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_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966639 | [S1] correlated with, uh, with each other. Uh, [LAUGHS] but- [S2] Yeah. Yeah, I see what, I see what you mean, yeah. So, um, we're, we're, we're, we've been, we've been going through this a little bit. There's all, I mean, there's a lot of, there's other papers which, which are definitely also interesting, like the, the gloss- [S1] Yeah, I, I really wanted to- [S2] ... the gloss one, the super interesting. Is there, yeah, is there one that you wanted to touch on particularly? | 24.84 | 3.262572 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966639.mp3 | [
10133,
4627,
12131,
821,
5686,
5941,
469,
3989,
5405,
2038,
4558,
4740,
422,
1263,
3878,
6154,
5185,
3172,
2798,
5007,
5713,
5728,
3586,
3915,
11228,
9719,
1388,
3814,
51,
5485,
4741,
916,
4997,
7362,
12233,
9168,
3439,
3895,
3815,
3347,
... | [
0.6759634613990784,
-5.626307487487793,
0.03391782194375992,
-0.10582918673753738,
0.5180732011795044,
-0.0332724004983902,
0.355693519115448,
-0.11554434150457382,
-0.07505165785551071,
-3.8614368438720703,
0.138558030128479,
-0.03567679226398468,
0.011549210175871849,
0.04084434732794762... | 621 | 24.84 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966640 | [S1] ... particularly? [S2] Well, I wanted to give, uh, for, you know, readers, uh, that are coming slightly outside of this field and moving into this, like, very rapidly moving field, kind of an overview of what are the questions that people are interested in. [S1] Yeah. [S2] Like, what are kind of the some of the interesting approaches that people are using, uh, to, um, to tackle these. And also encourage people to come in our field and, and, uh, and, and, | 28.08 | 3.158593 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966640.mp3 | [
9169,
10199,
7320,
11419,
9692,
10014,
5468,
8494,
5413,
3154,
5846,
12107,
6866,
8818,
7730,
4117,
4647,
9894,
8317,
12646,
10044,
12781,
12661,
12070,
7581,
6386,
8225,
2804,
8554,
11121,
7990,
7396,
11637,
8187,
12241,
5331,
5372,
7861,
... | [
0.4074578583240509,
-5.618268966674805,
0.13718527555465698,
-0.298043429851532,
0.46808743476867676,
0.022743478417396545,
0.5459803342819214,
-0.02260906994342804,
-0.021314147859811783,
-3.9359166622161865,
0.10248521715402603,
-0.16613976657390594,
-0.034867774695158005,
0.103361383080... | 702 | 28.08 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966641 | [S1] Um, so, uh, one of the things that I think is going to be interesting for, uh, for the future is, like, really taking evolution seriously. So we saw the, uh, uh, actually, maybe if you can scroll to, uh, where I show, um, Jess's, uh, Jess Thompson's, um, diagram of the different types of, um, of models and, and how they all fit together. It's at the very start. [S2] Yeah. [S1] It's at the intro. | 24.2 | 3.237674 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966641.mp3 | [
10072,
9426,
8923,
11124,
11563,
8501,
10613,
10614,
11063,
8798,
5734,
9247,
4885,
7088,
8600,
466,
11108,
11060,
10541,
10542,
10534,
7966,
7902,
5206,
21,
3163,
7209,
7269,
9134,
6134,
4462,
6070,
3822,
2668,
5173,
8050,
11420,
10668,
11... | [
0.3448101282119751,
-5.683995723724365,
0.02278035692870617,
-0.29036062955856323,
0.467542827129364,
-0.0010247814934700727,
0.5596100687980652,
-0.07762202620506287,
-0.24720968306064606,
-3.937666177749634,
0.08060799539089203,
-0.1717558056116104,
0.0021119844168424606,
0.1275638788938... | 605 | 24.2 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966642 | [S1] We really got into this accounts for neural activity, uh, part of, uh, so, you know, we can find models that can both perform a task which is biologically relevant and accounts for neural activity. [S2] Mm-hmm. [S1] I think this year was a big year for biological plausibility, and I don't want to say this is the last word because clearly there's way more work to be, uh, doing there. Uh, you're gonna have models which have, um, uh, realistic, uh, uh, biologically realistic kinds of gradient descent or, | 29.88 | 3.441719 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966642.mp3 | [
9433,
9497,
4243,
4636,
555,
2612,
11606,
11590,
11590,
11654,
8526,
5405,
731,
6177,
3106,
4640,
3597,
731,
8517,
12172,
5330,
12685,
6917,
11418,
11735,
11703,
10685,
10014,
3477,
11285,
6288,
3992,
11038,
6279,
202,
9647,
4007,
1370,
814... | [
0.3072362244129181,
-5.563983917236328,
0.10552138835191727,
-0.3106169104576111,
0.5060852766036987,
-0.03426038846373558,
0.5043781995773315,
-0.00625968910753727,
-0.14856864511966705,
-3.9118688106536865,
0.11138349771499634,
-0.25895771384239197,
-0.037817470729351044,
0.2413389384746... | 747 | 29.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966643 | [S1] ... evolution, uh, component to it. [S2] Yeah. [S1] So I think we're gonna start to see, and, and we can actually, like, see the process by which the brain can bootstrap itself into existence. [S2] Mm-hmm. [S1] Which I think is gonna teach us something about what it is to be human. And I'm sure there'll be TED talks and, and books and, and so forth. [S2] Yeah. [S1] But that's gonna take, like, another five, 10 years. | 21.08 | 3.103858 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966643.mp3 | [
4260,
5681,
9829,
12140,
11750,
1518,
949,
4389,
1756,
6761,
11637,
12781,
3425,
7990,
3669,
5963,
8342,
5830,
7570,
12672,
2981,
8351,
3094,
3094,
5646,
6213,
4226,
8731,
4782,
4462,
4398,
4332,
1781,
4773,
8818,
12138,
2863,
8223,
4651,
... | [
0.3728873133659363,
-5.766899108886719,
-0.016303230077028275,
-0.1503216177225113,
0.5159162878990173,
-0.061430782079696655,
0.6131001114845276,
-0.07506370544433594,
-0.21751518547534943,
-3.8836801052093506,
0.09414522349834442,
-0.16821074485778809,
-0.03209749236702919,
0.11416249722... | 527 | 21.08 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966645 | [S1] I, I think that, uh, uh, one, uh, one thing that we, uh, that we're, haven't, like, really taken seriously so far is the role of weak supervision, uh, from a parental perspective. [S2] Mm-hmm. [S1] But if you think of, like, uh, a parent and their baby, they're gonna point at things, they're gonna say, "This is this, this is that," and- [S2] Yeah. [S1] ... you know, it has had a, like, hands have had a, a, a huge role in our evolution as, uh, as Homo sapiens. Um, and it's even, like, thought | 29.88 | 3.419116 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966645.mp3 | [
4721,
3112,
11373,
10750,
11748,
10189,
6874,
9460,
10748,
12236,
6424,
6000,
11105,
9817,
12570,
3362,
11186,
11700,
6160,
6177,
10584,
8557,
11043,
10605,
11174,
10670,
10606,
10598,
10662,
3925,
6820,
7278,
10031,
6575,
6055,
3431,
3942,
8... | [
0.40411126613616943,
-5.528947353363037,
0.1147036924958229,
-0.228568896651268,
0.48677507042884827,
-0.0005111423088237643,
0.5284543037414551,
-0.02369174361228943,
-0.16917254030704498,
-3.9415528774261475,
0.07176321744918823,
-0.19471080601215363,
-0.03662960231304169,
0.128203243017... | 747 | 29.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966646 | [S1] ... that, um, uh, uh, uh, uh, sign language preceded the appearance of, uh, of voiced, uh, speech. [S2] Yeah. [S1] Uh, so that we probably have somewhere in our noggins, uh, some areas which are highly selective for hand gestures and which are used for a kind of weak supervision, uh, that's important for, uh, | 24.96 | 3.327386 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966646.mp3 | [
11133,
11766,
12290,
6152,
5760,
12262,
12783,
7159,
5103,
6647,
7446,
10376,
9369,
8258,
7532,
12085,
12470,
5678,
12573,
12388,
7204,
11412,
12628,
9242,
6226,
9424,
4709,
4463,
5630,
3062,
3575,
5623,
3062,
3445,
5661,
5679,
5670,
4127,
... | [
0.5682340264320374,
-5.613842010498047,
0.04802247881889343,
-0.19826620817184448,
0.5401743054389954,
-0.002164173172786832,
0.48132455348968506,
0.0003686507698148489,
-0.10712837427854538,
-3.9182229042053223,
0.026050006970763206,
-0.21729570627212524,
0.017408153042197227,
0.047624282... | 624 | 24.96 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966647 | [S1] Uh, for parents. [S2] Yeah. [S1] So understanding what happens with that peri- per- personal space and what, what happens as, uh, um, as we use tools, uh, is clearly important from, like, just this, that curiosity of how, uh, you know, we went from, uh, Australopithecus to, uh, to modern humans. And I think it's gonna teach us a lot about, um, yeah, what it means to be human. | 25.8 | 3.452587 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966647.mp3 | [
11354,
5817,
9481,
12114,
12633,
9428,
8565,
11510,
5559,
11631,
1449,
9144,
3900,
5462,
67,
10913,
2490,
10222,
9198,
5605,
982,
1390,
9326,
8709,
3050,
8144,
2960,
8046,
10550,
8430,
8621,
5286,
8867,
9894,
9518,
12532,
12598,
12398,
1233... | [
0.3915664255619049,
-5.675866603851318,
0.05062536522746086,
-0.2658359706401825,
0.4710462689399719,
-0.05786560848355293,
0.5398170948028564,
-0.05131630226969719,
-0.20958390831947327,
-3.911716938018799,
0.050803303718566895,
-0.2501600682735443,
0.018653176724910736,
0.116195894777774... | 645 | 25.8 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966648 | [S1] Awesome. Um, last question from my side. With, you're clearly interested in how the brain works, right? And, and, and see- [S2] Yeah. [S1] ... and seeing, you know, can we, can we make parallels between AI models, like deep models and, and brain areas and so on. Do you think that it is a necessity that we sort of | 21.4 | 2.994642 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966648.mp3 | [
8738,
3895,
7740,
8309,
4329,
2905,
3016,
3383,
1045,
8239,
9836,
9834,
9810,
9753,
8361,
6713,
9253,
11775,
11255,
11630,
8276,
5678,
4642,
4255,
2836,
3294,
5469,
11709,
11186,
10682,
4530,
11224,
4305,
12067,
755,
11124,
11251,
9574,
756... | [
0.8111924529075623,
-5.686675548553467,
0.03873392939567566,
-0.04933903366327286,
0.44976329803466797,
-0.010640476830303669,
0.23214605450630188,
-0.00030406820587813854,
-0.011406968347728252,
-3.733398675918579,
0.1367216855287552,
-0.13701316714286804,
0.07766881585121155,
0.003630538... | 535 | 21.4 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966649 | [S1] ... and certainly suboptimal. So how the retina is wired up is the classic example. It's wired up the wrong way around. Octopuses have, have it the right way around and it doesn't seem to bother them. [S2] [LAUGHS] [S1] So, uh, that's a, that's a clear example. Um, but maybe there's some thing that we can, | 19.64 | 3.104032 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966649.mp3 | [
11986,
12634,
9492,
9525,
11614,
3951,
1143,
4206,
3094,
8798,
2061,
8832,
4512,
9088,
3043,
10550,
8149,
9355,
7910,
2644,
9110,
10054,
5007,
4422,
7494,
6470,
4161,
1089,
3779,
6616,
9168,
9170,
10750,
4711,
5483,
6526,
5885,
5950,
5746,
... | [
0.35710370540618896,
-5.743293762207031,
-0.03541779890656471,
-0.18976525962352753,
0.6346353888511658,
0.028574123978614807,
0.5749945640563965,
0.016163842752575874,
-0.049050457775592804,
-3.9298617839813232,
0.00851250160485506,
-0.19578035175800323,
-0.04874458536505699,
0.0523759201... | 491 | 19.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966650 | [S1] ... that we can identify with, uh, with brains and that is gonna unlock the next generation of machine learning. Maybe it's spiking neural networks, for instance. [S2] Yeah. [S1] You know, people are demonstrating, like, you could get something which is, uh, like a thousand times or 10,000 times more energy efficient if you just used, uh, uh, these mixed signals spiking neural networks. So I, I don't know. | 22.64 | 3.166213 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966650.mp3 | [
6154,
7361,
6145,
4069,
12207,
8720,
11345,
3090,
52,
10581,
12612,
8912,
9204,
7685,
6220,
6155,
6168,
8226,
12726,
7101,
9581,
12549,
8192,
11221,
12270,
12083,
8732,
7944,
8180,
11036,
9712,
6128,
3374,
5943,
3446,
3958,
10159,
12758,
57... | [
0.35964134335517883,
-5.733741283416748,
0.013079194352030754,
-0.2156086564064026,
0.5065352320671082,
0.0012759241508319974,
0.5133655071258545,
-0.07305628806352615,
-0.17928288877010345,
-3.9233272075653076,
0.16671553254127502,
-0.22962982952594757,
-0.033442407846450806,
0.1079809293... | 566 | 22.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966651 | [S1] Yeah, and that would, I mean, a thousand times, 10,000 times, that is sort of the orders of magnitude you spoke about before when it came to, to data. [S2] Well, those are, so here I'm thinking about the, the energy efficiency. [S1] Sure. [S2] Uh, so I think like one recurrent, not, not super- | 17.2 | 3.114332 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966651.mp3 | [
5480,
6385,
8376,
5170,
4189,
12068,
4315,
9181,
8702,
1826,
9965,
6956,
9385,
7205,
6518,
6437,
7083,
7990,
6454,
6732,
7093,
5670,
9820,
3482,
11708,
11131,
6965,
1056,
3172,
2741,
7855,
8551,
11779,
8736,
8387,
6601,
4496,
5008,
5008,
... | [
0.5673335194587708,
-5.681784152984619,
-0.09634186327457428,
-0.06201283633708954,
0.5213974118232727,
0.03795770928263664,
0.40709879994392395,
-0.06548641622066498,
-0.11398971080780029,
-3.837991237640381,
-0.058507632464170456,
-0.15602527558803558,
0.01002959068864584,
0.040589779615... | 430 | 17.2 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966653 | [S1] ... machine. So, again, I encourage your, uh, your, your viewers to, uh, to get into this because you can still do things with one- [S2] Yeah. [S1] ... GTX 1080. [S2] That's awesome. [S1] [LAUGHS] | 9.08 | 3.011373 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966653.mp3 | [
11556,
7958,
8220,
11116,
11614,
12761,
12744,
5058,
12558,
12110,
12190,
9894,
8198,
6279,
3497,
8082,
354,
11062,
10542,
2863,
2726,
5199,
734,
9893,
9965,
9502,
9957,
10013,
9950,
12462,
12269,
6797,
12122,
11237,
11250,
11860,
8109,
12718... | [
0.5607796311378479,
-5.789758682250977,
0.02326863631606102,
-0.06904914230108261,
0.37147635221481323,
