WEBVTT

1
00:00:02.655 --> 00:00:03.175
<00:00:02.655>Hello <00:00:03.123>everyone.

2
00:00:03.395 --> 00:00:05.615
<00:00:03.395>And <00:00:03.595>as <00:00:03.795>I <00:00:03.995>mentioned <00:00:04.195>my <00:00:04.395>topic, <00:00:04.595>the <00:00:04.795>topic

3
00:00:05.635 --> 00:00:08.215
<00:00:05.635>of <00:00:05.835>my <00:00:06.035>today's <00:00:06.234>talk <00:00:06.435>is <00:00:06.635>multimodal <00:00:06.835>generative <00:00:07.035>ai.

4
00:00:08.875 --> 00:00:11.095
<00:00:08.875>And, <00:00:09.108>uh, <00:00:09.341>a <00:00:09.575>few <00:00:09.808>words <00:00:10.041>about <00:00:10.275>myself.

5
00:00:11.275 --> 00:00:14.135
<00:00:11.275>Um, <00:00:11.475>I'm <00:00:11.675>a <00:00:11.875>senior <00:00:12.075>data <00:00:12.275>scientist <00:00:12.475>at <00:00:12.675>Nvidia.

6
00:00:14.335 --> 00:00:15.455
<00:00:14.335>I <00:00:14.671>help <00:00:15.007>our <00:00:15.343>partners

7
00:00:15.455 --> 00:00:19.615
<00:00:15.455>and <00:00:15.630>customers <00:00:15.805>in <00:00:15.980>utilizing, <00:00:16.155>uh, <00:00:16.329>deep <00:00:16.505>learning, <00:00:16.680>uh, <00:00:16.855>software,

8
00:00:20.155 --> 00:00:23.375
<00:00:20.155>uh, <00:00:20.295>by <00:00:20.435>Nvidia, <00:00:20.575>but <00:00:20.715>also <00:00:20.855>hardware <00:00:20.995>in <00:00:21.135>the <00:00:21.275>best <00:00:21.415>possible <00:00:21.555>way.

9
00:00:24.175 --> 00:00:26.015
<00:00:24.175>Previously, <00:00:24.375>I <00:00:24.575>was <00:00:24.775>working <00:00:24.975>in <00:00:25.175>a <00:00:25.375>Japanese <00:00:25.575>company,

10
00:00:26.155 --> 00:00:28.055
<00:00:26.155>Conica <00:00:26.435>Minolta, <00:00:26.715>as <00:00:26.995>a <00:00:27.275>research, <00:00:27.555>uh,

11
00:00:28.335 --> 00:00:30.895
<00:00:28.335>engineer <00:00:28.568>in <00:00:28.801>deep <00:00:29.035>learning, <00:00:29.268>uh, <00:00:29.501>developing <00:00:29.735>algorithms

12
00:00:30.895 --> 00:00:32.455
<00:00:30.895>for <00:00:31.361>medical <00:00:31.828>image <00:00:32.295>processing.

13
00:00:33.035 --> 00:00:36.175
<00:00:33.035>And <00:00:33.209>I <00:00:33.385>got <00:00:33.559>my <00:00:33.735>PhD <00:00:33.910>from <00:00:34.084>RWTH <00:00:34.260>Rhan <00:00:34.434>University.

14
00:00:36.395 --> 00:00:39.695
<00:00:36.395>Um, <00:00:36.550>so <00:00:36.706>I <00:00:36.861>hope <00:00:37.017>there <00:00:37.172>are <00:00:37.328>some <00:00:37.483>data <00:00:37.639>scientists, <00:00:37.795>uh,

15
00:00:39.755 --> 00:00:41.095
<00:00:39.755>who <00:00:40.056>also <00:00:40.358>watching <00:00:40.659>the <00:00:40.961>stream.

16
00:00:41.675 --> 00:00:45.775
<00:00:41.675>And, <00:00:41.815>uh, <00:00:41.955>I <00:00:42.095>think <00:00:42.235>my <00:00:42.375>next <00:00:42.515>slide <00:00:42.654>demonstrates <00:00:42.794>how <00:00:42.934>it <00:00:43.074>feels

17
00:00:45.775 --> 00:00:47.295
<00:00:45.775>to <00:00:46.003>be <00:00:46.231>a <00:00:46.459>data <00:00:46.687>scientist <00:00:46.915>these <00:00:47.143>days.

18
00:00:47.835 --> 00:00:51.375
<00:00:47.835>Uh, <00:00:47.975>while <00:00:48.115>some <00:00:48.255>of <00:00:48.395>us <00:00:48.535>are <00:00:48.675>still <00:00:48.815>tuning <00:00:48.955>parameters <00:00:49.095>of <00:00:49.235>CNNs

19
00:00:51.435 --> 00:00:53.415
<00:00:51.435>and, <00:00:51.785>uh, <00:00:52.135>well <00:00:52.485>dating, <00:00:52.835>pitting

20
00:00:53.435 --> 00:00:56.695
<00:00:53.435>and, <00:00:53.635>uh, <00:00:53.835>stripes, <00:00:54.035>uh, <00:00:54.235>the <00:00:54.435>huge, <00:00:54.635>uh, <00:00:54.835>wave

21
00:00:56.915 --> 00:00:59.695
<00:00:56.915>or <00:00:57.195>rather <00:00:57.475>say <00:00:57.754>tornado <00:00:58.035>of, <00:00:58.315>uh,

22
00:01:00.005 --> 00:01:02.375
<00:01:00.005>another <00:01:00.355>modern <00:01:00.705>AI <00:01:01.055>techniques, <00:01:01.405>uh,

23
00:01:02.375 --> 00:01:04.335
<00:01:02.375>sweeps <00:01:02.655>out <00:01:02.935>everything <00:01:03.215>on <00:01:03.495>its <00:01:03.775>way

24
00:01:04.595 --> 00:01:06.415
<00:01:04.595>and <00:01:04.875>even <00:01:05.155>make <00:01:05.435>some <00:01:05.715>jobs <00:01:05.995>obsolete.

25
00:01:06.995 --> 00:01:09.375
<00:01:06.995>And <00:01:07.170>I <00:01:07.345>took <00:01:07.520>the <00:01:07.695>liberty <00:01:07.870>of <00:01:08.045>updating <00:01:08.220>this <00:01:08.395>slide

26
00:01:09.475 --> 00:01:11.815
<00:01:09.475>and <00:01:09.675>to <00:01:09.875>adding <00:01:10.074>the <00:01:10.274>word, <00:01:10.475>uh, <00:01:10.675>generative <00:01:10.875>ai.

27
00:01:12.115 --> 00:01:15.895
<00:01:12.115>So, <00:01:12.270>uh, <00:01:12.426>but <00:01:12.581>before <00:01:12.737>we <00:01:12.892>start <00:01:13.048>learning <00:01:13.203>about <00:01:13.359>generative <00:01:13.515>ai

28
00:01:15.895 --> 00:01:18.135
<00:01:15.895>and <00:01:16.175>in <00:01:16.455>particular, <00:01:16.735>multi-model, <00:01:17.015>generative <00:01:17.295>ai,

29
00:01:18.465 --> 00:01:20.655
<00:01:18.465>let's <00:01:18.698>take <00:01:18.931>a <00:01:19.165>look <00:01:19.398>into <00:01:19.631>the <00:01:19.865>retrospective.

30
00:01:21.435 --> 00:01:24.855
<00:01:21.435>So <00:01:21.635>before <00:01:21.835>we <00:01:22.035>had, <00:01:22.235>uh, <00:01:22.435>like <00:01:22.635>huge <00:01:22.835>variety

31
00:01:24.915 --> 00:01:26.535
<00:01:24.914>of <00:01:25.615>specialized <00:01:26.315>architectures.

32
00:01:26.915 --> 00:01:29.335
<00:01:26.914>We <00:01:27.148>had <00:01:27.381>convolutional <00:01:27.615>neural <00:01:27.848>networks, <00:01:28.081>we <00:01:28.315>have

33
00:01:29.435 --> 00:01:30.815
<00:01:29.435>for <00:01:30.056>image <00:01:30.676>processing.

34
00:01:31.155 --> 00:01:33.495
<00:01:31.155>Uh, <00:01:31.388>we <00:01:31.621>had <00:01:31.855>gans <00:01:32.088>for <00:01:32.321>image <00:01:32.555>generation.

35
00:01:33.555 --> 00:01:37.975
<00:01:33.555>We <00:01:33.755>had <00:01:33.955>recurrent <00:01:34.155>neural <00:01:34.355>networks <00:01:34.555>for <00:01:34.755>text <00:01:34.955>processing

36
00:01:37.975 --> 00:01:39.015
<00:01:37.975>and <00:01:38.443>speech <00:01:38.911>processing

37
00:01:39.015 --> 00:01:40.895
<00:01:39.015>and <00:01:39.365>reinforcement <00:01:39.715>learning <00:01:40.065>for <00:01:40.415>robotics.

38
00:01:41.555 --> 00:01:44.495
<00:01:41.555>So <00:01:41.710>typically <00:01:41.866>how <00:01:42.021>you <00:01:42.177>would <00:01:42.332>approach <00:01:42.488>this <00:01:42.643>task <00:01:42.799>is, <00:01:42.955>like,

39
00:01:44.515 --> 00:01:46.455
<00:01:44.515>for <00:01:44.795>example, <00:01:45.075>with <00:01:45.355>convolution <00:01:45.635>neural <00:01:45.915>networks,

40
00:01:46.475 --> 00:01:50.455
<00:01:46.475>you <00:01:46.602>would <00:01:46.729>collect <00:01:46.856>a <00:01:46.984>data <00:01:47.111>set <00:01:47.238>of <00:01:47.365>objects <00:01:47.493>you <00:01:47.620>want <00:01:47.747>to <00:01:47.875>detect

41
00:01:50.515 --> 00:01:52.895
<00:01:50.515>or <00:01:50.748>segment <00:01:50.981>with <00:01:51.215>this <00:01:51.448>convolution <00:01:51.681>neural <00:01:51.915>network,

42
00:01:53.235 --> 00:01:55.295
<00:01:53.235>and <00:01:53.468>you <00:01:53.701>would <00:01:53.935>label <00:01:54.168>this <00:01:54.401>data <00:01:54.635>set.

43
00:01:55.355 --> 00:01:57.735
<00:01:55.354>So <00:01:55.510>if <00:01:55.666>you <00:01:55.821>have <00:01:55.977>a <00:01:56.132>data <00:01:56.288>set <00:01:56.443>of <00:01:56.599>cats, <00:01:56.755>uh,

44
00:01:57.825 --> 00:02:00.575
<00:01:57.825>while <00:01:58.058>you'll <00:01:58.291>collect <00:01:58.525>photographs <00:01:58.758>of <00:01:58.991>different <00:01:59.225>cats,

45
00:02:00.675 --> 00:02:02.935
<00:02:00.675>you'll <00:02:00.908>draw <00:02:01.141>bond <00:02:01.375>in <00:02:01.608>boxes <00:02:01.841>around <00:02:02.075>cats,

46
00:02:03.075 --> 00:02:04.775
<00:02:03.075>and <00:02:03.355>then <00:02:03.635>you'll <00:02:03.915>train <00:02:04.195>your <00:02:04.475>volution

47
00:02:04.835 --> 00:02:06.735
<00:02:04.835>and <00:02:05.115>neural <00:02:05.395>network <00:02:05.675>to <00:02:05.955>detect <00:02:06.235>cats.

48
00:02:07.515 --> 00:02:11.335
<00:02:07.515>Um, <00:02:07.655>in <00:02:07.795>the <00:02:07.935>end, <00:02:08.074>you'll <00:02:08.215>get, <00:02:08.354>uh, <00:02:08.495>fairly, <00:02:08.634>fairly <00:02:08.775>good, <00:02:08.914>uh,

49
00:02:11.525 --> 00:02:15.055
<00:02:11.525>network, <00:02:11.700>which <00:02:11.875>is <00:02:12.050>able <00:02:12.225>to <00:02:12.400>detect <00:02:12.575>cats <00:02:12.750>with <00:02:12.925>like,

50
00:02:15.055 --> 00:02:16.655
<00:02:15.055>hopefully <00:02:15.755>high <00:02:16.455>confidence.

51
00:02:16.915 --> 00:02:19.735
<00:02:16.915>And, <00:02:17.090>uh, <00:02:17.265>it's, <00:02:17.440>it's <00:02:17.614>really <00:02:17.790>good <00:02:17.965>at <00:02:18.140>this <00:02:18.315>task,

52
00:02:19.875 --> 00:02:21.175
<00:02:19.875>but <00:02:20.167>only <00:02:20.460>at <00:02:20.752>this <00:02:21.045>task.

53
00:02:21.315 --> 00:02:24.095
<00:02:21.315>So <00:02:21.470>if <00:02:21.626>you <00:02:21.781>have <00:02:21.937>trained <00:02:22.092>this <00:02:22.248>network <00:02:22.403>to <00:02:22.559>detect <00:02:22.715>cats,

54
00:02:24.355 --> 00:02:27.135
<00:02:24.355>it <00:02:24.554>won't <00:02:24.755>automatically <00:02:24.954>be <00:02:25.155>able <00:02:25.355>to <00:02:25.554>detect <00:02:25.755>dogs.

55
00:02:27.795 --> 00:02:32.015
<00:02:27.795>And <00:02:27.935>you <00:02:28.075>can <00:02:28.215>imagine <00:02:28.355>this <00:02:28.495>approach <00:02:28.635>is <00:02:28.775>kind <00:02:28.915>of <00:02:29.055>not <00:02:29.195>sustainable

56
00:02:32.015 --> 00:02:35.055
<00:02:32.015>because, <00:02:32.190>uh, <00:02:32.364>if <00:02:32.540>you <00:02:32.714>want <00:02:32.890>to <00:02:33.065>develop <00:02:33.239>something <00:02:33.415>like

57
00:02:35.615 --> 00:02:37.735
<00:02:35.615>artificial <00:02:35.895>general <00:02:36.175>intelligence <00:02:36.455>where <00:02:36.735>you <00:02:37.015>want

58
00:02:37.735 --> 00:02:42.495
<00:02:37.735>to <00:02:37.875>identify <00:02:38.015>every <00:02:38.155>object <00:02:38.295>in <00:02:38.435>the <00:02:38.575>world, <00:02:38.715>um, <00:02:38.855>it's <00:02:38.995>hard <00:02:39.135>to,

59
00:02:42.555 --> 00:02:45.335
<00:02:42.555>to <00:02:42.695>create <00:02:42.835>such <00:02:42.975>a <00:02:43.115>data <00:02:43.255>set <00:02:43.395>to <00:02:43.535>train <00:02:43.675>such <00:02:43.815>a <00:02:43.955>network.

60
00:02:45.595 --> 00:02:47.815
<00:02:45.595>Uh, <00:02:45.828>you <00:02:46.061>understand <00:02:46.295>that. <00:02:46.528>And <00:02:46.761>I <00:02:46.995>think

61
00:02:47.815 --> 00:02:51.095
<00:02:47.815>that <00:02:48.048>the <00:02:48.281>sustainable <00:02:48.515>solution <00:02:48.748>has <00:02:48.981>been <00:02:49.215>found

62
00:02:51.395 --> 00:02:54.255
<00:02:51.395>and <00:02:51.628>it's <00:02:51.861>a <00:02:52.095>multimodal <00:02:52.328>large <00:02:52.561>language <00:02:52.795>models.

63
00:02:54.595 --> 00:02:57.015
<00:02:54.595>Um, <00:02:54.795>I <00:02:54.995>think <00:02:55.195>it's <00:02:55.395>the <00:02:55.595>future <00:02:55.795>of <00:02:55.995>AI

64
00:02:57.075 --> 00:02:58.215
<00:02:57.075>and <00:02:57.417>a <00:02:57.759>door <00:02:58.101>opener

65
00:02:58.235 --> 00:03:00.725
<00:02:58.235>to <00:02:58.515>something <00:02:58.795>like <00:02:59.075>artificial <00:02:59.355>general <00:02:59.635>intelligence.

66
00:03:01.225 --> 00:03:04.005
<00:03:01.225>And <00:03:01.424>the <00:03:01.625>idea <00:03:01.825>beneath <00:03:02.025>is <00:03:02.225>that, <00:03:02.424>uh, <00:03:02.625>well,

67
00:03:04.005 --> 00:03:05.485
<00:03:04.005>the <00:03:04.338>solution <00:03:04.671>is <00:03:05.004>based <00:03:05.337>on,

68
00:03:05.625 --> 00:03:07.765
<00:03:05.625>on <00:03:05.905>something <00:03:06.185>like <00:03:06.465>a <00:03:06.745>foundation <00:03:07.025>model

69
00:03:07.905 --> 00:03:09.805
<00:03:07.905>or <00:03:08.605>transformer <00:03:09.305>architecture.

70
00:03:10.585 --> 00:03:14.365
<00:03:10.585>And, <00:03:10.725>uh, <00:03:10.865>you <00:03:11.005>train <00:03:11.145>it <00:03:11.285>on <00:03:11.425>large <00:03:11.565>amounts <00:03:11.705>of <00:03:11.845>data, <00:03:11.985>like

71
00:03:14.745 --> 00:03:16.285
<00:03:14.745>to <00:03:14.976>the <00:03:15.207>scale <00:03:15.438>of <00:03:15.669>the <00:03:15.900>whole <00:03:16.131>internet.

72
00:03:16.705 --> 00:03:19.885
<00:03:16.704>And <00:03:16.860>once <00:03:17.016>you've <00:03:17.171>trained <00:03:17.327>it, <00:03:17.482>this <00:03:17.638>network <00:03:17.793>should <00:03:17.949>be <00:03:18.105>able

73
00:03:19.905 --> 00:03:23.965
<00:03:19.905>to <00:03:20.060>recognize <00:03:20.216>many <00:03:20.371>patterns <00:03:20.527>it <00:03:20.682>learned <00:03:20.838>from <00:03:20.993>the <00:03:21.149>whole <00:03:21.305>world

74
00:03:23.965 --> 00:03:26.245
<00:03:23.965>of <00:03:24.245>internet, <00:03:24.525>and <00:03:24.805>you <00:03:25.085>can <00:03:25.365>further

75
00:03:26.765 --> 00:03:28.365
<00:03:26.765>reuse <00:03:27.045>this <00:03:27.325>network <00:03:27.605>for <00:03:27.885>different <00:03:28.165>tasks.

76
00:03:28.505 --> 00:03:30.405
<00:03:28.505>So <00:03:28.680>let's <00:03:28.855>step <00:03:29.030>in <00:03:29.204>and <00:03:29.380>see <00:03:29.555>how <00:03:29.730>it <00:03:29.905>works.

77
00:03:31.185 --> 00:03:34.125
<00:03:31.185>Uh, <00:03:31.340>but <00:03:31.496>also, <00:03:31.651>I, <00:03:31.807>I'd <00:03:31.962>like <00:03:32.118>to <00:03:32.273>demonstrate <00:03:32.429>you <00:03:32.585>already

78
00:03:34.125 --> 00:03:37.285
<00:03:34.125>that <00:03:34.358>this <00:03:34.591>technology <00:03:34.825>already, <00:03:35.058>uh, <00:03:35.291>is <00:03:35.525>usable.

79
00:03:37.345 --> 00:03:38.725
<00:03:37.345>It <00:03:37.759>already <00:03:38.173>exists <00:03:38.587>today.

80
00:03:39.025 --> 00:03:43.005
<00:03:39.025>And, <00:03:39.180>uh, <00:03:39.336>with <00:03:39.491>this <00:03:39.647>report <00:03:39.802>by <00:03:39.958>Microsoft, <00:03:40.113>uh, <00:03:40.269>uh, <00:03:40.425>sparks

81
00:03:43.005 --> 00:03:46.605
<00:03:43.005>of <00:03:43.180>artificial <00:03:43.355>general <00:03:43.530>intelligence, <00:03:43.704>uh, <00:03:43.880>they, <00:03:44.055>they <00:03:44.230>were <00:03:44.405>able

82
00:03:46.625 --> 00:03:50.125
<00:03:46.625>to <00:03:46.765>demonstrate <00:03:46.905>that <00:03:47.045>their <00:03:47.185>GPT <00:03:47.325>four, <00:03:47.465>the, <00:03:47.605>the <00:03:47.745>G <00:03:47.885>PT <00:03:48.025>four

83
00:03:50.125 --> 00:03:52.205
<00:03:50.125>by <00:03:50.475>OpenAI <00:03:50.825>and <00:03:51.175>GPT <00:03:51.525>stands

84
00:03:52.225 --> 00:03:54.645
<00:03:52.225>for <00:03:52.691>generative <00:03:53.158>pre-trained <00:03:53.625>transformer.

85
00:03:54.665 --> 00:03:56.285
<00:03:54.665>So <00:03:55.015>it's <00:03:55.364>a <00:03:55.715>transformer <00:03:56.065>architecture

86
00:03:56.785 --> 00:03:59.285
<00:03:56.785>and <00:03:57.065>it's <00:03:57.345>a <00:03:57.625>multi-model, <00:03:57.905>transformer <00:03:58.185>architecture.

87
00:03:59.465 --> 00:04:02.325
<00:03:59.465>And, <00:03:59.640>uh, <00:03:59.815>with <00:03:59.990>this <00:04:00.165>gvt <00:04:00.340>four, <00:04:00.515>they <00:04:00.690>are <00:04:00.865>able

88
00:04:02.345 --> 00:04:06.765
<00:04:02.345>to <00:04:02.520>achieve <00:04:02.695>82% <00:04:02.870>accuracy <00:04:03.045>on <00:04:03.220>zero <00:04:03.395>shot, <00:04:03.570>uh, <00:04:03.745>tasks.

89
00:04:07.505 --> 00:04:08.925
<00:04:07.505>Uh, <00:04:07.760>what <00:04:08.016>is <00:04:08.271>zero <00:04:08.527>shot <00:04:08.783>tasks?

90
00:04:09.225 --> 00:04:11.485
<00:04:09.225>Uh, <00:04:09.505>zero <00:04:09.785>shot <00:04:10.065>means <00:04:10.345>that, <00:04:10.625>uh,

91
00:04:11.485 --> 00:04:14.485
<00:04:11.485>once <00:04:11.625>you <00:04:11.765>have <00:04:11.905>trained <00:04:12.045>your <00:04:12.185>model <00:04:12.325>on <00:04:12.465>your <00:04:12.605>whole <00:04:12.745>data <00:04:12.885>set,

92
00:04:15.105 --> 00:04:18.485
<00:04:15.105>you <00:04:15.280>don't <00:04:15.454>need <00:04:15.630>to <00:04:15.804>retrain <00:04:15.980>it <00:04:16.154>for <00:04:16.329>specific <00:04:16.505>objects.

93
00:04:18.485 --> 00:04:21.045
<00:04:18.485>So <00:04:18.625>you <00:04:18.765>don't <00:04:18.905>need <00:04:19.045>to <00:04:19.185>train <00:04:19.325>it <00:04:19.465>specifically <00:04:19.605>to <00:04:19.745>be <00:04:19.885>able

94
00:04:21.045 --> 00:04:22.805
<00:04:21.045>to <00:04:21.395>identify <00:04:21.745>cats <00:04:22.095>or <00:04:22.445>dogs.

95
00:04:23.305 --> 00:04:24.925
<00:04:23.305>It <00:04:23.655>already <00:04:24.005>has <00:04:24.355>this <00:04:24.705>knowledge

96
00:04:24.925 --> 00:04:27.285
<00:04:24.925>because <00:04:25.100>it <00:04:25.275>has <00:04:25.450>learned <00:04:25.625>it <00:04:25.800>from <00:04:25.975>the <00:04:26.150>whole <00:04:26.325>internet.

97
00:04:27.865 --> 00:04:29.085
<00:04:27.865>Uh, <00:04:28.139>this <00:04:28.414>means <00:04:28.688>zero <00:04:28.962>shot

98
00:04:29.145 --> 00:04:32.485
<00:04:29.145>and <00:04:29.378>imagine, <00:04:29.611>uh, <00:04:29.844>82% <00:04:30.078>on <00:04:30.311>a <00:04:30.544>network,

99
00:04:32.655 --> 00:04:34.885
<00:04:32.655>which <00:04:32.854>you <00:04:33.054>do <00:04:33.255>not <00:04:33.455>need <00:04:33.655>to <00:04:33.854>train <00:04:34.054>anymore

100
00:04:34.885 --> 00:04:36.245
<00:04:34.885>because <00:04:35.293>it's <00:04:35.701>been <00:04:36.109>pre-trained.

101
00:04:36.595 --> 00:04:39.325
<00:04:36.595>It's <00:04:36.750>already <00:04:36.906>very <00:04:37.061>useful. <00:04:37.217>It's <00:04:37.372>already <00:04:37.528>a <00:04:37.683>very <00:04:37.839>high <00:04:37.995>number.

102
00:04:40.225 --> 00:04:43.725
<00:04:40.225>And <00:04:40.425>here's <00:04:40.625>another <00:04:40.825>good <00:04:41.025>example <00:04:41.225>how <00:04:41.425>multimodality <00:04:41.625>helps.

103
00:04:44.025 --> 00:04:45.765
<00:04:44.025>So <00:04:44.375>in <00:04:44.724>this <00:04:45.075>diagram, <00:04:45.424>uh,

104
00:04:45.765 --> 00:04:48.525
<00:04:45.765>also <00:04:45.920>from <00:04:46.076>the <00:04:46.231>gbd <00:04:46.387>four <00:04:46.542>technical <00:04:46.698>report, <00:04:46.853>uh, <00:04:47.009>they're <00:04:47.164>showing

105
00:04:48.525 --> 00:04:53.405
<00:04:48.525>that, <00:04:48.664>uh, <00:04:48.804>tbt, <00:04:48.945>uh, <00:04:49.085>GPT <00:04:49.224>uh, <00:04:49.364>3.5, <00:04:49.505>uh, <00:04:49.645>known <00:04:49.784>as <00:04:49.924>Chad,

106
00:04:53.565 --> 00:04:57.405
<00:04:53.565>GPT, <00:04:53.705>the, <00:04:53.844>the <00:04:53.985>basic <00:04:54.125>version <00:04:54.265>of <00:04:54.405>Chad <00:04:54.545>GPT, <00:04:54.685>uh, <00:04:54.825>was <00:04:54.965>able

107
00:04:57.405 --> 00:05:00.005
<00:04:57.405>to <00:04:57.604>perform <00:04:57.804>quite <00:04:58.005>well <00:04:58.205>on <00:04:58.405>some <00:04:58.604>exams, <00:04:58.804>uh,

108
00:05:00.385 --> 00:05:03.005
<00:05:00.385>but <00:05:00.618>didn't <00:05:00.851>perform <00:05:01.085>well <00:05:01.318>on <00:05:01.551>another <00:05:01.784>exams.

109
00:05:03.385 --> 00:05:05.805
<00:05:03.385>And, <00:05:03.585>uh, <00:05:03.784>it <00:05:03.985>was <00:05:04.185>adding <00:05:04.385>one <00:05:04.585>extra <00:05:04.784>modality,

110
00:05:05.805 --> 00:05:06.885
<00:05:05.805>which <00:05:06.129>is, <00:05:06.453>uh, <00:05:06.777>vision.

111
00:05:07.705 --> 00:05:10.805
<00:05:07.705>Uh, <00:05:07.880>they <00:05:08.055>were <00:05:08.229>able <00:05:08.405>to <00:05:08.580>achieve <00:05:08.755>much <00:05:08.930>better <00:05:09.104>performance.

112
00:05:11.025 --> 00:05:14.645
<00:05:11.025>And <00:05:11.180>you <00:05:11.336>see <00:05:11.491>that <00:05:11.647>this <00:05:11.802>improvement <00:05:11.958>is <00:05:12.113>no <00:05:12.269>longer <00:05:12.424>incremental.

113
00:05:14.795 --> 00:05:19.645
<00:05:14.795>It's <00:05:14.970>substantial. <00:05:15.145>Uh, <00:05:15.320>another <00:05:15.495>great <00:05:15.670>example <00:05:15.845>of, <00:05:16.020>uh, <00:05:16.195>uh,

114
00:05:19.865 --> 00:05:22.765
<00:05:19.865>how <00:05:20.065>models <00:05:20.265>like <00:05:20.465>GBT <00:05:20.665>four <00:05:20.865>can <00:05:21.065>be <00:05:21.265>useful.

115
00:05:23.105 --> 00:05:27.445
<00:05:23.105>So, <00:05:23.260>uh, <00:05:23.416>in <00:05:23.571>here <00:05:23.727>they <00:05:23.882>have <00:05:24.038>uploaded <00:05:24.193>this <00:05:24.349>image, <00:05:24.505>uh,

116
00:05:27.485 --> 00:05:29.125
<00:05:27.485>where <00:05:27.765>you <00:05:28.045>see <00:05:28.325>a <00:05:28.605>smartphone <00:05:28.885>connected

117
00:05:29.125 --> 00:05:32.085
<00:05:29.125>with <00:05:29.325>AVGA <00:05:29.525>connector <00:05:29.725>being <00:05:29.925>charged <00:05:30.125>by <00:05:30.325>AVGA <00:05:30.525>connector.

118
00:05:32.745 --> 00:05:36.205
<00:05:32.745>Uh, <00:05:32.900>and <00:05:33.056>they've <00:05:33.211>asked <00:05:33.367>the <00:05:33.522>g <00:05:33.678>PT <00:05:33.833>four <00:05:33.989>to <00:05:34.145>answer

119
00:05:36.515 --> 00:05:38.125
<00:05:36.515>what <00:05:36.794>is <00:05:37.075>funny <00:05:37.354>about <00:05:37.635>this <00:05:37.914>image?

120
00:05:38.505 --> 00:05:41.045
<00:05:38.505>And, <00:05:38.680>uh, <00:05:38.855>this <00:05:39.030>network <00:05:39.205>was <00:05:39.380>able <00:05:39.555>to <00:05:39.730>say, <00:05:39.905>okay,

121
00:05:41.045 --> 00:05:42.925
<00:05:41.045>there <00:05:41.245>are <00:05:41.445>three <00:05:41.645>different <00:05:41.845>panels <00:05:42.045>in <00:05:42.245>this <00:05:42.445>image.

122
00:05:43.185 --> 00:05:47.525
<00:05:43.185>We <00:05:43.325>see <00:05:43.465>a <00:05:43.605>knife, <00:05:43.745>uh, <00:05:43.885>smartphone, <00:05:44.025>uh, <00:05:44.165>we <00:05:44.305>see <00:05:44.445>AVGA <00:05:44.585>connector.

123
00:05:47.665 --> 00:05:50.485
<00:05:47.665>And <00:05:47.945>typically <00:05:48.225>you <00:05:48.505>don't <00:05:48.785>charge <00:05:49.065>smartphones

124
00:05:50.485 --> 00:05:51.765
<00:05:50.485>with <00:05:51.061>AVGA <00:05:51.637>connector.

