2569,
6017,
9323,
12622,
9987,
10116,
11598,
5632,
8131,
10570,
8195,
11667,
12692,
5387,
2662,
8... | [
0.6141500473022461,
-5.400629997253418,
0.12878239154815674,
-0.11571364849805832,
0.5324649810791016,
-0.12179844081401825,
0.5371506214141846,
-0.0073394388891756535,
-0.09151153266429901,
-3.9209916591644287,
0.24981087446212769,
-0.11463813483715057,
-0.004344810266047716,
0.0854840055... | 701 | 28.04 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481731 | [S1] And that, and then, then we get technology, we get better, uh, better mechanisms of extracting the aluminum from the, from the aluminum oxide, from the bauxite. And this happens over and over again. In fact, that's just what we do. I mean, I think the number was last year in 2023, or maybe in 22, we had more elect, new electricity production from solar than from any other form. And, and- [S2] Yeah. [S1] ... and you've done an extraordinary job on battery production. | 28.04 | 3.093323 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481731.mp3 | [
9876,
11637,
11766,
12781,
7381,
7251,
12661,
12772,
11842,
8261,
3587,
5611,
12732,
11772,
4889,
9251,
12790,
11269,
5643,
3345,
11694,
12708,
7691,
10445,
5862,
603,
12451,
9314,
5733,
84,
4238,
2961,
12279,
12581,
3181,
10595,
9931,
12187,... | [
0.4022136628627777,
-5.661076545715332,
0.19439178705215454,
-0.1326889842748642,
0.26389241218566895,
-0.1927061825990677,
0.44452157616615295,
0.010719243437051773,
-0.034562572836875916,
-3.983431339263916,
0.17970576882362366,
-0.03329652175307274,
-0.09815915673971176,
0.2027355283498... | 701 | 28.04 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481732 | [S1] Yeah, and the battery production is growing, um, actually almost at, at, well, at several times the rate of vehicle production. [S2] Mm-hmm. [S1] So, um... | 11.44 | 3.106698 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481732.mp3 | [
4779,
2592,
6932,
11651,
11692,
4471,
3812,
9394,
6809,
3796,
4709,
11132,
12725,
10829,
7234,
9304,
11495,
7030,
6761,
8934,
9710,
1096,
3730,
6454,
10920,
3566,
11755,
6900,
11938,
4968,
11542,
6235,
8817,
2932,
6476,
8062,
6830,
12058,
7... | [
0.8169021606445312,
-5.665441989898682,
-0.14361624419689178,
-0.027425095438957214,
0.34013012051582336,
-0.061576198786497116,
0.43968698382377625,
-0.05857221782207489,
-0.18765363097190857,
-3.7881455421447754,
0.20191508531570435,
-0.027536578476428986,
0.03549293056130409,
0.00547230... | 286 | 11.44 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481733 | [S1] You know, in, in some cases, almost 10x the, the rate of vehicle production. Um, so, so, yes, the, there's massive demand for batteries and, and, you know, as, as the world uses more, uh, electricity, uh, th- there's actually a lot more capability that the grid has if you can buffer the energy, uh, than without it because m- the vast majority of elect- electrical grid- [S2] Sure, it's wasted. [S1] ... assume no buff- they assume no buffering. [S2] Yeah. | 27 | 2.967002 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481733.mp3 | [
7202,
9314,
10199,
8461,
886,
926,
7234,
4817,
9323,
11483,
8868,
12690,
12641,
8811,
9250,
9062,
6990,
5256,
1739,
12654,
8987,
3544,
11197,
5861,
11931,
5107,
12198,
10183,
3920,
2832,
3349,
6594,
1411,
11382,
7870,
8413,
8221,
5494,
5991... | [
0.5428128242492676,
-5.569400787353516,
0.08036838471889496,
-0.12437433004379272,
0.39828917384147644,
-0.1646723598241806,
0.46503493189811707,
-0.05910944938659668,
-0.16516324877738953,
-3.8400089740753174,
0.18456071615219116,
-0.03497067093849182,
0.06422217935323715,
0.0773911923170... | 675 | 27 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481734 | [S1] Yeah. Um, I was talking about another category, communications. Um, another area that you're revolutionizing. I think the number right now, I just was, uh, checking it earlier, it's like 6.9 billion smartphone users in 2023. Like 86, 86%. And that's what I got when I, when I Googled it. Um, I don't know what Google- [S2] It's 6.9, you don't say. [S1] Yeah. Uh- [S2] [LAUGHS] It's 6,900 million. | 26.72 | 3.111065 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481734.mp3 | [