-0.10467962920665741,
0.6098991632461548,
-0.07423078268766403,
-0.1798185408115387,
-4.011453628540039,
-0.08635418117046356,
-0.16095414757728577,
-0.1084979996085167,
0.00568232452496... | 227 | 9.08 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966654 | [S1] I don't think that naturally if you just train a neural network to solve a task, it's going to do it the same way that the brain does because I don't think that that's, that's really pointed out. I don't think that, uh, GPT-3 does things the same way that a human does in any sort of meaningful way. No way. [S2] Yeah. [S1] [LAUGHS] Uh, even though they're both very good in language. Uh- [S2] Yeah. Maybe GPT-4. [S1] Uh- [LAUGHS] | 25.68 | 3.37314 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966654.mp3 | [
12580,
12604,
12598,
4451,
8885,
5565,
12694,
8449,
10620,
7794,
8724,
8280,
2905,
11764,
9044,
9590,
11859,
6416,
11701,
11755,
11254,
7236,
11473,
6801,
1035,
5698,
7758,
10822,
11750,
12734,
12788,
9147,
9364,
12249,
10209,
348,
7973,
1041... | [
0.5356574058532715,
-5.50114107131958,
0.05492967367172241,
-0.22403642535209656,
0.612908661365509,
-0.07403198629617691,
0.4555680453777313,
-0.0677720159292221,
-0.08715900778770447,
-3.9655401706695557,
-0.11399015039205551,
-0.24669142067432404,
-0.027558695524930954,
-0.0128898266702... | 642 | 25.68 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966655 | [S1] Well, if you ask Gary Marcus, he'll say that there's no way. It'll never happen. [S2] Never. [S1] Uh, Neuro-symbolic AI all the way. Yeah. [S2] All right. Cool. Yeah. Um, for ev- to everyone, uh, follow Patrick, uh, many, he's mi- he's written papers, lots of papers. You're also the CTO of Neuromatch Academy, is that correct? | 21.56 | 3.153773 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966655.mp3 | [
9635,
9954,
3308,
8558,
7987,
8107,
12196,
3490,
6640,
9548,
11156,
10610,
10619,
3409,
4250,
9563,
11740,
8100,
6092,
8909,
10854,
8511,
8509,
5589,
3851,
12001,
4581,
9055,
3942,
3367,
865,
5393,
6480,
9702,
2686,
2645,
1434,
8097,
401,
... | [
0.702989399433136,
-5.683424949645996,
0.0331897996366024,
-0.10421159118413925,
0.3943340480327606,
-0.02392730489373207,
0.4188190996646881,
-0.04242071881890297,
-0.03180171921849251,
-3.8110530376434326,
0.10716966539621353,
-0.06895846873521805,
0.04304352402687073,
0.0867282226681709... | 539 | 21.56 |
bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966656 | [S1] ... year. Uh, I'm sure the announcement actually for, um, for the next, uh, version of, uh, Neuromatch Academy will happen pretty soon. [S2] Okay. [S1] Uh, so, uh, if you have people in the, in, in your, um, audience that are interested in that, I highly, uh, recommend to them to, uh, to do that. It's a great occasion to learn. And we already have, um, you know, the materials from last year online, so if you wanna get started on your learning, uh, you can do that today. | 29.84 | 3.118103 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p167_BV17h4y1j7aZ_p167_m4-dialogue_0966656.mp3 | [
6058,
3400,
4881,
12678,
6615,
4575,
462,
6486,
7402,
8292,
6951,
12533,
12070,
12590,
11950,
9838,
4444,
8866,
6761,
5161,
11804,
9206,
3062,
3006,
438,
4094,
4917,
8753,
9085,
11772,
9942,
11045,
12777,
12681,
3500,
8099,
11028,
2633,
127... | [
0.4599460959434509,
-5.467421531677246,
0.18084150552749634,
-0.26979461312294006,
0.4350109100341797,
-0.08273877948522568,
0.5836970210075378,
-0.029000671580433846,
-0.09121908992528915,
-3.910884141921997,
0.05898543819785118,
-0.2562299072742462,
-0.023756885901093483,
0.1264874488115... | 746 | 29.84 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721943 | [S1] Really fun project for me as well. [S2] It, it sounds like fun. [S1] Yeah, just, just trying to see, uh, whether you can output something like really, uh, realistic and, and really fun. Yeah. | 11.8 | 3.263531 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721943.mp3 | [
6818,
8232,
4612,
803,
11141,
11548,
7955,
12694,
4954,
9017,
436,
6462,
6390,
12326,
6370,
4977,
3884,
7980,
12134,
9168,
7604,
7019,
4963,
4016,
379,
8991,
8220,
11429,
9988,
12036,
6412,
1684,
12606,
2486,
4892,
2960,
1315,
2798,
11639,
... | [
0.7711265683174133,
-5.885841369628906,
-0.07626001536846161,
-0.051938459277153015,
0.5150131583213806,
-0.11095765233039856,
0.6276450753211975,
-0.08169475942850113,
-0.2646768391132355,
-3.869840145111084,
0.11583266407251358,
-0.15677040815353394,
-0.16347353160381317,
0.2384297698736... | 295 | 11.8 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721944 | [S1] Of course, yeah. So, um, the, um, the way that virtual home, uh, expresses this actions are, uh, via like this specific format where you put a square bracket, um, like, uh, for the action, atomic action, like grab, put open. [S2] Mm-hmm. [S1] And then you put, uh, I think it's a parenthesis, uh, or, or, yeah, something for, for the arguments. And, um, but the, the problem is, like, we, we can't just, like, | 28.32 | 2.868387 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721944.mp3 | [
12373,
8744,
8365,
8878,
3001,
7465,
5875,
10611,
847,
1972,
6040,
3777,
12548,
10651,
8574,
1461,
4828,
7153,
5512,
5534,
10940,
10356,
7845,
5910,
3795,
11817,
11134,
8052,
10614,
10606,
8047,
5999,
5935,
9374,
5670,
3623,
9294,
4685,
205... | [
0.7438966035842896,
-5.684598922729492,
0.07969485223293304,
-0.10875611007213593,
0.5974050164222717,
0.04471220821142197,
0.4344339966773987,
0.040373388677835464,
-0.10822620987892151,
-3.870227813720703,
0.08916813880205154,
-0.2573164403438568,
-0.12221874296665192,
0.3240948319435119... | 708 | 28.32 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721945 | [S1] expect language models to, to handle this because, um, I, I mean, even if we put an example in front, maybe they can do it, but it's definitely not the, the way that usually humans, um- [S2] Yeah. [S1] ... produce language. So, and after all, this language models are trained on human text, so we, we decide, like, maybe it's not the, the right way to carry this models. Maybe we just want to- [S2] Have you ever tried? [S1] Absolutely. [S2] Have you tried letting them output directly the syntax, or was it just like, "Yeah, it's not gonna work anyway"? | 29.72 | 3.286208 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721945.mp3 | [
8721,
6696,
9242,
12759,
11859,
6635,
3905,
9074,
12279,
3581,
7332,
9433,
4443,
37,
5542,
9148,
12499,
9821,
3436,
8535,
12033,
2380,
8222,
6006,
3182,
4758,
3282,
3263,
3302,
3409,
6209,
5520,
6632,
3884,
3351,
799,
662,
717,
2191,
3328... | [
0.8687304258346558,
-5.628979682922363,
0.10935854911804199,
-0.11621575057506561,
0.546183705329895,
0.08333225548267365,
0.38511422276496887,
-0.0007230194751173258,
-0.14842112362384796,
-3.9297800064086914,
0.15944813191890717,
-0.2040223628282547,
-0.07691000401973724,
0.2705781459808... | 743 | 29.72 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721946 | [S1] I tried briefly, but it's definitely not, uh, thoroughly investigated. Uh, and- [S2] Yeah. [S1] Like, it, like, in- intuition-wise, I think it's definitely- [S2] Sure, yeah. [S1] ... to, to, to use, like, natural language. But we, we did adopt for the, um, the most basic approach that we can, we can think of, which is, like, just define a straight-up, like, um, template for each atomic action. And actually, because this atomic actions are, | 27.2 | 3.008878 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721946.mp3 | [
8236,
7102,
12219,
11745,
11658,
11587,
11587,
9484,
5648,
11520,
12755,
890,
8059,
11753,
8906,
3730,
8353,
874,
2885,
8004,
8966,
3576,
4314,
9101,
10182,
7833,
2999,
10735,
7078,
4307,
12662,
12693,
5825,
2969,
3649,
11190,
12218,
6883,
... | [
0.6740664839744568,
-5.6417036056518555,
0.04512615129351616,
-0.06938908994197845,
0.6206698417663574,
-0.018518241122364998,
0.41880762577056885,
0.00870298221707344,
-0.1869788020849228,
-3.94286847114563,
0.058954715728759766,
-0.23973116278648376,
-0.13361404836177826,
0.2487460821866... | 680 | 27.2 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721947 | [S1] Yeah. [S2] Like this. [S1] And it's interesting because when the humans raid it, the humans would also skip a bunch of steps, right? If you, if you tell a human, "Go to the fridge and grab a glass of milk," the human will go like, "Oh, yeah, of course." Right? Which is, which is one of my, um, maybe this is jumping ahead a little bit, but one of the questions I had most when I read this was just, | 22.68 | 3.438905 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721947.mp3 | [
6160,
3716,
11733,
10686,
1324,
3691,
6178,
9355,
7283,
4835,
6817,
6291,
6228,
3163,
9372,
3238,
5557,
5036,
6291,
4202,
3436,
2397,
11709,
12373,
8660,
11612,
4362,
12396,
10005,
7945,
3032,
8676,
11602,
8730,
10945,
3554,
8092,
7089,
603... | [
0.8476184010505676,
-5.671880722045898,
-0.04321349039673805,
-0.0040102130733430386,
0.4756539762020111,
0.06302525848150253,
0.260779470205307,
-0.033557355403900146,
-0.0016181207029148936,
-3.746786117553711,
0.3579838275909424,
0.11555999517440796,
0.2528722286224365,
0.10232394933700... | 567 | 22.68 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721948 | [S1] Yeah, so that's a really good question. Actually, um, this granularity actually comes from the data set, uh, or mainly the vir- the virtual environment itself. [S2] Yeah. [S1] Because the way, um, because we essentially follow the format of, uh, virtual environment and also this data set they collected for, from humans, uh, of how to do, um, this really, uh, | 22.92 | 3.336777 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721948.mp3 | [
6161,
12116,
11081,
11114,
3574,
883,
3748,
6753,
4249,
9200,
8064,
8109,
10421,
10414,
1749,
9454,
8046,
9438,
5384,
11771,
11115,
5528,
3026,
10611,
8982,
3348,
732,
731,
8518,
11732,
6355,
10718,
11663,
6745,
12003,
3430,
5983,
6868,
572... | [
0.7337666153907776,
-5.77017879486084,
-0.013945992104709148,
-0.14212048053741455,
0.6636305451393127,
0.08991692960262299,
0.48186787962913513,
0.03730544447898865,
-0.10090664029121399,
-3.8374123573303223,
0.09735166281461716,
-0.3124111592769623,
-0.03417889401316643,
0.27795562148094... | 573 | 22.92 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721949 | [S1] And it's after that they build this environment based on the verbs used by, by those humans. So you can think of like this environment is really built on top of, uh, what humans says. Now, now- [S2] Yeah. [S1] ... the developers, um, who, uh, who just say like, okay, we want this granularity, we want this like walk, grab and those, et cetera. So, um, | 22.8 | 3.362267 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721949.mp3 | [
10017,
6248,
10989,
12277,
11292,
7365,
8857,
9825,
10931,
7721,
12437,
12556,
6680,
6185,
9243,
6431,
5712,
3689,
6162,
5478,
12125,
9541,
6347,
2696,
1688,
10925,
12189,
6936,
5513,
2443,
12597,
9211,
11705,
9002,
4088,
2923,
11109,
6227,
... | [
0.8483863472938538,
-5.7819318771362305,
0.01832788810133934,
-0.12370699644088745,
0.6414375305175781,
0.03016768954694271,
0.5112510919570923,
0.03306325152516365,
-0.2016681432723999,
-3.899839162826538,
0.20534725487232208,
-0.266513854265213,
-0.07761052250862122,
0.2222481518983841,
... | 570 | 22.8 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721950 | [S1] Because it's trained on human-produced text, and humans, after all, produce these actions. [S2] Yeah. [S1] So if you do, uh, do it careful enough, and then use some techniques to, uh, properly translate them, or doing something else, you can essentially get back something similar to what human produced in the beginning. [S2] Yeah. | 21.72 | 3.311681 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721950.mp3 | [
6193,
8224,
6626,
10117,
11411,
9705,
9075,
10536,
10536,
11635,
7454,
1361,
5529,
6032,
6016,
1425,
7203,
8369,
9763,
12255,
4354,
2832,
1729,
12617,
12778,
490,
12765,
9743,
4366,
3327,
3190,
4882,
12712,
10008,
9093,
5782,
10846,
4599,
7... | [
0.8162883520126343,
-5.7973103523254395,
0.04168148711323738,
0.024425407871603966,
0.6209627389907837,
0.05795138701796532,
0.4842572808265686,
0.03889338672161102,
-0.08025618642568588,
-3.9363536834716797,
0.14932040870189667,
-0.2615341246128082,
-0.20484288036823273,
0.217276990413665... | 543 | 21.72 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721951 | [S1] Uh, yeah, it's, I mean, you would, you would imagine that sort of the human-ness of how the environment was built would also be present a little bit in these language models, which makes, which makes sense. I don't have a better idea, like, of how to build an environment like this, so- [S2] Yeah, I think it's pretty- [S1] ... it seems pretty, pretty reasonable, yeah. | 19.32 | 3.242725 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721951.mp3 | [
9875,
12353,
10858,
6937,
12289,
9752,
11503,
9719,
11758,
7372,
3072,
8728,
11297,
11826,
9256,
7729,
7307,
11723,
10665,
11195,
1397,
7461,
6519,
11413,
5200,
3809,
11307,
9215,
9655,
8704,
5665,
5826,
2970,
2985,
2896,
3017,
705,
6730,
1... | [
0.9021322131156921,
-5.6315507888793945,
-0.10716143250465393,
0.07419758290052414,
0.5052431225776672,
0.07727543264627457,
0.27717679738998413,
-0.0817248597741127,
0.03133919835090637,
-3.819345474243164,
0.21254849433898926,
0.04766463115811348,
0.08488883078098297,
0.06166569516062736... | 483 | 19.32 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721952 | [S1] I think, uh, this is also what makes, like, language models particularly useful for- [S2] Yeah. [S1] ... for this task because these are basically just human tasks and language models are really good at, like, mimicking humans. Yeah. | 13.56 | 3.435217 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721952.mp3 | [
9738,
9760,
6692,
8126,
12710,
6424,
4083,
12683,
8978,
11419,
11877,
9876,
4371,
9772,
9895,
9894,
9134,
8551,
6063,
7181,
3136,
8678,
8982,
3520,
853,
3328,
9583,
7870,
7782,
4444,
8536,
403,
5367,
5231,
3276,
116,
8622,
11687,
3675,
26... | [