125
00:05:51.765 --> 00:05:54.845
<00:05:51.765>And <00:05:51.940>this <00:05:52.115>is <00:05:52.289>funny. <00:05:52.465>So <00:05:52.640>the <00:05:52.815>network <00:05:52.990>is <00:05:53.164>able

126
00:05:54.905 --> 00:05:57.405
<00:05:54.905>to <00:05:55.184>reason <00:05:55.465>about <00:05:55.744>humor, <00:05:56.025>uh, <00:05:56.304>humor.

127
00:05:57.405 --> 00:05:58.645
<00:05:57.405>This <00:05:57.683>is, <00:05:57.962>this <00:05:58.241>is <00:05:58.520>unbelievable.

128
00:05:58.715 --> 00:06:01.645
<00:05:58.715>Even <00:05:58.870>a <00:05:59.026>year <00:05:59.181>ago, <00:05:59.337>I <00:05:59.492>wouldn't <00:05:59.648>expect <00:05:59.803>this <00:05:59.959>to <00:06:00.115>happen.

129
00:06:01.795 --> 00:06:03.805
<00:06:01.795>This <00:06:02.028>is, <00:06:02.261>this <00:06:02.495>is <00:06:02.728>a <00:06:02.961>huge <00:06:03.195>breakthrough.

130
00:06:04.505 --> 00:06:07.765
<00:06:04.505>And, <00:06:04.680>uh, <00:06:04.855>with <00:06:05.030>Nvidia, <00:06:05.205>we <00:06:05.380>have <00:06:05.555>also, <00:06:05.730>uh, <00:06:05.905>we,

131
00:06:07.765 --> 00:06:09.645
<00:06:07.765>we <00:06:07.998>are <00:06:08.231>serving <00:06:08.465>also <00:06:08.698>model <00:06:08.931>for <00:06:09.164>question,

132
00:06:09.645 --> 00:06:10.925
<00:06:09.645>visual <00:06:10.221>question <00:06:10.797>answering.

133
00:06:10.985 --> 00:06:13.725
<00:06:10.985>And, <00:06:11.140>uh, <00:06:11.296>I, <00:06:11.451>at <00:06:11.607>home, <00:06:11.762>home <00:06:11.918>also <00:06:12.073>this <00:06:12.229>funny <00:06:12.385>setup

134
00:06:13.725 --> 00:06:16.205
<00:06:13.725>with <00:06:14.075>a <00:06:14.425>mini <00:06:14.775>Tesla <00:06:15.125>supercharger

135
00:06:16.225 --> 00:06:17.965
<00:06:16.225>and, <00:06:16.458>uh, <00:06:16.691>where <00:06:16.925>I'm <00:06:17.158>charging <00:06:17.391>my <00:06:17.625>phone.

136
00:06:18.105 --> 00:06:20.085
<00:06:18.105>And, <00:06:18.385>uh, <00:06:18.665>I've <00:06:18.945>asked <00:06:19.225>this <00:06:19.505>question,

137
00:06:20.115 --> 00:06:21.965
<00:06:20.115>what <00:06:20.395>is <00:06:20.675>unusual <00:06:20.955>about <00:06:21.235>this <00:06:21.515>image?

138
00:06:22.105 --> 00:06:25.845
<00:06:22.105>And <00:06:22.280>it <00:06:22.455>was <00:06:22.630>able <00:06:22.805>to <00:06:22.980>perfectly <00:06:23.155>describe <00:06:23.330>the <00:06:23.505>scene

139
00:06:25.945 --> 00:06:28.925
<00:06:25.945>and <00:06:26.120>explain, <00:06:26.295>uh, <00:06:26.469>what <00:06:26.645>is <00:06:26.820>unusual <00:06:26.995>about <00:06:27.170>this <00:06:27.344>setup.

140
00:06:29.985 --> 00:06:31.525
<00:06:29.985>Uh, <00:06:30.447>check <00:06:30.909>this <00:06:31.371>link.

141
00:06:31.625 --> 00:06:34.005
<00:06:31.625>Uh, <00:06:31.858>it's, <00:06:32.091>uh, <00:06:32.325>actually <00:06:32.558>everyone <00:06:32.791>can <00:06:33.025>play

142
00:06:34.085 --> 00:06:35.325
<00:06:34.085>with <00:06:34.364>it <00:06:34.643>in <00:06:34.921>the <00:06:35.200>playground.

143
00:06:35.865 --> 00:06:39.285
<00:06:35.865>Um, <00:06:36.145>so <00:06:36.425>multimodal <00:06:36.705>generative <00:06:36.985>AI <00:06:37.265>is,

144
00:06:39.345 --> 00:06:40.965
<00:06:39.345>uh, <00:06:39.811>changing <00:06:40.278>every <00:06:40.745>industry.

145
00:06:40.965 --> 00:06:43.045
<00:06:40.965>It's <00:06:41.198>not <00:06:41.431>like <00:06:41.665>only <00:06:41.898>one <00:06:42.131>specific <00:06:42.365>industry.

146
00:06:43.165 --> 00:06:46.125
<00:06:43.165>I <00:06:43.320>think <00:06:43.476>the <00:06:43.631>whole <00:06:43.787>industry <00:06:43.942>is <00:06:44.098>being <00:06:44.253>affected <00:06:44.409>by, <00:06:44.565>uh,

147
00:06:46.275 --> 00:06:47.325
<00:06:46.275>this <00:06:47.219>technology.

148
00:06:47.905 --> 00:06:50.365
<00:06:47.905>And <00:06:48.184>in <00:06:48.465>education, <00:06:48.744>for <00:06:49.025>example, <00:06:49.304>uh,

149
00:06:50.365 --> 00:06:53.605
<00:06:50.365>we <00:06:50.540>can <00:06:50.715>benefit <00:06:50.890>from <00:06:51.065>things <00:06:51.240>like <00:06:51.415>personalized <00:06:51.590>learning <00:06:51.765>where,

150
00:06:53.825 --> 00:06:56.325
<00:06:53.825>uh, <00:06:54.058>the, <00:06:54.291>uh, <00:06:54.525>AI <00:06:54.758>will <00:06:54.991>learn, <00:06:55.224>uh,

151
00:06:56.505 --> 00:06:59.005
<00:06:56.505>the <00:06:56.784>specific <00:06:57.065>learning <00:06:57.344>patterns <00:06:57.625>from <00:06:57.905>students

152
00:06:59.745 --> 00:07:03.325
<00:06:59.745>and <00:06:59.900>will <00:07:00.056>also <00:07:00.211>tailor, <00:07:00.367>uh, <00:07:00.522>study <00:07:00.678>plans <00:07:00.833>based <00:07:00.989>on <00:07:01.145>that.

153
00:07:03.665 --> 00:07:06.965
<00:07:03.665>Um, <00:07:03.840>you <00:07:04.015>can <00:07:04.190>create <00:07:04.365>enhanced <00:07:04.540>materials, <00:07:04.715>uh, <00:07:04.890>like, <00:07:05.065>uh,

154
00:07:06.965 --> 00:07:10.685
<00:07:06.965>you <00:07:07.198>can <00:07:07.431>automatically <00:07:07.664>generate <00:07:07.898>some <00:07:08.131>diagrams, <00:07:08.364>visualizations

155
00:07:10.745 --> 00:07:13.085
<00:07:10.745>of <00:07:11.025>heart, <00:07:11.305>uh, <00:07:11.585>to <00:07:11.865>understand <00:07:12.145>concepts.

156
00:07:13.185 --> 00:07:15.605
<00:07:13.185>Uh, <00:07:13.385>so <00:07:13.585>this <00:07:13.785>is <00:07:13.985>another <00:07:14.185>benefit <00:07:14.385>of, <00:07:14.585>uh,

157
00:07:15.735 --> 00:07:20.005
<00:07:15.735>generative <00:07:15.968>AI <00:07:16.201>in <00:07:16.435>education <00:07:16.668>and <00:07:16.901>also <00:07:17.135>brainstorming.

158
00:07:20.105 --> 00:07:24.005
<00:07:20.105>Uh, <00:07:20.245>who <00:07:20.385>of <00:07:20.525>you <00:07:20.665>have <00:07:20.805>ever <00:07:20.945>had <00:07:21.085>this? <00:07:21.225>So-called <00:07:21.365>Writer's <00:07:21.505>Block.

159
00:07:24.465 --> 00:07:25.605
<00:07:24.465>Um, <00:07:24.807>I <00:07:25.149>definitely <00:07:25.491>had

160
00:07:25.605 --> 00:07:28.165
<00:07:25.605>that <00:07:25.760>when <00:07:25.916>I <00:07:26.071>was <00:07:26.227>working <00:07:26.382>my <00:07:26.538>master <00:07:26.693>thesis <00:07:26.849>or <00:07:27.005>PhD,

161
00:07:28.345 --> 00:07:30.285
<00:07:28.345>and <00:07:28.625>then <00:07:28.905>also <00:07:29.185>writing <00:07:29.465>scientific <00:07:29.745>papers.

162
00:07:30.505 --> 00:07:32.405
<00:07:30.505>You <00:07:30.971>open <00:07:31.438>some <00:07:31.905>editor

163
00:07:32.745 --> 00:07:34.445
<00:07:32.745>and, <00:07:32.920>uh, <00:07:33.095>you <00:07:33.270>don't <00:07:33.445>know <00:07:33.620>what <00:07:33.795>to <00:07:33.970>write <00:07:34.145>there

164
00:07:34.445 --> 00:07:37.365
<00:07:34.445>because, <00:07:34.678>uh, <00:07:34.911>well, <00:07:35.145>because <00:07:35.378>it's <00:07:35.611>very <00:07:35.844>intimidating.

165
00:07:37.505 --> 00:07:39.325
<00:07:37.505>The, <00:07:37.784>the, <00:07:38.065>the <00:07:38.344>huge <00:07:38.625>white <00:07:38.905>page,

166
00:07:39.425 --> 00:07:43.685
<00:07:39.425>but <00:07:39.552>now <00:07:39.679>is <00:07:39.806>ai, <00:07:39.934>you <00:07:40.061>can, <00:07:40.188>uh, <00:07:40.315>ask <00:07:40.443>for <00:07:40.570>some <00:07:40.697>seeding <00:07:40.825>ideas,

167
00:07:43.775 --> 00:07:45.685
<00:07:43.775>which <00:07:43.974>you <00:07:44.174>then <00:07:44.375>can <00:07:44.575>expand <00:07:44.775>on <00:07:44.974>your <00:07:45.174>own.

168
00:07:45.685 --> 00:07:46.845
<00:07:45.685>So <00:07:45.893>I <00:07:46.102>think <00:07:46.311>it's <00:07:46.520>super <00:07:46.729>helpful,

169
00:07:47.505 --> 00:07:49.725
<00:07:47.505>and <00:07:47.645>I'm <00:07:47.784>using <00:07:47.925>it <00:07:48.065>a <00:07:48.205>lot <00:07:48.344>these <00:07:48.485>days, <00:07:48.625>I <00:07:48.765>must <00:07:48.905>admit.

170
00:07:50.225 --> 00:07:51.565
<00:07:50.225>Um, <00:07:50.828>in <00:07:51.431>healthcare

171
00:07:51.585 --> 00:07:54.045
<00:07:51.585>and <00:07:51.935>life <00:07:52.284>sciences, <00:07:52.635>uh, <00:07:52.984>multimodal,

172
00:07:54.045 --> 00:07:56.485
<00:07:54.045>generative <00:07:54.245>AI <00:07:54.445>is <00:07:54.645>being <00:07:54.845>used <00:07:55.045>in <00:07:55.245>drug <00:07:55.445>discovery

173
00:07:56.705 --> 00:07:57.965
<00:07:56.705>and <00:07:57.272>precision <00:07:57.839>medicine,

174
00:07:57.975 --> 00:08:01.405
<00:07:57.975>where <00:07:58.150>you <00:07:58.325>combine <00:07:58.500>data <00:07:58.675>from <00:07:58.850>different <00:07:59.025>sources, <00:07:59.200>like, <00:07:59.375>uh,

175
00:08:01.475 --> 00:08:05.325
<00:08:01.475>patient <00:08:01.708>records, <00:08:01.941>uh, <00:08:02.175>history <00:08:02.408>from <00:08:02.641>like <00:08:02.875>video,

176
00:08:05.325 --> 00:08:06.365
<00:08:05.325>grandma <00:08:05.793>had <00:08:06.261>cancer

177
00:08:06.705 --> 00:08:09.925
<00:08:06.705>or <00:08:07.055>whatsoever, <00:08:07.405>um, <00:08:07.755>images, <00:08:08.104>uh,

178
00:08:10.035 --> 00:08:12.125
<00:08:10.035>digital <00:08:10.501>pathology <00:08:10.968>slides, <00:08:11.435>uh,

179
00:08:12.225 --> 00:08:14.005
<00:08:12.225>all <00:08:12.505>this <00:08:12.785>can <00:08:13.065>be <00:08:13.345>combined <00:08:13.625>together

180
00:08:14.425 --> 00:08:18.405
<00:08:14.425>and, <00:08:14.580>uh, <00:08:14.736>we <00:08:14.891>can <00:08:15.047>tailor <00:08:15.202>also <00:08:15.358>your <00:08:15.513>treatment <00:08:15.669>plans, <00:08:15.825>uh,

181
00:08:18.695 --> 00:08:20.885
<00:08:18.695>based <00:08:19.045>on <00:08:19.395>your <00:08:19.745>specific <00:08:20.094>history.

182
00:08:21.445 --> 00:08:22.845
<00:08:21.445>I <00:08:21.760>can <00:08:22.075>also <00:08:22.390>incorporate, <00:08:22.705>uh,

183
00:08:22.845 --> 00:08:26.645
<00:08:22.845>scientific <00:08:23.078>literature <00:08:23.311>from <00:08:23.545>PubMed, <00:08:23.778>for <00:08:24.011>example, <00:08:24.245>uh,

184
00:08:26.825 --> 00:08:30.685
<00:08:26.825>in <00:08:27.058>genomics, <00:08:27.291>also <00:08:27.525>multimodal <00:08:27.758>multimodality <00:08:27.991>used, <00:08:28.224>uh,

185
00:08:30.695 --> 00:08:33.925
<00:08:30.695>quite <00:08:30.850>a <00:08:31.006>lot, <00:08:31.161>uh, <00:08:31.317>these <00:08:31.472>days, <00:08:31.628>uh, <00:08:31.783>in <00:08:31.939>protein <00:08:32.095>structures.

186
00:08:34.225 --> 00:08:36.485
<00:08:34.225>But <00:08:34.425>also <00:08:34.625>another <00:08:34.825>cool <00:08:35.025>use <00:08:35.225>case <00:08:35.425>is, <00:08:35.625>uh,

187
00:08:36.745 --> 00:08:39.805
<00:08:36.745>to <00:08:37.025>generate <00:08:37.304>synthetic <00:08:37.585>organisms <00:08:37.865>and <00:08:38.145>organs.

188
00:08:39.865 --> 00:08:42.965
<00:08:39.865>Uh, <00:08:40.039>uh, <00:08:40.215>it's <00:08:40.390>of <00:08:40.565>course <00:08:40.740>it's <00:08:40.914>very <00:08:41.090>early <00:08:41.265>development,

189
00:08:43.105 --> 00:08:46.085
<00:08:43.105>but, <00:08:43.245>uh, <00:08:43.385>I <00:08:43.525>think <00:08:43.664>there's <00:08:43.805>some <00:08:43.945>works, <00:08:44.085>uh, <00:08:44.225>going <00:08:44.365>on <00:08:44.505>there.

190
00:08:46.865 --> 00:08:47.965
<00:08:46.865>And, <00:08:47.195>uh, <00:08:47.525>in <00:08:47.855>design

191
00:08:47.965 --> 00:08:51.525
<00:08:47.965>and <00:08:48.120>engineering, <00:08:48.276>you <00:08:48.431>can <00:08:48.587>also <00:08:48.742>use, <00:08:48.898>uh, <00:08:49.053>AI <00:08:49.209>for <00:08:49.365>brainstorming

192
00:08:51.625 --> 00:08:53.325
<00:08:51.625>and <00:08:52.091>generating <00:08:52.558>some <00:08:53.025>ideas.

193
00:08:54.225 --> 00:08:58.125
<00:08:54.225>Uh, <00:08:54.400>for <00:08:54.575>example, <00:08:54.750>fashion <00:08:54.925>designers, <00:08:55.100>uh, <00:08:55.275>often <00:08:55.450>use, <00:08:55.625>uh,

194
00:08:58.655 --> 00:09:01.885
<00:08:58.655>multimodal <00:08:58.888>jive <00:08:59.121>AI <00:08:59.355>to <00:08:59.588>create <00:08:59.821>some <00:09:00.055>images

195
00:09:02.065 --> 00:09:03.485
<00:09:02.065>and <00:09:02.384>then <00:09:02.704>expand <00:09:03.023>on <00:09:03.343>them.

196
00:09:04.105 --> 00:09:06.725
<00:09:04.105>Uh, <00:09:04.305>you <00:09:04.505>can <00:09:04.705>also <00:09:04.905>use, <00:09:05.105>uh, <00:09:05.305>jive <00:09:05.505>AI

197
00:09:06.725 --> 00:09:10.845
<00:09:06.725>to <00:09:06.925>synthesize <00:09:07.125>large <00:09:07.325>complex <00:09:07.525>solutions <00:09:07.725>to <00:09:07.925>automate <00:09:08.125>some

198
00:09:11.405 --> 00:09:15.005
<00:09:11.405>mundane <00:09:11.605>tasks <00:09:11.805>and <00:09:12.005>also <00:09:12.204>to <00:09:12.405>generate <00:09:12.605>parametric <00:09:12.805>designs.

199
00:09:15.185 --> 00:09:17.845
<00:09:15.185>So <00:09:15.340>think <00:09:15.496>of <00:09:15.651>an <00:09:15.807>example <00:09:15.962>where <00:09:16.118>you <00:09:16.273>have <00:09:16.429>an <00:09:16.584>object,

200
00:09:18.085 --> 00:09:19.805
<00:09:18.085>a <00:09:18.285>three <00:09:18.485>D <00:09:18.685>shape, <00:09:18.885>like <00:09:19.085>a <00:09:19.285>CAD <00:09:19.485>model,

201
00:09:20.585 --> 00:09:23.725
<00:09:20.585>and, <00:09:20.785>uh, <00:09:20.985>you <00:09:21.185>want <00:09:21.385>to <00:09:21.585>modify <00:09:21.785>it, <00:09:21.985>uh,

202
00:09:23.755 --> 00:09:27.005
<00:09:23.755>depending <00:09:23.910>on <00:09:24.066>the <00:09:24.221>environment <00:09:24.377>where <00:09:24.532>this <00:09:24.688>object <00:09:24.843>will <00:09:24.999>be <00:09:25.155>used,

203
00:09:27.425 --> 00:09:29.125
<00:09:27.425>you <00:09:27.704>can <00:09:27.984>then <00:09:28.265>use <00:09:28.545>the <00:09:28.824>information,

204
00:09:29.395 --> 00:09:31.685
<00:09:29.395>some <00:09:29.675>meta <00:09:29.954>information <00:09:30.235>about <00:09:30.515>the <00:09:30.795>environment,

205
00:09:32.155 --> 00:09:33.325
<00:09:32.155>feed <00:09:32.418>it <00:09:32.681>into <00:09:32.944>the <00:09:33.208>network,

206
00:09:33.625 --> 00:09:36.645
<00:09:33.625>and <00:09:33.825>the <00:09:34.025>network <00:09:34.225>will <00:09:34.425>then <00:09:34.625>adjust <00:09:34.825>your, <00:09:35.025>uh,

207
00:09:37.145 --> 00:09:38.485
<00:09:37.145>CAD <00:09:37.748>model <00:09:38.351>accordingly.

208
00:09:38.625 --> 00:09:39.965
<00:09:38.625>So <00:09:38.926>it's, <00:09:39.228>it's <00:09:39.529>very <00:09:39.831>useful,

209
00:09:40.025 --> 00:09:41.805
<00:09:40.025>and <00:09:40.305>I <00:09:40.584>see <00:09:40.865>a <00:09:41.145>huge <00:09:41.425>demand

210
00:09:41.865 --> 00:09:43.845
<00:09:41.865>for <00:09:42.098>this <00:09:42.331>task <00:09:42.565>in <00:09:42.798>many <00:09:43.031>different <00:09:43.265>industries.

211
00:09:44.665 --> 00:09:46.805
<00:09:44.665>Um, <00:09:44.944>in <00:09:45.224>marketing, <00:09:45.505>media <00:09:45.785>and <00:09:46.064>entertainment.

212
00:09:47.105 --> 00:09:48.245
<00:09:47.105>Uh, <00:09:47.447>think <00:09:47.789>of, <00:09:48.131>uh,

213
00:09:48.275 --> 00:09:51.085
<00:09:48.275>such <00:09:48.625>things <00:09:48.975>like <00:09:49.324>personalized <00:09:49.675>advertisement

214
00:09:51.135 --> 00:09:54.445
<00:09:51.135>where <00:09:51.310>the <00:09:51.485>content <00:09:51.660>will <00:09:51.835>be <00:09:52.010>created <00:09:52.185>specifically <00:09:52.360>for <00:09:52.535>you,

215
00:09:55.005 --> 00:09:57.125
<00:09:55.005>although, <00:09:55.145>I <00:09:55.285>must <00:09:55.425>admit, <00:09:55.564>I <00:09:55.705>think <00:09:55.845>it's <00:09:55.985>a <00:09:56.125>little <00:09:56.265>bit <00:09:56.405>creepy.

216
00:09:57.425 --> 00:10:00.125
<00:09:57.425>Um, <00:09:57.625>then <00:09:57.824>you <00:09:58.025>can <00:09:58.224>think <00:09:58.425>of <00:09:58.625>content <00:09:58.824>creation

217
00:10:00.425 --> 00:10:02.285
<00:10:00.425>and <00:10:01.125>interactive <00:10:01.824>experiences.

218
00:10:03.145 --> 00:10:07.005
<00:10:03.145>Um, <00:10:03.378>interactive <00:10:03.611>experiences <00:10:03.845>means <00:10:04.078>things <00:10:04.311>like <00:10:04.545>VR,

219
00:10:07.135 --> 00:10:08.965
<00:10:07.135>where <00:10:07.368>you <00:10:07.601>will <00:10:07.835>be <00:10:08.068>playing <00:10:08.301>a <00:10:08.535>game,

220
00:10:09.305 --> 00:10:12.885
<00:10:09.305>and <00:10:09.444>then <00:10:09.584>the, <00:10:09.724>let's <00:10:09.864>say <00:10:10.005>non <00:10:10.145>playable <00:10:10.285>characters <00:10:10.425>will <00:10:10.564>be <00:10:10.704>able

221
00:10:12.885 --> 00:10:14.365
<00:10:12.885>to <00:10:13.329>communicate <00:10:13.773>with <00:10:14.217>you.

222
00:10:15.025 --> 00:10:17.965
<00:10:15.025>And, <00:10:15.165>uh, <00:10:15.305>in <00:10:15.444>the <00:10:15.584>next <00:10:15.725>example <00:10:15.865>in <00:10:16.005>the <00:10:16.145>video, <00:10:16.285>I'm <00:10:16.425>going

223
00:10:17.965 --> 00:10:21.525
<00:10:17.965>to <00:10:18.105>show <00:10:18.245>you <00:10:18.385>it's, <00:10:18.525>uh, <00:10:18.665>such <00:10:18.805>an <00:10:18.945>example <00:10:19.085>of, <00:10:19.225>uh, <00:10:19.365>Nvidia

224
00:10:21.525 --> 00:10:22.765
<00:10:21.525>and <00:10:21.896>our <00:10:22.269>partner <00:10:22.641>developed.

225
00:10:22.765 --> 00:10:25.885
<00:10:22.765>Uh, <00:10:22.920>it's <00:10:23.076>an <00:10:23.231>interactive <00:10:23.387>game, <00:10:23.542>so <00:10:23.698>let's <00:10:23.853>take <00:10:24.009>a <00:10:24.165>look.

226
00:10:36.475 --> 00:10:38.225
<00:10:36.475>We're <00:10:36.755>thrilled <00:10:37.035>to <00:10:37.315>showcase <00:10:37.595>an <00:10:37.875>immersive

227
00:10:38.405 --> 00:10:40.865
<00:10:38.405>and <00:10:38.638>dynamic <00:10:38.871>interaction <00:10:39.105>between <00:10:39.338>a <00:10:39.571>player <00:10:39.805>Kai

228
00:10:40.965 --> 00:10:43.185
<00:10:40.965>and <00:10:41.198>a <00:10:41.431>non <00:10:41.665>playable <00:10:41.898>character <00:10:42.131>named <00:10:42.365>Chin.

229
00:10:43.655 --> 00:10:46.025
<00:10:43.655>This <00:10:43.888>captivating <00:10:44.121>Robin <00:10:44.355>shop <00:10:44.588>scene <00:10:44.821>was <00:10:45.055>created

230
00:10:46.085 --> 00:10:49.025
<00:10:46.085>by <00:10:46.285>Nvidia <00:10:46.485>using <00:10:46.685>the <00:10:46.885>latest <00:10:47.085>RTX <00:10:47.285>rendering <00:10:47.485>technologies.

231
00:10:49.965 --> 00:10:53.385
<00:10:49.965>Nvidia <00:10:50.165>collaborated <00:10:50.365>with <00:10:50.565>Convey <00:10:50.765>an <00:10:50.965>Nvidia <00:10:51.165>inception <00:10:51.365>partner

232
00:10:53.805 --> 00:10:56.705
<00:10:53.805>to <00:10:53.944>showcase <00:10:54.084>how <00:10:54.224>developers <00:10:54.364>will <00:10:54.505>soon <00:10:54.645>be <00:10:54.785>able <00:10:54.925>to <00:10:55.064>use <00:10:55.204>Nvidia,

233
00:10:56.865 --> 00:10:57.945
<00:10:56.865>Reba <00:10:57.837>Nemo

234
00:10:58.205 --> 00:11:00.665
<00:10:58.205>and <00:10:58.405>audio <00:10:58.605>to <00:10:58.805>face <00:10:59.005>for <00:10:59.205>AI <00:10:59.405>powered <00:10:59.605>speech,

235
00:11:00.825 --> 00:11:02.505
<00:11:00.825>conversation <00:11:01.525>and <00:11:02.225>animation.

236
00:11:03.155 --> 00:11:05.785
<00:11:03.155>These <00:11:03.355>models <00:11:03.555>were <00:11:03.755>integrated <00:11:03.954>into <00:11:04.155>the <00:11:04.355>Convey <00:11:04.555>services

237
00:11:06.185 --> 00:11:08.105
<00:11:06.185>platform <00:11:06.385>and <00:11:06.584>fed <00:11:06.785>it <00:11:06.984>to <00:11:07.185>Unreal <00:11:07.385>Engine <00:11:07.584>five

238
00:11:08.105 --> 00:11:10.625
<00:11:08.105>and <00:11:08.338>Metahuman <00:11:08.571>to <00:11:08.805>bring <00:11:09.038>Jen <00:11:09.271>to <00:11:09.505>life.

239
00:11:11.165 --> 00:11:12.305
<00:11:11.165>Please <00:11:11.507>take <00:11:11.848>a <00:11:12.190>look.

240
00:11:38.445 --> 00:11:41.785
<00:11:38.445>Hey, <00:11:38.620>Jen, <00:11:38.795>how <00:11:38.970>are <00:11:39.145>you? <00:11:39.320>Unfortunately <00:11:39.495>not <00:11:39.670>so <00:11:39.845>good.

241
00:11:42.605 --> 00:11:45.705
<00:11:42.605>How <00:11:42.780>come <00:11:42.955>I'm <00:11:43.130>worried <00:11:43.305>about <00:11:43.480>the <00:11:43.655>crime <00:11:43.830>around <00:11:44.005>here?

242
00:11:45.775 --> 00:11:46.825
<00:11:45.775>It's <00:11:46.090>gotten <00:11:46.405>bad <00:11:46.720>lately.

243
00:11:47.685 --> 00:11:49.945
<00:11:47.685>My <00:11:47.885>ramen <00:11:48.084>shop <00:11:48.285>got <00:11:48.484>caught <00:11:48.685>in <00:11:48.885>the <00:11:49.084>crossfire.

244
00:11:50.765 --> 00:11:53.725
<00:11:50.765>Can <00:11:50.905>I <00:11:51.045>help? <00:11:51.185>If <00:11:51.324>you <00:11:51.465>want <00:11:51.605>to <00:11:51.745>do <00:11:51.885>something <00:11:52.025>about <00:11:52.165>this,

245
00:11:54.405 --> 00:11:56.325
<00:11:54.405>I <00:11:54.605>have <00:11:54.805>heard <00:11:55.005>rumors <00:11:55.204>that <00:11:55.405>the <00:11:55.605>powerful <00:11:55.805>crime,

246
00:11:56.325 --> 00:11:59.885
<00:11:56.325>Lord <00:11:56.452>Kuman <00:11:56.579>Aoki, <00:11:56.706>is <00:11:56.834>causing <00:11:56.961>all <00:11:57.088>sorts <00:11:57.215>of <00:11:57.343>chaos <00:11:57.470>in <00:11:57.597>the <00:11:57.725>city.

247
00:12:00.465 --> 00:12:01.805
<00:12:00.465>He <00:12:00.637>may <00:12:00.809>be <00:12:00.981>the <00:12:01.154>root <00:12:01.326>of <00:12:01.498>this <00:12:01.670>violence.

248
00:12:03.115 --> 00:12:04.965
<00:12:03.115>I'll <00:12:03.290>talk <00:12:03.465>to <00:12:03.640>him. <00:12:03.815>Where <00:12:03.990>can <00:12:04.165>I <00:12:04.340>find <00:12:04.515>him?

249
00:12:06.045 --> 00:12:08.005
<00:12:06.045>I <00:12:06.185>have <00:12:06.324>heard <00:12:06.464>he <00:12:06.604>hangs <00:12:06.745>out <00:12:06.885>in <00:12:07.025>the <00:12:07.165>underground <00:12:07.305>fight <00:12:07.444>clubs

250
00:12:08.025 --> 00:12:09.165
<00:12:08.025>on <00:12:08.281>the <00:12:08.538>city's <00:12:08.794>east <00:12:09.050>side.

251
00:12:10.225 --> 00:12:11.225
<00:12:10.225>Try <00:12:11.125>there.

252
00:12:11.955 --> 00:12:14.605
<00:12:11.955>Okay, <00:12:12.235>I'll <00:12:12.515>go. <00:12:12.795>Be <00:12:13.075>careful. <00:12:13.355>Kai.

253
00:12:28.715 --> 00:12:31.875
<00:12:28.715>I <00:12:28.870>think <00:12:29.026>it's <00:12:29.181>truly <00:12:29.337>fascinating, <00:12:29.492>uh, <00:12:29.648>to <00:12:29.803>watch <00:12:29.959>this <00:12:30.115>video.