6892,
6299,
12634,
6940,
10142,
5102,
380,
821,
6244,
6899,
9388,
8868,
8923,
9507,
6307,
12586,
7324,
11987,
12635,
12205,
8125,
8502,
12070,
8212,
10854,
6749,
4699,
3660,
3619,
12042,
8913,
9938,
9446,
9396,
7333,
9340,
6830,
6886,
7394,... | [
0.31197622418403625,
-5.694112300872803,
0.10271486639976501,
-0.19345852732658386,
0.32927206158638,
-0.13789184391498566,
0.4437139630317688,
-0.005466351751238108,
-0.15528565645217896,
-3.931891679763794,
0.16570813953876495,
-0.07803751528263092,
-0.04708823561668396,
0.15259902179241... | 668 | 26.72 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481735 | [S1] Okay. Uh, yes, that just went up. Um, so, you know, we, we still have to prove that it works and all, but, um, we, we're confident that even if the, if these early satellites don't work, we're confident from, from a physics standpoint that it can work. [S2] Yeah. [S1] Um, it, it's, it is, uh, a challenge because we have to emulate, uh, a self, self-tower on the ground in order for the phones to, | 26.64 | 2.947395 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481735.mp3 | [
9762,
9242,
8933,
6063,
6876,
7588,
8687,
2341,
4201,
9314,
6754,
8740,
11255,
4406,
3177,
9315,
6755,
10875,
5731,
4658,
11413,
9507,
8859,
10005,
11994,
9379,
11355,
6820,
6244,
9331,
9323,
9436,
9314,
10983,
11246,
8566,
4309,
9874,
1210... | [
0.639380693435669,
-5.511870861053467,
0.015673240646719933,
-0.06968198716640472,
0.27951574325561523,
-0.09052517265081406,
0.356655478477478,
0.002907998627051711,
-0.36352959275245667,
-3.818408966064453,
0.19459375739097595,
0.04283774271607399,
0.10489417612552643,
0.0946410000324249... | 666 | 26.64 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481737 | [S1] ... a mortality under the age of five was 42%, a couple hundred years ago. It was a coin flip of whether your kid survived. [S2] Oh. [S1] Um, and it's decreased now to under 5%, and it's gone down by 50% in the last 30 years. Um, so just childhood mortality and women dying in childbirth, all of these things, people don't think about when they're listening to all the news on, and all the issues. | 29.2 | 2.813934 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481737.mp3 | [
3104,
9313,
11502,
4327,
8230,
5742,
2933,
469,
6168,
11217,
7091,
12790,
9662,
6206,
2781,
8695,
9822,
2819,
12614,
10181,
4395,
11317,
5438,
9590,
11270,
3468,
8142,
3780,
7132,
10061,
9379,
12661,
10197,
9671,
9862,
5761,
10058,
4007,
58... | [
0.3426091969013214,
-5.6629838943481445,
0.11082476377487183,
-0.21150703728199005,
0.23721662163734436,
-0.1430283635854721,
0.4630242884159088,
-0.04809463396668434,
-0.1540229618549347,
-4.019983768463135,
0.21203982830047607,
0.04074077680706978,
-0.09995043277740479,
0.193222150206565... | 730 | 29.2 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481738 | [S1] And then life expectancy, my favorite subject, uh, has gone up from, you know, 30 years old to, uh, to 75 plus. I still disagree with you on longevity, though. [S2] That, that we should solve it or not, like we should, you think we should solve it? [S1] Well, I, I, listen, I'm not necessarily saying live forever, but I'd like to make it to 120, 150. Um, uh- [S2] Yeah, I, I sort of wonder if we should not solve it too soon. | 28.64 | 3.003282 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481738.mp3 | [
12498,
3616,
6831,
9511,
7819,
12206,
12261,
8717,
3724,
6253,
2653,
5477,
8122,
12241,
10010,
8048,
4510,
10843,
6634,
1418,
10528,
4015,
7159,
9899,
6737,
6043,
11869,
11396,
8673,
6608,
12685,
6158,
8366,
5951,
12716,
9037,
9008,
3067,
1... | [
0.6267008185386658,
-5.632920742034912,
0.03704003244638443,
-0.18179714679718018,
0.40855783224105835,
-0.0052230400033295155,
0.47836801409721375,
-0.013238268904387951,
-0.11652261018753052,
-4.022156715393066,
0.16604122519493103,
-0.07812990993261337,
-0.04395095258951187,