0.9824249148368835,
-5.85183048248291,
-0.03441658243536949,
0.026804478839039803,
0.6366844177246094,
-0.034331027418375015,
0.44158250093460083,
-0.06116629019379616,
-0.26614606380462646,
-3.9274396896362305,
0.20827268064022064,
-0.2603922486305237,
-0.03198353946208954,
0.132963448762... | 339 | 13.56 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721953 | [S1] Yeah. [S2] And you've noticed in the paper that very often they just produce plans that have nothing to do with the task description. They would just produce like a plan and that is according to the syntax of the examples that you give in the prompt, right? Um, but how can you explain that? Like even, even on the top here, like the, the large models, it's even better than humans at correctness. [S1] Mm-hmm. | 25.16 | 3.315841 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721953.mp3 | [
5672,
9360,
6659,
6411,
10643,
8557,
9508,
11246,
8679,
4902,
10317,
11414,
7179,
3019,
7454,
8745,
8737,
10866,
9753,
10275,
12454,
6852,
12675,
11156,
7829,
3355,
4704,
7203,
11652,
7958,
2985,
6357,
10828,
8623,
10485,
10859,
11030,
8514,
... | [
0.8797754049301147,
-5.578668594360352,
-0.01707238331437111,
0.04474708065390587,
0.46989908814430237,
0.07062038034200668,
0.30973660945892334,
-0.07290146499872208,
0.027791697531938553,
-3.808889389038086,
0.26785486936569214,
0.18767161667346954,
0.1962975412607193,
0.0758778229355812... | 629 | 25.16 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721954 | [S1] Yeah, so, uh, there were actually two questions, uh, that, like, uh, I think you, you raised. One is, uh, why this, like, smaller models, like, um, like when I say smaller, it's actually still pretty large. The large- [S2] Yeah. [S1] ... GPT, uh, GPT-2 model. So why do they produce, like, more executable plants? Uh, and the second question is, uh, why the GPT-3, the large GPT-3 model is actually better than human. So to answer the first question, um, | 29.96 | 3.158923 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721954.mp3 | [
10013,
11933,
12445,
9866,
6688,
4819,
11154,
10666,
5565,
1830,
4458,
8152,
5485,
10413,
10478,
7838,
4372,
11958,
8117,
11063,
6447,
3951,
6511,
4455,
3942,
4447,
4510,
6284,
9,
8273,
2050,
8156,
11178,
7970,
10586,
10673,
9202,
9436,
102... | [
0.7091339826583862,
-5.675645351409912,
0.05352119356393814,
-0.21925154328346252,
0.7070509791374207,
0.04139220342040062,
0.3731517195701599,
0.009165213443338871,
-0.24887719750404358,
-3.8488285541534424,
0.04155203700065613,
-0.3588705360889435,
-0.050614017993211746,
0.35353142023086... | 749 | 29.96 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721955 | [S1] Uh, there, so like, uh, I, you can totally imagine, like, it's super executable. But, uh, if you present them to humans, of course, like, humans will, will spot this and then say, "Okay, this is not correct." Because we, we, when we do human evaluations, we, we're trying to make it simple so that the, um, the, the error here is not too big because we, we don't ask, like, hundreds of humans to evaluate this. We only, uh, ask, got to ask ten, uh, evaluators- [S2] Mm-hmm. [S1] ... in this case. So, um, | 29.96 | 3.258555 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721955.mp3 | [
8736,
3113,
9749,
9135,
814,
1318,
8712,
5737,
5450,
12702,
11681,
8091,
974,
1492,
1878,
5025,
8600,
3027,
5951,
2614,
29,
613,
5332,
6078,
7150,
9923,
8810,
12179,
7393,
10834,
9957,
12525,
9901,
9390,
7274,
1561,
12290,
7169,
8786,
116... | [
0.7476412653923035,
-5.628249168395996,
0.1514415293931961,
-0.1467755287885666,
0.6428161859512329,
-0.00825441349297762,
0.3496337831020355,
0.0334036611020565,
-0.28832438588142395,
-3.90946888923645,
0.08113902062177658,
-0.37391263246536255,
-0.09109977632761002,
0.3599306344985962,
... | 749 | 29.96 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721956 | [S1] Uh, so that's why, like, th- this smaller models are now really good at scalability. [S2] Mm-hmm. [S1] And, and the second question that you ask is why, um, this, like, larger models, um, are actually better than humans. So, uh, we, I, actually, this is not a completely fair comparison if you just look at one axis. So, | 21.28 | 3.297755 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721956.mp3 | [
6697,
8224,
11638,
10551,
5999,
3511,
2031,
2405,
4317,
1875,
6641,
5448,
6503,
10991,
8367,
5158,
3287,
4245,
5648,
8685,
11772,
9812,
3353,
912,
3241,
5238,
11126,
11198,
11710,
12790,
10190,
7046,
7567,
4325,
2709,
8639,
10230,
6853,
113... | [
0.7961059212684631,
-5.749302387237549,
-0.03745337948203087,
-0.19711709022521973,
0.6854990124702454,
0.08917698264122009,
0.4546003043651581,
0.03941331058740616,
-0.1602509617805481,
-3.894834280014038,
0.08170843869447708,
-0.32581254839897156,
-0.0557171069085598,
0.23493045568466187... | 532 | 21.28 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721957 | [S1] ... cross source from, from, from Amazon, Turkish. So, um, this plans, they make sure that, like, these are executable plans. So, which means, um, that they're, they have- [S2] ... like here. [S1] Yeah. [S2] They'd be over here. | 16.56 | 3.310134 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721957.mp3 | [
9763,
9587,
2916,
5998,
5821,
5294,
4586,
7008,
3968,
5302,
7726,
10605,
3013,
3407,
3993,
3457,
5025,
6008,
2918,
8559,
3383,
3823,
6174,
4118,
1613,
2076,
3433,
9200,
5478,
8503,
1270,
3694,
6686,
1061,
3106,
4641,
3817,
3897,
4841,
136... | [
0.7873522639274597,
-5.8064775466918945,
-0.027424026280641556,
-0.053179267793893814,
0.6538490056991577,
0.06618398427963257,
0.4787382185459137,
-0.03186187148094177,
-0.10194359719753265,
-3.8789291381835938,
0.17955510318279266,
-0.313720166683197,
0.015177160501480103,
0.234821408987... | 414 | 16.56 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721958 | [S1] Yeah, but, but we, we don't want to put a spot right there, uh- [S2] Yeah. [S1] ... on the right because, uh, it's, it's hard to see because humans are, are a big, um, baseline and reference here. Uh, it's not a baseline that we're trying to beat. Of course, like- [S2] Yeah. [S1] ... is not there yet in terms of like, uh, at the same time outputting correct, uh, action plans and semantic- | 20.92 | 3.206911 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721958.mp3 | [
6705,
5656,
9555,
11082,
10675,
3006,
4133,
8106,
11190,
9006,
5729,
3573,
11060,
8294,
563,
8486,
8973,
1626,
5871,
11108,
5349,
10004,
1634,
5405,
12548,
2961,
10555,
7842,
5732,
10547,
11581,
8717,
2909,
5678,
10401,
4082,
4081,
818,
110... | [
0.7243174910545349,
-5.767364978790283,
-0.026117032393813133,
-0.10773412138223648,
0.649782657623291,
-0.026927705854177475,
0.3846733570098877,
0.006001085974276066,
-0.12438103556632996,
-3.8763091564178467,
0.09998500347137451,
-0.2869870662689209,
-0.06062548607587814,
0.251036614179... | 523 | 20.92 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721959 | [S1] What, something that, I mean, you know, the virtual home is maybe a test bed, right? It's not a- [S2] Yeah. [S1] I, I don't see this paper being about virtual home. It's more like, here is a model that outputs something, yet I need the output in some other form, right? In, in, in this | 18.48 | 3.336137 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721959.mp3 | [
8728,
6696,
7701,
116,
2741,
8046,
5431,
3870,
1625,
9249,
9249,
9242,
7251,
3441,
5984,
3025,
6015,
7934,
11300,
7726,
8988,
7144,
6440,
6645,
12182,
9862,
12382,
11783,
705,
6583,
5047,
4478,
3582,
7396,
10840,
8745,
10689,
6617,
9313,
... | [
0.8621620535850525,
-5.722487449645996,
-0.08273067325353622,
0.11962679773569107,
0.3871036171913147,
0.036676403135061264,
0.23988202214241028,
-0.07215073704719543,
0.011587858200073242,
-3.759000778198242,
0.1726294457912445,
0.23074471950531006,
0.2080892026424408,
0.080887071788311,
... | 462 | 18.48 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721960 | [S1] is a very general problem, has many applications. And if we could solve that bridge, that technically is, you know, is, is a big gain. That's exactly what you do. Uh, so how, how did you go about this? [S2] Yeah, so, uh, actually, uh, I just want to make sure that, uh, actually this paper just present a, a really, like, preliminary stuff. I, I don't think it solves anything particularly. I mean, it, it does, like, if this problem- [S1] Sure, but it's a big- | 28 | 3.336933 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721960.mp3 | [
6184,
5161,
12574,
12677,
12694,
9639,
3918,
2882,
450,
3410,
3009,
3423,
2414,
2340,
2533,
2405,
3121,
8229,
540,
361,
3946,
1002,
1513,
1515,
12189,
12201,
8500,
331,
2965,
12037,
7300,
11529,
10187,
9647,
10166,
9806,
5846,
3293,
787,
... | [
0.8811586499214172,
-5.551292896270752,
-0.017326749861240387,
0.013534493744373322,
0.5524575710296631,
0.07999803125858307,
0.2568460702896118,
-0.0350504033267498,
-0.09864829480648041,
-3.875778913497925,
0.18236520886421204,
-0.026604294776916504,
0.06981495767831802,
0.13028797507286... | 700 | 28 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721961 | [S1] about the semantic, uh, meaning for, for that sentence. So what we do is- [S2] Yeah. [S1] ... that we take the sentence outputted by GPT-3 or codecs, and then, uh, we just compare that against all the possible, um, admissible actions, allowed actions by the environments. [S2] Mm-hmm. [S1] And then we found the most similar one in terms of like, um, this distance in, in the embedding space. Um- [S2] Yeah. [S1] We actually used just cosine distance and, and found that to work decently well. [S2] Yeah. So, so- [S1] Yeah. | 29.92 | 3.130205 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721961.mp3 | [
7259,
9240,
8503,
3367,
4315,
11374,
6583,
7776,
6135,
9149,
7925,
8918,
8208,
4168,
2998,
3503,
1758,
6569,
8136,
3787,
5670,
11118,
12798,
12708,
8205,
8456,
5036,
12263,
7567,
11355,
12699,
2467,
4834,
9753,
9729,
12354,
9746,
5329,
9737... | [
0.8260990977287292,
-5.595976829528809,
0.10275355726480484,
-0.017276257276535034,
0.6080145239830017,
0.051653940230607986,
0.3613179922103882,
0.06774956732988358,
-0.1941271722316742,
-3.9061970710754395,
0.17460574209690094,
-0.3135702610015869,
-0.11756715178489685,
0.345329105854034... | 748 | 29.92 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721962 | [S1] And that's my kind of translated action. So here we have an example that would, that would translate like squeeze out a glob of lotion into pour lotion into right hand. [S2] Yeah. [S1] So this, it would map. [S2] Exactly. [S1] Yeah. Action into and pour, pour, it would be the verb, lotion, kind of the object and right hand also one of the objects. So maybe like there's two arguments to pour, um, | 26.92 | 3.063029 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721962.mp3 | [
7257,
9376,
11430,
7151,
7087,
7591,
5030,
7663,
7663,
7246,
6215,
6663,
3596,
11073,
6104,
6864,
10784,
6201,
6272,
6118,
7102,
5566,
7412,
8017,
2978,
6080,
6245,
8638,
9015,
4023,
5007,
12231,
7393,
6250,
12626,
6589,
7478,
10956,
1317,
... | [
1.0564366579055786,
-5.423051834106445,
0.024409053847193718,
0.08089136332273483,
0.24733422696590424,
0.04525085538625717,
0.18877148628234863,
-0.11014797538518906,
-0.04982445389032364,
-3.70749831199646,
0.2907870411872864,
0.27181869745254517,
0.21147100627422333,
0.12359504401683807... | 673 | 26.92 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721963 | [S1] ... to, like, the quick, quick options to respond, right? [S2] Yeah, yeah. [S1] And, and the, I think the first, I'm not sure how it is done now, but the first version of this, we were like, "Wow, this is, you know, you know, cool. It actually takes, you know, it takes into account the, the email message that was there." We always thought it was kind of like a, a, a language model, generative model somewhere. So I went to a talk and they were just like, | 23.52 | 3.37377 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721963.mp3 | [
12426,
3601,
8592,
5021,
11087,
1302,
2138,
2773,
10750,
9695,
12449,
4121,
3036,
4892,
8298,
11485,
3571,
3251,
2797,
11741,
10011,
10849,
6762,
12180,
7145,
4961,
9753,
6696,
9748,
9196,
3515,
179,
11238,
10196,
11944,
12380,
4068,
8747,
... | [
0.8313623666763306,
-5.563548564910889,
-0.041757866740226746,
0.07943075895309448,
0.5010662078857422,
0.07211518287658691,
0.2398482710123062,
-0.05921487882733345,
0.10316985100507736,
-3.739053964614868,
0.2943582832813263,
0.10064956545829773,
0.21060208976268768,
0.08148231357336044,... | 588 | 23.52 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721964 | [S1] pretty difficult to map a compounded action. Like you said, like, um, like two actions in, in one sentence into one invisible action. But this is partly mitigated by how you tune the temperature, the sampling parameter, uh, just the temperature for the GPT-3 or Codex models. [S2] Yeah. [S1] Because we found that if you, | 24.84 | 3.17434 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721964.mp3 | [
3113,
9202,
1000,
2891,
12034,
5377,
10593,
9225,
2953,
12181,
11477,
2928,
1907,
11428,
7153,
6333,
5239,
8918,
10944,
5584,
4065,
3946,
2719,
3725,
5669,
4135,
11998,
5110,
7540,
7540,
7549,
10682,
11508,
8233,
6514,
1896,
6450,
1329,
382... | [
0.8544318079948425,
-5.690187931060791,
0.04962008446455002,
-0.0644698515534401,
0.6552687287330627,
0.015987707301974297,
0.4581588804721832,
0.030519984662532806,
-0.08739330619573593,
-3.9554855823516846,
0.16889651119709015,
-0.3181256353855133,
-0.10651060193777084,
0.234770998358726... | 621 | 24.84 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721965 | [S1] If you try, like, all this, like, different settings, we, we, we, in the end, we found, like, usually you want to use, like, uh, lower temperature than, than what people mostly use for language generation, for example. [S2] Yeah. [S1] So, so that, like, each action, uh, is, uh, like, small enough and succinct enough. And then, and then after we translate this action, uh, so it's, so, so that it's easier for this bird model, uh, Roberto model to translate. | 27.08 | 3.354834 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721965.mp3 | [