254
00:12:32.455 --> 00:12:34.435
<00:12:32.455>Uh, <00:12:32.688>here's <00:12:32.921>another <00:12:33.155>good <00:12:33.388>example <00:12:33.621>of <00:12:33.855>of

255
00:12:34.655 --> 00:12:38.155
<00:12:34.655>how <00:12:34.829>this <00:12:35.005>technology <00:12:35.180>can <00:12:35.355>be <00:12:35.530>useful, <00:12:35.704>uh, <00:12:35.880>and, <00:12:36.055>um,

256
00:12:38.575 --> 00:12:39.915
<00:12:38.575>and <00:12:38.776>the <00:12:38.977>example <00:12:39.178>I'm <00:12:39.379>gonna <00:12:39.579>show <00:12:39.781>you.

257
00:12:40.055 --> 00:12:42.995
<00:12:40.055>Uh, <00:12:40.229>we <00:12:40.405>used <00:12:40.579>Omnibus <00:12:40.755>platform, <00:12:40.930>the <00:12:41.104>same <00:12:41.280>Omnis <00:12:41.454>platform,

258
00:12:43.045 --> 00:12:45.555
<00:12:43.045>which <00:12:43.200>was <00:12:43.356>used <00:12:43.511>to <00:12:43.667>build <00:12:43.822>the <00:12:43.978>Ram <00:12:44.133>shop <00:12:44.289>in <00:12:44.444>order

259
00:12:45.555 --> 00:12:49.515
<00:12:45.555>to <00:12:45.755>create <00:12:45.954>interactive, <00:12:46.155>uh, <00:12:46.354>floor <00:12:46.555>plan, <00:12:46.755>uh, <00:12:46.954>designs.

260
00:12:49.775 --> 00:12:50.875
<00:12:49.775>So <00:12:50.022>let's <00:12:50.270>take <00:12:50.517>a <00:12:50.765>look.

261
00:12:55.715 --> 00:12:58.295
<00:12:55.715>By <00:12:55.870>now, <00:12:56.026>you've <00:12:56.181>probably <00:12:56.337>heard <00:12:56.492>how <00:12:56.648>exciting <00:12:56.803>GPT <00:12:56.959>four <00:12:57.115>is.

262
00:12:58.555 --> 00:13:01.975
<00:12:58.555>Of <00:12:58.710>course, <00:12:58.866>omniverse <00:12:59.021>is <00:12:59.177>the <00:12:59.332>best <00:12:59.488>place <00:12:59.643>for <00:12:59.799>building, <00:12:59.954>sorry,

263
00:13:01.995 --> 00:13:03.655
<00:13:01.995>and <00:13:02.275>tools <00:13:02.555>for <00:13:02.835>three <00:13:03.115>degenerative <00:13:03.395>AI,

264
00:13:03.815 --> 00:13:05.175
<00:13:03.815>RTX <00:13:04.223>rendering <00:13:04.631>and <00:13:05.039>simulation.

265
00:13:05.355 --> 00:13:07.855
<00:13:05.355>So <00:13:05.471>what <00:13:05.588>makes <00:13:05.705>sense <00:13:05.821>to <00:13:05.938>combine <00:13:06.055>the <00:13:06.171>two <00:13:06.288>and <00:13:06.405>see <00:13:06.521>what <00:13:06.638>we <00:13:06.755>get?

266
00:13:08.435 --> 00:13:10.695
<00:13:08.435>We <00:13:08.668>already <00:13:08.901>have <00:13:09.135>partners <00:13:09.368>building <00:13:09.601>amazing <00:13:09.834>tools.

267
00:13:10.795 --> 00:13:12.735
<00:13:10.795>So <00:13:10.950>let's <00:13:11.106>start <00:13:11.261>off <00:13:11.417>by <00:13:11.572>using <00:13:11.728>one <00:13:11.883>from <00:13:12.039>Sync <00:13:12.194>Twin

268
00:13:12.735 --> 00:13:14.175
<00:13:12.735>that <00:13:12.951>makes <00:13:13.167>floor <00:13:13.383>plans <00:13:13.598>like <00:13:13.814>this <00:13:14.031>one.

269
00:13:14.595 --> 00:13:15.935
<00:13:14.595>We <00:13:14.896>built <00:13:15.198>our <00:13:15.499>own <00:13:15.800>extension

270
00:13:15.935 --> 00:13:17.895
<00:13:15.935>that <00:13:16.135>let's <00:13:16.334>us <00:13:16.535>tell <00:13:16.734>chat <00:13:16.935>GPT <00:13:17.135>the <00:13:17.334>type

271
00:13:17.895 --> 00:13:19.095
<00:13:17.895>area <00:13:18.111>we <00:13:18.327>want <00:13:18.543>this <00:13:18.759>to <00:13:18.975>be.

272
00:13:19.555 --> 00:13:21.695
<00:13:19.555>All <00:13:19.671>we <00:13:19.788>have <00:13:19.905>to <00:13:20.021>do <00:13:20.138>is <00:13:20.255>draw <00:13:20.371>around <00:13:20.488>the <00:13:20.604>area <00:13:20.721>based <00:13:20.838>on <00:13:20.954>Sync

273
00:13:21.695 --> 00:13:23.375
<00:13:21.695>Twin's <00:13:21.928>Blueprint, <00:13:22.161>and <00:13:22.395>it <00:13:22.628>builds <00:13:22.861>the <00:13:23.095>floor.

274
00:13:23.755 --> 00:13:25.255
<00:13:23.755>It <00:13:24.025>not <00:13:24.295>only <00:13:24.564>uses <00:13:24.835>that <00:13:25.105>description,

275
00:13:25.355 --> 00:13:26.655
<00:13:25.355>but <00:13:25.647>also <00:13:25.940>uses <00:13:26.232>Deep <00:13:26.525>SEARCHs

276
00:13:26.655 --> 00:13:29.495
<00:13:26.655>to <00:13:26.810>pull <00:13:26.966>in <00:13:27.121>the <00:13:27.277>perfectly <00:13:27.432>suited <00:13:27.588>sim <00:13:27.743>ready <00:13:27.899>library <00:13:28.055>asset.

277
00:13:30.115 --> 00:13:31.495
<00:13:30.115>If <00:13:30.270>we <00:13:30.425>want <00:13:30.580>to <00:13:30.736>change <00:13:30.891>one <00:13:31.046>of <00:13:31.201>the <00:13:31.357>items,

278
00:13:31.795 --> 00:13:32.895
<00:13:31.795>we <00:13:32.125>use <00:13:32.454>another <00:13:32.785>extension

279
00:13:32.895 --> 00:13:34.335
<00:13:32.895>that <00:13:33.080>lets <00:13:33.265>us <00:13:33.450>pick <00:13:33.635>through <00:13:33.820>the <00:13:34.005>search <00:13:34.191>result.

280
00:13:34.335 --> 00:13:37.535
<00:13:34.335>That <00:13:34.475>deep <00:13:34.615>search <00:13:34.755>gives <00:13:34.895>us <00:13:35.035>all <00:13:35.175>this <00:13:35.315>using <00:13:35.455>the <00:13:35.595>combined <00:13:35.735>power

281
00:13:37.535 --> 00:13:39.895
<00:13:37.535>of <00:13:37.735>GPT <00:13:37.935>and <00:13:38.135>generative <00:13:38.334>AI <00:13:38.535>tools <00:13:38.735>and <00:13:38.935>omniverse.

282
00:13:40.425 --> 00:13:42.935
<00:13:40.425>While <00:13:40.552>we're <00:13:40.679>at <00:13:40.806>it, <00:13:40.934>let's <00:13:41.061>use <00:13:41.188>chat <00:13:41.315>GPT <00:13:41.443>to <00:13:41.570>write <00:13:41.697>this <00:13:41.824>script

283
00:13:45.285 --> 00:13:46.285
<00:13:45.285>As <00:13:45.584>a <00:13:45.885>disclaimer. <00:13:46.185>I

284
00:13:46.285 --> 00:13:48.265
<00:13:46.285>didn't <00:13:46.485>use <00:13:46.685>chat <00:13:46.885>ability <00:13:47.084>to <00:13:47.285>write <00:13:47.485>my <00:13:47.685>script,

285
00:13:48.525 --> 00:13:50.665
<00:13:48.525>but <00:13:48.805>so <00:13:49.084>what <00:13:49.365>is <00:13:49.645>generative <00:13:49.925>ai?

286
00:13:50.765 --> 00:13:53.265
<00:13:50.765>Um, <00:13:50.965>generative <00:13:51.165>AI <00:13:51.365>is <00:13:51.564>a <00:13:51.765>type <00:13:51.965>of <00:13:52.165>AI

287
00:13:53.265 --> 00:13:54.825
<00:13:53.265>that <00:13:53.615>can <00:13:53.965>create <00:13:54.314>new <00:13:54.665>content

288
00:13:55.085 --> 00:13:57.585
<00:13:55.085>by <00:13:55.365>learning <00:13:55.645>patterns <00:13:55.925>from <00:13:56.205>existing <00:13:56.485>data,

289
00:13:57.765 --> 00:14:00.745
<00:13:57.765>and <00:13:57.939>then <00:13:58.115>using <00:13:58.290>this <00:13:58.465>knowledge <00:13:58.640>to <00:13:58.814>generate <00:13:58.990>some <00:13:59.165>new

290
00:14:00.805 --> 00:14:04.345
<00:14:00.805>and <00:14:00.979>unique <00:14:01.155>output <00:14:01.329>in <00:14:01.505>the <00:14:01.680>base <00:14:01.854>of <00:14:02.030>it <00:14:02.204>all.

291
00:14:04.725 --> 00:14:07.105
<00:14:04.725>We <00:14:04.925>have <00:14:05.125>a <00:14:05.325>large <00:14:05.525>language <00:14:05.725>model, <00:14:05.925>also <00:14:06.125>known

292
00:14:07.185 --> 00:14:08.345
<00:14:07.185>as <00:14:07.533>a <00:14:07.881>foundation <00:14:08.229>model,

293
00:14:09.165 --> 00:14:11.785
<00:14:09.165>and, <00:14:09.398>uh, <00:14:09.631>it's <00:14:09.865>been <00:14:10.098>trained <00:14:10.331>on <00:14:10.564>data

294
00:14:11.845 --> 00:14:14.825
<00:14:11.845>of <00:14:12.078>different <00:14:12.311>sources <00:14:12.545>like <00:14:12.778>text, <00:14:13.011>image, <00:14:13.245>audio,

295
00:14:15.115 --> 00:14:16.545
<00:14:15.115>three <00:14:15.436>D, <00:14:15.758>and <00:14:16.080>structured <00:14:16.401>data.

296
00:14:17.125 --> 00:14:20.105
<00:14:17.125>Uh, <00:14:17.358>well <00:14:17.591>previously <00:14:17.825>conventionally, <00:14:18.058>uh, <00:14:18.291>this, <00:14:18.525>uh,

297
00:14:20.105 --> 00:14:22.105
<00:14:20.105>foundation <00:14:20.305>models <00:14:20.505>have <00:14:20.705>been <00:14:20.905>trained <00:14:21.105>on <00:14:21.305>text <00:14:21.505>data,

298
00:14:22.165 --> 00:14:25.065
<00:14:22.165>but <00:14:22.365>now <00:14:22.564>with <00:14:22.765>multimodality, <00:14:22.964>we <00:14:23.165>are <00:14:23.365>expanding <00:14:23.564>that

299
00:14:25.165 --> 00:14:26.825
<00:14:25.165>to <00:14:25.515>the <00:14:25.865>new, <00:14:26.214>new <00:14:26.564>modalities.

300
00:14:27.765 --> 00:14:31.425
<00:14:27.765>And, <00:14:27.920>uh, <00:14:28.076>the, <00:14:28.231>basically <00:14:28.387>the <00:14:28.542>way <00:14:28.698>it's <00:14:28.853>been <00:14:29.009>trained, <00:14:29.165>uh,

301
00:14:31.425 --> 00:14:33.025
<00:14:31.425>this <00:14:31.775>foundation <00:14:32.125>model, <00:14:32.474>it's, <00:14:32.824>uh,

302
00:14:33.025 --> 00:14:34.505
<00:14:33.025>it's <00:14:33.358>the <00:14:33.691>next <00:14:34.024>word <00:14:34.357>predictor.

303
00:14:34.525 --> 00:14:38.025
<00:14:34.525>So <00:14:34.680>we <00:14:34.836>have <00:14:34.991>this <00:14:35.147>huge <00:14:35.302>transformer <00:14:35.458>model, <00:14:35.613>which <00:14:35.769>is <00:14:35.925>capable

304
00:14:38.245 --> 00:14:40.705
<00:14:38.245>of <00:14:38.525>doing <00:14:38.805>one <00:14:39.085>thing <00:14:39.365>only <00:14:39.645>predict

305
00:14:40.735 --> 00:14:42.945
<00:14:40.735>what <00:14:41.015>comes <00:14:41.295>next <00:14:41.575>in <00:14:41.855>the <00:14:42.135>sentence.

306
00:14:43.605 --> 00:14:48.465
<00:14:43.605>And <00:14:43.760>this <00:14:43.916>task <00:14:44.071>can <00:14:44.227>be <00:14:44.382>further <00:14:44.538>adapted, <00:14:44.693>uh, <00:14:44.849>with, <00:14:45.005>uh,

307
00:14:48.895 --> 00:14:49.905
<00:14:48.895>well <00:14:49.349>different <00:14:49.804>tasks.

308
00:14:50.165 --> 00:14:53.965
<00:14:50.165>Uh, <00:14:50.292>you <00:14:50.419>can <00:14:50.546>further <00:14:50.674>customize <00:14:50.801>your <00:14:50.928>model <00:14:51.055>in <00:14:51.183>order <00:14:51.310>to <00:14:51.437>be <00:14:51.564>able

309
00:14:53.965 --> 00:14:57.765
<00:14:53.965>to <00:14:54.165>perform <00:14:54.365>different <00:14:54.565>tasks <00:14:54.765>like <00:14:54.965>code <00:14:55.165>generation <00:14:55.365>translation,

310
00:14:58.255 --> 00:15:00.765
<00:14:58.255>texts <00:14:58.488>and, <00:14:58.721>and <00:14:58.955>image <00:14:59.188>and <00:14:59.421>video <00:14:59.655>summarization.

311
00:15:01.305 --> 00:15:03.605
<00:15:01.305>You <00:15:01.538>can <00:15:01.771>do <00:15:02.005>image <00:15:02.238>captioning, <00:15:02.471>uh, <00:15:02.704>virtual

312
00:15:04.205 --> 00:15:07.205
<00:15:04.205>question <00:15:04.438>ans <00:15:04.671>visual <00:15:04.905>question <00:15:05.138>answering <00:15:05.371>sentiment <00:15:05.605>analysis

313
00:15:07.265 --> 00:15:08.325
<00:15:07.265>and <00:15:07.742>object <00:15:08.219>recognition,

314
00:15:08.325 --> 00:15:10.885
<00:15:08.325>like <00:15:08.605>the <00:15:08.885>typical <00:15:09.165>computer <00:15:09.445>vision <00:15:09.725>task.

315
00:15:11.185 --> 00:15:14.725
<00:15:11.185>So, <00:15:11.359>uh, <00:15:11.535>you <00:15:11.709>can <00:15:11.885>basically <00:15:12.060>take <00:15:12.234>your <00:15:12.410>transformer <00:15:12.584>model

316
00:15:14.865 --> 00:15:18.365
<00:15:14.865>and <00:15:15.098>then <00:15:15.331>customize <00:15:15.565>it, <00:15:15.798>applying <00:15:16.031>different <00:15:16.265>techniques,

317
00:15:19.465 --> 00:15:22.005
<00:15:19.465>and <00:15:19.745>to <00:15:20.025>create <00:15:20.305>this <00:15:20.585>foundation <00:15:20.865>model

318
00:15:22.105 --> 00:15:24.885
<00:15:22.105>or <00:15:22.305>to <00:15:22.505>pre-train <00:15:22.705>it, <00:15:22.905>uh, <00:15:23.105>from <00:15:23.305>scratch, <00:15:23.505>uh,

319
00:15:24.945 --> 00:15:27.005
<00:15:24.945>you <00:15:25.295>need <00:15:25.645>basically <00:15:25.995>streaming <00:15:26.345>components.

320
00:15:27.225 --> 00:15:29.925
<00:15:27.225>The <00:15:27.400>first <00:15:27.575>one <00:15:27.750>is, <00:15:27.925>uh, <00:15:28.100>large <00:15:28.275>amounts <00:15:28.450>of <00:15:28.625>data.

321
00:15:30.145 --> 00:15:31.405
<00:15:30.145>And <00:15:30.428>as <00:15:30.712>I've <00:15:30.995>already <00:15:31.279>mentioned

322
00:15:31.405 --> 00:15:33.725
<00:15:31.405>before, <00:15:31.605>you <00:15:31.805>need <00:15:32.005>the <00:15:32.204>whole <00:15:32.405>scale <00:15:32.605>of <00:15:32.805>internet

323
00:15:33.785 --> 00:15:36.285
<00:15:33.785>to <00:15:34.064>train <00:15:34.344>a <00:15:34.625>meaningful <00:15:34.905>foundation <00:15:35.185>model.

324
00:15:37.065 --> 00:15:40.525
<00:15:37.065>And, <00:15:37.298>uh, <00:15:37.531>such, <00:15:37.765>uh, <00:15:37.998>foundation <00:15:38.231>models, <00:15:38.465>uh,

325
00:15:40.525 --> 00:15:42.725
<00:15:40.525>usually <00:15:40.875>contain <00:15:41.225>billions <00:15:41.574>of <00:15:41.925>parameters

326
00:15:42.725 --> 00:15:44.965
<00:15:42.725>because <00:15:42.958>you <00:15:43.191>need <00:15:43.425>to <00:15:43.658>reflect <00:15:43.891>all <00:15:44.125>this,

327
00:15:45.145 --> 00:15:46.685
<00:15:45.145>uh, <00:15:45.607>the <00:15:46.069>whole <00:15:46.531>information.

328
00:15:46.805 --> 00:15:48.885
<00:15:46.805>You <00:15:46.979>need <00:15:47.155>to <00:15:47.329>somehow <00:15:47.505>encode <00:15:47.680>it <00:15:47.854>in <00:15:48.030>the <00:15:48.204>network.

329
00:15:49.145 --> 00:15:52.085
<00:15:49.145>And, <00:15:49.345>uh, <00:15:49.545>it <00:15:49.745>needs <00:15:49.944>many <00:15:50.145>connections, <00:15:50.345>many <00:15:50.545>neurons,

330
00:15:52.485 --> 00:15:53.565
<00:15:52.485>billions <00:15:52.971>of <00:15:53.457>parameters.

331
00:15:53.565 --> 00:15:54.805
<00:15:53.565>So <00:15:53.844>this <00:15:54.123>models <00:15:54.401>are <00:15:54.680>large

332
00:15:54.905 --> 00:15:57.805
<00:15:54.905>and <00:15:55.045>they <00:15:55.185>require <00:15:55.324>a <00:15:55.464>lot <00:15:55.605>like, <00:15:55.745>uh, <00:15:55.885>a <00:15:56.025>lot <00:15:56.165>of <00:15:56.305>compute

333
00:15:57.865 --> 00:15:59.045
<00:15:57.865>and <00:15:58.396>huge <00:15:58.926>infrastructure.

334
00:15:59.865 --> 00:16:02.525
<00:15:59.865>And <00:16:00.040>you <00:16:00.215>also <00:16:00.390>need <00:16:00.565>to <00:16:00.740>rely <00:16:00.915>on <00:16:01.090>unsupervised <00:16:01.265>learning

335
00:16:02.745 --> 00:16:04.125
<00:16:02.745>and <00:16:03.366>unsupervised <00:16:03.987>learning.

336
00:16:04.125 --> 00:16:07.885
<00:16:04.125>It's <00:16:04.300>also <00:16:04.475>clear <00:16:04.650>because <00:16:04.825>you <00:16:05.000>cannot <00:16:05.175>label <00:16:05.350>the <00:16:05.525>whole,

337
00:16:08.185 --> 00:16:09.565
<00:16:08.185>uh, <00:16:08.598>internet <00:16:09.013>of <00:16:09.427>data.

338
00:16:10.345 --> 00:16:13.445
<00:16:10.345>Um, <00:16:10.485>so <00:16:10.625>how <00:16:10.765>do, <00:16:10.905>how <00:16:11.045>do <00:16:11.185>you, <00:16:11.325>how <00:16:11.465>does <00:16:11.605>it <00:16:11.745>work?

339
00:16:13.745 --> 00:16:14.885
<00:16:13.745>And, <00:16:14.001>uh, <00:16:14.258>as <00:16:14.514>I <00:16:14.771>mentioned

340
00:16:14.885 --> 00:16:17.685
<00:16:14.885>before, <00:16:15.060>in <00:16:15.235>the <00:16:15.410>basis <00:16:15.585>of <00:16:15.760>it, <00:16:15.935>we <00:16:16.110>have <00:16:16.285>the,

341
00:16:17.795 --> 00:16:19.565
<00:16:17.795>so-called <00:16:18.495>transform <00:16:19.194>architecture.

342
00:16:20.105 --> 00:16:21.845
<00:16:20.105>And <00:16:20.571>the <00:16:21.038>simplified, <00:16:21.505>uh,

343
00:16:22.045 --> 00:16:24.605
<00:16:22.045>transform <00:16:22.245>architecture <00:16:22.444>I'm <00:16:22.645>showing <00:16:22.844>you <00:16:23.045>on <00:16:23.245>this <00:16:23.444>slide,

344
00:16:24.745 --> 00:16:26.325
<00:16:24.745>uh, <00:16:25.025>actually <00:16:25.305>comes <00:16:25.585>from <00:16:25.865>the <00:16:26.145>paper.

345
00:16:26.595 --> 00:16:31.285
<00:16:26.595>Attention <00:16:26.735>is <00:16:26.875>all <00:16:27.015>you <00:16:27.155>need, <00:16:27.295>uh, <00:16:27.435>published <00:16:27.575>in, <00:16:27.715>uh, <00:16:27.855>2017, <00:16:27.995>um,

346
00:16:31.825 --> 00:16:33.205
<00:16:31.825>by <00:16:32.239>researchers <00:16:32.653>from <00:16:33.067>Google.

347
00:16:34.345 --> 00:16:37.485
<00:16:34.345>And, <00:16:34.545>uh, <00:16:34.745>it, <00:16:34.945>it <00:16:35.145>had <00:16:35.345>two <00:16:35.545>major <00:16:35.745>components,

348
00:16:37.485 --> 00:16:39.525
<00:16:37.485>which <00:16:37.835>are <00:16:38.185>encoder <00:16:38.535>and <00:16:38.885>decoder.

349
00:16:39.835 --> 00:16:43.325
<00:16:39.835>Encoder <00:16:40.035>was <00:16:40.235>designed <00:16:40.435>to <00:16:40.635>understand <00:16:40.835>the <00:16:41.035>input <00:16:41.235>text,

350
00:16:43.905 --> 00:16:47.325
<00:16:43.905>and <00:16:44.079>the <00:16:44.255>quota <00:16:44.430>was <00:16:44.605>designed <00:16:44.780>to <00:16:44.954>generate <00:16:45.130>new <00:16:45.305>text.

351
00:16:48.145 --> 00:16:51.605
<00:16:48.145>And, <00:16:48.345>uh, <00:16:48.545>it <00:16:48.745>relies <00:16:48.944>on <00:16:49.145>inputs, <00:16:49.345>uh, <00:16:49.545>of,

352
00:16:51.705 --> 00:16:53.245
<00:16:51.705>uh, <00:16:52.167>a <00:16:52.629>specific <00:16:53.091>format.

353
00:16:53.505 --> 00:16:55.485
<00:16:53.505>So <00:16:53.660>what <00:16:53.816>do <00:16:53.971>we, <00:16:54.127>what <00:16:54.282>do <00:16:54.438>we <00:16:54.593>need <00:16:54.749>to <00:16:54.905>input?

354
00:16:55.545 --> 00:16:58.165
<00:16:55.545>In <00:16:55.719>the <00:16:55.895>transformer <00:16:56.069>model, <00:16:56.245>you <00:16:56.420>know, <00:16:56.594>that <00:16:56.770>neural <00:16:56.944>networks,

355
00:16:58.165 --> 00:17:00.885
<00:16:58.165>they <00:16:58.339>work <00:16:58.515>with <00:16:58.689>numbers, <00:16:58.865>they <00:16:59.040>don't <00:16:59.214>work <00:16:59.390>with <00:16:59.564>text.

356
00:17:01.265 --> 00:17:05.045
<00:17:01.265>And <00:17:01.381>in <00:17:01.498>order <00:17:01.615>to <00:17:01.731>work <00:17:01.848>with <00:17:01.965>text, <00:17:02.081>we <00:17:02.198>need <00:17:02.314>to <00:17:02.431>process <00:17:02.548>it, <00:17:02.665>uh,

357
00:17:05.225 --> 00:17:07.125
<00:17:05.224>and <00:17:05.504>transform <00:17:05.784>it <00:17:06.064>into <00:17:06.344>some <00:17:06.625>numbers.

358
00:17:07.465 --> 00:17:10.565
<00:17:07.464>And <00:17:07.572>then <00:17:07.680>a <00:17:07.788>num, <00:17:07.895>there <00:17:08.003>is <00:17:08.111>a <00:17:08.218>number <00:17:08.326>of <00:17:08.434>stages <00:17:08.541>we <00:17:08.649>need <00:17:08.757>to <00:17:08.865>go

359
00:17:10.565 --> 00:17:14.245
<00:17:10.565>through <00:17:10.765>in <00:17:10.965>order <00:17:11.165>to <00:17:11.365>transform <00:17:11.565>text <00:17:11.765>into <00:17:11.965>the,

360
00:17:14.245 --> 00:17:17.125
<00:17:14.244>so-called <00:17:14.524>vector, <00:17:14.804>uh, <00:17:15.084>vector <00:17:15.364>of <00:17:15.645>embeddings.

361
00:17:17.855 --> 00:17:22.765
<00:17:17.855>Embeddings <00:17:18.055>is <00:17:18.255>basically, <00:17:18.454>um, <00:17:18.655>numeric <00:17:18.855>representation <00:17:19.055>of, <00:17:19.255>uh,

362
00:17:22.995 --> 00:17:24.045
<00:17:22.994>some <00:17:23.940>concepts.

363
00:17:24.665 --> 00:17:26.525
<00:17:24.665>And <00:17:24.898>the <00:17:25.131>power <00:17:25.365>of <00:17:25.598>embeddings <00:17:25.831>is <00:17:26.065>that

364
00:17:26.585 --> 00:17:31.205
<00:17:26.585>by <00:17:26.740>using <00:17:26.896>these <00:17:27.051>embeddings, <00:17:27.207>you <00:17:27.362>can <00:17:27.518>actually <00:17:27.673>map, <00:17:27.829>uh, <00:17:27.985>images.

365
00:17:31.465 --> 00:17:34.405
<00:17:31.465>You <00:17:31.639>can <00:17:31.814>map <00:17:31.990>text, <00:17:32.165>you <00:17:32.340>can <00:17:32.514>map <00:17:32.689>different <00:17:32.865>sources,

366
00:17:34.675 --> 00:17:38.445
<00:17:34.675>data <00:17:34.850>from <00:17:35.024>different <00:17:35.200>sources <00:17:35.375>to <00:17:35.550>the <00:17:35.725>same <00:17:35.899>conceptual <00:17:36.075>meaning.

367
00:17:38.625 --> 00:17:42.685
<00:17:38.625>So <00:17:38.858>they <00:17:39.091>reflect <00:17:39.325>the <00:17:39.558>semantic <00:17:39.791>meaning <00:17:40.025>of

368
00:17:43.275 --> 00:17:47.005
<00:17:43.275>some <00:17:43.508>concept <00:17:43.741>without <00:17:43.975>using <00:17:44.208>the <00:17:44.441>specific <00:17:44.675>word.

369
00:17:47.545 --> 00:17:50.565
<00:17:47.545>And <00:17:47.745>this <00:17:47.945>embeddings, <00:17:48.145>uh, <00:17:48.345>are <00:17:48.545>stored <00:17:48.745>in <00:17:48.945>the,

370
00:17:50.565 --> 00:17:51.805
<00:17:50.565>so-called <00:17:51.123>embedding <00:17:51.681>space.

371
00:17:52.225 --> 00:17:55.405
<00:17:52.225>And <00:17:52.399>if <00:17:52.574>you <00:17:52.750>dig <00:17:52.925>deeper <00:17:53.100>in <00:17:53.274>this <00:17:53.449>embedding <00:17:53.625>space,

372
00:17:55.785 --> 00:17:59.605
<00:17:55.785>you <00:17:56.065>find <00:17:56.345>like <00:17:56.625>very, <00:17:56.905>um, <00:17:57.185>similarities

373
00:17:59.605 --> 00:18:00.885
<00:17:59.605>between <00:18:00.181>different <00:18:00.757>words

374
00:18:01.025 --> 00:18:04.125
<00:18:01.025>and <00:18:01.258>concepts, <00:18:01.491>uh, <00:18:01.725>all <00:18:01.958>represented <00:18:02.191>in <00:18:02.425>numbers.

375
00:18:04.265 --> 00:18:06.325
<00:18:04.265>So <00:18:04.615>it's <00:18:04.965>a <00:18:05.315>multi-dimensional <00:18:05.665>vector.

376
00:18:07.265 --> 00:18:11.445
<00:18:07.265>Uh, <00:18:07.392>but <00:18:07.519>to <00:18:07.646>get <00:18:07.774>to <00:18:07.901>this <00:18:08.028>embeddings, <00:18:08.155>uh, <00:18:08.283>we <00:18:08.410>need <00:18:08.537>to <00:18:08.665>go

377
00:18:11.445 --> 00:18:14.805
<00:18:11.445>through <00:18:11.645>another <00:18:11.845>stage, <00:18:12.044>uh, <00:18:12.245>which <00:18:12.445>is <00:18:12.645>called <00:18:12.845>tokenization,

378
00:18:15.175 --> 00:18:17.565
<00:18:15.175>where <00:18:15.408>you <00:18:15.641>take <00:18:15.875>a <00:18:16.108>sequence <00:18:16.341>of <00:18:16.575>words

379
00:18:18.385 --> 00:18:20.605
<00:18:18.385>and <00:18:18.735>some <00:18:19.085>of <00:18:19.435>these <00:18:19.785>words

380
00:18:20.705 --> 00:18:22.445
<00:18:20.705>or <00:18:20.938>some <00:18:21.171>parts <00:18:21.405>of <00:18:21.638>these <00:18:21.871>words, <00:18:22.105>uh,

381
00:18:22.445 --> 00:18:23.765
<00:18:22.445>they're <00:18:22.841>represented <00:18:23.237>by <00:18:23.633>numbers.

382
00:18:24.345 --> 00:18:26.365
<00:18:24.345>And <00:18:24.578>these <00:18:24.811>numbers, <00:18:25.045>uh, <00:18:25.278>it's <00:18:25.511>very <00:18:25.745>simple.