0.191374748... | 716 | 28.64 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481739 | [S1] ... which are quite, quite helpful. The, the, the cells in our body all age, age at basically, almost exactly the same speed. [S2] Mm-hmm. [S1] Um, like, what, like, I've not seen anyone who has an old left arm and a young right arm. [S2] [LAUGHS] That's correct. [S1] I mean, I've never seen that, not even once. [S2] Yes. [S1] But, so how, how are the cells communicating? And how, what is keeping them, what is synchronizing their behavior? Um, | 26.64 | 2.827838 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481739.mp3 | [
6363,
11930,
6822,
4535,
6336,
12680,
3374,
2918,
789,
6433,
9371,
9243,
6682,
9755,
9971,
4266,
2862,
4708,
9086,
4519,
4438,
6752,
8820,
1258,
8063,
7591,
4435,
8502,
5366,
3742,
5360,
1460,
5751,
5342,
3795,
8809,
11998,
9908,
12533,
8... | [
0.5481727123260498,
-5.515572547912598,
0.004907405003905296,
-0.035289864987134933,
0.23610803484916687,
0.05726614221930504,
0.4323263168334961,
-0.06043277308344841,
-0.4245128929615021,
-3.8570122718811035,
0.11900020390748978,
-0.10372690111398697,
0.17541402578353882,
0.2364542037248... | 666 | 26.64 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481743 | [S1] ... two and a half percent, two and a percent, two and a half percent have an aneurysm, 14.4% either have metabolic disease, coronary disease, uh, neurodegenerative disease, and- [S2] Okay. [S1] ... and so your body is incredibly good. | 15.88 | 2.822103 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481743.mp3 | [
4139,
11137,
12768,
8405,
8020,
11781,
8623,
4959,
10162,
10158,
12149,
11134,
4010,
6761,
6073,
8086,
5098,
8528,
4005,
12647,
12783,
11340,
7306,
6680,
7399,
822,
6419,
10211,
8621,
8980,
9502,
8865,
1019,
8094,
5968,
3428,
12077,
6886,
4... | [
0.3166489899158478,
-5.969777584075928,
0.14238940179347992,
0.02988799661397934,
0.26786941289901733,
-0.1402146816253662,
0.5105261206626892,
-0.04941176623106003,
-0.13698126375675201,
-3.9343788623809814,
0.07545990496873856,
0.1811889111995697,
0.0005351079744286835,
0.161046460270881... | 397 | 15.88 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481745 | [S1] Um, well, none that I know of. [S2] [LAUGHS] Yeah. [S1] Well, they might, I mean, they might be, probably are a few schools that are doing it, but probably 99% of schools are not. | 9.68 | 3.045368 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481745.mp3 | [
6292,
9811,
9491,
9874,
8859,
11261,
9079,
6502,
6446,
8566,
11566,
9014,
11124,
8853,
5669,
5677,
6369,
5930,
2741,
8044,
10476,
7869,
5294,
3221,
3785,
6305,
12570,
12393,
12434,
8979,
9113,
9433,
11419,
9813,
8411,
9771,
12626,
9298,
119... | [
0.6229205131530762,
-5.767107963562012,
-0.14123395085334778,
-0.033462926745414734,
0.3748718202114105,
0.06640351563692093,
0.5557252168655396,
-0.07484778016805649,
-0.2679169178009033,
-3.8783419132232666,
0.11147250235080719,
-0.1450842320919037,
0.15391281247138977,
0.021985312923789... | 242 | 9.68 |
bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481747 | [S1] Um, I mean, it's a way of giving every child on the planet the best education. I mean, you, you funded back years ago, if you remember, the Global Learning X Prize that we did. We did a, uh, that we, we demoed in Tanzania with AIs on tablets. It was the earliest days of, of AI. Uh, Imad Mustafino was one of the, uh, winners of that competition who went on the hunt to create stability. [S2] Yeah. [S1] Um, I mean, the, the challenges, I don't think the educational systems are gonna give up, uh, | 29.96 | 3.00347 | 24,000 | audio/en/bilibili_data_1650386044_BV1cK411x7Yh_BV1cK411x7Yh_m4-dialogue_0481747.mp3 | [
8336,
7211,
12206,
8055,
9014,
6903,
3894,
3614,
6239,
4191,
4686,
2133,
6243,
10027,
8804,
9258,
12053,
9579,
9307,
10004,
5795,
11411,
12634,
9251,
6755,
9773,
10750,
9943,
10468,
12613,
11220,
12445,
6609,
474,
5806,
2676,
11189,
12184,
... | [
0.617484450340271,
-5.601032733917236,
0.18340617418289185,
-0.15366500616073608,
0.43432480096817017,
-0.052776940166950226,
0.23552091419696808,
-0.03700026497244835,
-0.17544259130954742,
-3.998598575592041,
0.24890054762363434,
-0.14834696054458618,
-0.11183495819568634,
0.146967992186... | 749 | 29.96 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.