6177,
8795,
9877,
6794,
12165,
3419,
6632,
11086,
8407,
7880,
12753,
3561,
2868,
10548,
9210,
12241,
11667,
9540,
10116,
11717,
12118,
5778,
86,
8174,
9710,
9819,
6939,
8272,
6816,
5640,
1819,
8289,
10998,
7926,
7903,
6153,
355,
11695,
3350... | [
0.9442831873893738,
-5.6273369789123535,
0.017726905643939972,
-0.09789220243692398,
0.6769168376922607,
0.011658141389489174,
0.43611857295036316,
0.020343443378806114,
-0.2277238965034485,
-3.933183193206787,
0.08053980767726898,
-0.250290185213089,
-0.01360019389539957,
0.24636170268058... | 677 | 27.08 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721966 | [S1] How am I gonna do based on this, like, action already- [S2] Yeah. [S1] ... performed? Um, th- so, yeah, like you said, th- uh, like you pointed, this is the third, uh, sub-figure here. Um, | 11.36 | 3.367181 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721966.mp3 | [
6443,
7025,
7036,
8381,
7861,
4447,
12533,
12789,
9310,
8430,
9663,
6502,
9301,
11867,
9305,
11929,
3942,
8213,
5413,
6823,
11360,
7809,
3009,
10595,
10521,
10394,
7766,
95,
3158,
598,
3230,
4707,
6249,
3499,
12222,
12678,
10062,
1450,
8665... | [
0.9576979279518127,
-5.884413242340088,
-0.0992211177945137,
-0.019031643867492676,
0.6465244293212891,
0.07344220578670502,
0.5825424194335938,
-0.03209168463945389,
-0.16693620383739471,
-3.927607297897339,
0.09841997921466827,
-0.22443926334381104,
-0.07106044143438339,
0.13742063939571... | 284 | 11.36 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721967 | [S1] If you were to let it generate all at once and then you translate each action individually, they almost like lose connection to each other, right? [S2] Exactly, yeah. [S1] So this, it, this here might mitigate some of this, this stuff, right? If I have a compound action, like go to the fridge and grab a glass, and the closest, I hope that the closest sentence is the go to fridge, right? [S2] Yeah. [S1] Um, the, the language model might still recover and recognize, "Aha, I haven't, you know, grabbed, uh, haven't grabbed the glass | 29.64 | 3.332193 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721967.mp3 | [
5664,
12329,
12783,
9623,
3938,
8625,
4456,
9477,
9093,
11094,
10966,
8470,
7894,
7958,
7958,
7886,
2661,
5879,
11694,
5525,
2892,
10264,
6705,
11145,
5034,
8479,
2694,
3206,
5918,
724,
4642,
11314,
5681,
4630,
93,
93,
2829,
12279,
12790,
... | [
0.826043426990509,
-5.3790693283081055,
0.04158879071474075,
-0.01643434539437294,
0.44472646713256836,
0.09654445946216583,
0.24144110083580017,
-0.004361298866569996,
0.006345425732433796,
-3.804229736328125,
0.31342875957489014,
0.097361721098423,
0.22712621092796326,
0.1642116308212280... | 741 | 29.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721968 | [S1] how you structure the prompt, whether you- [S2] Yeah. [S1] ... you add, like, multiple things there, and then, um, or, or you change the template here. Because I just defined this template from day one, like, task something, step one, something- [S2] Yeah. [S1] ... step two, something. Maybe there is a better template. Maybe you want to add some instruction there to make it better. And, um, so I, like, | 20.88 | 3.199158 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721968.mp3 | [
7186,
4177,
10089,
9657,
7033,
6973,
8894,
10421,
7830,
4309,
9866,
5766,
2711,
3944,
6488,
7496,
9633,
2908,
2982,
1270,
11995,
12697,
10209,
5534,
3727,
5378,
6015,
3617,
6265,
9201,
1506,
5991,
2871,
3310,
8438,
11366,
5678,
10066,
3408,... | [
0.8539193868637085,
-5.74171781539917,
-0.04543545842170715,
-0.008940874598920345,
0.6218336224555969,
-0.007991126738488674,
0.4299095571041107,
-0.005544605664908886,
-0.25670456886291504,
-3.9169528484344482,
0.13834133744239807,
-0.35029298067092896,
-0.13008899986743927,
0.2743690311... | 522 | 20.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721969 | [S1] Mm-hmm. I mean, it's, yeah. So, sorry, the, the, the recording has stopped on the screen side, but we can, we can still see it. [S2] Okay. [S1] Um, yeah, so, yeah, I was, I was, I was quite interested in, in that, in the sense of, of the prompt structuring because I know that can also, uh, make a big difference, but I also like the sort of approach of not having too many moving parts, you know, in, uh, in one single, | 28.12 | 3.218769 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721969.mp3 | [
8720,
3122,
6147,
11030,
10964,
11548,
11414,
2082,
11347,
9313,
12338,
7713,
6705,
6153,
9764,
7203,
12555,
7468,
6875,
9794,
9470,
8188,
6181,
10603,
11721,
12420,
5699,
1683,
12013,
11228,
9802,
5329,
482,
2977,
3033,
3392,
7193,
4645,
7... | [
0.871046781539917,
-5.484719276428223,
0.009369333274662495,
0.0017027411377057433,
0.4367944896221161,
0.14258521795272827,
0.17699094116687775,
-0.09136262536048889,
-0.0966280996799469,
-3.7257561683654785,
0.2787089943885803,
0.07752928137779236,
0.28500986099243164,
0.104527048766613,... | 703 | 28.12 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721970 | [S1] Because- [S2] You can't restrict the vocabulary dynamically. [S1] Yes. So, uh, I've only done this on smaller model, uh, relatively smaller models like, uh, the GBT Neo, uh, and then I think I might have tried on GBTJ as well, which is a 6 billion per ampere model. And, uh, it actually turns out that they don't do really well with, uh, if you really just constrain the vocabulary that way. Uh, and- [S2] Yeah. [S1] ... specifically just the, just the beam search constraining the, | 26.2 | 2.978801 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721970.mp3 | [
6683,
10601,
9542,
6858,
6053,
6438,
9634,
11244,
7453,
2074,
220,
8148,
10140,
8665,
3473,
8787,
3841,
2019,
2893,
8149,
11236,
7508,
10850,
11747,
4704,
5120,
5589,
11181,
803,
8118,
3180,
11428,
7592,
9198,
11253,
8701,
4645,
7977,
12741... | [
0.9377697110176086,
-5.663862705230713,
0.08947200328111649,
-0.1022489070892334,
0.640984058380127,
0.027877286076545715,
0.3450183868408203,
0.049869660288095474,
-0.021552981808781624,
-3.8892879486083984,
0.051884617656469345,
-0.31208130717277527,
-0.011302727274596691,
0.264288485050... | 655 | 26.2 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721971 | [S1] Uh, vocabulary can generate. But, uh, so my hypo- so this now thoroughly tested because it's now invested on larger models as well. [S2] Mm-hmm. [S1] Um, but my intuition why it doesn't work so well is that this la- language models are really trained on human text, so it really- | 18.64 | 3.269923 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721971.mp3 | [
7193,
5744,
6676,
9143,
3382,
3887,
1782,
1781,
4214,
1186,
816,
3374,
3684,
11675,
5048,
6973,
8125,
6741,
5930,
6431,
340,
4358,
12441,
4451,
4887,
9740,
10888,
9675,
11774,
9966,
10893,
3942,
275,
5069,
4940,
7816,
7064,
12385,
8716,
1... | [
0.8719837665557861,
-5.810280799865723,
-0.0742780864238739,
-0.0236054640263319,
0.64507657289505,
0.0815896987915039,
0.4694797098636627,
0.0017250131350010633,
-0.1339544653892517,
-3.8682892322540283,
0.12330207973718643,
-0.26723361015319824,
-0.07407301664352417,
0.2742324769496918,
... | 466 | 18.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721972 | [S1] It's, it, they're really used to how humans, um, speak a certain language, uh, in this case. [S2] Yeah. [S1] English. So, like, people don't speak things in this way. Step one, something. [S2] Yeah. [S1] Two, something. Step three, something. So, that's why if you really constrain the models this way, a lot of the, um, the, the, the world knowledge encoded in this models are, | 22.88 | 3.036449 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721972.mp3 | [
6681,
5744,
7189,
12646,
6861,
5464,
2960,
3968,
7427,
8453,
4308,
8019,
9483,
8865,
4129,
10006,
11606,
1162,
10,
3089,
5217,
6657,
5680,
8320,
12262,
11243,
2885,
7436,
3356,
8018,
11474,
5441,
11153,
11971,
7489,
6793,
4163,
9794,
12613,... | [
0.8191725611686707,
-5.764402866363525,
-0.03679982200264931,
-0.062388043850660324,
0.6016712188720703,
0.10385515540838242,
0.42342042922973633,
-0.017237037420272827,
-0.18654677271842957,
-3.8780670166015625,
0.08912293612957001,
-0.1924491822719574,
-0.0067478991113603115,
0.223234370... | 572 | 22.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721973 | [S1] So- [S2] Yeah. [S1] It's, uh, it's actually imagining, like, people just, like, asking the docstring here, but instead of letting, keep generating the code, uh, we kinda just stop here. So, okay. [S2] Yeah. [S1] Finish the docstring for us. That's enough. So, um, so, yeah, so it's actually doing some, this kind of docstring, uh, it generates this docstring thing. And I, the reason I think the smaller codecs model are actually | 26.64 | 3.297082 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721973.mp3 | [
6170,
3171,
9941,
12525,
12462,
4698,
3633,
1305,
8088,
452,
9015,
5359,
3311,
2663,
4325,
3691,
12164,
11220,
11758,
6063,
7594,
10218,
7026,
1081,
4655,
11759,
8613,
4885,
12196,
11669,
6416,
2778,
2905,
2952,
6233,
11826,
5344,
7193,
944... | [
0.6700606346130371,
-5.685357093811035,
0.0367896668612957,
-0.11807820200920105,
0.626650869846344,
-0.017076784744858742,
0.4047543704509735,
0.014010046608746052,
-0.2187385857105255,
-3.877657175064087,
0.04719560593366623,
-0.3426283597946167,
-0.04568710923194885,
0.29528045654296875... | 666 | 26.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721974 | [S1] consists of doctrine and, and, and, and the code. So it not only, not only can handle code really well, it can also generate really realistic doctrines. So, and, and people in doctrine, they don't write in, um, like- [S2] Yeah, they don't write a novel. [S1] Yeah, so- | 18.08 | 3.074639 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721974.mp3 | [
6748,
9640,
6461,
11885,
8917,
8152,
322,
12142,
11724,
3876,
5456,
3794,
5943,
5374,
3234,
1449,
8728,
5597,
7934,
8053,
9316,
9580,
3480,
10049,
4528,
2908,
12556,
11926,
9766,
7308,
9966,
9198,
7151,
8781,
8774,
4677,
9353,
12321,
12437,... | [
0.869388997554779,
-5.729149341583252,
-0.07191532850265503,
0.04122094437479973,
0.6184346675872803,
0.05912765488028526,
0.2673144042491913,
0.06052269786596298,
-0.2292269915342331,
-3.8301949501037598,
0.025174515321850777,
-0.2502462565898895,
0.015055197291076183,
0.17382444441318512... | 452 | 18.08 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721975 | [S1] Yeah, so you write something really step by step and have more structure in it. So that's my intuition why, uh, actually does really well with this task. So you can really process this sequential, um, like logical reasoning better than- [S2] Yeah. [S1] ... the same, uh, size GPT- GPT-3 model. But of course, if you use a larger model, that potentially be more helpful. [S2] Mm-hmm. [S1] Yeah. | 25.48 | 3.374109 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721975.mp3 | [
7201,
4176,
9484,
11211,
11253,
3005,
4389,
9685,
8621,
3548,
9900,
7974,
7982,
3874,
6378,
4394,
10907,
9436,
11685,
3413,
3412,
10722,
8561,
12634,
6032,
8179,
8432,
10276,
3944,
9131,
9537,
11926,
4828,
283,
9095,
11238,
6292,
9043,
1267... | [
0.9422005414962769,
-5.6464152336120605,
0.03234102204442024,
-0.11331897228956223,
0.7039747834205627,
0.008174869231879711,
0.36874642968177795,
0.02346210554242134,
-0.1464046984910965,
-3.9221444129943848,
0.07162120193243027,
-0.32779446244239807,
-0.05520383641123772,
0.2201469242572... | 637 | 25.48 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721976 | [S1] came to my attention right here is this top row right here, which I found hilarious. So this is, the task is complete Amazon Turk surveys. [S2] [LAUGHS] [S1] So the, the, the four steps, apparently, that you need to do is walk to home office, sit on chair, switch on computer, | 20.32 | 3.287665 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721976.mp3 | [
6688,
6753,
12114,
9708,
12764,
11669,
11669,
5653,
5678,
4673,
7312,
5440,
7147,
10966,
3653,
3108,
8558,
6455,
6455,
6391,
9078,
12645,
12677,
11726,
9615,
6945,
12377,
9812,
7324,
4772,
4707,
7355,
4820,
9859,
7249,
10928,
7208,
5656,
11... | [
1.0105258226394653,
-5.629255771636963,
-0.09288579970598221,
0.0937008336186409,
0.4174383580684662,
0.0520334467291832,
0.2559551000595093,
-0.07702741771936417,
-0.007569487206637859,
-3.7629387378692627,
0.29231274127960205,
0.2863869071006775,
0.18394596874713898,
0.05734356492757797,... | 508 | 20.32 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721977 | [S1] So, so, like I said, this tests are generated by, um, by, by cross source, uh, from humans. [S2] Yeah. [S1] And this, uh, the humans here happen to be Amazon Turkers. So- [S2] Yeah. [S1] ... one of them decided that, okay, if you want me to generate- [S2] [LAUGHS] [S1] ... like generate some tests, I will say, like, just complete surveys on Amazon, uh, Amazon. [S2] Yeah. [S1] So, they decided to put one of this here. And we found this hilarious too. So, um, uh, | 24.88 | 3.315456 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721977.mp3 | [
8921,
3869,
8908,
8181,
8183,
9574,
7328,
3474,
6620,
3860,
8565,
10484,
7909,
3285,
4193,
12508,
5809,
7263,
5027,
5592,
8053,
5363,
7907,
3795,
6763,
6434,
6884,
5852,
10748,
12572,
10018,
9076,
12516,
6489,
3025,
11698,
11771,
4887,
2560... | [
0.6775036454200745,
-5.69859504699707,
0.01687241718173027,
-0.12736733257770538,
0.5556808710098267,
0.054579537361860275,
0.5298488140106201,
0.016379905864596367,
-0.1902722716331482,
-3.890202045440674,
0.04756588116288185,
-0.2592381536960602,
-0.0522020123898983,
0.32654863595962524,... | 622 | 24.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721978 | [S1] Like, like I said, so, so this language model, so they c- can't really handle, um, anything that you wanted to, um, to- [S2] Yeah. [S1] ... generate, so, w- because we, we did put a, um, example, uh, in the front. So, uh, I think in this case, the example ha- happens to be something related to computer, and, um, the, the models actually happen to reason | 22.84 | 3.272762 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721978.mp3 | [