383
00:18:26.395 --> 00:18:29.685
<00:18:26.395>They're <00:18:26.550>coming <00:18:26.706>from <00:18:26.861>some <00:18:27.017>dictionary, <00:18:27.172>so <00:18:27.328>it's <00:18:27.483>like <00:18:27.639>a <00:18:27.795>map.

384
00:18:29.865 --> 00:18:31.685
<00:18:29.865>Uh, <00:18:30.040>those <00:18:30.215>of <00:18:30.390>you <00:18:30.565>who <00:18:30.740>are <00:18:30.915>developers, <00:18:31.090>you <00:18:31.265>know

385
00:18:31.685 --> 00:18:34.685
<00:18:31.685>what <00:18:31.885>map <00:18:32.085>means, <00:18:32.284>where <00:18:32.485>you <00:18:32.685>have <00:18:32.885>some, <00:18:33.085>uh,

386
00:18:35.005 --> 00:18:37.885
<00:18:35.005>frequently <00:18:35.238>used <00:18:35.471>structures <00:18:35.705>or <00:18:35.938>frequently <00:18:36.171>used <00:18:36.405>words,

387
00:18:38.265 --> 00:18:41.325
<00:18:38.265>and <00:18:38.420>they're <00:18:38.576>basically <00:18:38.731>being <00:18:38.887>mapped <00:18:39.042>on <00:18:39.198>some <00:18:39.353>in <00:18:39.509>inte <00:18:39.665>number.

388
00:18:42.145 --> 00:18:46.485
<00:18:42.145>And <00:18:42.320>first, <00:18:42.495>if <00:18:42.670>you <00:18:42.845>have <00:18:43.020>your, <00:18:43.195>uh, <00:18:43.370>sentence, <00:18:43.545>you

389
00:18:47.085 --> 00:18:48.565
<00:18:47.085>transform <00:18:47.418>it <00:18:47.751>into <00:18:48.084>the <00:18:48.417>tokens.

390
00:18:48.785 --> 00:18:50.365
<00:18:48.785>And, <00:18:49.251>uh, <00:18:49.718>tokenization <00:18:50.185>works.

391
00:18:50.505 --> 00:18:52.645
<00:18:50.505>Uh, <00:18:50.738>you <00:18:50.971>can <00:18:51.205>either <00:18:51.438>engineer <00:18:51.671>the <00:18:51.905>step

392
00:18:53.105 --> 00:18:54.885
<00:18:53.105>or <00:18:53.338>you <00:18:53.571>can <00:18:53.805>also <00:18:54.038>learn <00:18:54.271>these <00:18:54.505>tokens.

393
00:18:55.425 --> 00:18:59.885
<00:18:55.425>And, <00:18:55.565>uh, <00:18:55.705>once <00:18:55.845>you've <00:18:55.985>tokenized <00:18:56.125>the <00:18:56.264>text, <00:18:56.405>uh, <00:18:56.544>you <00:18:56.685>use, <00:18:56.825>uh,

394
00:19:00.205 --> 00:19:02.725
<00:19:00.205>a <00:19:00.438>neural <00:19:00.671>network <00:19:00.905>to <00:19:01.138>get <00:19:01.371>token <00:19:01.605>embeddings.

395
00:19:02.725 --> 00:19:06.045
<00:19:02.725>So <00:19:02.958>embeddings <00:19:03.191>are <00:19:03.425>learned <00:19:03.658>from <00:19:03.891>these <00:19:04.125>tokens.

396
00:19:06.785 --> 00:19:11.485
<00:19:06.785>Um, <00:19:06.940>in <00:19:07.096>order <00:19:07.251>to <00:19:07.407>understand <00:19:07.562>also <00:19:07.718>the, <00:19:07.873>uh, <00:19:08.029>relations <00:19:08.185>between,

397
00:19:11.705 --> 00:19:14.685
<00:19:11.705>uh, <00:19:11.845>the, <00:19:11.985>the <00:19:12.125>semantic <00:19:12.264>structure <00:19:12.405>of <00:19:12.544>your <00:19:12.685>se <00:19:12.824>of <00:19:12.965>your <00:19:13.105>sentence,

398
00:19:14.865 --> 00:19:17.445
<00:19:14.865>we <00:19:15.098>also <00:19:15.331>need <00:19:15.565>the, <00:19:15.798>so-called <00:19:16.031>positional <00:19:16.265>embeddings.

399
00:19:18.105 --> 00:19:22.085
<00:19:18.105>And <00:19:18.280>for <00:19:18.455>that, <00:19:18.630>uh, <00:19:18.805>we <00:19:18.980>include, <00:19:19.155>uh, <00:19:19.330>our <00:19:19.505>positions

400
00:19:22.085 --> 00:19:25.365
<00:19:22.085>of <00:19:22.240>our <00:19:22.396>words <00:19:22.551>in <00:19:22.707>the <00:19:22.862>sentence <00:19:23.018>correspondent <00:19:23.173>to <00:19:23.329>the <00:19:23.485>tokens.

401
00:19:26.025 --> 00:19:29.245
<00:19:26.025>And <00:19:26.258>we <00:19:26.491>apply <00:19:26.725>some <00:19:26.958>mathematical <00:19:27.191>transformation <00:19:27.425>functions

402
00:19:29.245 --> 00:19:31.045
<00:19:29.245>using <00:19:29.594>sinus <00:19:29.945>and <00:19:30.294>co <00:19:30.645>sinus.

403
00:19:31.065 --> 00:19:33.725
<00:19:31.065>And <00:19:31.205>I <00:19:31.345>didn't <00:19:31.485>put <00:19:31.625>it, <00:19:31.765>uh, <00:19:31.905>on <00:19:32.045>the <00:19:32.185>slide <00:19:32.325>to, <00:19:32.465>not

404
00:19:33.725 --> 00:19:35.925
<00:19:33.725>to <00:19:33.899>intimidate <00:19:34.074>you, <00:19:34.250>but <00:19:34.425>you <00:19:34.600>can <00:19:34.774>look <00:19:34.949>it <00:19:35.125>up.

405
00:19:36.185 --> 00:19:39.925
<00:19:36.185>Uh, <00:19:36.385>and <00:19:36.585>you <00:19:36.784>create <00:19:36.985>the, <00:19:37.185>so-called <00:19:37.385>positional <00:19:37.585>encodings.

406
00:19:40.745 --> 00:19:42.845
<00:19:40.745>And <00:19:40.945>then <00:19:41.145>once <00:19:41.344>you <00:19:41.544>have <00:19:41.745>the <00:19:41.945>token <00:19:42.145>embeddings

407
00:19:42.845 --> 00:19:45.405
<00:19:42.845>and <00:19:43.078>positional <00:19:43.311>embeddings, <00:19:43.545>you <00:19:43.778>sum <00:19:44.011>them <00:19:44.245>up,

408
00:19:45.865 --> 00:19:48.045
<00:19:45.865>and <00:19:46.098>then <00:19:46.331>you <00:19:46.565>get <00:19:46.798>the <00:19:47.031>input <00:19:47.265>embeddings.

409
00:19:48.115 --> 00:19:52.165
<00:19:48.115>This <00:19:48.270>is <00:19:48.426>essentially <00:19:48.581>what <00:19:48.737>a <00:19:48.892>transformer <00:19:49.048>gets <00:19:49.203>as <00:19:49.359>an <00:19:49.515>input,

410
00:19:52.705 --> 00:19:56.085
<00:19:52.705>and <00:19:52.845>it <00:19:52.985>keeps <00:19:53.125>operating <00:19:53.264>in <00:19:53.405>the <00:19:53.544>embed <00:19:53.685>on <00:19:53.824>the <00:19:53.965>embedding <00:19:54.105>level.

411
00:19:56.705 --> 00:19:58.965
<00:19:56.705>So <00:19:56.985>now <00:19:57.264>it <00:19:57.544>looks <00:19:57.824>like <00:19:58.105>that.

412
00:19:59.145 --> 00:20:03.525
<00:19:59.145>So <00:19:59.285>we <00:19:59.425>have <00:19:59.565>our <00:19:59.705>encoder, <00:19:59.845>uh, <00:19:59.985>which, <00:20:00.125>uh, <00:20:00.264>gets <00:20:00.405>the <00:20:00.545>embeddings

413
00:20:03.525 --> 00:20:04.805
<00:20:03.525>and <00:20:04.101>positional <00:20:04.677>encoding.

414
00:20:05.505 --> 00:20:08.205
<00:20:05.505>Um, <00:20:05.738>and <00:20:05.971>it <00:20:06.205>comes <00:20:06.438>into <00:20:06.671>the <00:20:06.905>encoder.

415
00:20:08.835 --> 00:20:10.765
<00:20:08.835>Encoder <00:20:09.185>has <00:20:09.535>a <00:20:09.885>number <00:20:10.235>of,

416
00:20:11.145 --> 00:20:12.685
<00:20:11.145>and <00:20:11.376>both <00:20:11.607>actually <00:20:11.838>a <00:20:12.069>decoder <00:20:12.300>as <00:20:12.531>well.

417
00:20:12.875 --> 00:20:16.045
<00:20:12.875>Both <00:20:13.050>have <00:20:13.225>a <00:20:13.400>number <00:20:13.575>of, <00:20:13.750>uh, <00:20:13.925>layers <00:20:14.100>in <00:20:14.275>them.

418
00:20:16.905 --> 00:20:20.605
<00:20:16.905>And <00:20:17.138>what <00:20:17.371>makes <00:20:17.605>transformer <00:20:17.838>architecture <00:20:18.071>so <00:20:18.305>unique

419
00:20:20.705 --> 00:20:23.765
<00:20:20.705>and <00:20:20.879>so, <00:20:21.054>uh, <00:20:21.230>useful <00:20:21.405>for <00:20:21.580>gen <00:20:21.754>ai <00:20:21.929>are <00:20:22.105>the,

420
00:20:23.765 --> 00:20:25.085
<00:20:23.765>so-called <00:20:24.359>attention <00:20:24.953>layers.

421
00:20:25.775 --> 00:20:27.925
<00:20:25.775>These <00:20:26.055>attention <00:20:26.335>layers <00:20:26.615>allow <00:20:26.895>the <00:20:27.175>model

422
00:20:27.985 --> 00:20:31.045
<00:20:27.985>to <00:20:28.185>focus <00:20:28.385>on <00:20:28.584>specific <00:20:28.784>parts <00:20:28.985>of <00:20:29.185>the <00:20:29.385>input

423
00:20:31.225 --> 00:20:32.525
<00:20:31.225>by <00:20:31.615>assigning <00:20:32.005>different <00:20:32.395>weights

424
00:20:32.525 --> 00:20:34.565
<00:20:32.525>to <00:20:32.805>different <00:20:33.085>parts <00:20:33.365>of <00:20:33.645>the <00:20:33.925>input.

425
00:20:35.145 --> 00:20:37.765
<00:20:35.145>Uh, <00:20:35.378>if <00:20:35.611>you <00:20:35.845>compare <00:20:36.078>it <00:20:36.311>with, <00:20:36.545>uh,

426
00:20:38.145 --> 00:20:41.685
<00:20:38.145>the <00:20:38.300>recurrent <00:20:38.456>neural <00:20:38.611>networks <00:20:38.767>used, <00:20:38.922>uh, <00:20:39.078>well <00:20:39.233>in <00:20:39.389>the <00:20:39.545>past,

427
00:20:41.685 --> 00:20:42.805
<00:20:41.685>well, <00:20:41.937>they're <00:20:42.189>still <00:20:42.441>being <00:20:42.693>used,

428
00:20:42.905 --> 00:20:46.125
<00:20:42.905>but, <00:20:43.060>uh, <00:20:43.216>the, <00:20:43.371>the <00:20:43.527>mechanism <00:20:43.682>of <00:20:43.838>recurrent <00:20:43.993>neural <00:20:44.149>network <00:20:44.305>is

429
00:20:46.125 --> 00:20:49.045
<00:20:46.125>that <00:20:46.325>you <00:20:46.525>process <00:20:46.725>a <00:20:46.925>sentence <00:20:47.125>word <00:20:47.325>by <00:20:47.525>word,

430
00:20:49.465 --> 00:20:54.205
<00:20:49.465>and <00:20:49.620>you <00:20:49.776>kind <00:20:49.931>of <00:20:50.087>need <00:20:50.242>this <00:20:50.398>sequential <00:20:50.553>relationship <00:20:50.709>at, <00:20:50.865>uh,

431
00:20:54.235 --> 00:20:55.885
<00:20:54.235>with <00:20:54.701>the <00:20:55.168>transformer <00:20:55.635>architecture,

432
00:20:56.065 --> 00:20:57.885
<00:20:56.065>you <00:20:56.345>can <00:20:56.625>look <00:20:56.905>at <00:20:57.185>different <00:20:57.465>parts

433
00:20:57.905 --> 00:20:59.565
<00:20:57.905>of <00:20:58.371>the <00:20:58.838>sentence <00:20:59.305>simultaneously,

434
00:20:59.825 --> 00:21:02.565
<00:20:59.825>so <00:21:00.025>you <00:21:00.225>can <00:21:00.425>paralyze <00:21:00.625>this <00:21:00.825>process, <00:21:01.025>uh, <00:21:01.225>thanks

435
00:21:02.565 --> 00:21:03.845
<00:21:02.565>to <00:21:02.949>the <00:21:03.333>attention <00:21:03.717>mechanism.

436
00:21:04.185 --> 00:21:06.445
<00:21:04.185>And <00:21:04.360>if <00:21:04.534>you <00:21:04.710>can <00:21:04.885>paralyze <00:21:05.060>this <00:21:05.235>process, <00:21:05.409>it <00:21:05.585>means

437
00:21:06.445 --> 00:21:08.725
<00:21:06.445>that <00:21:06.585>you <00:21:06.725>can <00:21:06.865>also <00:21:07.004>run <00:21:07.145>it <00:21:07.284>very <00:21:07.425>well <00:21:07.564>on <00:21:07.705>the <00:21:07.845>GPUs.

438
00:21:08.905 --> 00:21:12.525
<00:21:08.905>So <00:21:09.080>transformers, <00:21:09.254>uh, <00:21:09.430>and <00:21:09.605>GPUs <00:21:09.780>are <00:21:09.955>really <00:21:10.129>good <00:21:10.305>friends.

439
00:21:13.465 --> 00:21:16.605
<00:21:13.465>And, <00:21:13.665>uh, <00:21:13.865>once <00:21:14.064>I <00:21:14.264>process <00:21:14.465>this <00:21:14.665>information, <00:21:14.865>uh,

440
00:21:17.145 --> 00:21:20.525
<00:21:17.145>the <00:21:17.300>input <00:21:17.456>information <00:21:17.611>with <00:21:17.767>the <00:21:17.922>encoder <00:21:18.078>you <00:21:18.233>put <00:21:18.389>into <00:21:18.545>theder,

441
00:21:20.545 --> 00:21:22.925
<00:21:20.545>and <00:21:20.745>the <00:21:20.945>coder <00:21:21.145>also <00:21:21.345>gets <00:21:21.545>the <00:21:21.745>input <00:21:21.945>from

442
00:21:22.925 --> 00:21:24.285
<00:21:22.925>the <00:21:23.537>previous <00:21:24.149>iteration.

443
00:21:24.905 --> 00:21:26.165
<00:21:24.905>Uh, <00:21:25.188>why <00:21:25.472>is <00:21:25.755>it <00:21:26.039>necessary?

444
00:21:26.505 --> 00:21:27.605
<00:21:26.505>Is <00:21:27.000>because, <00:21:27.495>um,

445
00:21:28.235 --> 00:21:31.605
<00:21:28.235>when <00:21:28.375>we <00:21:28.514>are <00:21:28.655>training <00:21:28.794>such <00:21:28.935>a <00:21:29.074>network <00:21:29.215>in <00:21:29.354>the <00:21:29.495>unsupervised <00:21:29.635>way,

446
00:21:32.225 --> 00:21:35.445
<00:21:32.225>uh, <00:21:32.425>we <00:21:32.625>want <00:21:32.824>to <00:21:33.024>compare <00:21:33.225>the, <00:21:33.425>our <00:21:33.625>prediction,

447
00:21:35.545 --> 00:21:38.125
<00:21:35.545>our <00:21:35.778>predicted <00:21:36.011>word, <00:21:36.245>uh, <00:21:36.478>with <00:21:36.711>the <00:21:36.945>word,

448
00:21:38.125 --> 00:21:39.645
<00:21:38.125>which <00:21:38.398>is <00:21:38.672>actually <00:21:38.945>in <00:21:39.219>the <00:21:39.493>sentence.

449
00:21:40.305 --> 00:21:43.885
<00:21:40.305>So <00:21:40.432>by <00:21:40.559>doing <00:21:40.686>that, <00:21:40.814>we, <00:21:40.941>we <00:21:41.068>have, <00:21:41.195>well, <00:21:41.323>we <00:21:41.450>have <00:21:41.577>this <00:21:41.705>kind

450
00:21:43.885 --> 00:21:45.565
<00:21:43.885>of <00:21:44.351>labeled, <00:21:44.818>uh, <00:21:45.285>automatically

451
00:21:45.565 --> 00:21:47.485
<00:21:45.565>because <00:21:45.765>we <00:21:45.965>know <00:21:46.165>from <00:21:46.365>the <00:21:46.565>text <00:21:46.765>we <00:21:46.965>have,

452
00:21:47.505 --> 00:21:48.565
<00:21:47.505>we <00:21:47.743>know <00:21:47.982>what <00:21:48.220>comes <00:21:48.459>next.

453
00:21:48.665 --> 00:21:51.365
<00:21:48.665>So <00:21:48.805>we <00:21:48.945>can <00:21:49.085>use <00:21:49.225>it <00:21:49.365>to <00:21:49.504>calculate <00:21:49.645>the, <00:21:49.784>uh, <00:21:49.925>the <00:21:50.065>objection,

454
00:21:51.545 --> 00:21:54.685
<00:21:51.545>the <00:21:51.825>objection <00:21:52.105>part <00:21:52.385>objective <00:21:52.665>function, <00:21:52.945>uh,

455
00:21:54.785 --> 00:21:56.445
<00:21:54.785>and <00:21:55.018>then <00:21:55.251>we <00:21:55.485>can <00:21:55.718>get <00:21:55.951>the <00:21:56.185>probabilities.

456
00:21:56.745 --> 00:22:01.165
<00:21:56.745>Um, <00:21:56.900>and <00:21:57.056>in <00:21:57.211>fact, <00:21:57.367>this <00:21:57.522>actually <00:21:57.678>looks, <00:21:57.833>uh, <00:21:57.989>like <00:21:58.145>this.

457
00:22:01.265 --> 00:22:04.445
<00:22:01.265>So <00:22:01.405>it's <00:22:01.545>not <00:22:01.685>just <00:22:01.825>a <00:22:01.965>single <00:22:02.105>encoder, <00:22:02.245>not <00:22:02.385>a <00:22:02.525>single <00:22:02.665>decoder,

458
00:22:04.445 --> 00:22:06.365
<00:22:04.445>but <00:22:04.678>a <00:22:04.911>stack <00:22:05.145>of <00:22:05.378>encoders <00:22:05.611>and <00:22:05.845>decoders.

459
00:22:06.665 --> 00:22:08.805
<00:22:06.665>And <00:22:06.898>now <00:22:07.131>you <00:22:07.365>understand <00:22:07.598>where <00:22:07.831>the <00:22:08.065>billions

460
00:22:08.805 --> 00:22:09.965
<00:22:08.805>of <00:22:09.153>parameters <00:22:09.501>come <00:22:09.849>from.

461
00:22:10.025 --> 00:22:11.965
<00:22:10.025>So <00:22:10.258>it's <00:22:10.491>a, <00:22:10.725>it's <00:22:10.958>a <00:22:11.191>massive <00:22:11.425>architecture.

462
00:22:12.705 --> 00:22:15.325
<00:22:12.705>And <00:22:12.860>in <00:22:13.016>the <00:22:13.171>end, <00:22:13.327>uh, <00:22:13.482>in <00:22:13.638>order <00:22:13.793>to <00:22:13.949>process, <00:22:14.105>uh,

463
00:22:15.325 --> 00:22:18.485
<00:22:15.325>the <00:22:15.480>final <00:22:15.636>output <00:22:15.791>output, <00:22:15.947>uh, <00:22:16.102>we <00:22:16.258>need <00:22:16.413>to <00:22:16.569>hand <00:22:16.725>the,

464
00:22:18.485 --> 00:22:19.725
<00:22:18.485>so-called <00:22:19.043>linear <00:22:19.601>layer,

465
00:22:19.825 --> 00:22:22.245
<00:22:19.825>and <00:22:20.025>then <00:22:20.225>the <00:22:20.425>soft <00:22:20.625>mac, <00:22:20.825>soft <00:22:21.025>max <00:22:21.225>layer

466
00:22:22.385 --> 00:22:24.285
<00:22:22.385>to <00:22:22.851>output <00:22:23.318>the <00:22:23.785>probabilities.

467
00:22:25.025 --> 00:22:26.285
<00:22:25.025>That's, <00:22:25.592>uh, <00:22:26.159>essentially

468
00:22:26.305 --> 00:22:29.125
<00:22:26.305>how <00:22:26.505>a <00:22:26.705>transformer <00:22:26.905>architecture <00:22:27.105>looks <00:22:27.305>like <00:22:27.505>these <00:22:27.705>days.

469
00:22:29.265 --> 00:22:31.325
<00:22:29.265>We <00:22:29.545>have <00:22:29.825>different <00:22:30.105>kinds <00:22:30.385>of <00:22:30.665>transformers.

470
00:22:31.325 --> 00:22:33.565
<00:22:31.325>There <00:22:31.675>are <00:22:32.025>encoder <00:22:32.375>only <00:22:32.725>transformers.

471
00:22:33.565 --> 00:22:35.485
<00:22:33.565>There <00:22:33.915>are <00:22:34.265>decoder <00:22:34.615>only <00:22:34.965>transformers.

472
00:22:35.745 --> 00:22:36.885
<00:22:35.745>So <00:22:36.001>we <00:22:36.258>have <00:22:36.514>like <00:22:36.771>B

473
00:22:36.985 --> 00:22:40.205
<00:22:36.985>and <00:22:37.218>GPT <00:22:37.451>transformers, <00:22:37.685>uh, <00:22:37.918>popular <00:22:38.151>these <00:22:38.385>days.

474
00:22:40.345 --> 00:22:44.005
<00:22:40.345>So, <00:22:40.500>um, <00:22:40.656>there, <00:22:40.811>there <00:22:40.967>are <00:22:41.122>different <00:22:41.278>ways <00:22:41.433>to <00:22:41.589>implement <00:22:41.745>those.

475
00:22:44.145 --> 00:22:45.845
<00:22:44.145>But <00:22:44.425>this, <00:22:44.705>uh, <00:22:44.985>this <00:22:45.264>is, <00:22:45.545>uh,

476
00:22:45.865 --> 00:22:48.405
<00:22:45.865>the <00:22:46.145>original <00:22:46.425>transformer <00:22:46.705>architecture <00:22:46.985>is <00:22:47.265>suggested

477
00:22:48.405 --> 00:22:49.405
<00:22:48.405>by <00:22:48.705>the <00:22:49.004>original <00:22:49.305>orders.

478
00:22:50.545 --> 00:22:52.685
<00:22:50.545>Um, <00:22:50.825>when <00:22:51.105>we <00:22:51.385>talk <00:22:51.665>about, <00:22:51.945>well,

479
00:22:52.685 --> 00:22:54.405
<00:22:52.685>and <00:22:52.918>transformers <00:22:53.151>are <00:22:53.385>used <00:22:53.618>for, <00:22:53.851>of <00:22:54.085>course,

480
00:22:54.465 --> 00:22:56.005
<00:22:54.465>for <00:22:55.158>text <00:22:55.851>processing,

481
00:22:56.345 --> 00:22:59.285
<00:22:56.345>and, <00:22:56.500>uh, <00:22:56.656>we <00:22:56.811>can <00:22:56.967>also <00:22:57.122>use <00:22:57.278>them <00:22:57.433>for <00:22:57.589>image <00:22:57.745>processing.

482
00:22:59.285 --> 00:23:01.285
<00:22:59.285>So <00:22:59.518>what <00:22:59.751>is <00:22:59.985>the <00:23:00.218>difference <00:23:00.451>between, <00:23:00.685>uh,

483
00:23:01.335 --> 00:23:04.525
<00:23:01.335>generative <00:23:01.568>language <00:23:01.801>models <00:23:02.035>versus <00:23:02.268>generative <00:23:02.501>image <00:23:02.735>models?

484
00:23:05.185 --> 00:23:07.925
<00:23:05.185>Uh, <00:23:05.418>when <00:23:05.651>talking <00:23:05.885>about <00:23:06.118>generative <00:23:06.351>language <00:23:06.585>models,

485
00:23:08.185 --> 00:23:11.725
<00:23:08.185>we <00:23:08.360>basically <00:23:08.534>learn <00:23:08.710>from <00:23:08.885>the <00:23:09.060>patterns <00:23:09.235>in <00:23:09.409>language, <00:23:09.585>uh,

486
00:23:11.725 --> 00:23:13.365
<00:23:11.725>through <00:23:12.074>all <00:23:12.425>this <00:23:12.774>data, <00:23:13.125>uh,

487
00:23:13.585 --> 00:23:17.885
<00:23:13.585>and, <00:23:13.712>uh, <00:23:13.839>with <00:23:13.966>some <00:23:14.094>text, <00:23:14.221>we <00:23:14.348>want <00:23:14.475>to <00:23:14.603>predict <00:23:14.730>what <00:23:14.857>comes <00:23:14.985>next.

488
00:23:17.985 --> 00:23:20.045
<00:23:17.985>As, <00:23:18.264>as <00:23:18.544>I've <00:23:18.824>already <00:23:19.104>mentioned, <00:23:19.385>um,

489
00:23:20.475 --> 00:23:24.045
<00:23:20.475>when <00:23:20.649>we <00:23:20.824>talk <00:23:21.000>about <00:23:21.175>generative <00:23:21.350>AI <00:23:21.524>for <00:23:21.699>images, <00:23:21.875>uh,

490
00:23:24.195 --> 00:23:27.085
<00:23:24.195>there's <00:23:24.395>another <00:23:24.595>technique <00:23:24.794>we <00:23:24.995>need <00:23:25.195>to <00:23:25.395>know <00:23:25.595>of,

491
00:23:27.225 --> 00:23:29.085
<00:23:27.225>and <00:23:27.504>it's <00:23:27.784>called, <00:23:28.064>uh, <00:23:28.344>diffusion <00:23:28.625>model.

492
00:23:29.185 --> 00:23:33.725
<00:23:29.185>So <00:23:29.325>given <00:23:29.465>some <00:23:29.605>text, <00:23:29.745>we <00:23:29.885>want <00:23:30.024>to <00:23:30.165>create <00:23:30.304>some <00:23:30.445>image, <00:23:30.585>uh,

493
00:23:33.955 --> 00:23:36.405
<00:23:33.955>some <00:23:34.188>artistic <00:23:34.421>quality <00:23:34.655>image <00:23:34.888>of <00:23:35.121>a <00:23:35.355>photograph.

494
00:23:37.185 --> 00:23:39.445
<00:23:37.185>And <00:23:37.340>for <00:23:37.496>that, <00:23:37.651>we <00:23:37.807>need <00:23:37.962>to <00:23:38.118>rely <00:23:38.273>on <00:23:38.429>the <00:23:38.585>technique,

495
00:23:39.445 --> 00:23:40.605
<00:23:39.445>which <00:23:39.793>is <00:23:40.141>called <00:23:40.489>diffusion.

496
00:23:41.025 --> 00:23:43.805
<00:23:41.025>And <00:23:41.180>in <00:23:41.336>the <00:23:41.491>rest <00:23:41.647>of <00:23:41.802>my <00:23:41.958>presentation, <00:23:42.113>I <00:23:42.269>will, <00:23:42.425>uh,

497
00:23:44.035 --> 00:23:46.605
<00:23:44.035>tell <00:23:44.235>you <00:23:44.435>about <00:23:44.635>how <00:23:44.835>this <00:23:45.035>diffusion <00:23:45.235>technique <00:23:45.435>works.

498
00:23:47.585 --> 00:23:50.525
<00:23:47.585>So <00:23:47.785>what <00:23:47.985>actually, <00:23:48.185>uh, <00:23:48.385>what <00:23:48.585>was <00:23:48.785>my <00:23:48.985>journey

499
00:23:50.635 --> 00:23:51.685
<00:23:50.635>with <00:23:51.107>diffusion <00:23:51.580>models?

500
00:23:51.945 --> 00:23:55.565
<00:23:51.945>Um, <00:23:52.120>I've, <00:23:52.294>I <00:23:52.470>learned <00:23:52.645>about <00:23:52.820>this <00:23:52.995>like <00:23:53.169>last <00:23:53.345>year,

501
00:23:55.865 --> 00:23:59.565
<00:23:55.865>and, <00:23:56.005>uh, <00:23:56.145>there <00:23:56.285>was <00:23:56.425>this <00:23:56.565>app <00:23:56.705>called <00:23:56.845>Lens <00:23:56.985>ai, <00:23:57.125>uh, <00:23:57.265>which

502
00:23:59.655 --> 00:24:02.885
<00:23:59.655>where <00:23:59.830>you <00:24:00.004>can <00:24:00.180>basically <00:24:00.355>upload <00:24:00.530>dozen <00:24:00.705>of <00:24:00.879>your <00:24:01.055>photographs,

503
00:24:03.065 --> 00:24:06.325
<00:24:03.065>and <00:24:03.265>then, <00:24:03.465>uh, <00:24:03.665>it'll <00:24:03.865>generate <00:24:04.065>beautiful <00:24:04.265>artistic <00:24:04.465>portraits

504
00:24:06.325 --> 00:24:08.405
<00:24:06.325>of <00:24:06.500>yourself, <00:24:06.675>and <00:24:06.850>you <00:24:07.025>need <00:24:07.200>to <00:24:07.375>pay <00:24:07.550>for <00:24:07.725>that.

505
00:24:08.625 --> 00:24:12.245
<00:24:08.625>Um, <00:24:08.780>I've, <00:24:08.936>I've <00:24:09.091>done <00:24:09.247>that, <00:24:09.402>and <00:24:09.558>I <00:24:09.713>was <00:24:09.869>so <00:24:10.025>fascinated.

506
00:24:12.365 --> 00:24:14.645
<00:24:12.365>I <00:24:12.520>wanted <00:24:12.676>to <00:24:12.831>learn <00:24:12.987>how <00:24:13.142>this <00:24:13.298>technology <00:24:13.453>works <00:24:13.609>in <00:24:13.765>order

507
00:24:14.645 --> 00:24:16.045
<00:24:14.645>to <00:24:14.802>be <00:24:14.960>able <00:24:15.117>to <00:24:15.275>do <00:24:15.432>it <00:24:15.590>on <00:24:15.747>my <00:24:15.905>own.

508
00:24:16.185 --> 00:24:20.205
<00:24:16.185>So, <00:24:16.301>um, <00:24:16.418>now <00:24:16.534>I <00:24:16.651>know <00:24:16.768>how <00:24:16.885>it <00:24:17.001>works, <00:24:17.118>and <00:24:17.235>that <00:24:17.351>was <00:24:17.468>my <00:24:17.585>motivation.