6675,
6160,
6236,
8029,
11259,
7493,
11361,
12052,
7586,
7649,
7584,
4764,
9257,
10408,
8232,
12355,
9753,
6169,
3358,
11198,
12379,
7140,
12214,
3425,
10640,
5555,
11706,
4436,
8016,
5595,
10422,
8990,
3778,
770,
5413,
5844,
5584,
3035,
10... | [
0.8550583124160767,
-5.730362892150879,
-0.05401430279016495,
-0.008326640352606773,
0.6621048450469971,
0.0783720463514328,
0.44921961426734924,
-0.0032727133948355913,
-0.18633075058460236,
-3.90053391456604,
0.1337209939956665,
-0.23181390762329102,
0.003707835217937827,
0.2460391521453... | 571 | 22.84 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721979 | [S1] So, so, uh, in terms of, uh, so a bit background here for, for the virtual environment. So it comes in two versions. One is the, I think the, they call the evolving graph version, which is a pure, like you said, a state machine, a, a Python, like written in Python. So it just goes in and then checks which, uh, whether the actions can be parsed and then- [S2] Yeah. [S1] ... satisfy the, the common sense constraint. And they are the version they implement | 29.16 | 3.158007 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721979.mp3 | [
5594,
10549,
8468,
11739,
5556,
8422,
5982,
2202,
6179,
7402,
8804,
6244,
9316,
8371,
7185,
4749,
1313,
720,
3353,
11571,
7860,
10862,
8341,
8341,
7829,
7828,
5205,
662,
1115,
9257,
5809,
8739,
11190,
7981,
8558,
9006,
6503,
8551,
8614,
4... | [
0.7392672896385193,
-5.628685474395752,
0.09420160949230194,
-0.08950125426054001,
0.6807747483253479,
0.011769495904445648,
0.33903494477272034,
0.03207407891750336,
-0.05187218636274338,
-3.9307656288146973,
0.028087319806218147,
-0.28027182817459106,
-0.025733232498168945,
0.33241286873... | 729 | 29.16 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721981 | [S1] Uh, after the posting of this paper, they, they tried to ask the, uh, GPT-3 model, how do I start a company? [S2] [LAUGHS] [S1] So, yeah, they, they do have a lot of knowledge for this, and as long as you can provide a set of actions that are, um, necessary to complete these tasks, | 17.16 | 3.340186 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721981.mp3 | [
9369,
5665,
6966,
3446,
3383,
3757,
1123,
9251,
12726,
11766,
6967,
3883,
10072,
7983,
3202,
3759,
8481,
4473,
5822,
5231,
4893,
5528,
6996,
12510,
3390,
6254,
4186,
347,
10159,
788,
5576,
6178,
11632,
7542,
10199,
4879,
8433,
3949,
10645,
... | [
0.7464563846588135,
-5.798789024353027,
-0.02112961746752262,
-0.05289778858423233,
0.6202359795570374,
-0.03252953663468361,
0.518204927444458,
0.03103616274893284,
-0.1241600513458252,
-3.9092626571655273,
0.09800247848033905,
-0.3741280734539032,
-0.08786267787218094,
0.2640593349933624... | 429 | 17.16 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721982 | [S1] And you know you can make a clip classify things by just evaluating a bunch of different- [S2] Yeah. [S1] ... strings with it. So I like try to, I try to evaluate the strings, "Go left, go up, go right, go down." Or like, "Pac-Man should go left, Pac-Man should go up." But it never worked out. Um, so if you can, if you could get something like this running, this would be amazing. [S2] Yeah. [S1] Like, [LAUGHS] uh, with, maybe with your knowledge, maybe Pac-Man isn't the right environment because Clip was trained on whatever picture scrape from Instagram. | 29.88 | 3.252093 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721982.mp3 | [
5672,
5160,
12135,
11759,
10094,
8709,
2370,
11653,
6726,
10573,
11638,
7845,
6862,
10115,
3414,
10914,
7659,
12503,
3109,
11758,
12254,
7385,
12195,
4075,
12451,
10225,
3306,
10734,
10006,
7793,
1771,
10909,
4592,
3237,
8550,
11710,
11133,
1... | [
0.8985422849655151,
-5.336024284362793,
0.10385817289352417,
-0.030525436624884605,
0.39255061745643616,
0.005447407253086567,
0.24841879308223724,
-0.018057815730571747,
-0.019894033670425415,
-3.8943138122558594,
0.24539771676063538,
0.11930625140666962,
0.18824611604213715,
0.1648416370... | 747 | 29.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721983 | [S1] Excellent. All right. Thank you. And I hope to see you again. [S2] [LAUGHS] Yeah, I'm like, I, it would be a honor to [LAUGHS] always to, to be here. Yeah. [S1] Excellent. All right. Bye-bye. [S2] Yeah. See ya. | 15 | 2.933289 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p168_BV17h4y1j7aZ_p168_m4-dialogue_0721983.mp3 | [
9828,
11766,
10150,
11362,
7084,
3032,
466,
5815,
2709,
813,
4124,
6149,
9242,
5161,
3521,
3820,
12337,
7208,
9794,
5657,
4759,
9971,
5643,
9910,
3383,
3373,
1363,
6062,
11710,
7598,
7493,
12432,
8817,
5889,
1834,
4328,
5865,
6432,
4534,
... | [
1.111018180847168,
-5.721179485321045,
-0.11664321273565292,
0.10900687426328659,
0.46109792590141296,
0.01637893170118332,
0.49616220593452454,
-0.0934654101729393,
-0.04226372390985489,
-3.8876824378967285,
0.1671454757452011,
-0.10576074570417404,
-0.18270806968212128,
0.089852504432201... | 375 | 15 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441414 | [S1] Welcome everyone. Today I have with me here Andrey Shmogunov. Is that approximately correct, Andrey? [S2] [LAUGHS] Approximately correct, yeah. Thank you. Thanks for having me. [S1] Thank you. So you're, you're one of the authors of the Hyper-transformer, uh, paper, and this is a, a pretty cool paper I found, uh, | 19.96 | 3.171058 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441414.mp3 | [
6184,
560,
7708,
627,
8109,
11130,
7908,
7392,
11682,
8556,
10404,
11763,
12142,
235,
283,
11598,
6341,
116,
8495,
3318,
822,
7180,
3590,
5637,
1100,
2073,
5753,
5137,
11224,
4509,
3655,
8001,
12663,
12781,
12718,
12181,
12099,
6607,
4789,
... | [
0.9204923510551453,
-5.643076419830322,
-0.2101612240076065,
0.13997747004032135,
0.36435821652412415,
0.00532927829772234,
0.2306813895702362,
-0.0746343806385994,
0.012088282965123653,
-3.7947781085968018,
0.3164755702018738,
0.15095792710781097,
-0.07282935082912445,
0.04828127101063728... | 499 | 19.96 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441415 | [S1] is already good enough, embedding model, then, um, in principle, all you need to do is look at the final embeddings produced for different images, and kind of based on that, figure out how you need to assign labels to essentially these embeddings. [S2] Mm-hmm. [S1] So in practice, as we've seen, all that matters for, especially for very large models that, you know, can, uh, have a very large embedding inside, is to just generate the final layer. But once you get into the land of smaller models, | 28.72 | 2.816372 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441415.mp3 | [
6185,
5681,
9244,
12622,
4355,
840,
5423,
2622,
3285,
3364,
8157,
11758,
7296,
6538,
6982,
11363,
2917,
10526,
11459,
3519,
3621,
7244,
3884,
6902,
4413,
6253,
3434,
9201,
1465,
3888,
1265,
7200,
12087,
3598,
6746,
9207,
2399,
11656,
7446,
... | [
1.4566985368728638,
-5.276408672332764,
-0.1225353330373764,
0.1063166931271553,
0.6222379207611084,
-0.2507466971874237,
0.563601553440094,
0.019340405240654945,
0.0058227903209626675,
-3.814290761947632,
0.5499803423881531,
-0.029123013839125633,
-0.3378589153289795,
0.0995580330491066,
... | 718 | 28.72 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441416 | [S1] Yeah. [S2] ... process. [S1] So, is it, is it fair to say that you're, so this, this, what you call support set, that is essentially the data set | 9.64 | 3.131845 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441416.mp3 | [
3946,
1747,
10117,
8182,
4331,
3178,
6317,
6380,
6388,
3764,
3809,
9504,
4323,
4187,
4819,
3928,
6937,
9890,
4653,
4707,
6307,
7251,
3553,
3026,
11126,
7924,
5270,
1814,
9825,
12387,
9895,
9062,
3917,
7265,
9878,
9946,
8880,
9830,
12702,
... | [
0.9015756845474243,
-5.834453105926514,
-0.1817466765642166,
0.22798366844654083,
0.27006781101226807,
-0.040989190340042114,
0.2714889645576477,
-0.10896111279726028,
-0.14644846320152283,
-3.8286941051483154,
0.18104398250579834,
0.13026507198810577,
0.05086876451969147,
0.08805448561906... | 241 | 9.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441417 | [S1] of the, the few shot task, right? It's, it's like here are ten images of dogs and cats with corresponding labels, uh, which in, this is a diagram of, of your architecture in general. [S2] Okay. [S1] So this is the support set with the samples and the labels. And then you make, you make use of lots of signals throughout the network such that, uh, as you said, you make sure you first build the first layer and then based on that build the second layer. [S2] Okay. [S1] So if we, if we quickly walk through it, one core | 29.88 | 3.396965 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441417.mp3 | [
3625,
8227,
12223,
7990,
8054,
11054,
3043,
489,
489,
3432,
6185,
5785,
5136,
3037,
6055,
1374,
1886,
3502,
2331,
3625,
1026,
3046,
5027,
9080,
4024,
2537,
7363,
9157,
7813,
5189,
3853,
12778,
7667,
7640,
5494,
5310,
3326,
9896,
10944,
40... | [
0.8410271406173706,
-5.451140403747559,
-0.0062120528891682625,
0.12211857736110687,
0.31851619482040405,
0.0013681385898962617,
0.09451832622289658,
0.024160416796803474,
-0.13730941712856293,
-3.7943460941314697,
0.33844974637031555,
0.16659004986286163,
0.04331836476922035,
0.1303164958... | 747 | 29.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441418 | [S1] component is this, uh, image feature extractor that is a trained, uh, let's say a convnet that is applied to each image individually and just extract some sort of a feature map. And this feature map is then given to every single computation layer in your, in your set, right? [S2] Yeah. [S1] So your main model is this, this transformer thing here. [S2] Mm-hmm. | 24.52 | 3.338886 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441418.mp3 | [
4196,
9641,
5877,
8231,
4139,
10920,
1981,
11454,
10348,
11037,
8844,
5935,
9318,
9228,
5440,
4697,
9454,
12183,
10143,
1883,
8145,
705,
2625,
7535,
10079,
409,
5537,
2889,
3536,
9242,
10029,
7284,
6830,
6892,
4776,
2134,
4535,
4079,
4833,
... | [
0.8825263381004333,
-5.609006881713867,
-0.027319220826029778,
0.1178915798664093,
0.31230831146240234,
-0.10161904990673065,
0.13074997067451477,
0.019544225186109543,
-0.12048814445734024,
-3.743720531463623,
0.24835087358951569,
0.18645833432674408,
0.1003665179014206,
0.110800653696060... | 613 | 24.52 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441419 | [S1] Correct. [S2] That it takes in, as you can see, it takes in these, these, uh, embeddings of the support set. It takes in the labels, uh, obviously, right? It needs to know what it needs to classify how. And it gener- it takes in this thing right here. And I think in the first layer, this is kind of the same as these image embeddings. It's, it's another embedding, right? It's sort of- [S1] It's another embedding- [S2] A signal. [S1] ... smaller. Yeah, but it's- [S2] Yeah. [S1] It's, it's basically- | 29.88 | 3.291779 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441419.mp3 | [
4140,
11988,
6364,
3821,
2972,
3476,
6817,
6686,
3072,
6118,
4543,
10686,
12661,
8849,
6237,
8384,
724,
5725,
9363,
9736,
8739,
10182,
11800,
5728,
2593,
8648,
5115,
11765,
12187,
12164,
6826,
11684,
3561,
1488,
9249,
9591,
11162,
12199,
21... | [
0.9524767994880676,
-5.453127384185791,
-0.015141692943871021,
0.09621754288673401,
0.4827651381492615,
-0.06798211485147476,
-0.023644546046853065,
0.05150788649916649,
0.046118784695863724,
-3.8041183948516846,
0.316598504781723,
0.18349695205688477,
0.0779377818107605,
0.037892427295446... | 747 | 29.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441420 | [S1] It's another embedding- [S2] A signal. [S1] ... smaller. Yeah, but it's- [S2] Yeah. [S1] It's, it's basically produced from the same images, essentially. [S2] I guess we'll, we'll, we'll come, like this is in subsequent layers, this will actually be different. So what we do is the transformer here, it will produce the weights of the first layer. And as you said, we don't just produce | 21.64 | 3.132143 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441420.mp3 | [
6404,
5935,
9042,
5857,
2864,
4265,
6939,
9050,
8842,
3794,
3408,
2660,
7716,
5771,
9013,
8884,
8892,
4194,
3438,
8940,
2957,
5845,
8797,
4130,
6573,
11763,
8916,
4380,
6143,
4577,
5539,
2905,
6108,
12247,
12435,
4884,
5781,
3951,
3118,
2... | [
0.9922854900360107,
-5.604620933532715,
-0.12318569421768188,
0.18354563415050507,
0.4279736280441284,
-0.03215212747454643,
0.24394194781780243,
0.008498422801494598,
-0.054432813078165054,
-3.836357355117798,
0.32927459478378296,
0.24818158149719238,
0.05980927124619484,
0.07050098478794... | 541 | 21.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441421 | [S1] ... the first layer and the second and the third in one batch. But what is, what seems to be really important is now we actually forward propagate, uh, I need a different color here. [S2] Okay. [S1] We, we forward propagate the, the support set through the weights we've just generated. [S2] Okay. [S1] And that will give us the next layer's representation. And then that can be used again by the transformer to generate the next layer's weights. [S2] Yeah. [S1] Along with the embeddings of the original images, along with the labels and so on. So, | 29.88 | 3.269941 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441421.mp3 | [
6704,
8704,
8182,
4586,
8696,
500,
10669,
10646,
8150,
5092,
10721,
5577,
6849,
3293,
8119,
10125,
9166,
5477,
3029,
4716,
11370,
11315,
5298,
11801,
12214,
10733,
12647,
8718,
466,
4470,
10720,
8144,
6118,
11245,
9199,
9953,
11923,
9189,
8... | [
0.898984968662262,
-5.451369285583496,
0.08253820985555649,
0.08134981989860535,
0.35424986481666565,
-0.000024114739062497392,
0.0654682144522667,
0.044207070022821426,
-0.029923908412456512,
-3.8010804653167725,
0.4298117160797119,
0.12526650726795197,