509
00:24:20.225 --> 00:24:22.925
<00:24:20.225>But <00:24:20.399>you <00:24:20.574>also <00:24:20.750>may <00:24:20.925>have <00:24:21.100>heard <00:24:21.274>of, <00:24:21.449>uh, <00:24:21.625>uh,

510
00:24:22.925 --> 00:24:25.005
<00:24:22.925>networks <00:24:23.274>like <00:24:23.625>Dali <00:24:23.975>by <00:24:24.325>OpenAI.

511
00:24:25.105 --> 00:24:27.165
<00:24:25.105>Uh, <00:24:25.455>it <00:24:25.805>was <00:24:26.155>somewhat <00:24:26.505>revolutionary

512
00:24:27.165 --> 00:24:31.525
<00:24:27.165>because, <00:24:27.305>uh, <00:24:27.445>they <00:24:27.585>made <00:24:27.725>it <00:24:27.865>so <00:24:28.004>convenient <00:24:28.145>to <00:24:28.284>use <00:24:28.425>for <00:24:28.565>everyone

513
00:24:32.065 --> 00:24:35.045
<00:24:32.065>by <00:24:32.298>publishing <00:24:32.531>the, <00:24:32.765>the, <00:24:32.998>on <00:24:33.231>the <00:24:33.465>website,

514
00:24:35.105 --> 00:24:38.085
<00:24:35.105>and <00:24:35.260>everybody <00:24:35.416>can <00:24:35.571>reach <00:24:35.727>it, <00:24:35.882>and <00:24:36.038>everybody <00:24:36.193>can <00:24:36.349>go <00:24:36.505>there

515
00:24:38.105 --> 00:24:40.085
<00:24:38.105>and <00:24:38.455>prompt, <00:24:38.805>uh, <00:24:39.155>the <00:24:39.505>network,

516
00:24:40.425 --> 00:24:43.325
<00:24:40.425>and <00:24:40.625>then <00:24:40.825>it'll <00:24:41.024>give <00:24:41.225>you <00:24:41.425>some <00:24:41.625>nice <00:24:41.825>images.

517
00:24:43.705 --> 00:24:46.685
<00:24:43.705>Um, <00:24:43.860>I <00:24:44.016>think <00:24:44.171>also <00:24:44.327>the <00:24:44.482>creators <00:24:44.638>of, <00:24:44.793>uh, <00:24:44.949>this <00:24:45.105>approach,

518
00:24:46.785 --> 00:24:49.845
<00:24:46.785>mid <00:24:46.925>journey, <00:24:47.065>uh, <00:24:47.205>they <00:24:47.345>took <00:24:47.485>it <00:24:47.625>to <00:24:47.765>the <00:24:47.905>very <00:24:48.045>new <00:24:48.185>level.

519
00:24:50.225 --> 00:24:54.365
<00:24:50.225>And <00:24:50.365>the, <00:24:50.504>the <00:24:50.645>art, <00:24:50.784>the <00:24:50.925>results, <00:24:51.064>uh, <00:24:51.205>the, <00:24:51.344>you <00:24:51.485>can <00:24:51.625>get

520
00:24:54.365 --> 00:24:58.045
<00:24:54.365>with <00:24:54.520>the <00:24:54.676>network <00:24:54.831>are <00:24:54.987>really <00:24:55.142>amazing, <00:24:55.298>like <00:24:55.453>very <00:24:55.609>fine <00:24:55.765>detail,

521
00:24:58.115 --> 00:25:01.285
<00:24:58.115>very <00:24:58.255>artistic, <00:24:58.395>and <00:24:58.535>I've <00:24:58.675>played <00:24:58.815>with <00:24:58.955>it <00:24:59.095>a <00:24:59.235>couple <00:24:59.375>of <00:24:59.515>times,

522
00:25:01.585 --> 00:25:03.285
<00:25:01.585>and, <00:25:01.865>uh, <00:25:02.145>it <00:25:02.425>sometimes <00:25:02.705>feels <00:25:02.985>like

523
00:25:03.345 --> 00:25:04.805
<00:25:03.345>as <00:25:03.564>if <00:25:03.783>it <00:25:04.002>was <00:25:04.221>reading <00:25:04.440>your <00:25:04.659>mind.

524
00:25:04.985 --> 00:25:06.685
<00:25:04.985>Uh, <00:25:05.334>it's, <00:25:05.685>uh, <00:25:06.034>it's <00:25:06.385>fascinating.

525
00:25:07.385 --> 00:25:10.045
<00:25:07.385>Um, <00:25:07.618>also <00:25:07.851>there <00:25:08.085>is <00:25:08.318>a <00:25:08.551>popular <00:25:08.785>architecture

526
00:25:10.045 --> 00:25:11.205
<00:25:10.045>called <00:25:10.567>Stable <00:25:11.089>Diffusion.

527
00:25:11.625 --> 00:25:13.605
<00:25:11.625>And, <00:25:11.975>uh, <00:25:12.325>its <00:25:12.675>creators, <00:25:13.025>uh,

528
00:25:13.715 --> 00:25:16.765
<00:25:13.715>they <00:25:13.995>also <00:25:14.274>revolutionized <00:25:14.554>the <00:25:14.834>field <00:25:15.115>by

529
00:25:17.455 --> 00:25:18.765
<00:25:17.455>publishing <00:25:18.044>their <00:25:18.634>model

530
00:25:18.825 --> 00:25:20.685
<00:25:18.825>and <00:25:19.025>the <00:25:19.225>weights, <00:25:19.425>uh, <00:25:19.625>on <00:25:19.825>the <00:25:20.025>hugging <00:25:20.225>face.

531
00:25:21.385 --> 00:25:25.405
<00:25:21.385>And <00:25:21.585>that <00:25:21.785>allowed <00:25:21.985>to, <00:25:22.185>like <00:25:22.385>many <00:25:22.585>researchers, <00:25:22.785>experiment

532
00:25:25.405 --> 00:25:26.805
<00:25:25.405>with <00:25:25.720>it, <00:25:26.035>train <00:25:26.350>it, <00:25:26.665>uh,

533
00:25:26.905 --> 00:25:29.925
<00:25:26.905>and <00:25:27.080>basically <00:25:27.254>everyone <00:25:27.430>can <00:25:27.605>train <00:25:27.780>the <00:25:27.955>diffusion <00:25:28.129>models <00:25:28.305>on

534
00:25:29.925 --> 00:25:31.605
<00:25:29.925>their <00:25:30.274>own, <00:25:30.625>having <00:25:30.975>their <00:25:31.325>models.

535
00:25:31.825 --> 00:25:34.725
<00:25:31.825>So, <00:25:31.965>um, <00:25:32.105>I've <00:25:32.245>played <00:25:32.385>with <00:25:32.525>it <00:25:32.665>also, <00:25:32.805>of <00:25:32.945>course, <00:25:33.085>a <00:25:33.225>lot.

536
00:25:35.025 --> 00:25:36.165
<00:25:35.025>And <00:25:35.367>the <00:25:35.709>question <00:25:36.051>is,

537
00:25:36.345 --> 00:25:37.885
<00:25:36.345>can <00:25:36.576>you <00:25:36.807>actually <00:25:37.038>fine <00:25:37.269>tune <00:25:37.500>these <00:25:37.731>models?

538
00:25:38.065 --> 00:25:39.285
<00:25:38.065>Can <00:25:38.339>you <00:25:38.614>modify <00:25:38.888>these <00:25:39.163>models?

539
00:25:39.505 --> 00:25:42.605
<00:25:39.505>Uh, <00:25:39.705>how <00:25:39.905>can <00:25:40.105>you <00:25:40.305>create, <00:25:40.505>uh, <00:25:40.705>objects, <00:25:40.905>uh,

540
00:25:42.745 --> 00:25:44.685
<00:25:42.745>or <00:25:42.978>can <00:25:43.211>you <00:25:43.445>do <00:25:43.678>a <00:25:43.911>style <00:25:44.145>transfer?

541
00:25:44.825 --> 00:25:46.165
<00:25:44.825>And <00:25:45.066>the <00:25:45.307>the <00:25:45.548>answer <00:25:45.789>is <00:25:46.031>yes.

542
00:25:46.265 --> 00:25:48.645
<00:25:46.265>So <00:25:46.498>you <00:25:46.731>can <00:25:46.965>either <00:25:47.198>train <00:25:47.431>them, <00:25:47.665>uh,

543
00:25:48.705 --> 00:25:51.845
<00:25:48.705>to <00:25:48.938>incorporate <00:25:49.171>some <00:25:49.405>specific <00:25:49.638>object <00:25:49.871>or <00:25:50.105>subject.

544
00:25:52.185 --> 00:25:56.005
<00:25:52.185>Uh, <00:25:52.312>so <00:25:52.439>if <00:25:52.566>you <00:25:52.694>want <00:25:52.821>a <00:25:52.948>person, <00:25:53.075>specific <00:25:53.203>person <00:25:53.330>to <00:25:53.457>be <00:25:53.585>there,

545
00:25:56.105 --> 00:25:58.885
<00:25:56.105>you <00:25:56.260>just <00:25:56.416>train, <00:25:56.571>you <00:25:56.727>know, <00:25:56.882>in <00:25:57.038>a <00:25:57.193>couple <00:25:57.349>of <00:25:57.505>portraits.

546
00:25:59.145 --> 00:26:02.085
<00:25:59.145>Or <00:25:59.285>if <00:25:59.425>you <00:25:59.565>want, <00:25:59.705>if <00:25:59.845>you're, <00:25:59.985>for <00:26:00.125>example, <00:26:00.264>in <00:26:00.405>the <00:26:00.545>marketing

547
00:26:02.085 --> 00:26:04.685
<00:26:02.085>and <00:26:02.260>creating <00:26:02.435>a <00:26:02.610>marketing <00:26:02.785>campaign, <00:26:02.960>you <00:26:03.135>take <00:26:03.310>an <00:26:03.485>object

548
00:26:04.905 --> 00:26:08.885
<00:26:04.905>and, <00:26:05.060>uh, <00:26:05.216>uh, <00:26:05.371>also <00:26:05.527>generate <00:26:05.682>various <00:26:05.838>scenes <00:26:05.993>with <00:26:06.149>this <00:26:06.305>object.

549
00:26:09.145 --> 00:26:13.045
<00:26:09.145>Or <00:26:09.285>if <00:26:09.425>you <00:26:09.565>want, <00:26:09.705>if <00:26:09.845>you <00:26:09.985>like <00:26:10.125>some, <00:26:10.264>uh, <00:26:10.405>particular <00:26:10.545>style

550
00:26:13.385 --> 00:26:16.645
<00:26:13.385>and <00:26:13.501>want <00:26:13.618>to, <00:26:13.735>to <00:26:13.851>apply <00:26:13.968>it <00:26:14.085>on <00:26:14.201>your <00:26:14.318>images, <00:26:14.435>you <00:26:14.551>can <00:26:14.668>also <00:26:14.785>use,

551
00:26:16.745 --> 00:26:19.125
<00:26:16.745>uh, <00:26:17.024>uh, <00:26:17.304>diffusion <00:26:17.584>models <00:26:17.864>for <00:26:18.145>that.

552
00:26:20.255 --> 00:26:22.755
<00:26:20.255>Uh, <00:26:20.535>so <00:26:20.815>I've <00:26:21.095>run <00:26:21.375>some, <00:26:21.655>uh,

553
00:26:22.755 --> 00:26:24.555
<00:26:22.755>fine <00:26:23.105>tuning <00:26:23.455>experiment, <00:26:23.805>uh, <00:26:24.155>myself.

554
00:26:24.855 --> 00:26:26.275
<00:26:24.855>And, <00:26:25.494>um, <00:26:26.133>I,

555
00:26:26.435 --> 00:26:28.875
<00:26:26.435>I <00:26:26.575>grew <00:26:26.715>up <00:26:26.855>watching <00:26:26.995>Lion <00:26:27.135>King <00:26:27.274>movie, <00:26:27.415>and <00:26:27.554>I <00:26:27.695>love <00:26:27.835>it.

556
00:26:29.495 --> 00:26:31.395
<00:26:29.495>And, <00:26:29.844>uh, <00:26:30.195>I <00:26:30.544>went <00:26:30.895>online

557
00:26:31.775 --> 00:26:35.515
<00:26:31.775>and <00:26:31.975>just <00:26:32.175>browsed <00:26:32.375>for <00:26:32.575>images <00:26:32.775>from <00:26:32.975>Lion <00:26:33.175>King,

558
00:26:35.615 --> 00:26:38.675
<00:26:35.615>and <00:26:35.790>this <00:26:35.965>was <00:26:36.140>the <00:26:36.315>entire <00:26:36.490>data <00:26:36.665>set <00:26:36.840>I've <00:26:37.015>used

559
00:26:38.695 --> 00:26:40.115
<00:26:38.695>for <00:26:39.120>my <00:26:39.547>frontier <00:26:39.973>experiment.

560
00:26:40.455 --> 00:26:45.125
<00:26:40.455>And <00:26:40.595>I <00:26:40.735>trained <00:26:40.875>a <00:26:41.014>model <00:26:41.155>called, <00:26:41.294>uh, <00:26:41.435>textual <00:26:41.574>inversion, <00:26:41.715>uh, <00:26:41.855>on,

561
00:26:45.215 --> 00:26:46.965
<00:26:45.215>based <00:26:45.495>on <00:26:45.774>top <00:26:46.054>of <00:26:46.334>stable <00:26:46.615>diffusion

562
00:26:47.745 --> 00:26:50.525
<00:26:47.745>to <00:26:47.945>create <00:26:48.145>images <00:26:48.344>in <00:26:48.544>the <00:26:48.745>style <00:26:48.945>of <00:26:49.145>lion.

563
00:26:51.185 --> 00:26:54.365
<00:26:51.185>And <00:26:51.360>that <00:26:51.534>is <00:26:51.710>the <00:26:51.885>result <00:26:52.060>I've <00:26:52.235>basically <00:26:52.409>prompted, <00:26:52.585>uh,

564
00:26:54.465 --> 00:26:56.965
<00:26:54.465>the <00:26:54.620>model <00:26:54.776>to <00:26:54.931>give <00:26:55.087>me <00:26:55.242>an <00:26:55.398>image, <00:26:55.553>give <00:26:55.709>me <00:26:55.865>images

565
00:26:57.025 --> 00:26:58.205
<00:26:57.025>of <00:26:57.556>pink <00:26:58.087>panter.

566
00:26:58.345 --> 00:27:01.645
<00:26:58.345>And, <00:26:58.520>uh, <00:26:58.695>uh, <00:26:58.870>it <00:26:59.045>was, <00:26:59.220>it <00:26:59.395>was <00:26:59.570>really <00:26:59.745>fascinating.

567
00:27:01.785 --> 00:27:06.045
<00:27:01.785>So <00:27:01.912>I've <00:27:02.039>trained <00:27:02.166>this <00:27:02.294>model <00:27:02.421>for <00:27:02.548>just <00:27:02.675>one <00:27:02.803>hour <00:27:02.930>on <00:27:03.057>my, <00:27:03.185>uh,

568
00:27:06.045 --> 00:27:09.645
<00:27:06.045>single <00:27:06.220>A <00:27:06.395>6,000 <00:27:06.570>GPU, <00:27:06.745>and <00:27:06.920>that's <00:27:07.095>what <00:27:07.270>I <00:27:07.445>got.

569
00:27:09.825 --> 00:27:12.005
<00:27:09.825>Uh, <00:27:10.000>of <00:27:10.175>course, <00:27:10.350>some <00:27:10.525>of <00:27:10.700>them <00:27:10.875>are <00:27:11.050>funny, <00:27:11.225>some

570
00:27:12.005 --> 00:27:14.445
<00:27:12.005>of <00:27:12.205>them <00:27:12.405>have <00:27:12.605>five <00:27:12.805>legs <00:27:13.005>or <00:27:13.205>funny <00:27:13.405>tails.

571
00:27:14.545 --> 00:27:17.365
<00:27:14.545>But, <00:27:14.778>uh, <00:27:15.011>this <00:27:15.245>style <00:27:15.478>is <00:27:15.711>clearly <00:27:15.945>there.

572
00:27:17.505 --> 00:27:19.525
<00:27:17.505>So <00:27:17.855>it's, <00:27:18.205>it's <00:27:18.555>really <00:27:18.905>impressive.

573
00:27:19.745 --> 00:27:22.605
<00:27:19.745>Uh, <00:27:19.919>with <00:27:20.094>longer <00:27:20.270>training, <00:27:20.445>with <00:27:20.620>a <00:27:20.794>bigger <00:27:20.969>data <00:27:21.145>set,

574
00:27:22.605 --> 00:27:24.165
<00:27:22.605>you <00:27:22.885>can <00:27:23.165>achieve <00:27:23.445>even <00:27:23.725>better <00:27:24.005>results.

575
00:27:24.985 --> 00:27:29.285
<00:27:24.985>Um, <00:27:25.140>the <00:27:25.296>history <00:27:25.451>of, <00:27:25.607>uh, <00:27:25.762>generative <00:27:25.918>AI <00:27:26.073>in <00:27:26.229>images, <00:27:26.385>uh,

576
00:27:29.285 --> 00:27:31.405
<00:27:29.285>dates <00:27:29.635>back <00:27:29.985>to <00:27:30.335>2018, <00:27:30.685>uh,

577
00:27:31.945 --> 00:27:36.885
<00:27:31.945>and, <00:27:32.120>uh, <00:27:32.294>the <00:27:32.470>CVPR <00:27:32.645>in <00:27:32.820>2018, <00:27:32.995>Nvidia <00:27:33.169>has <00:27:33.345>introduced,

578
00:27:37.145 --> 00:27:38.525
<00:27:37.145>uh, <00:27:37.559>the <00:27:37.973>style <00:27:38.387>gun.

579
00:27:39.205 --> 00:27:40.725
<00:27:39.205>I <00:27:39.433>wasn't <00:27:39.660>working <00:27:39.889>at <00:27:40.117>Nvidia <00:27:40.344>back <00:27:40.572>then,

580
00:27:40.725 --> 00:27:42.885
<00:27:40.725>but <00:27:40.925>I <00:27:41.125>was <00:27:41.324>very <00:27:41.524>fascinated <00:27:41.725>by <00:27:41.925>this <00:27:42.125>approach.

581
00:27:42.985 --> 00:27:47.645
<00:27:42.985>Uh, <00:27:43.112>with <00:27:43.239>style <00:27:43.366>GaN, <00:27:43.494>uh, <00:27:43.621>you <00:27:43.748>were <00:27:43.875>able <00:27:44.003>to <00:27:44.130>generate, <00:27:44.257>uh, <00:27:44.385>um,

582
00:27:47.915 --> 00:27:51.325
<00:27:47.915>naturally <00:27:48.115>looking <00:27:48.315>portraits <00:27:48.514>of <00:27:48.715>people <00:27:48.915>who <00:27:49.115>never <00:27:49.315>existed.

583
00:27:51.625 --> 00:27:53.285
<00:27:51.625>Uh, <00:27:51.858>there <00:27:52.091>was <00:27:52.325>already <00:27:52.558>a <00:27:52.791>big <00:27:53.025>breakthrough.

584
00:27:53.785 --> 00:27:56.885
<00:27:53.785>And <00:27:54.018>then <00:27:54.251>they've <00:27:54.485>introduced <00:27:54.718>Gagan <00:27:54.951>for <00:27:55.185>landscapes.

585
00:27:56.915 --> 00:27:59.245
<00:27:56.915>They've <00:27:57.195>improved <00:27:57.475>the <00:27:57.754>initial <00:27:58.034>style <00:27:58.315>again.

586
00:28:00.025 --> 00:28:02.085
<00:28:00.025>And, <00:28:00.491>uh, <00:28:00.958>last <00:28:01.425>year,

587
00:28:02.085 --> 00:28:04.605
<00:28:02.085>you <00:28:02.285>remember <00:28:02.485>the <00:28:02.685>approaches <00:28:02.885>like <00:28:03.085>Dolly <00:28:03.285>have <00:28:03.485>appeared.

588
00:28:04.635 --> 00:28:05.925
<00:28:04.635>Then <00:28:04.925>we <00:28:05.215>also <00:28:05.505>had <00:28:05.796>Nerf.

589
00:28:06.025 --> 00:28:08.125
<00:28:06.025>And <00:28:06.305>Nerf <00:28:06.585>is <00:28:06.865>the <00:28:07.145>generative <00:28:07.425>AI

590
00:28:08.145 --> 00:28:10.125
<00:28:08.145>for <00:28:08.425>three <00:28:08.705>D <00:28:08.985>reconstruction <00:28:09.264>from <00:28:09.545>images.

591
00:28:11.105 --> 00:28:15.805
<00:28:11.105>Um, <00:28:11.305>and, <00:28:11.505>uh, <00:28:11.705>now <00:28:11.905>we <00:28:12.105>facing <00:28:12.305>the <00:28:12.505>emergence

592
00:28:15.905 --> 00:28:18.485
<00:28:15.905>of <00:28:16.138>new <00:28:16.371>approaches <00:28:16.605>like <00:28:16.838>for <00:28:17.071>video <00:28:17.305>generation

593
00:28:18.505 --> 00:28:21.445
<00:28:18.505>for <00:28:18.660>three <00:28:18.816>D <00:28:18.971>generation <00:28:19.127>with <00:28:19.282>approaches <00:28:19.438>like <00:28:19.593>Magic <00:28:19.749>three <00:28:19.905>D

594
00:28:21.445 --> 00:28:23.645
<00:28:21.445>or <00:28:21.911>edify <00:28:22.378>edify <00:28:22.845>video.

595
00:28:24.105 --> 00:28:27.485
<00:28:24.105>So, <00:28:24.280>uh, <00:28:24.455>the, <00:28:24.630>yeah, <00:28:24.805>the <00:28:24.980>progress, <00:28:25.155>the <00:28:25.330>progress <00:28:25.505>is,

596
00:28:27.505 --> 00:28:28.645
<00:28:27.505>uh, <00:28:27.847>really <00:28:28.189>mind <00:28:28.531>blowing.

597
00:28:29.505 --> 00:28:33.285
<00:28:29.505>And <00:28:29.660>now <00:28:29.816>is <00:28:29.971>the <00:28:30.127>question, <00:28:30.282>how <00:28:30.438>does, <00:28:30.593>uh, <00:28:30.749>this, <00:28:30.905>uh,

598
00:28:33.285 --> 00:28:35.685
<00:28:33.285>text <00:28:33.635>image <00:28:33.985>generation <00:28:34.335>actually <00:28:34.685>work?

599
00:28:36.105 --> 00:28:39.005
<00:28:36.105>And <00:28:36.260>the <00:28:36.416>thing <00:28:36.571>is, <00:28:36.727>uh, <00:28:36.882>it's <00:28:37.038>not <00:28:37.193>a <00:28:37.349>single <00:28:37.505>model,

600
00:28:39.275 --> 00:28:40.765
<00:28:39.275>it's <00:28:39.610>a <00:28:39.945>pipeline <00:28:40.280>of <00:28:40.616>models.

601
00:28:40.905 --> 00:28:43.685
<00:28:40.905>So <00:28:41.105>we <00:28:41.305>usually <00:28:41.504>have, <00:28:41.705>uh, <00:28:41.905>three <00:28:42.105>main <00:28:42.305>components.

602
00:28:44.345 --> 00:28:47.645
<00:28:44.345>The <00:28:44.545>first <00:28:44.745>component <00:28:44.945>is, <00:28:45.145>uh, <00:28:45.345>the, <00:28:45.545>so-called, <00:28:45.745>uh,

603
00:28:47.955 --> 00:28:50.445
<00:28:47.955>clip <00:28:48.235>text <00:28:48.514>encoder <00:28:48.794>and <00:28:49.074>image <00:28:49.355>encoder.

604
00:28:50.745 --> 00:28:53.565
<00:28:50.745>And, <00:28:50.945>uh, <00:28:51.145>this <00:28:51.344>is <00:28:51.544>our <00:28:51.745>large <00:28:51.945>language <00:28:52.145>model.

605
00:28:53.755 --> 00:28:56.165
<00:28:53.755>This <00:28:53.955>is <00:28:54.155>the, <00:28:54.355>the <00:28:54.555>biggest <00:28:54.755>part, <00:28:54.955>uh, <00:28:55.155>of,

606
00:28:56.225 --> 00:28:57.245
<00:28:56.225>uh, <00:28:56.531>this <00:28:56.837>whole <00:28:57.142>approach.

607
00:28:57.705 --> 00:29:01.365
<00:28:57.705>If <00:28:57.832>you <00:28:57.959>want <00:28:58.086>to <00:28:58.214>have <00:28:58.341>the <00:28:58.468>full <00:28:58.595>control <00:28:58.723>on, <00:28:58.850>uh, <00:28:58.977>the <00:28:59.105>generation,

608
00:29:01.665 --> 00:29:05.045
<00:29:01.665>you <00:29:01.898>need <00:29:02.131>to <00:29:02.365>understand <00:29:02.598>natural <00:29:02.831>language, <00:29:03.065>uh,

609
00:29:05.045 --> 00:29:06.965
<00:29:05.045>which <00:29:05.278>is <00:29:05.511>being <00:29:05.745>entered <00:29:05.978>by <00:29:06.211>the <00:29:06.445>user,

610
00:29:07.425 --> 00:29:09.085
<00:29:07.425>and <00:29:07.625>then <00:29:07.825>you <00:29:08.024>need <00:29:08.225>to <00:29:08.425>take <00:29:08.625>this <00:29:08.825>prompt

611
00:29:09.305 --> 00:29:11.845
<00:29:09.305>and <00:29:09.480>you <00:29:09.655>need <00:29:09.830>to <00:29:10.005>pass <00:29:10.180>it <00:29:10.355>to <00:29:10.530>your <00:29:10.705>generator.

612
00:29:12.625 --> 00:29:14.565
<00:29:12.625>And <00:29:12.905>then <00:29:13.185>the <00:29:13.465>generator <00:29:13.745>part <00:29:14.025>consists

613
00:29:14.565 --> 00:29:16.925
<00:29:14.565>of <00:29:14.765>two <00:29:14.965>other <00:29:15.165>components, <00:29:15.365>which <00:29:15.565>are <00:29:15.765>auto <00:29:15.965>encoder

614
00:29:17.265 --> 00:29:18.605
<00:29:17.265>and <00:29:17.667>the <00:29:18.069>unit <00:29:18.471>architecture.

615
00:29:18.825 --> 00:29:21.165
<00:29:18.825>And, <00:29:18.965>uh, <00:29:19.105>let <00:29:19.245>me <00:29:19.385>go <00:29:19.525>through <00:29:19.665>them, <00:29:19.805>uh, <00:29:19.945>one <00:29:20.085>by <00:29:20.225>one.

616
00:29:21.785 --> 00:29:25.005
<00:29:21.785>So <00:29:21.940>those <00:29:22.096>of <00:29:22.251>you <00:29:22.407>who've <00:29:22.562>been <00:29:22.718>in <00:29:22.873>data <00:29:23.029>science, <00:29:23.185>uh,

617
00:29:25.195 --> 00:29:28.245
<00:29:25.195>long <00:29:25.370>enough, <00:29:25.544>uh, <00:29:25.720>you <00:29:25.895>all <00:29:26.070>know <00:29:26.245>the <00:29:26.419>architecture <00:29:26.595>unit,

618
00:29:28.875 --> 00:29:32.285
<00:29:28.875>it's <00:29:29.050>been <00:29:29.225>super <00:29:29.400>useful <00:29:29.575>for <00:29:29.750>the <00:29:29.925>tasks <00:29:30.100>like <00:29:30.275>segmentation.

619
00:29:32.585 --> 00:29:34.685
<00:29:32.585>Uh, <00:29:32.785>when <00:29:32.985>I <00:29:33.185>was <00:29:33.385>working <00:29:33.585>on <00:29:33.785>medical <00:29:33.985>imaging,

620
00:29:34.845 --> 00:29:36.885
<00:29:34.845>I <00:29:35.045>was <00:29:35.245>using <00:29:35.445>it <00:29:35.645>a <00:29:35.845>lot <00:29:36.045>for, <00:29:36.245>uh,

621
00:29:36.885 --> 00:29:39.285
<00:29:36.885>digital <00:29:37.235>pathology <00:29:37.585>segmentation, <00:29:37.935>for <00:29:38.285>example.

622
00:29:39.665 --> 00:29:41.045
<00:29:39.665>And <00:29:39.975>typically <00:29:40.286>how <00:29:40.596>it <00:29:40.907>works,

623
00:29:41.305 --> 00:29:44.285
<00:29:41.305>you <00:29:41.538>have <00:29:41.771>this <00:29:42.005>cascaded <00:29:42.238>architecture <00:29:42.471>where, <00:29:42.705>uh,

624
00:29:44.505 --> 00:29:47.645
<00:29:44.505>you <00:29:44.680>take <00:29:44.855>your <00:29:45.030>original <00:29:45.205>image, <00:29:45.380>you <00:29:45.555>scale <00:29:45.730>it <00:29:45.905>down,

625
00:29:48.145 --> 00:29:49.405
<00:29:48.145>and <00:29:48.428>then <00:29:48.712>scale <00:29:48.995>it <00:29:49.279>up.

626
00:29:49.825 --> 00:29:51.325
<00:29:49.825>And <00:29:50.162>then <00:29:50.500>between <00:29:50.837>this <00:29:51.175>up

627
00:29:51.325 --> 00:29:54.605
<00:29:51.325>and <00:29:51.500>down <00:29:51.675>branches, <00:29:51.850>you <00:29:52.025>have <00:29:52.200>the <00:29:52.375>so-called <00:29:52.550>Skip <00:29:52.725>connections

628
00:29:54.605 --> 00:29:58.005
<00:29:54.605>to <00:29:54.780>preserve <00:29:54.955>the <00:29:55.130>information <00:29:55.305>from, <00:29:55.480>uh, <00:29:55.655>the <00:29:55.830>previous <00:29:56.005>layers

629
00:29:58.505 --> 00:30:00.885
<00:29:58.505>and, <00:29:58.705>and <00:29:58.905>pass <00:29:59.105>them <00:29:59.305>forward <00:29:59.505>in <00:29:59.705>the <00:29:59.905>network.

630
00:30:01.505 --> 00:30:03.605
<00:30:01.505>Uh, <00:30:01.705>so <00:30:01.905>this <00:30:02.105>architecture <00:30:02.305>turned <00:30:02.505>to, <00:30:02.705>turned <00:30:02.905>out

631
00:30:03.605 --> 00:30:07.605
<00:30:03.605>to <00:30:03.780>be <00:30:03.955>very <00:30:04.130>robust <00:30:04.305>for <00:30:04.480>the <00:30:04.655>assigning <00:30:04.830>labels, <00:30:05.005>uh,

632
00:30:07.825 --> 00:30:09.285
<00:30:07.825>to <00:30:09.139>pixels.

633
00:30:10.065 --> 00:30:13.565
<00:30:10.065>And, <00:30:10.220>uh, <00:30:10.376>in <00:30:10.531>the <00:30:10.687>input <00:30:10.842>you <00:30:10.998>would <00:30:11.153>typically <00:30:11.309>have, <00:30:11.465>uh,

634
00:30:13.565 --> 00:30:15.685
<00:30:13.565>your <00:30:13.798>image, <00:30:14.031>and <00:30:14.265>then <00:30:14.498>also <00:30:14.731>some <00:30:14.965>mask.

635
00:30:15.825 --> 00:30:19.365
<00:30:15.825>And <00:30:15.980>you <00:30:16.136>would <00:30:16.291>then <00:30:16.447>learn <00:30:16.602>how <00:30:16.758>to <00:30:16.913>predict <00:30:17.069>this <00:30:17.225>mask

636
00:30:19.465 --> 00:30:20.605
<00:30:19.465>and <00:30:19.721>then <00:30:19.978>see <00:30:20.234>an <00:30:20.491>image.

637
00:30:21.135 --> 00:30:23.805
<00:30:21.135>Turns <00:30:21.310>out <00:30:21.485>that <00:30:21.660>this <00:30:21.835>network <00:30:22.010>is <00:30:22.185>also <00:30:22.360>very <00:30:22.535>good

638
00:30:23.805 --> 00:30:24.965
<00:30:23.805>at <00:30:24.327>predicting <00:30:24.849>noise.

639
00:30:25.785 --> 00:30:27.365
<00:30:25.785>And, <00:30:26.065>uh, <00:30:26.345>if <00:30:26.625>you, <00:30:26.905>for <00:30:27.185>example,

640
00:30:27.455 --> 00:30:29.005
<00:30:27.455>enter <00:30:27.734>an <00:30:28.013>image <00:30:28.292>in <00:30:28.571>this <00:30:28.850>network,

641
00:30:29.505 --> 00:30:33.365
<00:30:29.505>and, <00:30:29.645>uh, <00:30:29.785>as <00:30:29.925>a <00:30:30.065>mask, <00:30:30.205>uh, <00:30:30.345>you <00:30:30.485>generate <00:30:30.625>some <00:30:30.765>random <00:30:30.905>noise,

642
00:30:33.905 --> 00:30:36.965
<00:30:33.905>you <00:30:34.060>can <00:30:34.216>then <00:30:34.371>predict, <00:30:34.527>you <00:30:34.682>can <00:30:34.838>then <00:30:34.993>use <00:30:35.149>this <00:30:35.305>architecture

643
00:30:36.965 --> 00:30:38.845
<00:30:36.965>to <00:30:37.165>learn <00:30:37.365>this <00:30:37.564>noise <00:30:37.764>and <00:30:37.965>predict <00:30:38.165>this <00:30:38.365>noise.

644
00:30:39.625 --> 00:30:42.565
<00:30:39.625>Now, <00:30:39.780>you <00:30:39.936>may <00:30:40.091>wonder <00:30:40.247>why <00:30:40.402>we <00:30:40.558>are <00:30:40.713>talking <00:30:40.869>about <00:30:41.025>predicting

645
00:30:42.565 --> 00:30:44.485
<00:30:42.565>noise <00:30:42.798>when <00:30:43.031>all <00:30:43.265>we <00:30:43.498>want <00:30:43.731>is <00:30:43.965>actually

646
00:30:44.485 --> 00:30:46.245
<00:30:44.485>to <00:30:44.951>generate <00:30:45.418>beautiful <00:30:45.885>images.

647
00:30:46.665 --> 00:30:50.085
<00:30:46.665>And <00:30:46.805>the <00:30:46.945>answer <00:30:47.085>is, <00:30:47.225>if <00:30:47.365>you <00:30:47.504>are <00:30:47.645>able <00:30:47.784>to <00:30:47.925>predict <00:30:48.065>noise,

648
00:30:50.665 --> 00:30:53.765
<00:30:50.665>you <00:30:50.840>can <00:30:51.014>then <00:30:51.190>gradually <00:30:51.365>subtract <00:30:51.540>this <00:30:51.715>noise <00:30:51.889>from <00:30:52.065>your

649
00:30:54.005 --> 00:30:57.005
<00:30:54.005>original <00:30:54.180>image, <00:30:54.355>and <00:30:54.530>then <00:30:54.705>you <00:30:54.880>will <00:30:55.055>arrive <00:30:55.230>at <00:30:55.405>a

650
00:30:57.005 --> 00:30:58.045
<00:30:57.005>nice <00:30:57.317>and <00:30:57.629>beautiful <00:30:57.941>image.

651
00:30:58.065 --> 00:30:59.525
<00:30:58.065>So <00:30:58.252>let <00:30:58.440>me <00:30:58.628>show <00:30:58.815>you <00:30:59.003>how <00:30:59.191>it <00:30:59.379>works.

652
00:31:00.145 --> 00:31:03.085
<00:31:00.145>You <00:31:00.300>start, <00:31:00.456>uh, <00:31:00.611>like <00:31:00.767>you <00:31:00.922>start <00:31:01.078>with <00:31:01.233>this <00:31:01.389>like <00:31:01.545>very,

653
00:31:03.085 --> 00:31:05.525
<00:31:03.085>very <00:31:03.240>noisy <00:31:03.396>image, <00:31:03.551>like <00:31:03.707>I'm <00:31:03.862>showing <00:31:04.018>you <00:31:04.173>in <00:31:04.329>the, <00:31:04.485>uh,

654
00:31:05.575 --> 00:31:09.765
<00:31:05.575>upper <00:31:05.750>left <00:31:05.925>corner, <00:31:06.100>and <00:31:06.275>then <00:31:06.450>you <00:31:06.625>reduce <00:31:06.800>it <00:31:06.975>iteratively,

655
00:31:09.765 --> 00:31:12.045
<00:31:09.765>and <00:31:09.965>then <00:31:10.165>you <00:31:10.365>get <00:31:10.565>this <00:31:10.765>nice <00:31:10.965>image, <00:31:11.165>uh,

656
00:31:12.145 --> 00:31:14.485
<00:31:12.145>in <00:31:12.495>the <00:31:12.845>down <00:31:13.195>right <00:31:13.545>corner.

657
00:31:15.425 --> 00:31:17.965
<00:31:15.425>Uh, <00:31:15.658>the <00:31:15.891>training <00:31:16.125>works, <00:31:16.358>uh, <00:31:16.591>as <00:31:16.825>follows.

658
00:31:17.985 --> 00:31:19.405
<00:31:17.985>So <00:31:18.304>you <00:31:18.624>take <00:31:18.943>an <00:31:19.263>image

659
00:31:19.865 --> 00:31:22.605
<00:31:19.865>and <00:31:20.065>you <00:31:20.265>need, <00:31:20.465>yeah, <00:31:20.665>some <00:31:20.865>millions <00:31:21.065>of <00:31:21.265>images

660
00:31:22.705 --> 00:31:24.285
<00:31:22.705>to <00:31:22.905>train <00:31:23.105>a <00:31:23.304>model <00:31:23.504>like <00:31:23.705>that <00:31:23.905>from <00:31:24.105>scratch.

661
00:31:25.225 --> 00:31:27.005
<00:31:25.225>And <00:31:25.504>you <00:31:25.784>generate <00:31:26.064>some <00:31:26.344>random <00:31:26.625>noise

662
00:31:27.225 --> 00:31:29.485
<00:31:27.225>and, <00:31:27.504>uh, <00:31:27.784>the <00:31:28.064>different, <00:31:28.344>uh, <00:31:28.625>approaches

663
00:31:29.485 --> 00:31:32.805
<00:31:29.485>to <00:31:29.659>generate <00:31:29.834>this <00:31:30.010>noise, <00:31:30.185>the <00:31:30.360>different <00:31:30.534>nice <00:31:30.709>noise <00:31:30.885>schedulers.

664
00:31:33.625 --> 00:31:38.525
<00:31:33.625>Um, <00:31:33.765>and, <00:31:33.905>uh, <00:31:34.045>once <00:31:34.185>you've <00:31:34.325>generated <00:31:34.465>a <00:31:34.605>noise, <00:31:34.745>you <00:31:34.885>peak <00:31:35.025>the,

665
00:31:38.875 --> 00:31:40.645
<00:31:38.875>some <00:31:39.155>random <00:31:39.435>amount <00:31:39.715>of <00:31:39.995>this <00:31:40.275>noise,

666
00:31:41.305 --> 00:31:44.365
<00:31:41.305>and <00:31:41.460>you <00:31:41.616>add <00:31:41.771>this, <00:31:41.927>uh, <00:31:42.082>noise <00:31:42.238>to <00:31:42.393>the <00:31:42.549>original <00:31:42.705>image.

667
00:31:44.825 --> 00:31:46.325
<00:31:44.825>And <00:31:45.017>then <00:31:45.210>you <00:31:45.403>also, <00:31:45.596>of <00:31:45.789>course, <00:31:45.982>you <00:31:46.175>know

668
00:31:46.345 --> 00:31:48.445
<00:31:46.345>how <00:31:46.520>much <00:31:46.695>noise <00:31:46.870>you've <00:31:47.045>generated, <00:31:47.220>what <00:31:47.395>is <00:31:47.570>the <00:31:47.745>pattern,

669
00:31:48.785 --> 00:31:50.165
<00:31:48.785>you <00:31:49.033>know <00:31:49.281>how <00:31:49.530>to <00:31:49.778>predict <00:31:50.027>that.

670
00:31:50.305 --> 00:31:52.165
<00:31:50.305>So <00:31:50.585>makes <00:31:50.865>it <00:31:51.145>straightforward <00:31:51.425>for <00:31:51.705>unit.

671
00:31:52.905 --> 00:31:54.725
<00:31:52.905>And <00:31:53.138>once <00:31:53.371>you've <00:31:53.605>trained <00:31:53.838>this <00:31:54.071>image, <00:31:54.305>um,

672
00:31:55.075 --> 00:31:58.925
<00:31:55.075>inference <00:31:55.230>process <00:31:55.386>is, <00:31:55.541>uh, <00:31:55.697>is <00:31:55.852>not <00:31:56.008>like <00:31:56.163>typical <00:31:56.319>for <00:31:56.475>example,

673
00:31:58.925 --> 00:32:01.245
<00:31:58.925>segmentation, <00:31:59.125>where <00:31:59.325>you <00:31:59.524>just <00:31:59.725>take <00:31:59.925>image, <00:32:00.125>run <00:32:00.325>it

674
00:32:01.245 --> 00:32:03.805
<00:32:01.245>through <00:32:01.419>the <00:32:01.594>network, <00:32:01.770>and <00:32:01.945>you <00:32:02.120>get <00:32:02.294>your <00:32:02.469>segmentation <00:32:02.645>mask.

675
00:32:04.395 --> 00:32:07.445
<00:32:04.395>Inference <00:32:04.595>in <00:32:04.795>diffusion <00:32:04.995>is <00:32:05.195>also <00:32:05.395>an <00:32:05.595>iterative <00:32:05.795>process.

676
00:32:08.345 --> 00:32:10.505
<00:32:08.345>So <00:32:08.545>you <00:32:08.745>start <00:32:08.945>with <00:32:09.145>a <00:32:09.345>very <00:32:09.545>noisy <00:32:09.745>image,

677
00:32:10.725 --> 00:32:12.625
<00:32:10.725>and, <00:32:10.958>uh, <00:32:11.191>with <00:32:11.425>like <00:32:11.658>something <00:32:11.891>like <00:32:12.125>50

678
00:32:12.845 --> 00:32:15.705
<00:32:12.845>or <00:32:13.045>a <00:32:13.245>hundred <00:32:13.445>iterations, <00:32:13.645>you <00:32:13.845>deno <00:32:14.045>it <00:32:14.245>basically.

679
00:32:16.245 --> 00:32:17.785
<00:32:16.245>So <00:32:16.476>it's, <00:32:16.706>it's <00:32:16.937>a <00:32:17.169>really <00:32:17.400>fun <00:32:17.631>approach.

680
00:32:17.845 --> 00:32:21.505
<00:32:17.845>But <00:32:17.972>so <00:32:18.099>far, <00:32:18.226>what <00:32:18.354>I've <00:32:18.481>told <00:32:18.608>you, <00:32:18.735>uh, <00:32:18.863>yeah, <00:32:18.990>uh, <00:32:19.117>it's, <00:32:19.245>uh,

681
00:32:21.615 --> 00:32:23.385
<00:32:21.615>it's <00:32:21.965>computationally <00:32:22.315>heavy, <00:32:22.665>uh, <00:32:23.015>right,

682
00:32:23.385 --> 00:32:25.265
<00:32:23.385>so <00:32:23.585>if <00:32:23.785>you <00:32:23.985>would <00:32:24.185>work <00:32:24.385>on <00:32:24.585>original <00:32:24.785>images

683
00:32:25.725 --> 00:32:28.945
<00:32:25.725>and <00:32:25.880>the <00:32:26.036>researchers <00:32:26.191>came <00:32:26.347>out, <00:32:26.502>came <00:32:26.658>up <00:32:26.813>with <00:32:26.969>this, <00:32:27.125>uh,

684
00:32:28.945 --> 00:32:31.545
<00:32:28.945>cool <00:32:29.178>trick, <00:32:29.411>uh, <00:32:29.645>to <00:32:29.878>use, <00:32:30.111>uh, <00:32:30.345>the,

685
00:32:31.545 --> 00:32:35.105
<00:32:31.545>so-called <00:32:31.745>Auto <00:32:31.945>Enco <00:32:32.145>Variational <00:32:32.345>Auto <00:32:32.545>Quarter <00:32:32.745>Architecture <00:32:32.945>in

686
00:32:35.105 --> 00:32:38.105
<00:32:35.105>order <00:32:35.280>to <00:32:35.455>reduce, <00:32:35.630>uh, <00:32:35.805>your <00:32:35.980>image <00:32:36.155>in, <00:32:36.330>in <00:32:36.505>order

687
00:32:38.105 --> 00:32:42.085
<00:32:38.105>to <00:32:38.245>encode <00:32:38.385>your <00:32:38.525>image <00:32:38.665>into <00:32:38.805>some, <00:32:38.945>uh, <00:32:39.085>smaller, <00:32:39.225>uh, <00:32:39.365>space, <00:32:39.505>uh,

688
00:32:42.085 --> 00:32:43.725
<00:32:42.085>which <00:32:42.435>we <00:32:42.785>call <00:32:43.135>latent <00:32:43.485>space.

689
00:32:44.585 --> 00:32:47.805
<00:32:44.585>And, <00:32:44.785>uh, <00:32:44.985>basically <00:32:45.185>we <00:32:45.385>have <00:32:45.585>our <00:32:45.785>input <00:32:45.985>image,

690
00:32:48.105 --> 00:32:50.525
<00:32:48.105>and <00:32:48.305>we <00:32:48.505>have <00:32:48.705>two <00:32:48.905>networks <00:32:49.105>encoder <00:32:49.305>and <00:32:49.505>quarter.

691
00:32:51.305 --> 00:32:53.405
<00:32:51.305>The <00:32:51.538>encoder <00:32:51.771>architecture <00:32:52.005>takes <00:32:52.238>the <00:32:52.471>original <00:32:52.705>image,

692
00:32:53.585 --> 00:32:56.165
<00:32:53.585>it <00:32:53.818>compresses <00:32:54.051>it <00:32:54.285>to <00:32:54.518>some <00:32:54.751>smaller <00:32:54.985>resolution.

693
00:32:57.425 --> 00:33:00.285
<00:32:57.425>And, <00:32:57.580>uh, <00:32:57.736>and <00:32:57.891>by <00:32:58.047>compressing, <00:32:58.202>it's <00:32:58.358>not <00:32:58.513>just <00:32:58.669>scaling <00:32:58.825>down,

694
00:33:00.345 --> 00:33:03.605
<00:33:00.345>but <00:33:00.500>basically <00:33:00.656>modifying <00:33:00.811>your <00:33:00.967>image <00:33:01.122>so <00:33:01.278>that <00:33:01.433>it <00:33:01.589>gets <00:33:01.745>smaller.

695
00:33:04.225 --> 00:33:05.525
<00:33:04.225>And <00:33:04.458>then <00:33:04.693>you <00:33:04.927>have <00:33:05.161>the <00:33:05.395>decoder,

696
00:33:05.575 --> 00:33:08.205
<00:33:05.575>which <00:33:05.855>then <00:33:06.135>takes <00:33:06.415>your <00:33:06.695>compressed <00:33:06.975>image,

697
00:33:08.305 --> 00:33:11.685
<00:33:08.305>and <00:33:08.460>then <00:33:08.616>it <00:33:08.771>in, <00:33:08.927>like, <00:33:09.082>it <00:33:09.238>increases <00:33:09.393>its, <00:33:09.549>uh, <00:33:09.705>size

698
00:33:11.745 --> 00:33:14.005
<00:33:11.745>and <00:33:11.978>resolution, <00:33:12.211>and <00:33:12.445>you <00:33:12.678>get <00:33:12.911>the, <00:33:13.145>the,

699
00:33:14.265 --> 00:33:15.685
<00:33:14.265>the <00:33:14.691>original <00:33:15.117>output <00:33:15.543>image.

700
00:33:15.945 --> 00:33:19.005
<00:33:15.945>And <00:33:16.072>I <00:33:16.199>can <00:33:16.326>imagine <00:33:16.454>that <00:33:16.581>training <00:33:16.708>a <00:33:16.835>network <00:33:16.963>like <00:33:17.090>this <00:33:17.217>might <00:33:17.345>be

701
00:33:19.365 --> 00:33:22.285
<00:33:19.365>straightforward <00:33:19.565>because, <00:33:19.765>uh, <00:33:19.965>you <00:33:20.165>have <00:33:20.365>your <00:33:20.565>input <00:33:20.765>image,

702
00:33:22.385 --> 00:33:26.005
<00:33:22.385>and <00:33:22.618>ideally, <00:33:22.851>um, <00:33:23.085>when <00:33:23.318>you're <00:33:23.551>training, <00:33:23.785>uh,

703
00:33:26.475 --> 00:33:29.325
<00:33:26.475>your <00:33:26.630>output <00:33:26.786>image <00:33:26.941>should <00:33:27.097>be <00:33:27.252>identical <00:33:27.408>to <00:33:27.563>your <00:33:27.719>input <00:33:27.875>image.

704
00:33:29.325 --> 00:33:31.485
<00:33:29.325>So <00:33:29.500>that's <00:33:29.675>how <00:33:29.850>you <00:33:30.025>update <00:33:30.200>the <00:33:30.375>parameters <00:33:30.550>of <00:33:30.725>encoder

705
00:33:31.485 --> 00:33:36.445
<00:33:31.485>and <00:33:31.659>decoder <00:33:31.834>and <00:33:32.010>diffusion <00:33:32.185>operates <00:33:32.360>on <00:33:32.534>this <00:33:32.709>latent <00:33:32.885>space,

706
00:33:36.595 --> 00:33:40.285
<00:33:36.595>meaning <00:33:36.750>that, <00:33:36.906>uh, <00:33:37.061>you <00:33:37.217>generate <00:33:37.372>noise <00:33:37.528>in <00:33:37.683>this <00:33:37.839>latent <00:33:37.995>space.

707
00:33:40.465 --> 00:33:42.285
<00:33:40.465>You <00:33:40.698>predict <00:33:40.931>noise <00:33:41.165>in <00:33:41.398>this <00:33:41.631>latent <00:33:41.865>space,

708
00:33:42.345 --> 00:33:44.805
<00:33:42.345>and <00:33:42.545>you <00:33:42.745>subtract <00:33:42.945>noise <00:33:43.145>in <00:33:43.345>this <00:33:43.545>latent <00:33:43.745>space,

709
00:33:45.385 --> 00:33:47.885
<00:33:45.385>and <00:33:45.585>then <00:33:45.785>you <00:33:45.985>use <00:33:46.185>this <00:33:46.385>decoder <00:33:46.585>architecture <00:33:46.785>to

710
00:33:48.425 --> 00:33:50.965
<00:33:48.425>get <00:33:48.705>this <00:33:48.985>bigger <00:33:49.264>size <00:33:49.544>nice <00:33:49.825>image,

711
00:33:51.785 --> 00:33:54.405
<00:33:51.785>and <00:33:52.018>it <00:33:52.251>speeds <00:33:52.485>up <00:33:52.718>your <00:33:52.951>training <00:33:53.185>dramatically.

712
00:33:54.745 --> 00:33:57.965
<00:33:54.745>And <00:33:55.024>not <00:33:55.304>every, <00:33:55.584>uh, <00:33:55.864>diffusion, <00:33:56.145>uh,

713
00:33:58.265 --> 00:34:00.445
<00:33:58.265>latent <00:33:58.615>diffusion <00:33:58.965>approach <00:33:59.315>uses <00:33:59.665>that,

714
00:34:00.625 --> 00:34:04.725
<00:34:00.625>but <00:34:00.752>I <00:34:00.879>think <00:34:01.006>it's <00:34:01.134>a <00:34:01.261>very, <00:34:01.388>uh, <00:34:01.515>a <00:34:01.643>very, <00:34:01.770>uh, <00:34:01.897>useful <00:34:02.025>step.

715
00:34:05.865 --> 00:34:08.605
<00:34:05.865>And <00:34:06.020>so <00:34:06.176>how <00:34:06.331>would <00:34:06.487>I've <00:34:06.642>presented, <00:34:06.798>you <00:34:06.953>would <00:34:07.109>work, <00:34:07.265>um,

716
00:34:09.305 --> 00:34:13.565
<00:34:09.304>on <00:34:09.460>any <00:34:09.616>kinds <00:34:09.771>of <00:34:09.927>images, <00:34:10.082>depending <00:34:10.238>like, <00:34:10.393>uh, <00:34:10.549>what <00:34:10.705>kind

717
00:34:13.565 --> 00:34:15.965
<00:34:13.565>of <00:34:13.740>images <00:34:13.915>you <00:34:14.090>have, <00:34:14.264>uh, <00:34:14.440>trained <00:34:14.615>your <00:34:14.790>network <00:34:14.965>on.

718
00:34:16.175 --> 00:34:17.765
<00:34:16.175>It'll <00:34:16.641>generate <00:34:17.108>random <00:34:17.575>images,

719
00:34:18.345 --> 00:34:20.205
<00:34:18.344>but <00:34:18.520>if <00:34:18.694>you <00:34:18.870>want <00:34:19.044>to <00:34:19.219>be <00:34:19.395>in <00:34:19.569>control <00:34:19.745>of

720
00:34:20.205 --> 00:34:22.965
<00:34:20.205>what <00:34:20.404>the <00:34:20.605>image <00:34:20.804>generates, <00:34:21.005>you <00:34:21.205>need <00:34:21.404>this <00:34:21.605>component.

721
00:34:23.185 --> 00:34:25.005
<00:34:23.185>Uh, <00:34:23.465>text <00:34:23.745>encoder <00:34:24.025>and <00:34:24.304>image <00:34:24.585>encoder.

722
00:34:25.005 --> 00:34:26.925
<00:34:25.005>And <00:34:25.238>clip <00:34:25.471>clip <00:34:25.705>is <00:34:25.938>just <00:34:26.171>one <00:34:26.405>example.

723
00:34:27.035 --> 00:34:30.085
<00:34:27.034>It's <00:34:27.268>a <00:34:27.501>popular <00:34:27.734>architecture <00:34:27.968>trained <00:34:28.201>by <00:34:28.435>OpenAI,

724
00:34:30.225 --> 00:34:31.485
<00:34:30.225>and <00:34:30.603>everybody <00:34:30.981>uses <00:34:31.359>it.

725
00:34:31.585 --> 00:34:35.085
<00:34:31.585>Uh, <00:34:31.760>they've <00:34:31.935>trained <00:34:32.110>on <00:34:32.284>400 <00:34:32.460>million <00:34:32.635>pairs <00:34:32.810>of <00:34:32.985>images.

726
00:34:35.705 --> 00:34:39.925
<00:34:35.705>And <00:34:35.880>the <00:34:36.054>idea <00:34:36.230>here <00:34:36.404>is <00:34:36.580>to <00:34:36.755>bring <00:34:36.929>our <00:34:37.105>images

727
00:34:40.425 --> 00:34:43.485
<00:34:40.425>and <00:34:40.625>text <00:34:40.825>in <00:34:41.025>the <00:34:41.225>same <00:34:41.425>embeds <00:34:41.625>embedding <00:34:41.825>space.

728
00:34:44.085 --> 00:34:46.085
<00:34:44.085>Remember <00:34:44.240>when <00:34:44.396>I <00:34:44.551>was <00:34:44.707>talking <00:34:44.862>about <00:34:45.018>embeddings, <00:34:45.173>I <00:34:45.329>told <00:34:45.485>you

729
00:34:46.085 --> 00:34:50.965
<00:34:46.085>that, <00:34:46.240>uh, <00:34:46.396>this <00:34:46.551>is <00:34:46.707>a <00:34:46.862>unifying, <00:34:47.018>the <00:34:47.173>unifying, <00:34:47.329>uh, <00:34:47.485>language,

730
00:34:51.015 --> 00:34:54.325
<00:34:51.014>let's <00:34:51.190>call <00:34:51.364>it <00:34:51.540>like <00:34:51.714>that, <00:34:51.889>uh, <00:34:52.065>for <00:34:52.239>different <00:34:52.415>modalities.

731
00:34:54.905 --> 00:34:58.165
<00:34:54.905>And, <00:34:55.105>uh, <00:34:55.305>originally, <00:34:55.505>images <00:34:55.705>and <00:34:55.905>texts <00:34:56.105>are <00:34:56.305>different.

732
00:34:58.505 --> 00:35:02.365
<00:34:58.505>And <00:34:58.645>our <00:34:58.785>goal <00:34:58.925>here <00:34:59.065>is <00:34:59.205>to <00:34:59.345>basically <00:34:59.485>create <00:34:59.625>an <00:34:59.765>image <00:34:59.905>encoder,

733
00:35:02.365 --> 00:35:04.485
<00:35:02.365>text <00:35:02.564>encoder, <00:35:02.765>and <00:35:02.964>bring <00:35:03.165>them <00:35:03.365>in <00:35:03.564>into <00:35:03.765>the,

734
00:35:04.625 --> 00:35:07.285
<00:35:04.625>the <00:35:04.825>very <00:35:05.025>same <00:35:05.225>space, <00:35:05.425>the <00:35:05.625>very <00:35:05.825>same <00:35:06.025>domain.

735
00:35:07.905 --> 00:35:11.525
<00:35:07.905>Uh, <00:35:08.032>so <00:35:08.159>how <00:35:08.286>it <00:35:08.414>works, <00:35:08.541>uh, <00:35:08.668>we <00:35:08.795>have, <00:35:08.923>uh, <00:35:09.050>an <00:35:09.177>image <00:35:09.305>encoder.

736
00:35:11.525 --> 00:35:14.325
<00:35:11.525>Typically, <00:35:11.700>we <00:35:11.875>use <00:35:12.050>vision <00:35:12.225>transformer <00:35:12.400>or <00:35:12.575>beat <00:35:12.750>for <00:35:12.925>that.

737
00:35:14.665 --> 00:35:16.725
<00:35:14.665>Um, <00:35:14.898>and <00:35:15.131>we <00:35:15.365>have <00:35:15.598>a <00:35:15.831>text <00:35:16.065>encoder.

738
00:35:17.425 --> 00:35:20.645
<00:35:17.425>And, <00:35:17.625>uh, <00:35:17.825>initially <00:35:18.025>we <00:35:18.225>extract <00:35:18.425>the <00:35:18.625>embeddings <00:35:18.825>from

739
00:35:20.645 --> 00:35:21.685
<00:35:20.645>both <00:35:20.957>images <00:35:21.269>and <00:35:21.581>text.

740
00:35:21.905 --> 00:35:24.805
<00:35:21.905>And <00:35:22.105>these <00:35:22.305>400 <00:35:22.505>million <00:35:22.705>images, <00:35:22.905>uh, <00:35:23.105>image <00:35:23.305>pairs,

741
00:35:24.805 --> 00:35:26.565
<00:35:24.805>we <00:35:25.085>just <00:35:25.365>crawl <00:35:25.645>from <00:35:25.924>the <00:35:26.205>internet.

742
00:35:26.985 --> 00:35:30.805
<00:35:26.985>And <00:35:27.140>the <00:35:27.296>text <00:35:27.451>comes <00:35:27.607>from <00:35:27.762>the <00:35:27.918>metadata <00:35:28.073>from <00:35:28.229>the <00:35:28.385>HTML.

743
00:35:30.905 --> 00:35:34.405
<00:35:30.905>So <00:35:31.060>this <00:35:31.216>is <00:35:31.371>basically <00:35:31.527>how <00:35:31.682>this <00:35:31.838>supervised <00:35:31.993>peers <00:35:32.149>were <00:35:32.305>created.

744
00:35:35.545 --> 00:35:40.205
<00:35:35.545>And <00:35:35.685>then <00:35:35.825>you <00:35:35.965>take <00:35:36.105>this <00:35:36.245>image <00:35:36.385>and, <00:35:36.525>uh, <00:35:36.665>and <00:35:36.805>text <00:35:36.945>embeddings,

745
00:35:40.205 --> 00:35:42.765
<00:35:40.205>and <00:35:40.438>you <00:35:40.671>compare <00:35:40.904>them <00:35:41.138>using <00:35:41.371>cosign <00:35:41.605>similarity.

746
00:35:43.305 --> 00:35:47.605
<00:35:43.305>And <00:35:43.432>if <00:35:43.559>they're <00:35:43.686>similar, <00:35:43.814>you <00:35:43.941>give, <00:35:44.068>uh, <00:35:44.195>well, <00:35:44.323>you <00:35:44.450>give <00:35:44.577>it, <00:35:44.705>uh,

747
00:35:47.605 --> 00:35:50.405
<00:35:47.605>something, <00:35:47.760>uh, <00:35:47.916>like <00:35:48.071>one, <00:35:48.227>and <00:35:48.382>if <00:35:48.538>they <00:35:48.693>are <00:35:48.849>not <00:35:49.005>similar,

748
00:35:50.665 --> 00:35:53.005
<00:35:50.665>and <00:35:50.840>you <00:35:51.015>need <00:35:51.190>to <00:35:51.365>use <00:35:51.540>also <00:35:51.715>different <00:35:51.890>examples, <00:35:52.065>uh,

749
00:35:53.005 --> 00:35:54.925
<00:35:53.005>it's <00:35:53.285>called <00:35:53.565>like <00:35:53.845>the <00:35:54.125>negative <00:35:54.405>pairs,

750
00:35:54.925 --> 00:35:57.285
<00:35:54.925>because <00:35:55.158>you <00:35:55.391>use <00:35:55.625>this, <00:35:55.858>so-called <00:35:56.091>Contrastive <00:35:56.325>loss.