0.031135911121964455,
0.12090409547... | 747 | 29.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441423 | [S1] ... This looks, it looks, it's two layers right here, and it looks already quite complicated, right? Here is like an entire transformer, right? [S2] Mm-hmm. [S1] Um, that, and then, uh, that transformer generates a set of weights, right? And then I forward propagate a signal through the weights that were generated by using that signal as an input. [S2] Mm-hmm. [S1] Right? So I'm, I'm imagining the computation graph here gets, | 27.68 | 3.350174 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441423.mp3 | [
9363,
9370,
9949,
9492,
9435,
9890,
12442,
9939,
7267,
1555,
2954,
12195,
11161,
4963,
5609,
3545,
2896,
704,
4625,
8361,
9756,
6375,
2709,
2644,
3422,
10413,
10901,
9385,
11356,
9708,
6120,
2888,
3776,
5673,
8746,
12262,
12244,
2265,
4136,... | [
0.8644829988479614,
-5.582186698913574,
-0.013486387208104134,
0.014698367565870285,
0.4364038109779358,
-0.028848091140389442,
0.18284422159194946,
0.040208958089351654,
0.13755162060260773,
-3.8566715717315674,
0.2886788845062256,
0.13852033019065857,
0.03818545863032341,
0.0213261879980... | 692 | 27.68 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441424 | [S1] Pretty, pretty iffy, quite freak, like quite fast. And then there is another transformer. [S2] [LAUGHS] [S1] And then I'm back, prop through all of this back, right? How does, how, like, what's the concerns with, like, stability here in, in, like, how do, how big does the computational graph get? Is this a problem? | 19.12 | 2.883291 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441424.mp3 | [
5673,
6642,
1913,
5522,
10586,
11171,
6723,
3469,
10054,
10118,
4866,
10793,
6193,
7381,
9833,
8225,
11803,
9843,
2608,
11698,
3042,
10642,
11550,
5890,
7431,
7557,
6688,
11502,
11229,
11676,
12638,
6633,
6000,
1898,
4870,
7362,
9329,
12388,
... | [
0.7763431668281555,
-5.675063610076904,
-0.1391022950410843,
0.07856187969446182,
0.2832910120487213,
0.03981025889515877,
0.27058932185173035,
-0.021701524034142494,
0.07898389548063278,
-3.8566946983337402,
0.267472505569458,
0.12810583412647247,
0.08253157883882523,
0.06318476051092148,... | 478 | 19.12 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441425 | [S1] ... right, in this graph. And so I- [S2] Yeah. [S1] ... I assume it will still be, like, it should be still proportional to the number of layers. But it is true that, like, when you generate the final layer, you essentially have to back propagate through all of the transformers that you have, right? Like, if you have multiple layers- [S2] Yeah. [S1] ... in each transformer, you have to back propagate through all of them. But in practice, this thing was surprisingly stable to train, actually. That was one of the things that surprised me. | 24.76 | 2.839083 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441425.mp3 | [
5648,
2601,
3372,
2966,
6526,
6989,
6738,
5824,
10087,
7310,
5719,
3144,
4526,
2023,
1758,
6032,
4816,
12012,
4587,
2965,
4093,
7669,
7540,
6461,
4340,
3505,
7473,
8804,
9518,
11631,
9340,
6766,
1824,
6237,
9455,
8715,
3425,
8576,
5932,
7... | [
1.5691580772399902,
-5.348242282867432,
-0.2372657209634781,
0.1444203406572342,
0.652441680431366,
-0.18218007683753967,
0.6749507784843445,
0.03934149816632271,
0.02424284629523754,
-3.867039442062378,
0.5777460336685181,
-0.04830328747630119,
-0.3675459325313568,
0.08589291572570801,
... | 619 | 24.76 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441427 | [S1] in like a BERT model when I want to classify something. I have one token per output that I want to do. They, they, these have different embeddings, so like, they're like addresses of the weights that I want to output. [S2] Mm-hmm. [S1] And, uh, yeah, this, this whole thing, it's, it's, uh, | 15.64 | 3.370516 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441427.mp3 | [
11914,
6929,
6758,
12118,
10053,
8197,
3795,
5477,
6646,
7322,
12123,
6647,
532,
5664,
8684,
5494,
2900,
5508,
3347,
7785,
11728,
6812,
10790,
8044,
10870,
7989,
9492,
5487,
10934,
7894,
1172,
5429,
10087,
7942,
6583,
7015,
2669,
3190,
9831... | [
1.214316725730896,
-5.5629730224609375,
-0.19206397235393524,
0.09373193234205246,
0.4508565068244934,
-0.15090076625347137,
0.15136846899986267,
0.04137988016009331,
0.1375805139541626,
-3.7787938117980957,
0.26419419050216675,
0.008812358602881432,
0.0030615669675171375,
-0.1137185022234... | 391 | 15.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441428 | [S1] ... then there's just, just as transformer. But y- you have, you, you already said with respect to like the last layer, that this is implementable. But you also make the case that if I have a two-layer transformer, I can implement like a nearest neighbor algorithm. Is, um, like, do, do you wanna maybe- [S2] Yeah. [S1] ... just briefly, the, what's the idea behind- [S2] [LAUGHS] [S1] ... how, how does, how does a two-layer transformer implement nearest neighbor? | 28.6 | 3.389906 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441428.mp3 | [
8793,
3633,
3712,
5031,
7614,
12366,
7685,
8262,
4115,
5313,
10087,
12734,
5461,
2957,
4945,
1672,
11977,
4053,
11063,
3561,
848,
3088,
2880,
4961,
6177,
4701,
4645,
4722,
4160,
7745,
10128,
1375,
3948,
6024,
5120,
3033,
11766,
11243,
11182... | [
0.9250755310058594,
-5.529598236083984,
-0.0049015977419912815,
-0.00986643135547638,
0.3428615927696228,
-0.07383643090724945,
0.15806753933429718,
0.015187634155154228,
0.07946911454200745,
-3.8020853996276855,
0.2766440808773041,
0.1406143456697464,
0.01464857067912817,
0.08561866730451... | 715 | 28.6 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441429 | [S1] And then at the second- [S2] It's interesting that the, that, that, the, the, the, the weights, they don't care at all about the unlabeled samples. Like they learn to ignore- [S1] Yeah. [S2] ... the unlabeled samples. That's pretty interesting. [S1] [LAUGHS] | 12.4 | 3.280151 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441429.mp3 | [
9232,
12142,
12213,
8796,
3666,
3948,
10675,
8620,
8869,
4257,
9454,
12150,
8873,
3443,
11503,
6875,
402,
6025,
2897,
9069,
10021,
8392,
921,
11107,
12051,
6228,
12107,
3553,
8029,
6561,
3363,
5968,
12196,
9940,
7261,
3035,
11197,
9526,
114... | [
0.8693853616714478,
-5.812838554382324,
-0.12463635206222534,
0.10732954740524292,
0.381793349981308,
-0.006835398264229298,
0.3584643602371216,
-0.05770046263933182,
0.0922674834728241,
-3.7748029232025146,
0.1699882596731186,
0.014858192764222622,
0.0063058664090931416,
0.086341097950935... | 310 | 12.4 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441432 | [S1] Yeah, I think actually that's a wonderful idea. Actually, as a matter of fact, what, what we do right now is we say, "Oh, we think that's what's happening," and then we look at the attention masks and we see that, yes, that's mostly what's happening. But you're absolutely right, that if we were certain that we wanted to restrict the flow in, of information in a particular way, we could very well manipulate the, basically the masking of each self-attention layer. [S2] Mm-hmm. [S1] And this way, very carefully restrict how the computation should actually be. | 29.88 | 2.947683 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441432.mp3 | [
4779,
5681,
6220,
9930,
9994,
11602,
11104,
12001,
4333,
3518,
1261,
9261,
11646,
12629,
3808,
6504,
11673,
11858,
11939,
2476,
9599,
8620,
3170,
9554,
12754,
10200,
8484,
3658,
9044,
9669,
1237,
1233,
3436,
7534,
12213,
11571,
11644,
12378,
... | [
1.50787353515625,
-5.234848976135254,
-0.09280849248170853,
0.08142275363206863,
0.6272925734519958,
-0.2712464928627014,
0.6340591311454773,
0.03585723415017128,
0.05530214309692383,
-3.892742395401001,
0.6336116790771484,
-0.10780659317970276,
-0.3536927103996277,
0.19733008742332458,
... | 747 | 29.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441433 | [S1] Yeah. To be honest, uh, it's very difficult to interpret them, and I think I would rather not, uh, go into too much because it, it, we, we really have a hard time understanding what these mean. But I think to some degree- [S2] Yeah. [S1] ... one thing to observe is that, first of all, we discussed several ways of generating weights. And one of them, it, it's, it's all- [S2] Mm-hmm. [S1] ... it all ends up being how you take the outputs produced by a transformer and how you combine them into single convolution filters. | 29.56 | 2.827714 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441433.mp3 | [
9881,
6806,
7436,
10009,
10003,
12183,
10135,
7510,
4591,
1014,
4015,
3950,
1390,
3738,
6819,
9380,
9908,
6308,
7268,
1187,
6802,
6176,
2578,
6878,
9581,
12717,
12781,
11053,
10085,
9498,
9898,
11273,
8147,
3604,
8361,
4948,
12545,
11649,
7... | [
1.4194912910461426,
-5.3456926345825195,
-0.13852842152118683,
0.047943584620952606,
0.6727785468101501,
-0.21636749804019928,
0.6144035458564758,
0.07111703604459763,
0.06028972938656807,
-3.840365171432495,
0.5478352904319763,
-0.10856636613607407,
-0.30398833751678467,
0.173931837081909... | 739 | 29.56 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441434 | [S1] ... What's happening. [S2] So, yeah, you, you find that in the small models, you, you, you fare better generating all the weights than if you, uh, and in the, in the larger models, the strategy is essentially to only train the model to produce the last layer and then use regular back prop through that generated layer- [S1] Mm-hmm. [S2] ... to essentially learn the lower layers. And that might be | 24.48 | 3.329974 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441434.mp3 | [
6740,
9299,
8990,
8853,
8926,
2204,
8924,
6308,
5848,
3416,
6891,
9005,
6436,
6379,
4194,
6950,
4323,
9428,
6739,
6667,
9803,
12354,
12288,
3600,
3601,
5304,
8210,
3922,
2896,
9639,
11045,
1237,
5673,
7318,
10052,
10710,
9135,
1461,
2317,
... | [
0.920060932636261,
-5.52853536605835,
-0.10376390069723129,
0.017181212082505226,
0.4159982204437256,
-0.035174157470464706,
0.2108106166124344,
-0.026443330571055412,
-0.03889824450016022,
-3.780911922454834,
0.21451281011104584,
0.0677575021982193,
0.052827175706624985,
0.059652023017406... | 612 | 24.48 |
bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441435 | [S1] I think I copied the paper twice, I'm sorry. [S2] [LAUGHS] [S1] Um, so this, the, the, yeah, you're going, you're going to generate for each of these weight tokens, you're going to generate some sort of an output, uh, which you can interpret directly. Is it also possible to interpret this output as, let's say, the embedding | 20.08 | 3.209534 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p169_BV17h4y1j7aZ_p169_m4-dialogue_0441435.mp3 | [
7216,
7776,
10926,
11262,
4940,
4317,
4440,
6138,
12049,
9243,
7067,
6078,
5086,
6240,
11619,
380,
7926,
8749,
7720,
5736,
2380,
4358,
3616,
521,
1399,
6696,
10401,
5044,
7598,
10054,
7243,
8296,
1461,
284,
194,
8241,
11521,
1961,
35,
267... | [
1.042985439300537,
-5.605748176574707,
-0.16133072972297668,
0.11402083188295364,
0.5652390122413635,
0.0312519371509552,
0.1648373007774353,
-0.04710384085774422,
0.05864627659320831,
-3.8186793327331543,
0.3806505799293518,
0.22851510345935822,
0.054278939962387085,
-0.027905112132430077... | 502 | 20.08 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758501 | [S1] Welcome to the show and, uh, thanks for being here. [S2] Thank you. Thank you for having me. I'm excited to be here. [S1] So the, the goal of today, it's a little bit different because I'm a total newb at security. Most of the audience of this channel is not, is into machine learning. Maybe some know about security, some know about, um, | 21.36 | 3.255997 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758501.mp3 | [
6169,
5744,
11415,
51,
5494,
11253,
11005,
7956,
9441,
11547,
3572,
11054,
8798,
8231,
8295,
8231,
12374,
9224,
6168,
11433,
7744,
5520,
7657,
12630,
3343,
3278,
3214,
734,
3231,
3287,
2261,
3750,
3422,
4761,
1561,
2907,
7446,
9859,
7428,
... | [
0.8178518414497375,
-5.747809410095215,
-0.061425868421792984,
-0.06989362835884094,
0.6616123914718628,
0.02949465438723564,
0.2434031069278717,
-0.008457439951598644,
-0.09167752414941788,
-3.7878799438476562,
0.20705828070640564,
0.02488812431693077,
0.05988048017024994,
0.0492848008871... | 534 | 21.36 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758502 | [S1] ... the, the censorship apparatus that's in place around the world and what people do about it, I think most won't. So today I'll be asking mostly noob-ish questions and, um- [S2] Let's do it. [S1] ... we'll have you here to guide us through everything, to guide us through, like, what's happening in this world. So maybe you first can start off a little bit. How did you | 22.76 | 3.405873 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758502.mp3 | [
9818,
9553,
11410,
7405,
6157,
979,
12118,
4998,
4997,
10181,
2322,
3098,
129,
5014,
4435,
5489,
2944,
6567,
9206,
12131,
10765,
913,
6616,
8622,
2956,
3923,
10090,
10185,
11743,
4950,
8424,
3818,
9778,
9214,
9149,
7348,
8296,
8881,
5043,
... | [
0.7758394479751587,
-5.7187581062316895,
-0.017294954508543015,
-0.10178013145923615,
0.5382451415061951,
0.010895652696490288,
0.30890268087387085,
-0.012678312137722969,
-0.05630967393517494,
-3.8077476024627686,
0.29193341732025146,
0.018391963094472885,
0.07018794864416122,
0.061356931... | 569 | 22.76 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758503 | [S1] ... still discovering all sorts of new ways you can apply these techniques. Um, and that was one of my motivations initially, actually, of bringing this censorship, 'cause this project was really a, the entire field of censorship's first foray into using AI and ML-like techniques. [S2] And if you, if you talk about censorship, what do you mean exactly by that? [S1] Yeah, so there's a t- | 22.64 | 2.970643 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758503.mp3 | [
8619,
8161,
6464,
6115,
10364,
69,
2953,
3374,
8152,
12506,
6629,
8118,
2917,
3385,
2958,
4237,
8487,
7806,
5174,
3936,
6601,
10494,
2958,
8849,
2944,
1302,
5122,
8332,
11615,
142,
2798,
10094,
4934,
7502,
3779,
1344,
4997,
7243,
1893,
33... | [
0.8325424790382385,
-5.657095432281494,
-0.11396369338035583,
-0.19002416729927063,
0.6603766083717346,
-0.1174459159374237,
0.3990328311920166,
-0.021159963682293892,
-0.05584384873509407,
-3.972188949584961,
0.23558250069618225,
-0.01334262266755104,
-0.31563475728034973,
0.1747082918882... | 566 | 22.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758504 | [S1] ... that web traffic crosses through the, the border of the country. It is scanned, parsed, and inspected by some machines that physically reside in the network called middle boxes, called, because they're in the middle of the network. And these middle boxes examine your requests and they say, "Is this something we should allow or not?" And if the answer's no, they either inject traffic to take down your connection or they drop your connection or they do something to disrupt what's going on. [S2] Mm-hmm. [S1] And you'll notice everything I just said there, there's no human in the loop. There's, there's no, like, human content review or anything like this. | 29.8 | 2.809218 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758504.mp3 | [
4316,
1034,
2804,
11764,
11556,
2658,
12114,
10217,
3045,
12287,
4542,
4376,
2985,
7087,
6801,
12461,
9594,
2979,
7974,
7931,
6900,
7592,
1481,
8358,
3480,
3392,
3153,
3208,
5990,
10510,
6278,
787,
3374,
5999,
871,
3879,
6421,
3210,
3082,
... | [
0.676882803440094,
-5.502530574798584,
-0.09304014593362808,
-0.33422887325286865,
0.8525396585464478,
-0.1703891009092331,
0.35024380683898926,
0.011835429817438126,
-0.1420021951198578,
-3.9644830226898193,
0.30985113978385925,
0.007999804802238941,
-0.2873878479003906,
0.227655515074729... | 745 | 29.8 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758505 | [S1] I guess it's, it's a little bit like if everyone encrypts, like with HTTPS nowadays, everyone does it, so you can't conceivably block HTTPS, you know, just, just because you don't like some traffic. But if there's a new type of encryption, uh, you, you can probably- [S2] Exactly correct. [S1] It's, it's probably only the people that have something to hide that use that type of encryption. So is, is like, | 23.72 | 3.404375 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758505.mp3 | [
8752,
11583,
9590,
12379,
9692,
11693,
11755,
10676,
3419,
8025,
393,
1420,
7210,
12143,
11678,
11614,
8833,
9360,
8746,
8257,
3026,
3417,
3920,
3776,
3632,
8804,
11751,
12182,
10257,
401,
987,
11630,
5275,
8101,
11051,
5825,
7854,
5892,
36... | [
0.7805511355400085,
-5.692116737365723,
-0.08774083107709885,
-0.11235616356134415,
0.5676640272140503,
-0.04914914816617966,
0.26374518871307373,
-0.03845364600419998,
-0.0875590518116951,
-3.8178646564483643,
0.26327985525131226,
-0.03117748536169529,
0.09701380878686905,
0.0824910700321... | 593 | 23.72 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758506 | [S1] ... the development of the field, like what has been tried before and what has been, let's say, thwarted, or what's the cat and mouse game looked like in the past? I imagine different things like there's Tor, there is, you know, all, all kinds of things. There is probably things that everyone installs on their end, like VPNs and tunnels and so on. Like what's- [S2] Yeah. [S1] ... what's been the general development, uh, over the years? | 25.84 | 3.372838 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758506.mp3 | [
5657,
2601,
8673,
11613,
5128,
6101,
6294,
2978,
3051,
3059,
11245,
11252,
11571,
5339,
10603,
8438,
1633,
3616,
5719,
10675,
8941,
11278,
8578,
8062,
5942,
3570,
4690,
3045,
6431,
4081,
9136,
9201,
9200,
4587,
12675,
12547,
10181,
4558,
12... | [
0.6574094295501709,
-5.619838237762451,
-0.046721894294023514,
-0.06724946200847626,
0.44823530316352844,
-0.028949478641152382,
0.2541424036026001,
-0.04398278146982193,
-0.05613052472472191,
-3.7864444255828857,
0.28510454297065735,
0.021421577781438828,
0.13653747737407684,
0.1618641167... | 646 | 25.84 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758509 | [S1] VPNs in, in many places work, in many places they don't. Uh, there's a country in the news recently, um, be- they were in the news because they rolled out a new law that forced their citizens to swear on the Quran that they would not use a VPN in order to get internet access installed in their homes. Um, so just like a crazy sentence to say out loud. [S2] Yeah. [S1] Um, but in China, for example, these VPNs, they, many of them work most of the time. [S2] Mm-hmm. [S1] Um, but what researchers have noticed is that around the time politically sensitive events are happening, or political, uh, | 29.4 | 2.903606 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758509.mp3 | [
6163,
3162,
7188,
860,
5611,
11714,
11650,
10841,
9200,
11723,
11211,
11252,
11251,
11555,
11347,
10900,
11413,
8532,
5956,
2945,
3521,
8128,
8640,
6600,
9953,
9826,
3697,
9757,
9647,
8606,
10133,
4436,
12388,
12700,
12236,
8707,
8206,
8270,
... | [
0.7333511710166931,
-5.536483287811279,
-0.08010967820882797,
-0.29155993461608887,
0.6568416953086853,
-0.12705720961093903,
0.4028034508228302,
0.009668556042015553,
-0.0028956555761396885,
-3.970777750015259,
0.30770403146743774,
0.04405653104186058,
-0.3717629909515381,
0.3075821101665... | 735 | 29.4 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758510 | [S1] ... such as elections, things like this, uh, a lot of VPNs will just mysteriously stop working. [S2] Yeah. [S1] And then after the event, they'll mysteriously start working again. Um, and it kinda points to this broader idea that some of these countries may be sitting on more censorship capability than they deploy on a daily basis. [S2] Mm-hmm. [S1] Um, and they, they have more power than they use. Um, so this cat and mouse game may even be, like, the cat may even be stronger than, than we think it is. [S2] Yeah. | 21.32 | 2.922897 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758510.mp3 | [
6881,
5776,
786,
4702,
10057,
4050,
11038,
4040,
5302,
2652,
6070,
7542,
11939,
10201,
6098,
11356,
7763,
3291,
3344,
3498,
9942,
12388,
336,
2781,
7151,
12434,
4697,
3942,
11180,
11492,
6576,
1489,
9041,
3920,
11994,
9961,
12518,
12661,
94... | [
0.8078051805496216,
-5.5994181632995605,
-0.18524010479450226,
-0.20392315089702606,
0.6192687153816223,
-0.14180761575698853,
0.30967235565185547,
0.026145175099372864,
0.020320724695920944,
-3.971651077270508,
0.24039681255817413,
-0.033140119165182114,
-0.32691675424575806,
0.1503245234... | 533 | 21.32 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758519 | [S1] ... tap- [S2] How does, how does this, you want to go a bit more into de- I mean, it sounds great at the, at the surface, but there's a reason, right? [S1] [LAUGHS] [S2] We need security researchers probing, making sense, and there's a reason that's the bottleneck. If I were just to be like, "Well, you know, fuzz a bit," um, it's probably not gonna work. So what, what does, what does Geneva do, uh, that allows it to even be successful where maybe humans t- take a long time or wouldn't be successful? | 29.4 | 3.348743 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758519.mp3 | [
6250,
8769,
8672,
5998,
9589,
4339,
9258,
6731,
3868,
7005,
4459,
11709,
7924,
9792,
5539,
10553,
5972,
5576,
2945,
3458,
6177,
9762,
6895,
4979,
12222,
8059,
8798,
10432,
11125,
10997,
2969,
265,
530,
2900,
12207,
8551,
858,
8153,
3392,
... | [
0.711111843585968,
-5.53690767288208,
-0.03477802872657776,
-0.02380063757300377,
0.4271656274795532,
-0.025927217677235603,
0.2906361222267151,
0.004542579874396324,
-0.08195763826370239,
-3.896449089050293,
0.267988920211792,
0.128678560256958,
0.16212309896945953,
0.0792357325553894,
... | 735 | 29.4 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758525 | [S1] Um, so we build out these building blocks and then allow it to compose these things together in trees. [S2] Yeah. So like syntax, like, uh, you, you give it a syntax and it can perform, it can assemble a little program out of this syntax. One, like one we see right here. [S1] That's exactly correct. [S2] Can you walk us through what this particular thing does? | 23.72 | 3.17615 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758525.mp3 | [
6705,
4098,
4847,
4462,
4405,
6972,
6453,
3100,
3685,
9885,
9843,
9817,
6168,
544,
9874,
12754,
12513,
3681,
3456,
2860,
10524,
9477,
5722,
6122,
10537,
7923,
1667,
3502,
11707,
7925,
6233,
3425,
12550,
4292,
1408,
8361,
2533,
8375,
5158,
... | [
0.8759613633155823,
-5.68680477142334,
-0.11844988912343979,
-0.03941929712891579,
0.5521088242530823,
-0.0822705626487732,
0.38348129391670227,
-0.033343687653541565,
-0.052448585629463196,
-3.867582082748413,
0.2560428977012634,
0.015215813182294369,
-0.03413622826337814,
0.1493931412696... | 593 | 23.72 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758530 | [S1] ... through. [S2] I see. [S1] So that's this idea of a strategy. [S2] So that connection in the, in the mind of the sensor is already established as, "Here's a server, here's a client," and it, it kinda keeps that state for subsequent packages. [S1] More or less. [S2] Yeah. [S1] Yeah, that's exactly it. [S2] Mm-hmm. [S1] Yeah. So that's, this is an example of just one strategy in one of these programs that, so Geneva built this program itself and it built this through the process of evolution. | 23.72 | 3.074706 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758530.mp3 | [
12061,
3856,
4064,
2803,
2884,
853,
3237,
6314,
12386,
9314,
11795,
9250,
11404,
12018,
9306,
8723,
6690,
12171,
12450,
4413,
10085,
1312,
2904,
867,
4502,
7492,
7436,
9385,
9770,
9761,
10835,
12570,
12114,
9946,
7332,
4369,
834,
3806,
1073... | [
0.7905217409133911,
-5.661736965179443,
-0.08672787249088287,
-0.14565010368824005,
0.5979823470115662,
-0.10487566888332367,
0.39326342940330505,
-0.006846040952950716,
-0.03968106582760811,
-3.8423333168029785,
0.17770177125930786,
-0.0629095807671547,
-0.11962777376174927,
0.16179451346... | 593 | 23.72 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758532 | [S1] Our best guess as to what, our, our best, uh, implementation, I should say, as to what these sensors looked like based on what previous researchers found. [S2] Mm-hmm. [S1] And then trained Geneva against these mock sensors and also trained it against the great firewall and, and real sensors where we could. [S2] Mm-hmm. [S1] Um, and we found it was very quickly, it was able to reproduce basically the entire field. [S2] Yeah. [S1] Um, every strategy a human had come up with, this, this also found, and it found them each pretty quickly. [S2] Cool. [S1] Um, so it, it's really shown the power of automated approaches and AI/ML. [S2] Yeah. | 28.08 | 2.855724 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758532.mp3 | [
9954,
4211,
10869,
5438,
8118,
8069,
3148,
3114,
6050,
12278,
11259,
10028,
5608,
2888,
3137,
531,
11987,
9686,
11757,
12278,
4599,
4022,
4972,
6129,
8144,
2947,
12526,
11631,
1441,
5960,
8256,
3554,
2662,
2734,
9079,
9135,
7006,
3216,
9369... | [
0.6750162839889526,
-5.53823709487915,
-0.05352839455008507,
-0.3230801522731781,
0.6374803781509399,
-0.13493451476097107,
0.37616509199142456,
0.026806600391864777,
-0.018536195158958435,
-3.957095146179199,
0.3175765573978424,
0.05033869668841362,
-0.2662252187728882,
0.2204470038414001... | 702 | 28.08 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758533 | [S1] Yeah. So you have, you have, uh, let's, let's get back a little bit. You have this syntax, right? [S2] Sure. [S1] That you can build, uh, trees from, which are valid programs in Geneva. This will modify the traffic somehow. Now, to say that most of this traffic will just not even be traffic, probably, like it will, like the connection will be somehow bad. Uh, some of it will, | 22.64 | 3.337311 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758533.mp3 | [
9547,
11229,
3958,
1128,
6308,
3763,
3755,
8802,
9242,
7857,
5153,
9235,
9235,
4124,
4643,
4645,
4132,
6185,
7202,
9235,
9243,
7778,
6705,
6165,
4316,
2977,
472,
8566,
7923,
7917,
7903,
1691,
9892,
6185,
9805,
10052,
8581,
10446,
5902,
49... | [
0.7318312525749207,
-5.744078159332275,
-0.09982609003782272,
-0.029454730451107025,
0.45141446590423584,
-0.0002758621994871646,
0.3266761004924774,
-0.01645512320101261,
-0.07513731718063354,
-3.816187858581543,
0.28840017318725586,
0.0865732952952385,
0.25338566303253174,
0.135698094964... | 566 | 22.64 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758534 | [S1] And, I mean, even if, if, like, sorry to interrupt, but if the fitness- [S2] No, you're good. [S1] Even, like, if, if the fit, I guess the fitness, is it anything else than zero, like, okay, maybe some connections don't even work to, like, the server next to you. You can discard those. But other than that, the fitness is either doesn't reach the target or does reach the target. And if it does, you've kind of won, right? Like, how can you even get a meaningful signal? Is there a fitness in between zero and one? | 28.72 | 3.188392 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758534.mp3 | [
12427,
8814,
6646,
9742,
6539,
6707,
12734,
9590,
7717,
12693,
12226,
7254,
6866,
12013,
11715,
9493,
2985,
11635,
10771,
3917,
12461,
11237,
8540,
5851,
3369,
1962,
9250,
9762,
12654,
11669,
9486,
3504,
1393,
7274,
9756,
726,
8182,
9207,
1... | [
0.6947270631790161,
-5.621864318847656,
0.01076270081102848,
-0.017701568081974983,
0.5861132740974426,
-0.05080970749258995,
0.3075495958328247,
-0.025908933952450752,
-0.06812534481287003,
-3.8390235900878906,
0.2409064769744873,
0.10240013897418976,
0.09395784139633179,