751
00:35:58.105 --> 00:36:00.325
<00:35:58.105>Um, <00:35:58.338>you <00:35:58.571>give <00:35:58.805>something <00:35:59.038>closer <00:35:59.271>to <00:35:59.505>zero,

752
00:36:00.905 --> 00:36:03.605
<00:36:00.905>and <00:36:01.060>based <00:36:01.216>on <00:36:01.371>this <00:36:01.527>values, <00:36:01.682>like <00:36:01.838>of <00:36:01.993>the <00:36:02.149>objective <00:36:02.305>function,

753
00:36:03.785 --> 00:36:06.925
<00:36:03.785>you <00:36:04.065>then <00:36:04.345>update <00:36:04.625>your <00:36:04.904>encoder <00:36:05.185>parameters,

754
00:36:07.705 --> 00:36:11.045
<00:36:07.705>and <00:36:07.845>then <00:36:07.985>idea <00:36:08.125>that <00:36:08.265>you <00:36:08.404>arrive <00:36:08.545>at <00:36:08.685>something <00:36:08.825>that <00:36:08.965>your <00:36:09.105>images

755
00:36:11.185 --> 00:36:14.485
<00:36:11.185>and <00:36:11.465>text <00:36:11.745>will <00:36:12.025>produce <00:36:12.305>similar <00:36:12.585>embeddings,

756
00:36:14.655 --> 00:36:16.725
<00:36:14.655>which <00:36:14.935>you <00:36:15.215>can <00:36:15.495>then <00:36:15.775>use <00:36:16.055>interchangeably.

757
00:36:17.705 --> 00:36:20.565
<00:36:17.705>Um, <00:36:17.860>this <00:36:18.016>is <00:36:18.171>how <00:36:18.327>the <00:36:18.482>full, <00:36:18.638>full <00:36:18.793>picture <00:36:18.949>looks <00:36:19.105>like.

758
00:36:21.105 --> 00:36:23.845
<00:36:21.105>You <00:36:21.338>have <00:36:21.571>your <00:36:21.805>image <00:36:22.038>latent, <00:36:22.271>um, <00:36:22.505>produced

759
00:36:23.945 --> 00:36:25.805
<00:36:23.945>by <00:36:24.295>your <00:36:24.645>variational <00:36:24.995>auto <00:36:25.345>encoder.

760
00:36:26.065 --> 00:36:29.325
<00:36:26.065>You <00:36:26.220>use <00:36:26.376>it <00:36:26.531>as <00:36:26.687>an <00:36:26.842>input <00:36:26.998>to <00:36:27.153>your <00:36:27.309>unit <00:36:27.465>together

761
00:36:29.325 --> 00:36:30.405
<00:36:29.325>with <00:36:29.649>the <00:36:29.973>noise <00:36:30.297>amount.

762
00:36:30.705 --> 00:36:33.005
<00:36:30.705>And <00:36:30.938>then <00:36:31.171>you <00:36:31.404>include <00:36:31.638>this <00:36:31.871>text <00:36:32.105>embeddings,

763
00:36:33.125 --> 00:36:36.285
<00:36:33.125>which <00:36:33.280>you <00:36:33.436>get <00:36:33.591>from <00:36:33.747>clip, <00:36:33.902>and <00:36:34.058>you <00:36:34.213>combine <00:36:34.369>the <00:36:34.525>soul

764
00:36:36.425 --> 00:36:38.045
<00:36:36.425>by <00:36:36.891>in <00:36:37.358>introducing <00:36:37.825>the

765
00:36:38.045 --> 00:36:40.445
<00:36:38.045>so <00:36:38.245>called <00:36:38.445>cross <00:36:38.645>attention <00:36:38.845>layers <00:36:39.045>to <00:36:39.245>all <00:36:39.445>units.

766
00:36:40.625 --> 00:36:41.925
<00:36:40.625>So <00:36:41.015>cross <00:36:41.405>attention <00:36:41.795>means

767
00:36:41.925 --> 00:36:45.125
<00:36:41.925>that <00:36:42.100>we <00:36:42.275>are <00:36:42.450>using <00:36:42.625>different <00:36:42.800>modalities <00:36:42.975>in <00:36:43.150>our <00:36:43.325>architecture.

768
00:36:45.505 --> 00:36:47.845
<00:36:45.505>And <00:36:45.660>of <00:36:45.816>course, <00:36:45.971>it <00:36:46.127>has <00:36:46.282>all <00:36:46.438>still <00:36:46.593>all <00:36:46.749>the <00:36:46.905>skip

769
00:36:47.845 --> 00:36:49.125
<00:36:47.845>connections <00:36:48.421>and <00:36:48.997>everything.

770
00:36:49.345 --> 00:36:50.845
<00:36:49.345>And <00:36:49.537>in <00:36:49.730>the <00:36:49.923>end, <00:36:50.116>you <00:36:50.309>predict <00:36:50.502>your <00:36:50.694>noise,

771
00:36:50.925 --> 00:36:52.485
<00:36:50.925>which <00:36:51.205>you <00:36:51.485>subtract <00:36:51.765>from <00:36:52.045>the <00:36:52.325>image,

772
00:36:52.505 --> 00:36:54.645
<00:36:52.505>and <00:36:52.785>then, <00:36:53.065>uh, <00:36:53.345>you <00:36:53.625>generate, <00:36:53.905>uh,

773
00:36:55.475 --> 00:36:57.885
<00:36:55.475>beautiful <00:36:55.755>images <00:36:56.035>in <00:36:56.315>the <00:36:56.595>controlled <00:36:56.875>environment.

774
00:36:58.425 --> 00:37:00.725
<00:36:58.425>And, <00:36:58.580>uh, <00:36:58.736>you <00:36:58.891>can <00:36:59.047>also <00:36:59.202>do <00:36:59.358>image <00:36:59.513>to <00:36:59.669>image <00:36:59.825>generation.

775
00:37:00.785 --> 00:37:04.005
<00:37:00.785>And <00:37:00.924>in <00:37:01.065>this <00:37:01.205>example, <00:37:01.345>I'm <00:37:01.484>showing <00:37:01.625>you, <00:37:01.765>um, <00:37:01.904>it's <00:37:02.045>an <00:37:02.185>image

776
00:37:04.005 --> 00:37:06.325
<00:37:04.005>of <00:37:04.180>a <00:37:04.355>pink <00:37:04.530>pan, <00:37:04.705>which <00:37:04.880>I <00:37:05.055>painted <00:37:05.230>myself, <00:37:05.405>uh,

777
00:37:06.825 --> 00:37:08.765
<00:37:06.825>please <00:37:07.105>don't <00:37:07.384>judge <00:37:07.665>my <00:37:07.944>drawing <00:37:08.225>skills.

778
00:37:09.185 --> 00:37:12.805
<00:37:09.185>Uh, <00:37:09.360>and <00:37:09.535>I <00:37:09.710>wanted <00:37:09.884>to, <00:37:10.060>uh, <00:37:10.235>actually <00:37:10.410>improve <00:37:10.585>it.

779
00:37:13.025 --> 00:37:16.005
<00:37:13.025>And <00:37:13.152>I've <00:37:13.279>asked <00:37:13.406>a <00:37:13.534>stable <00:37:13.661>user <00:37:13.788>network <00:37:13.915>to <00:37:14.043>give <00:37:14.170>me <00:37:14.297>an <00:37:14.425>image

780
00:37:16.005 --> 00:37:17.725
<00:37:16.005>of <00:37:16.180>a <00:37:16.355>ping <00:37:16.530>puncher, <00:37:16.705>and <00:37:16.880>that's <00:37:17.055>what <00:37:17.230>I <00:37:17.405>got.

781
00:37:18.345 --> 00:37:20.485
<00:37:18.345>Uh, <00:37:18.578>I <00:37:18.811>thought, <00:37:19.044>well, <00:37:19.278>that's <00:37:19.511>pretty, <00:37:19.745>uh,

782
00:37:20.545 --> 00:37:22.725
<00:37:20.545>and <00:37:20.700>then <00:37:20.856>I <00:37:21.011>took <00:37:21.167>a <00:37:21.322>look <00:37:21.478>at <00:37:21.633>my <00:37:21.789>original <00:37:21.945>drawing

783
00:37:22.785 --> 00:37:24.645
<00:37:22.785>and <00:37:22.960>I <00:37:23.134>said, <00:37:23.310>Hmm, <00:37:23.484>maybe <00:37:23.660>it, <00:37:23.835>it's <00:37:24.009>a <00:37:24.185>cat.

784
00:37:25.185 --> 00:37:26.885
<00:37:25.185>And <00:37:25.465>it <00:37:25.745>generated <00:37:26.025>me <00:37:26.305>a <00:37:26.585>cat.

785
00:37:27.305 --> 00:37:29.325
<00:37:27.305>But <00:37:27.480>then <00:37:27.654>somebody <00:37:27.830>else <00:37:28.004>took <00:37:28.180>a <00:37:28.355>look <00:37:28.529>at <00:37:28.705>this

786
00:37:29.325 --> 00:37:31.365
<00:37:29.325>and <00:37:29.524>said, <00:37:29.725>no, <00:37:29.924>it <00:37:30.125>looks, <00:37:30.325>it <00:37:30.524>looks <00:37:30.725>different.

787
00:37:31.365 --> 00:37:33.685
<00:37:31.365>It <00:37:31.520>looks <00:37:31.676>like <00:37:31.831>a <00:37:31.987>bear. <00:37:32.142>And <00:37:32.298>I <00:37:32.453>got <00:37:32.609>a <00:37:32.765>bear.

788
00:37:33.985 --> 00:37:37.045
<00:37:33.985>But <00:37:34.140>then <00:37:34.296>I, <00:37:34.451>I, <00:37:34.607>I <00:37:34.762>went <00:37:34.918>crazy <00:37:35.073>and <00:37:35.229>started <00:37:35.385>experimenting

789
00:37:37.045 --> 00:37:40.245
<00:37:37.045>and <00:37:37.185>got <00:37:37.325>me <00:37:37.465>a <00:37:37.605>monkey, <00:37:37.745>a <00:37:37.885>cow, <00:37:38.025>and <00:37:38.165>even <00:37:38.305>a <00:37:38.445>pig.

790
00:37:40.385 --> 00:37:43.165
<00:37:40.385>So <00:37:40.512>you <00:37:40.639>can <00:37:40.766>imagine <00:37:40.894>I <00:37:41.021>had <00:37:41.148>lots <00:37:41.275>of <00:37:41.403>fun <00:37:41.530>doing <00:37:41.657>this <00:37:41.785>exercise.

791
00:37:44.145 --> 00:37:47.605
<00:37:44.145>Um, <00:37:44.345>there <00:37:44.545>many, <00:37:44.745>uh, <00:37:44.945>image <00:37:45.145>generation <00:37:45.345>use <00:37:45.545>cases

792
00:37:48.145 --> 00:37:51.645
<00:37:48.145>and, <00:37:48.378>uh, <00:37:48.611>for <00:37:48.845>the, <00:37:49.078>so-called <00:37:49.311>unconditioned <00:37:49.545>generation,

793
00:37:51.735 --> 00:37:54.445
<00:37:51.735>where <00:37:51.890>we <00:37:52.046>don't <00:37:52.201>use, <00:37:52.357>for <00:37:52.512>example, <00:37:52.668>a <00:37:52.823>network <00:37:52.979>like <00:37:53.135>clip.

794
00:37:54.665 --> 00:37:56.565
<00:37:54.665>Um, <00:37:54.898>it's, <00:37:55.131>it <00:37:55.365>can <00:37:55.598>still <00:37:55.831>be <00:37:56.065>useful.

795
00:37:56.785 --> 00:37:58.485
<00:37:56.785>You <00:37:57.134>can <00:37:57.484>generate <00:37:57.835>human <00:37:58.185>faces.

796
00:37:58.825 --> 00:38:01.365
<00:37:58.825>Uh, <00:37:59.000>this <00:37:59.174>is <00:37:59.350>very <00:37:59.524>useful <00:37:59.700>for <00:37:59.875>creation <00:38:00.049>of <00:38:00.225>avatars,

797
00:38:01.365 --> 00:38:05.325
<00:38:01.365>for <00:38:01.540>example, <00:38:01.714>um, <00:38:01.890>you <00:38:02.064>can <00:38:02.240>generate <00:38:02.415>synthetic <00:38:02.589>assets, <00:38:02.765>uh,

798
00:38:05.325 --> 00:38:08.485
<00:38:05.325>in <00:38:05.452>two <00:38:05.579>D <00:38:05.706>or <00:38:05.834>in <00:38:05.961>three <00:38:06.088>D <00:38:06.215>in <00:38:06.343>order <00:38:06.470>to <00:38:06.597>populate, <00:38:06.725>uh,

799
00:38:08.485 --> 00:38:10.685
<00:38:08.485>virtual <00:38:08.718>worlds, <00:38:08.951>for <00:38:09.185>example, <00:38:09.418>if <00:38:09.651>you <00:38:09.885>want

800
00:38:10.685 --> 00:38:11.765
<00:38:10.685>to <00:38:11.009>create <00:38:11.333>synthetic <00:38:11.656>data.

801
00:38:12.385 --> 00:38:15.365
<00:38:12.385>Um, <00:38:12.540>and <00:38:12.696>you <00:38:12.851>can <00:38:13.007>also <00:38:13.162>use <00:38:13.318>it <00:38:13.473>for <00:38:13.629>super <00:38:13.785>resolution.

802
00:38:15.545 --> 00:38:18.765
<00:38:15.545>And <00:38:15.700>then <00:38:15.856>for <00:38:16.011>condition <00:38:16.167>generation, <00:38:16.322>you <00:38:16.478>can <00:38:16.633>use <00:38:16.789>text <00:38:16.945>image,

803
00:38:19.025 --> 00:38:22.405
<00:38:19.025>uh, <00:38:19.165>to, <00:38:19.305>you <00:38:19.445>can <00:38:19.585>use, <00:38:19.725>uh, <00:38:19.865>it <00:38:20.005>for <00:38:20.145>selective <00:38:20.285>image <00:38:20.425>editing.

804
00:38:22.585 --> 00:38:25.285
<00:38:22.585>And <00:38:22.785>in <00:38:22.985>painting, <00:38:23.185>um, <00:38:23.385>let's <00:38:23.585>say <00:38:23.785>you <00:38:23.985>want

805
00:38:25.285 --> 00:38:27.365
<00:38:25.285>to <00:38:25.518>remove <00:38:25.751>your <00:38:25.984>ex-boyfriend <00:38:26.218>from <00:38:26.451>your <00:38:26.685>photograph,

806
00:38:27.365 --> 00:38:29.485
<00:38:27.365>you <00:38:27.540>remove <00:38:27.714>him <00:38:27.890>and <00:38:28.064>then <00:38:28.240>use <00:38:28.415>something <00:38:28.589>like <00:38:28.765>that

807
00:38:30.025 --> 00:38:31.605
<00:38:30.025>to <00:38:30.491>restore <00:38:30.958>the <00:38:31.425>background.

808
00:38:31.985 --> 00:38:34.645
<00:38:31.985>And <00:38:32.125>of <00:38:32.265>course, <00:38:32.405>you <00:38:32.545>can <00:38:32.685>use <00:38:32.825>text <00:38:32.965>guided <00:38:33.105>image <00:38:33.245>to <00:38:33.385>image.

809
00:38:35.145 --> 00:38:37.605
<00:38:35.145>So <00:38:35.345>this, <00:38:35.545>this <00:38:35.745>is <00:38:35.945>a <00:38:36.145>super <00:38:36.345>fun <00:38:36.545>technology.

810
00:38:37.985 --> 00:38:41.325
<00:38:37.985>And, <00:38:38.160>uh, <00:38:38.335>with <00:38:38.510>Nvidia <00:38:38.685>have <00:38:38.860>also <00:38:39.035>our <00:38:39.210>own <00:38:39.385>development.

811
00:38:41.425 --> 00:38:44.445
<00:38:41.425>Uh, <00:38:41.625>we <00:38:41.825>have <00:38:42.025>our <00:38:42.225>own <00:38:42.425>edify <00:38:42.625>model, <00:38:42.825>um,

812
00:38:44.575 --> 00:38:46.165
<00:38:44.575>which <00:38:44.808>I'm <00:38:45.041>showing <00:38:45.274>you <00:38:45.508>in <00:38:45.741>this <00:38:45.975>image.

813
00:38:46.265 --> 00:38:50.325
<00:38:46.265>So <00:38:46.420>our <00:38:46.576>edify <00:38:46.731>approach, <00:38:46.887>uh, <00:38:47.042>focuses <00:38:47.198>on <00:38:47.353>creation <00:38:47.509>of, <00:38:47.665>uh,

814
00:38:50.325 --> 00:38:54.565
<00:38:50.325>super <00:38:50.605>resolution, <00:38:50.884>uh, <00:38:51.165>artistic <00:38:51.444>images, <00:38:51.725>uh,

815
00:38:54.565 --> 00:38:56.405
<00:38:54.565>which <00:38:54.845>are <00:38:55.125>ideally <00:38:55.405>free <00:38:55.685>of <00:38:55.965>bias.

816
00:38:56.785 --> 00:39:00.245
<00:38:56.785>Uh, <00:38:57.018>and, <00:38:57.251>um, <00:38:57.484>we <00:38:57.718>are <00:38:57.951>also <00:38:58.185>generating,

817
00:39:00.245 --> 00:39:01.485
<00:39:00.245>working <00:39:00.617>on <00:39:00.989>generating <00:39:01.361>videos

818
00:39:02.105 --> 00:39:05.965
<00:39:02.105>and <00:39:02.280>even <00:39:02.455>three <00:39:02.630>d <00:39:02.805>uh, <00:39:02.980>based <00:39:03.155>on <00:39:03.330>textural <00:39:03.505>prompts.

819
00:39:06.385 --> 00:39:09.725
<00:39:06.385>And <00:39:06.540>you <00:39:06.696>can <00:39:06.851>then <00:39:07.007>take <00:39:07.162>this, <00:39:07.318>well, <00:39:07.473>this <00:39:07.629>models, <00:39:07.785>uh,

820
00:39:09.825 --> 00:39:12.965
<00:39:09.825>we <00:39:09.980>call <00:39:10.136>it, <00:39:10.291>uh, <00:39:10.447>we <00:39:10.602>call <00:39:10.758>the <00:39:10.913>service <00:39:11.069>Picasso, <00:39:11.225>uh,

821
00:39:12.965 --> 00:39:16.205
<00:39:12.965>where <00:39:13.105>we <00:39:13.245>integrate <00:39:13.385>this <00:39:13.525>edify <00:39:13.665>model, <00:39:13.805>and <00:39:13.945>there <00:39:14.085>is <00:39:14.225>an <00:39:14.365>API,

822
00:39:16.265 --> 00:39:20.445
<00:39:16.265>and <00:39:16.404>you <00:39:16.545>can <00:39:16.685>then <00:39:16.825>get <00:39:16.964>the <00:39:17.105>outputs <00:39:17.245>of, <00:39:17.384>uh, <00:39:17.525>the <00:39:17.665>generations

823
00:39:20.585 --> 00:39:25.325
<00:39:20.585>and <00:39:20.725>bring <00:39:20.865>them <00:39:21.005>in <00:39:21.145>some <00:39:21.285>visual <00:39:21.425>editors <00:39:21.565>or, <00:39:21.705>or <00:39:21.845>Omnis <00:39:21.985>platform.

824
00:39:25.705 --> 00:39:28.565
<00:39:25.705>And <00:39:25.938>you <00:39:26.171>can <00:39:26.404>generate <00:39:26.638>complex, <00:39:26.871>uh, <00:39:27.105>scenes.

825
00:39:28.705 --> 00:39:31.325
<00:39:28.705>You <00:39:28.880>can <00:39:29.055>build <00:39:29.230>this <00:39:29.404>complex <00:39:29.580>scenes <00:39:29.755>like <00:39:29.930>this <00:39:30.105>one.

826
00:39:32.185 --> 00:39:35.685
<00:39:32.185>Um, <00:39:32.360>I <00:39:32.535>think <00:39:32.710>the, <00:39:32.884>the <00:39:33.060>possibilities <00:39:33.235>are <00:39:33.410>really <00:39:33.585>endless,

827
00:39:35.785 --> 00:39:37.965
<00:39:35.785>and <00:39:35.960>I'm <00:39:36.134>super <00:39:36.310>excited <00:39:36.484>to <00:39:36.660>see <00:39:36.835>what <00:39:37.009>comes <00:39:37.185>next.

828
00:39:44.245 --> 00:39:45.245
<00:39:44.245>All <00:39:45.145>right.

829
00:39:46.755 --> 00:39:47.945
<00:39:46.755>Hello <00:39:47.290>again, <00:39:47.826>Katia.

830
00:39:48.275 --> 00:39:50.145
<00:39:48.275>Thank <00:39:48.508>you <00:39:48.741>so <00:39:48.975>much <00:39:49.208>for <00:39:49.441>the <00:39:49.675>presentation.

831
00:39:50.425 --> 00:39:55.065
<00:39:50.425>I <00:39:50.705>loved <00:39:50.985>the <00:39:51.265>images. <00:39:51.545>Nice <00:39:51.825>idea.

832
00:39:55.685 --> 00:39:57.585
<00:39:55.685>Fun. <00:39:55.884>Yeah, <00:39:56.085>it's <00:39:56.285>so <00:39:56.485>much <00:39:56.685>fun <00:39:56.884>creating <00:39:57.085>them.

833
00:39:59.265 --> 00:40:01.905
<00:39:59.265>I <00:39:59.545>can <00:39:59.825>imagine. <00:40:00.105>Alright, <00:40:00.384>um, <00:40:00.665>Katya,

834
00:40:02.275 --> 00:40:04.225
<00:40:02.275>since <00:40:02.508>we <00:40:02.741>still <00:40:02.975>have <00:40:03.208>some <00:40:03.441>time <00:40:03.675>left,

835
00:40:04.435 --> 00:40:06.305
<00:40:04.435>let's <00:40:04.575>jump <00:40:04.715>into <00:40:04.855>a <00:40:04.995>quick <00:40:05.134>Q <00:40:05.275>and <00:40:05.415>a, <00:40:05.555>if <00:40:05.695>you're <00:40:05.835>ready.

836
00:40:06.965 --> 00:40:10.545
<00:40:06.965>Sounds <00:40:07.245>good. <00:40:07.525>Then <00:40:07.805>let <00:40:08.085>me <00:40:08.365>see.

837
00:40:10.915 --> 00:40:15.545
<00:40:10.915>First <00:40:11.115>question, <00:40:11.315>where <00:40:11.515>should <00:40:11.715>I <00:40:11.915>start <00:40:12.115>learning <00:40:12.315>machine

838
00:40:15.745 --> 00:40:16.985
<00:40:15.745>learning <00:40:16.117>as <00:40:16.489>a <00:40:16.861>beginner?

839
00:40:17.765 --> 00:40:20.625
<00:40:17.765>Oh, <00:40:17.964>uh, <00:40:18.165>there's <00:40:18.365>so <00:40:18.565>many <00:40:18.765>sources <00:40:18.964>there, <00:40:19.165>there's

840
00:40:20.625 --> 00:40:22.185
<00:40:20.625>so <00:40:20.975>many <00:40:21.325>possibilities <00:40:21.675>these <00:40:22.025>days.

841
00:40:22.405 --> 00:40:26.225
<00:40:22.405>Um, <00:40:22.638>I <00:40:22.871>personally, <00:40:23.105>um, <00:40:23.338>like <00:40:23.571>the, <00:40:23.805>uh,

842
00:40:26.525 --> 00:40:28.465
<00:40:26.525>the <00:40:26.805>Coursera <00:40:27.085>courses, <00:40:27.365>they're <00:40:27.645>really <00:40:27.925>good

843
00:40:28.725 --> 00:40:30.585
<00:40:28.725>or <00:40:29.075>they're <00:40:29.424>good <00:40:29.775>YouTube <00:40:30.125>materials,

844
00:40:30.585 --> 00:40:32.785
<00:40:30.585>like <00:40:30.785>NYU <00:40:30.985>has <00:40:31.185>a <00:40:31.385>good <00:40:31.585>deep <00:40:31.785>learning <00:40:31.985>course,

845
00:40:33.125 --> 00:40:36.825
<00:40:33.125>and, <00:40:33.252>uh, <00:40:33.379>of <00:40:33.506>course <00:40:33.634>not <00:40:33.761>of, <00:40:33.888>not <00:40:34.015>all <00:40:34.143>of <00:40:34.270>them <00:40:34.397>are, <00:40:34.525>um,

846
00:40:37.235 --> 00:40:38.625
<00:40:37.235>about <00:40:37.860>generative <00:40:38.486>ai,

847
00:40:38.725 --> 00:40:41.545
<00:40:38.725>but <00:40:38.852>I <00:40:38.979>think <00:40:39.106>even <00:40:39.234>if <00:40:39.361>you <00:40:39.488>want <00:40:39.615>to <00:40:39.743>work <00:40:39.870>in <00:40:39.997>generative <00:40:40.125>ai,

848
00:40:41.545 --> 00:40:42.985
<00:40:41.545>you <00:40:41.761>still <00:40:41.977>need <00:40:42.193>to <00:40:42.409>know <00:40:42.625>the <00:40:42.841>basics.

849
00:40:43.125 --> 00:40:44.745
<00:40:43.125>And <00:40:43.405>for <00:40:43.685>that, <00:40:43.965>yeah, <00:40:44.245>this, <00:40:44.525>this,

850
00:40:44.775 --> 00:40:46.345
<00:40:44.775>this <00:40:45.125>resources <00:40:45.475>are <00:40:45.825>really <00:40:46.175>good.

851
00:40:46.485 --> 00:40:49.345
<00:40:46.485>And <00:40:46.640>Nvidia, <00:40:46.796>we <00:40:46.951>also <00:40:47.107>have <00:40:47.262>our <00:40:47.418>Deep <00:40:47.573>Learning <00:40:47.729>Institute <00:40:47.885>where,

852
00:40:49.805 --> 00:40:52.425
<00:40:49.805>um, <00:40:49.960>some, <00:40:50.116>some <00:40:50.271>of <00:40:50.427>the <00:40:50.582>courses <00:40:50.738>are <00:40:50.893>for <00:40:51.049>free, <00:40:51.205>some

853
00:40:52.425 --> 00:40:54.025
<00:40:52.425>of <00:40:52.625>the <00:40:52.825>courses, <00:40:53.025>uh, <00:40:53.225>you <00:40:53.425>need <00:40:53.625>to <00:40:53.825>pay,

854
00:40:54.085 --> 00:40:56.945
<00:40:54.085>but, <00:40:54.260>uh, <00:40:54.435>it's <00:40:54.610>usually <00:40:54.785>also <00:40:54.960>very <00:40:55.135>good <00:40:55.310>content, <00:40:55.485>uh,

855
00:40:57.405 --> 00:40:59.865
<00:40:57.405>and, <00:40:57.605>uh, <00:40:57.805>yeah, <00:40:58.005>highly <00:40:58.205>recommend <00:40:58.405>also <00:40:58.605>to <00:40:58.805>check.

856
00:41:01.745 --> 00:41:06.295
<00:41:01.745>Right. <00:41:01.900>Thank <00:41:02.056>you, <00:41:02.211>Katia. <00:41:02.367>Let's <00:41:02.522>jump <00:41:02.678>to <00:41:02.833>the <00:41:02.989>next <00:41:03.145>question.

857
00:41:08.595 --> 00:41:10.295
<00:41:08.595>How <00:41:08.944>do <00:41:09.294>you <00:41:09.645>address <00:41:09.995>security

858
00:41:10.555 --> 00:41:15.095
<00:41:10.555>and <00:41:10.754>robustness <00:41:10.955>concerns <00:41:11.154>in <00:41:11.355>multimodal <00:41:11.555>generative <00:41:11.754>AI <00:41:11.955>models,

859
00:41:15.585 --> 00:41:18.855
<00:41:15.585>especially <00:41:15.865>when <00:41:16.145>applied <00:41:16.425>in <00:41:16.705>sensitive <00:41:16.985>contexts?

860
00:41:19.735 --> 00:41:20.735
<00:41:19.735>A <00:41:19.960>very, <00:41:20.185>very <00:41:20.410>good <00:41:20.635>question.

861
00:41:20.955 --> 00:41:24.135
<00:41:20.955>Um, <00:41:21.095>and <00:41:21.235>I <00:41:21.375>think <00:41:21.515>it's <00:41:21.654>important <00:41:21.795>to <00:41:21.935>rely <00:41:22.075>on <00:41:22.215>the <00:41:22.355>technology,

862
00:41:24.135 --> 00:41:25.615
<00:41:24.135>which <00:41:24.579>is <00:41:25.023>called <00:41:25.466>guardrails.

863
00:41:25.875 --> 00:41:29.575
<00:41:25.875>So <00:41:26.015>guardrails <00:41:26.155>is <00:41:26.295>something <00:41:26.435>which, <00:41:26.575>uh, <00:41:26.715>helps <00:41:26.855>you <00:41:26.995>to <00:41:27.135>check <00:41:27.275>if,

864
00:41:29.575 --> 00:41:32.055
<00:41:29.575>for <00:41:29.774>example, <00:41:29.975>the <00:41:30.174>user <00:41:30.375>is <00:41:30.575>asking <00:41:30.774>the <00:41:30.975>network

865
00:41:32.155 --> 00:41:36.125
<00:41:32.155>to <00:41:32.388>generate <00:41:32.621>something <00:41:32.855>inappropriate, <00:41:33.088>um, <00:41:33.321>the <00:41:33.555>or.

866
00:41:36.515 --> 00:41:41.245
<00:41:36.515>Yeah. <00:41:36.654>Uh, <00:41:36.795>and <00:41:36.935>it's <00:41:37.075>also <00:41:37.214>important <00:41:37.355>to <00:41:37.495>know <00:41:37.634>the <00:41:37.775>sources, <00:41:37.915>like

867
00:41:41.465 --> 00:41:43.525
<00:41:41.465>how <00:41:41.620>you <00:41:41.776>train <00:41:41.931>your <00:41:42.087>day, <00:41:42.242>how <00:41:42.398>you <00:41:42.553>train <00:41:42.709>your <00:41:42.865>network.

868
00:41:43.905 --> 00:41:46.645
<00:41:43.905>That's <00:41:44.045>why <00:41:44.185>you <00:41:44.325>need <00:41:44.465>to <00:41:44.605>be <00:41:44.745>very <00:41:44.885>careful <00:41:45.025>in <00:41:45.165>using <00:41:45.305>networks,

869
00:41:46.695 --> 00:41:48.205
<00:41:46.695>which <00:41:47.034>you <00:41:47.374>download <00:41:47.714>from <00:41:48.054>web.

870
00:41:48.385 --> 00:41:51.205
<00:41:48.385>Uh, <00:41:48.585>and, <00:41:48.785>uh, <00:41:48.985>of <00:41:49.185>course <00:41:49.385>for <00:41:49.585>commercial <00:41:49.785>applications,

871
00:41:51.275 --> 00:41:53.845
<00:41:51.275>it's <00:41:51.508>crucial <00:41:51.741>to <00:41:51.975>really <00:41:52.208>know <00:41:52.441>your <00:41:52.675>data.