0.07944090664386... | 718 | 28.72 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758535 | [S1] Yeah. So, and, and part of what makes Geneva work is we've kind of shoehorned our way into getting fitness between zero and one. [S2] Yeah. [S1] Um, and specifically what we do is, is, uh, rule out those strategies that break your own connection. Um, so that, that's kind of how we've gotten between zero and one. 'Cause it's not, it's not technically zero or one, it's almost negative one, zero, one. [S2] Yeah. [S1] And negative one is Geneva shooting itself in the foot. [S2] Yeah. | 20.8 | 2.821573 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758535.mp3 | [
7341,
7208,
4163,
12613,
9157,
12261,
12268,
12271,
2420,
4406,
3946,
5593,
5579,
10549,
10461,
1750,
12597,
11246,
9485,
5187,
8268,
8776,
6761,
8731,
12142,
11852,
5635,
2562,
5673,
6514,
2866,
7985,
8490,
7874,
10610,
8283,
2867,
10486,
... | [
0.7205654978752136,
-5.740525245666504,
-0.1206107959151268,
-0.18134434521198273,
0.6952124834060669,
-0.09887237846851349,
0.2791687846183777,
0.015746209770441055,
-0.09947659075260162,
-3.895880699157715,
0.2669699490070343,
0.07661079615354538,
-0.3264288008213043,
0.17246408760547638... | 520 | 20.8 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758538 | [S1] Mm-hmm. Um, and, and we, we see here a, a few ways you can mutate. I think this, this just essentially comes down to changing the syntax to syntax tree in some form. Um- [S2] Yep. [S1] And these are- [S2] Basically, you can, yeah, you can imagine all the different ways you could, you could take these programs and mix them around. If you can think of that, Geneva can probably do it. | 21.28 | 3.492543 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758538.mp3 | [
6802,
8407,
8915,
10974,
4692,
6689,
12321,
8729,
8378,
9753,
5713,
12560,
6745,
5664,
6811,
6680,
11315,
6737,
7268,
12403,
4770,
6225,
12617,
12122,
3681,
6672,
6712,
4640,
8282,
9382,
9151,
11190,
10677,
11054,
7708,
8751,
7257,
9250,
92... | [
0.9770177006721497,
-5.625787258148193,
-0.07886196672916412,
-0.03483102470636368,
0.6568044424057007,
-0.17003104090690613,
0.2698584794998169,
-0.02810562402009964,
-0.012809754349291325,
-3.8682806491851807,
0.21198374032974243,
0.043777868151664734,
-0.04536569118499756,
0.03410198166... | 532 | 21.28 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758542 | [S1] Uh, and you could imagine a sensor that's a little more sophisticated than that. So something we've kept an eye on is, is at the end of the future if either the sensor starts rolling out AI/ML techniques, or if the sensor, um, starts hunting for traffic that looks very abnormal. [S2] Mm-hmm. [S1] And you could imagine encoding additional, uh, bits into the fitness function such that you could encourage Geneva to make this strategy blend in with normal traffic. [S2] Yeah. [S1] I want this to look as normal as possible but still get through, things like this. So you could imagine all sorts of modifications to the fitness function to make, um, an algorithm | 29.8 | 2.963069 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758542.mp3 | [
9893,
3172,
7333,
6812,
12379,
5289,
7245,
7415,
1975,
4531,
1901,
6660,
9931,
8518,
6292,
9194,
9567,
5524,
2143,
10982,
11261,
11248,
10672,
11938,
9025,
11712,
11686,
8262,
6071,
4012,
8088,
3044,
7095,
11295,
3923,
9680,
6101,
1295,
284... | [
0.48183584213256836,
-5.5983171463012695,
-0.035720083862543106,
-0.26539310812950134,
0.746940016746521,
-0.09855975955724716,
0.394502192735672,
0.014409398660063744,
-0.03418051078915596,
-3.9423487186431885,
0.222418412566185,
-0.04819256439805031,
-0.2895307242870331,
0.30703777074813... | 745 | 29.8 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758543 | [S1] ... it, but, you know, in malware, it's like just bits, and I flip, like, you know, very small number of bits. There's nothing like how the human sees it and how the machine sees it. It's- [S2] Totally. [S1] It's so weird. Um, but, yeah, I th- I think, I think it's, it's pretty cool, and you got some attention in the media, uh, and the, the articles usually go something like, uh, | 23.96 | 3.340144 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758543.mp3 | [
12692,
12038,
1410,
2965,
11109,
11749,
9869,
8456,
9380,
12725,
12595,
12086,
6948,
8808,
3061,
11126,
2256,
7250,
11726,
9422,
10678,
7974,
7334,
12637,
12549,
7749,
5734,
5678,
11109,
7678,
6973,
11130,
10411,
7955,
8990,
2740,
7599,
5116,... | [
0.7119515538215637,
-5.697813510894775,
-0.0762149915099144,
0.0034705311991274357,
0.5342110991477966,
-0.006938542705029249,
0.30594536662101746,
-0.042356714606285095,
-0.007990997284650803,
-3.8156423568725586,
0.3003932535648346,
0.09498480707406998,
0.09240449965000153,
0.06459775567... | 599 | 23.96 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758545 | [S1] Cool. And the, just, just saying the code for Geneva is available. It's on GitHub. Um, you know, anyone can, anyone can, I guess, look it up. Your builds fail right now. I'll, I'll just have to tell you. I'm sorry. Um, but, uh- [S2] Yeah, we're, we're switching between CI systems and having to finish the migration. [S1] [LAUGHS] | 17.96 | 3.287301 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758545.mp3 | [
6283,
9368,
9314,
5681,
8845,
7005,
10037,
8308,
12470,
9846,
7332,
5808,
11292,
4466,
2595,
10348,
5717,
4306,
11446,
12277,
11933,
8205,
2624,
3997,
12035,
11588,
11140,
9158,
7405,
11306,
7278,
4757,
3368,
3840,
9612,
8053,
9363,
713,
96... | [
0.6347101330757141,
-5.740915298461914,
0.005196972284466028,
-0.10526430606842041,
0.6718980073928833,
-0.09283576160669327,
0.4526509642601013,
-0.07359756529331207,
0.04879940301179886,
-3.8377649784088135,
0.0966300293803215,
-0.05504268407821655,
-0.056934915482997894,
0.1109120026230... | 449 | 17.96 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758546 | [S1] There's definitely possibilities. If you think of something like, you know, AlphaGo or something like this, that's a, it's a discrete game, but also- [S2] Right. [S1] ... you know, they, they work with neural networks that, for example, uh, when you build your tree, your modifications that guide that somehow, that, you know, ha, have an idea which of the modifications might lead to a better algorithm, to a worse algorithm, and so on. Do you see any sort of, uh, involvement that could happen there? | 29.12 | 3.45556 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758546.mp3 | [
9745,
6833,
7317,
8811,
4308,
4123,
3202,
12790,
10730,
5524,
6467,
1168,
5674,
5760,
5604,
10737,
6506,
7080,
8347,
8805,
2089,
3286,
11098,
12741,
6169,
10857,
8297,
5162,
9203,
6524,
10485,
5430,
4462,
8170,
6096,
1419,
10206,
1043,
5737... | [
0.7281749844551086,
-5.505130290985107,
0.051631636917591095,
-0.05475796014070511,
0.4488210082054138,
-0.019960204139351845,
0.30444517731666565,
0.013694074004888535,
-0.04496080428361893,
-3.8220105171203613,
0.33936572074890137,
0.1063884049654007,
0.11108703911304474,
0.0909424275159... | 728 | 29.12 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758549 | [S1] Um, so I think building towards this idea of, can we do federated training across a, a large, a large population? Can we do this in a way that protects users? Can we make the algorithm more efficient so it needs, it needs less connections to figure things out? [S2] Mm-hmm. [S1] Um, all sorts of things like this I think are, um, really good goals to shoot for and as more people, viewers, try this out. As more people, like, jump into this space and play with this, um, these are some of the problems they're gonna be building towards. | 25.56 | 2.813063 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758549.mp3 | [
4128,
6823,
6957,
4462,
4468,
4396,
3110,
6444,
6544,
2790,
10678,
12182,
6296,
5619,
11681,
11348,
6690,
12635,
3995,
6748,
6757,
5854,
3293,
3092,
2594,
7713,
6114,
11170,
10346,
5762,
5392,
12067,
11302,
9035,
6626,
2748,
7869,
10700,
11... | [
0.9406667351722717,
-5.5593695640563965,
-0.12396009266376495,
-0.23870307207107544,
0.6194337606430054,
-0.1572885811328888,
0.4098471999168396,
0.007948040962219238,
-0.011823865585029125,
-3.9363770484924316,
0.3295212686061859,
0.04448183998465538,
-0.2580816447734833,
0.24125452339649... | 639 | 25.56 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758552 | [S1] But when you apply that fix, you open yourself up to the entire space of attacks of maybe I can sneak a letter in there that you think belongs halfway through the message, but it actually belongs in the beginning, or it actually belongs to the end, or it actually doesn't belong in that, uh, at all. Um, and so you have, this is one example that we've seen in the wild where, um, this idea of, uh, I have, I need to fix a limitation, and by fixing the limitation, I've opened myself up to- [S2] Mm-hmm. [S1] ... a dozen other potential attacks. So that definitely exists. | 26.6 | 2.894797 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758552.mp3 | [
4722,
8225,
1012,
10539,
605,
8622,
8772,
12684,
10509,
3750,
8369,
6568,
2722,
7922,
11568,
11184,
12258,
8010,
10740,
4591,
3986,
6576,
5936,
9197,
11759,
2341,
6251,
12692,
4584,
4480,
9474,
8518,
718,
9974,
10404,
5773,
8304,
3436,
8708... | [
0.7986462116241455,
-5.5255208015441895,
-0.020825635641813278,
-0.27604082226753235,
0.6494817733764648,
-0.21859495341777802,
0.43953245878219604,
0.05891433358192444,
0.08545003831386566,
-3.877166986465454,
0.3532451093196869,
-0.013432888314127922,
-0.31805384159088135,
0.269002944231... | 665 | 26.6 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758555 | [S1] I would love to. Yeah, so an area of work that we went, that we started exploring a year or two ago, uh, something we noticed for a lot of these sensors is, um, when you interact with them as a user, like they need to respond to you, they need to send you some traffic, right? Like if I'm, if I'm trying to request some resource and that resource is forbidden, maybe the sensor sends me a block page. [S2] Mm-hmm. [S1] And that block page says, "Hey, you're not allowed to access this." And the thing is that, that communication there, uh, what's going on is my request can often be much smaller than the size of the block page I get back. [S2] Mm-hmm. | 29.88 | 2.832215 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758555.mp3 | [
6298,
6940,
6804,
6804,
6821,
3966,
9085,
12781,
85,
117,
4375,
6216,
659,
2586,
2910,
11133,
10973,
937,
4905,
8448,
6040,
4573,
2005,
733,
711,
6805,
7269,
4692,
5745,
8834,
9996,
12750,
12189,
12717,
7671,
4534,
5108,
9192,
6592,
3556,... | [
0.6973893642425537,
-5.524351119995117,
-0.01629173383116722,
-0.27584367990493774,
0.7320960164070129,
-0.12526771426200867,
0.40310004353523254,
-0.022441525012254715,
0.052594613283872604,
-3.9943313598632812,
0.28240665793418884,
0.07315783947706223,
-0.35709676146507263,
0.23906138539... | 747 | 29.88 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758556 | [S1] So as an attacker, this opens up the space of, "Hey, maybe I can use the sensor to watch an attack at somebody else by making a request for forbidden things, pretending to be someone else, and then letting them send that huge response at that other person." [S2] Mm-hmm. [S1] Um, and this is a, this is an idea of a reflected attack or an amplification attack. [S2] Mm-hmm. [S1] Because as an attacker, I can make a tiny request and then get a bigger request out of it, so I'm amplifying my traffic. [S2] Yeah. [S1] So amplification attack. Um, so we started exploring whether we could do this | 29.92 | 3.062167 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758556.mp3 | [
12570,
9331,
5740,
9326,
9317,
6257,
4369,
8657,
8117,
10606,
5486,
2846,
2694,
1814,
9883,
9893,
9765,
12653,
11740,
12197,
11622,
8470,
2949,
3008,
8550,
9247,
5974,
10030,
8336,
2960,
10233,
9211,
11696,
11640,
9962,
12058,
7148,
8086,
5... | [
0.6796230673789978,
-5.511667251586914,
-0.06476554274559021,
-0.23075522482395172,
0.801497220993042,
-0.14362724125385284,
0.38830989599227905,
0.026345428079366684,
-0.004523833282291889,
-3.995664119720459,
0.2373868077993393,
0.09290416538715363,
-0.28452110290527344,
0.18290734291076... | 748 | 29.92 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758557 | [S1] And the, so the, the, who does it hurt more, the sensors or the final recipients of the, the attack? [S2] Yeah, so in this case, the, the weight is beared by both, but the brunt of the impact will be felt by the victim. [S1] Yeah. [S2] Um, so this line of work, it, it mucks with the sensor, but really, really the, some of the, I wanna say the, the purpose or, uh, | 24.32 | 3.214052 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758557.mp3 | [
4145,
5160,
11942,
8694,
12719,
11341,
4097,
459,
7268,
4630,
2624,
6053,
12035,
12612,
11717,
7118,
6185,
8738,
11809,
6193,
9244,
4965,
11484,
9980,
9973,
6768,
7010,
5536,
3025,
11127,
5295,
10398,
3310,
1187,
5730,
11891,
8803,
9307,
52... | [
0.7682501077651978,
-5.720666885375977,
-0.07472839951515198,
-0.13084644079208374,
0.7800875902175903,
-0.06635569036006927,
0.4201221168041229,
-0.06889677792787552,
-0.04876023530960083,
-3.9096412658691406,
0.2570120692253113,
0.06686564534902573,
-0.15956848859786987,
0.09672962129116... | 608 | 24.32 |
bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758567 | [S1] ... of if I never publish this, there's often no incentive for, for this issue to be patched. [S2] Yeah. [S1] Like if, if there's no, if there's no downside to the network. | 8.28 | 2.971973 | 24,000 | audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p170_BV17h4y1j7aZ_p170_m4-dialogue_0758567.mp3 | [
4714,
8209,
6821,
11255,
9071,
3951,
3887,
3366,
3157,
2716,
2977,
5480,
6834,
8216,
12069,
12693,
5864,
2933,
11069,
12684,
6148,
11405,
12782,
12070,
2730,
10653,
2948,
5786,
4530,
7036,
8054,
6182,
8489,
5286,
5981,
12754,
4480,
705,
592... | [
0.8796035647392273,
-5.881635665893555,
-0.06839099526405334,
-0.06751997023820877,
0.45830124616622925,
-0.20719420909881592,
0.48644426465034485,
-0.042527198791503906,
0.010159293189644814,
-3.868030548095703,
0.20409268140792847,
0.06932903081178665,
-0.3635208010673523,
0.178723141551... | 207 | 8.28 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.