872
00:41:53.985 --> 00:41:57.205
<00:41:53.985>And <00:41:54.140>of <00:41:54.296>course, <00:41:54.451>with <00:41:54.607>400 <00:41:54.762>million <00:41:54.918>pair, <00:41:55.073>it's <00:41:55.229>really <00:41:55.385>hard

873
00:41:57.205 --> 00:41:59.205
<00:41:57.205>to <00:41:57.555>check, <00:41:57.904>but, <00:41:58.255>uh, <00:41:58.605>it's,

874
00:41:59.315 --> 00:42:01.525
<00:41:59.315>it's <00:41:59.548>highly <00:41:59.781>advised <00:42:00.015>to <00:42:00.248>invest <00:42:00.481>in <00:42:00.715>that.

875
00:42:04.465 --> 00:42:08.085
<00:42:04.465>Thanks, <00:42:04.745>Katya. <00:42:05.025>We <00:42:05.305>got <00:42:05.585>another <00:42:05.865>question.

876
00:42:10.075 --> 00:42:11.615
<00:42:10.075>Our <00:42:10.768>dev <00:42:11.461>strategies

877
00:42:11.795 --> 00:42:14.775
<00:42:11.795>for <00:42:12.075>mitigating <00:42:12.355>potential <00:42:12.635>risks <00:42:12.915>in <00:42:13.195>the

878
00:42:14.775 --> 00:42:15.935
<00:42:14.775>context <00:42:15.123>of <00:42:15.471>these <00:42:15.819>models.

879
00:42:16.775 --> 00:42:18.655
<00:42:16.775>Hmm. <00:42:16.975>So <00:42:17.175>I <00:42:17.375>think <00:42:17.575>I've <00:42:17.775>already <00:42:17.975>mentioned <00:42:18.175>that

880
00:42:18.725 --> 00:42:20.295
<00:42:18.725>with <00:42:19.075>like <00:42:19.424>knowing <00:42:19.775>your <00:42:20.125>data

881
00:42:20.715 --> 00:42:23.655
<00:42:20.715>and <00:42:20.948>applying <00:42:21.181>guardrails <00:42:21.415>like <00:42:21.648>this, <00:42:21.881>this, <00:42:22.115>uh,

882
00:42:23.655 --> 00:42:26.405
<00:42:23.655>the <00:42:23.888>best <00:42:24.121>strategies, <00:42:24.355>um, <00:42:24.588>so <00:42:24.821>far <00:42:25.055>and,

883
00:42:26.625 --> 00:42:29.685
<00:42:26.625>and <00:42:26.780>knowing <00:42:26.936>the <00:42:27.091>where <00:42:27.247>the <00:42:27.402>models <00:42:27.558>come <00:42:27.713>from, <00:42:27.869>make <00:42:28.025>sure

884
00:42:29.685 --> 00:42:32.045
<00:42:29.685>that <00:42:29.840>you <00:42:29.996>have <00:42:30.151>the <00:42:30.307>legal <00:42:30.462>rights <00:42:30.618>to <00:42:30.773>use <00:42:30.929>your <00:42:31.085>models.

885
00:42:32.395 --> 00:42:34.205
<00:42:32.395>It's <00:42:32.745>also <00:42:33.095>equally <00:42:33.445>important. <00:42:33.795>Uh,

886
00:42:38.505 --> 00:42:39.965
<00:42:38.505>Thanks <00:42:38.767>for <00:42:39.030>that <00:42:39.293>answer <00:42:39.556>as <00:42:39.819>well.

887
00:42:40.375 --> 00:42:41.565
<00:42:40.375>Let's <00:42:40.589>jump <00:42:40.803>to <00:42:41.017>the <00:42:41.231>next <00:42:41.446>one.

888
00:42:42.945 --> 00:42:45.395
<00:42:42.945>What <00:42:43.145>do <00:42:43.345>you <00:42:43.545>see <00:42:43.745>as <00:42:43.945>the <00:42:44.145>future <00:42:44.345>trends

889
00:42:45.455 --> 00:42:49.995
<00:42:45.455>and <00:42:45.688>potential <00:42:45.921>developments, <00:42:46.154>uh, <00:42:46.388>multimodal <00:42:46.621>generative <00:42:46.855>ai?

890
00:42:50.755 --> 00:42:53.075
<00:42:50.755>Hmm. <00:42:51.105>So <00:42:51.455>in <00:42:51.805>my <00:42:52.155>presentation,

891
00:42:53.155 --> 00:42:55.115
<00:42:53.155>I <00:42:53.388>gave <00:42:53.621>you <00:42:53.855>already <00:42:54.088>some, <00:42:54.321>uh, <00:42:54.555>hints.

892
00:42:55.265 --> 00:42:57.555
<00:42:55.265>Like, <00:42:55.545>uh, <00:42:55.825>we'll <00:42:56.105>do <00:42:56.384>video <00:42:56.665>generation,

893
00:42:57.695 --> 00:42:59.915
<00:42:57.695>we <00:42:57.928>will <00:42:58.161>do <00:42:58.395>three <00:42:58.628>d <00:42:58.861>uh, <00:42:59.095>generation.

894
00:43:00.385 --> 00:43:03.035
<00:43:00.385>This <00:43:00.585>is <00:43:00.785>kind <00:43:00.985>of <00:43:01.185>work <00:43:01.385>in <00:43:01.585>progress, <00:43:01.785>uh,

895
00:43:03.095 --> 00:43:06.315
<00:43:03.095>and, <00:43:03.270>uh, <00:43:03.444>our <00:43:03.620>research <00:43:03.794>has <00:43:03.970>demonstrated <00:43:04.145>already <00:43:04.319>good <00:43:04.495>results,

896
00:43:06.535 --> 00:43:09.155
<00:43:06.535>but <00:43:06.768>in <00:43:07.001>order <00:43:07.234>to <00:43:07.468>make <00:43:07.701>it <00:43:07.935>robuster,

897
00:43:09.355 --> 00:43:11.555
<00:43:09.355>I <00:43:09.530>think <00:43:09.705>we <00:43:09.880>still <00:43:10.055>need <00:43:10.230>a <00:43:10.405>couple <00:43:10.580>of <00:43:10.755>iterations,

898
00:43:11.735 --> 00:43:12.915
<00:43:11.735>uh, <00:43:12.000>on <00:43:12.266>working <00:43:12.531>on <00:43:12.797>that.

899
00:43:13.495 --> 00:43:15.635
<00:43:13.495>Um, <00:43:13.728>what <00:43:13.961>I <00:43:14.194>also <00:43:14.428>see <00:43:14.661>that, <00:43:14.895>uh,

900
00:43:15.635 --> 00:43:18.635
<00:43:15.635>generative <00:43:15.868>AI <00:43:16.101>is <00:43:16.335>coming <00:43:16.568>into <00:43:16.801>video <00:43:17.035>processing,

901
00:43:18.855 --> 00:43:22.315
<00:43:18.855>and, <00:43:18.962>uh, <00:43:19.070>I <00:43:19.178>think <00:43:19.285>in <00:43:19.393>the, <00:43:19.501>in <00:43:19.608>the <00:43:19.716>similar <00:43:19.824>way, <00:43:19.931>we <00:43:20.039>are <00:43:20.147>able <00:43:20.255>to

902
00:43:23.215 --> 00:43:26.355
<00:43:23.215>ask <00:43:23.355>chat <00:43:23.495>GPT <00:43:23.635>everything <00:43:23.775>about <00:43:23.915>text, <00:43:24.055>text, <00:43:24.195>we'll <00:43:24.335>be <00:43:24.475>also <00:43:24.615>able

903
00:43:26.375 --> 00:43:30.675
<00:43:26.375>to <00:43:26.530>do <00:43:26.686>that <00:43:26.841>with <00:43:26.997>videos <00:43:27.152>like, <00:43:27.308>uh, <00:43:27.463>imagine <00:43:27.619>large <00:43:27.775>videos,

904
00:43:30.975 --> 00:43:34.115
<00:43:30.975>and <00:43:31.174>then <00:43:31.375>you <00:43:31.575>ask <00:43:31.775>like <00:43:31.975>what, <00:43:32.174>uh, <00:43:32.375>this

905
00:43:34.635 --> 00:43:37.555
<00:43:34.635>BBC <00:43:34.868>reporter <00:43:35.101>was <00:43:35.335>wearing <00:43:35.568>on <00:43:35.801>that <00:43:36.035>day.

906
00:43:37.815 --> 00:43:41.435
<00:43:37.815>Um, <00:43:37.922>I <00:43:38.030>think <00:43:38.138>in <00:43:38.245>the <00:43:38.353>future <00:43:38.461>we'll <00:43:38.568>be <00:43:38.676>able <00:43:38.784>to <00:43:38.891>do <00:43:38.999>that <00:43:39.107>as <00:43:39.215>well.

907
00:43:44.555 --> 00:43:47.705
<00:43:44.555>Right. <00:43:44.788>Thanks <00:43:45.021>for <00:43:45.254>that <00:43:45.488>outlook <00:43:45.721>of <00:43:45.955>yours.

908
00:43:48.605 --> 00:43:49.905
<00:43:48.605>We <00:43:48.897>got <00:43:49.190>some <00:43:49.482>more <00:43:49.775>questions.

909
00:43:52.055 --> 00:43:53.305
<00:43:52.055>What <00:43:53.180>limitations

910
00:43:53.445 --> 00:43:57.105
<00:43:53.445>or <00:43:53.645>ethical <00:43:53.845>concerns <00:43:54.045>should <00:43:54.245>be <00:43:54.445>considered <00:43:54.645>when <00:43:54.845>using

911
00:43:57.175 --> 00:43:58.625
<00:43:57.175>this <00:43:57.501>mo <00:43:57.827>these <00:43:58.153>models? <00:43:58.480>Mm-Hmm.

912
00:43:59.005 --> 00:44:00.665
<00:43:59.005>So <00:43:59.471>the <00:43:59.938>models <00:44:00.405>hallucinate

913
00:44:00.765 --> 00:44:05.585
<00:44:00.765>and, <00:44:00.940>uh, <00:44:01.115>sometimes <00:44:01.290>they <00:44:01.464>create <00:44:01.640>something, <00:44:01.815>uh, <00:44:01.990>well, <00:44:02.165>weird.

914
00:44:06.525 --> 00:44:08.945
<00:44:06.525>Uh, <00:44:06.758>this <00:44:06.991>is <00:44:07.225>one <00:44:07.458>issue <00:44:07.691>then, <00:44:07.925>um,

915
00:44:09.615 --> 00:44:11.385
<00:44:09.615>they <00:44:09.848>are <00:44:10.081>not, <00:44:10.314>uh, <00:44:10.548>free <00:44:10.781>of <00:44:11.015>bias.

916
00:44:11.805 --> 00:44:16.425
<00:44:11.805>And, <00:44:11.932>uh, <00:44:12.059>as <00:44:12.186>in <00:44:12.314>the <00:44:12.441>in <00:44:12.568>network, <00:44:12.695>uh, <00:44:12.823>the <00:44:12.950>edify <00:44:13.077>I've <00:44:13.205>presented

917
00:44:16.425 --> 00:44:19.185
<00:44:16.425>to <00:44:16.600>you, <00:44:16.775>uh, <00:44:16.950>we <00:44:17.125>actually <00:44:17.300>working <00:44:17.475>hard <00:44:17.650>on <00:44:17.825>trying

918
00:44:19.185 --> 00:44:20.345
<00:44:19.185>to <00:44:19.533>eliminate <00:44:19.881>this <00:44:20.229>bias.

919
00:44:20.535 --> 00:44:22.945
<00:44:20.535>Like, <00:44:20.768>for <00:44:21.001>example, <00:44:21.234>somebody <00:44:21.468>asks <00:44:21.701>like, <00:44:21.935>Hey,

920
00:44:23.085 --> 00:44:25.585
<00:44:23.085>can <00:44:23.285>you <00:44:23.485>generate <00:44:23.685>me <00:44:23.885>an <00:44:24.085>image <00:44:24.285>of <00:44:24.485>ACEO?

921
00:44:26.125 --> 00:44:29.145
<00:44:26.125>And <00:44:26.280>then <00:44:26.436>what <00:44:26.591>the <00:44:26.747>typical <00:44:26.902>network <00:44:27.058>crawled <00:44:27.213>from <00:44:27.369>the <00:44:27.525>internet

922
00:44:29.145 --> 00:44:31.605
<00:44:29.145>trained <00:44:29.285>on <00:44:29.425>the <00:44:29.565>data <00:44:29.705>crawled <00:44:29.845>from <00:44:29.985>the <00:44:30.125>internet <00:44:30.265>would <00:44:30.405>give <00:44:30.545>you

923
00:44:31.635 --> 00:44:34.485
<00:44:31.635>like <00:44:31.810>a <00:44:31.985>bunch <00:44:32.160>of <00:44:32.335>guys, <00:44:32.510>white <00:44:32.685>guys <00:44:32.860>in <00:44:33.035>suits.

924
00:44:34.585 --> 00:44:37.165
<00:44:34.585>And, <00:44:34.725>uh, <00:44:34.865>of <00:44:35.005>course, <00:44:35.145>that's <00:44:35.285>not <00:44:35.425>what <00:44:35.565>you <00:44:35.705>want <00:44:35.845>to <00:44:35.985>have.

925
00:44:37.465 --> 00:44:38.725
<00:44:37.465>And, <00:44:37.748>uh, <00:44:38.032>I <00:44:38.315>think <00:44:38.599>this,

926
00:44:38.725 --> 00:44:40.925
<00:44:38.725>these <00:44:39.005>are <00:44:39.285>very, <00:44:39.565>very <00:44:39.845>important <00:44:40.125>things.

927
00:44:41.225 --> 00:44:42.765
<00:44:41.225>The <00:44:41.571>second <00:44:41.917>thing <00:44:42.264>is <00:44:42.611>attribution.

928
00:44:42.905 --> 00:44:46.165
<00:44:42.905>Uh, <00:44:43.045>like <00:44:43.185>all <00:44:43.325>these <00:44:43.465>artists <00:44:43.605>who <00:44:43.745>publish <00:44:43.885>their <00:44:44.025>works <00:44:44.165>on, <00:44:44.305>uh,

929
00:44:46.795 --> 00:44:49.485
<00:44:46.795>well, <00:44:47.261>stock <00:44:47.728>images <00:44:48.195>websites

930
00:44:49.585 --> 00:44:51.685
<00:44:49.585>or <00:44:49.785>whatsoever, <00:44:49.985>they <00:44:50.185>need <00:44:50.385>to <00:44:50.585>get <00:44:50.785>the <00:44:50.985>attribution

931
00:44:51.685 --> 00:44:55.165
<00:44:51.685>because <00:44:51.825>they <00:44:51.965>art, <00:44:52.105>uh, <00:44:52.245>has <00:44:52.384>been <00:44:52.525>used <00:44:52.665>to <00:44:52.805>train <00:44:52.945>these <00:44:53.085>networks.

932
00:44:55.165 --> 00:44:56.445
<00:44:55.165>And <00:44:55.395>now <00:44:55.625>it's <00:44:55.856>used <00:44:56.086>to <00:44:56.317>generate,

933
00:44:56.665 --> 00:44:59.845
<00:44:56.665>and <00:44:56.865>sometimes, <00:44:57.065>uh, <00:44:57.265>situations <00:44:57.465>happen <00:44:57.665>that <00:44:57.865>artists, <00:44:58.065>they

934
00:45:00.485 --> 00:45:02.845
<00:45:00.485>recognize <00:45:00.718>their <00:45:00.951>artistic <00:45:01.185>style <00:45:01.418>in <00:45:01.651>these <00:45:01.885>networks,

935
00:45:02.905 --> 00:45:05.085
<00:45:02.905>and, <00:45:03.138>uh, <00:45:03.371>of <00:45:03.605>course <00:45:03.838>it <00:45:04.071>causes <00:45:04.305>trouble.

936
00:45:05.225 --> 00:45:06.805
<00:45:05.225>So <00:45:05.691>this <00:45:06.158>one <00:45:06.625>has

937
00:45:06.805 --> 00:45:08.925
<00:45:06.805>to <00:45:07.038>be <00:45:07.271>particularly <00:45:07.504>careful <00:45:07.738>about <00:45:07.971>these <00:45:08.205>things.

938
00:45:10.855 --> 00:45:13.555
<00:45:10.855>Yes, <00:45:11.135>for <00:45:11.415>sure. <00:45:11.695>Um, <00:45:11.975>maybe <00:45:12.255>a

939
00:45:14.155 --> 00:45:15.755
<00:45:14.155>personal <00:45:14.435>follow <00:45:14.715>up <00:45:14.995>question <00:45:15.275>from <00:45:15.555>me.

940
00:45:15.755 --> 00:45:19.915
<00:45:15.755>Maybe <00:45:15.910>you <00:45:16.066>happen <00:45:16.221>to <00:45:16.377>know <00:45:16.532>who <00:45:16.688>is <00:45:16.843>legally <00:45:16.999>responsible <00:45:17.155>when,

941
00:45:20.055 --> 00:45:24.275
<00:45:20.055>uh, <00:45:20.194>something <00:45:20.335>illegal <00:45:20.475>is, <00:45:20.615>uh, <00:45:20.754>generated <00:45:20.895>by <00:45:21.035>the <00:45:21.174>ai, <00:45:21.315>like <00:45:21.455>the,

942
00:45:24.575 --> 00:45:28.635
<00:45:24.575>the <00:45:24.750>user <00:45:24.924>or <00:45:25.100>the <00:45:25.274>company <00:45:25.450>which <00:45:25.625>provided <00:45:25.799>the <00:45:25.975>model,

943
00:45:29.175 --> 00:45:30.915
<00:45:29.175>Uh, <00:45:29.408>the <00:45:29.641>company <00:45:29.875>which <00:45:30.108>provided <00:45:30.341>the <00:45:30.575>model,

944
00:45:31.135 --> 00:45:32.135
<00:45:31.135>The <00:45:31.435>company. <00:45:31.735>Okay. <00:45:32.035>Mm-Hmm.

945
00:45:32.135 --> 00:45:36.395
<00:45:32.135>Thank <00:45:32.368>you. <00:45:32.601>We <00:45:32.835>got <00:45:33.068>some <00:45:33.301>more <00:45:33.535>questions.

946
00:45:37.545 --> 00:45:40.955
<00:45:37.545>What <00:45:37.720>is <00:45:37.895>the <00:45:38.070>best <00:45:38.245>programming <00:45:38.420>language <00:45:38.595>for <00:45:38.770>AI <00:45:38.945>development?

947
00:45:41.835 --> 00:45:44.375
<00:45:41.835>Um, <00:45:42.185>that's <00:45:42.535>an <00:45:42.885>interesting <00:45:43.235>question.

948
00:45:44.595 --> 00:45:48.735
<00:45:44.595>So, <00:45:44.750>uh, <00:45:44.906>these <00:45:45.061>days <00:45:45.217>most <00:45:45.372>of <00:45:45.528>AI <00:45:45.683>developers <00:45:45.839>use <00:45:45.995>Python.

949
00:45:49.635 --> 00:45:52.855
<00:45:49.635>Uh, <00:45:49.790>but, <00:45:49.946>uh, <00:45:50.101>Python <00:45:50.257>is <00:45:50.412>very <00:45:50.568>good <00:45:50.723>for <00:45:50.879>training, <00:45:51.035>uh,

950
00:45:52.855 --> 00:45:55.175
<00:45:52.855>because <00:45:53.055>you <00:45:53.255>can <00:45:53.455>also <00:45:53.655>prototype <00:45:53.855>quickly <00:45:54.055>with <00:45:54.255>Python.

951
00:45:55.805 --> 00:45:59.135
<00:45:55.805>When <00:45:56.004>it <00:45:56.205>comes <00:45:56.404>to <00:45:56.605>deployment, <00:45:56.805>uh, <00:45:57.004>it's <00:45:57.205>different.

952
00:45:59.315 --> 00:46:02.975
<00:45:59.315>Uh, <00:45:59.431>we <00:45:59.548>have <00:45:59.665>also <00:45:59.781>things <00:45:59.898>like <00:46:00.015>JAKs, <00:46:00.131>uh, <00:46:00.248>we <00:46:00.365>have <00:46:00.481>c <00:46:00.598>plus <00:46:00.715>plus.

953
00:46:03.355 --> 00:46:07.735
<00:46:03.355>And, <00:46:03.510>uh, <00:46:03.666>for, <00:46:03.821>for <00:46:03.977>example, <00:46:04.132>for <00:46:04.288>robotic <00:46:04.443>applications, <00:46:04.599>I <00:46:04.755>think

954
00:46:07.735 --> 00:46:09.615
<00:46:07.735>that <00:46:07.935>nothing <00:46:08.135>is <00:46:08.335>better <00:46:08.535>than <00:46:08.735>c <00:46:08.935>plus <00:46:09.135>plus,

955
00:46:09.705 --> 00:46:12.215
<00:46:09.705>where <00:46:09.985>you <00:46:10.265>basically <00:46:10.545>go <00:46:10.825>like <00:46:11.105>very,

956
00:46:12.635 --> 00:46:14.415
<00:46:12.635>you <00:46:12.915>work <00:46:13.195>very <00:46:13.475>closely <00:46:13.755>with <00:46:14.035>hardware.

957
00:46:14.635 --> 00:46:17.335
<00:46:14.635>And, <00:46:14.985>uh, <00:46:15.335>the <00:46:15.685>question <00:46:16.035>is,

958
00:46:17.335 --> 00:46:19.055
<00:46:17.335>it <00:46:17.490>depends <00:46:17.646>on <00:46:17.801>which <00:46:17.957>side <00:46:18.112>do <00:46:18.268>you <00:46:18.423>want <00:46:18.579>to <00:46:18.735>work.

959
00:46:19.155 --> 00:46:20.215
<00:46:19.155>If <00:46:19.393>you <00:46:19.632>want <00:46:19.870>to <00:46:20.109>develop

960
00:46:20.355 --> 00:46:21.375
<00:46:20.355>and <00:46:20.661>train <00:46:20.967>your <00:46:21.273>networks,

961
00:46:21.405 --> 00:46:23.855
<00:46:21.405>sometimes <00:46:21.605>you <00:46:21.805>don't <00:46:22.005>even <00:46:22.205>need <00:46:22.405>a <00:46:22.605>programming <00:46:22.805>language

962
00:46:23.855 --> 00:46:26.695
<00:46:23.855>because, <00:46:24.030>uh, <00:46:24.205>there <00:46:24.380>are <00:46:24.555>already <00:46:24.730>so <00:46:24.905>many <00:46:25.080>frameworks, <00:46:25.255>there's

963
00:46:26.695 --> 00:46:30.455
<00:46:26.695>so <00:46:26.835>many <00:46:26.975>tools <00:46:27.115>which <00:46:27.255>you <00:46:27.395>can <00:46:27.535>use <00:46:27.675>without <00:46:27.815>writing <00:46:27.955>even <00:46:28.095>a

964
00:46:30.455 --> 00:46:31.455
<00:46:30.455>single <00:46:30.755>line <00:46:31.055>of <00:46:31.355>quotes.

965
00:46:31.455 --> 00:46:33.895
<00:46:31.455>Like <00:46:31.595>we <00:46:31.735>had <00:46:31.875>Nvidia, <00:46:32.015>we <00:46:32.154>develop <00:46:32.295>our <00:46:32.435>TAO <00:46:32.575>toolkit, <00:46:32.715>for <00:46:32.855>example,

966
00:46:33.995 --> 00:46:37.695
<00:46:33.995>for <00:46:34.228>image <00:46:34.461>training, <00:46:34.694>um, <00:46:34.928>for <00:46:35.161>language <00:46:35.395>models,

967
00:46:37.695 --> 00:46:39.735
<00:46:37.695>we <00:46:37.975>have <00:46:38.255>our <00:46:38.535>NIMO <00:46:38.815>framework, <00:46:39.095>uh,

968
00:46:39.745 --> 00:46:41.695
<00:46:39.745>which <00:46:40.025>is <00:46:40.305>a <00:46:40.585>Python, <00:46:40.865>uh, <00:46:41.145>framework.

969
00:46:42.035 --> 00:46:43.615
<00:46:42.035>Uh, <00:46:42.384>so <00:46:42.734>it's, <00:46:43.085>it <00:46:43.435>depends.

970
00:46:43.795 --> 00:46:46.015
<00:46:43.795>But <00:46:44.075>if <00:46:44.355>you <00:46:44.635>are <00:46:44.915>just <00:46:45.195>starting,

971
00:46:46.195 --> 00:46:48.015
<00:46:46.195>if <00:46:46.395>you're <00:46:46.595>just <00:46:46.795>in <00:46:46.995>the <00:46:47.195>beginning, <00:46:47.395>then, <00:46:47.595>uh,

972
00:46:48.175 --> 00:46:49.695
<00:46:48.175>I <00:46:48.448>would <00:46:48.722>just <00:46:48.995>say <00:46:49.269>try <00:46:49.543>Python

973
00:46:49.695 --> 00:46:52.615
<00:46:49.695>because <00:46:49.850>it's <00:46:50.006>so <00:46:50.161>easy, <00:46:50.317>it's <00:46:50.472>so <00:46:50.628>easy <00:46:50.783>to <00:46:50.939>learn, <00:46:51.095>uh,

974
00:46:52.615 --> 00:46:54.175
<00:46:52.615>very <00:46:53.081>intuitive, <00:46:53.548>and, <00:46:54.015>uh,

975
00:46:54.725 --> 00:46:57.015
<00:46:54.725>it's <00:46:54.924>being <00:46:55.125>widely <00:46:55.325>used <00:46:55.525>in <00:46:55.725>data <00:46:55.924>science <00:46:56.125>community.

976
00:46:59.155 --> 00:47:02.975
<00:46:59.155>Thank <00:46:59.505>you, <00:46:59.855>Katia. <00:47:00.205>Next <00:47:00.555>one.

977
00:47:03.755 --> 00:47:07.735
<00:47:03.755>Is <00:47:03.895>it <00:47:04.035>difficult <00:47:04.175>to <00:47:04.315>get <00:47:04.455>an <00:47:04.595>entry <00:47:04.735>level <00:47:04.875>job <00:47:05.015>in <00:47:05.155>ai?

978
00:47:08.395 --> 00:47:11.175
<00:47:08.395>Um, <00:47:08.628>I <00:47:08.861>would <00:47:09.095>say <00:47:09.328>it's <00:47:09.561>not <00:47:09.795>difficult

979
00:47:11.175 --> 00:47:12.935
<00:47:11.175>because <00:47:11.455>it's <00:47:11.735>such <00:47:12.015>a <00:47:12.295>hot <00:47:12.575>topic.

980
00:47:13.085 --> 00:47:16.255
<00:47:13.085>It's, <00:47:13.285>it's, <00:47:13.485>uh, <00:47:13.685>everyone <00:47:13.885>is <00:47:14.085>searching <00:47:14.285>in <00:47:14.485>search

981
00:47:16.255 --> 00:47:18.175
<00:47:16.255>for <00:47:16.605>good <00:47:16.955>specialists <00:47:17.305>in <00:47:17.655>ai.

982
00:47:18.315 --> 00:47:22.455
<00:47:18.315>And, <00:47:18.470>uh, <00:47:18.626>if <00:47:18.781>you <00:47:18.937>do <00:47:19.092>some <00:47:19.248>experimentation <00:47:19.403>with <00:47:19.559>AI <00:47:19.715>already,

983
00:47:22.475 --> 00:47:26.135
<00:47:22.475>if <00:47:22.630>you <00:47:22.786>have <00:47:22.941>some <00:47:23.097>hands-on <00:47:23.252>experience, <00:47:23.408>uh, <00:47:23.563>it'll <00:47:23.719>be, <00:47:23.875>uh,

984
00:47:26.395 --> 00:47:28.015
<00:47:26.395>fairly <00:47:26.595>easy <00:47:26.795>for <00:47:26.995>you <00:47:27.195>to <00:47:27.395>get <00:47:27.595>a <00:47:27.795>job.

985
00:47:28.435 --> 00:47:32.245
<00:47:28.435>And <00:47:28.562>of <00:47:28.689>course <00:47:28.816>you <00:47:28.944>need <00:47:29.071>to <00:47:29.198>be <00:47:29.325>able <00:47:29.453>to <00:47:29.580>demonstrate, <00:47:29.707>uh, <00:47:29.835>uh,

986
00:47:32.245 --> 00:47:34.605
<00:47:32.245>that, <00:47:32.478>that <00:47:32.711>you <00:47:32.944>have <00:47:33.178>this <00:47:33.411>capability, <00:47:33.645>oh,

987
00:47:34.605 --> 00:47:38.405
<00:47:34.605>it's <00:47:34.745>getting <00:47:34.885>really <00:47:35.025>sunny <00:47:35.165>here, <00:47:35.305>um, <00:47:35.445>that, <00:47:35.585>uh, <00:47:35.725>you <00:47:35.865>need <00:47:36.005>to,

988
00:47:38.465 --> 00:47:40.845
<00:47:38.465>for <00:47:38.745>example, <00:47:39.025>have <00:47:39.305>some <00:47:39.585>GitHub <00:47:39.865>projects

989
00:47:40.945 --> 00:47:43.445
<00:47:40.945>or <00:47:41.225>do <00:47:41.505>some <00:47:41.785>Kaggle, <00:47:42.065>uh, <00:47:42.345>projects.

990
00:47:43.665 --> 00:47:44.805
<00:47:43.665>Uh, <00:47:43.921>it <00:47:44.178>really <00:47:44.434>helps <00:47:44.691>you,

991
00:47:47.305 --> 00:47:48.305
<00:47:47.305>But <00:47:47.754>thank <00:47:48.205>you.

992
00:47:48.795 --> 00:47:51.445
<00:47:48.795>Okay, <00:47:48.970>so <00:47:49.145>you <00:47:49.320>got <00:47:49.495>all <00:47:49.670>the <00:47:49.845>sun <00:47:50.020>in <00:47:50.195>Munich.

993
00:47:52.425 --> 00:47:54.125
<00:47:52.425>Oh, <00:47:53.125>that's <00:47:53.825>after

994
00:47:54.525 --> 00:47:55.525
<00:47:54.525>I <00:47:54.705>look <00:47:54.885>out <00:47:55.065>the <00:47:55.245>window. <00:47:55.425>It's

995
00:47:55.525 --> 00:47:58.365
<00:47:55.525>gray <00:47:55.991>and <00:47:56.458>rainy. <00:47:56.925>Gosh.

996
00:48:00.555 --> 00:48:03.485
<00:48:00.555>Alright, <00:48:01.021>what's <00:48:01.488>next <00:48:01.955>question.

997
00:48:04.825 --> 00:48:09.525
<00:48:04.825>Can <00:48:05.024>you <00:48:05.225>recommend <00:48:05.424>smaller <00:48:05.625>l <00:48:05.825>LLMs <00:48:06.024>to <00:48:06.225>experiment

998
00:48:09.525 --> 00:48:11.125
<00:48:09.525>with <00:48:10.225>consumer <00:48:10.925>GPOs?

999
00:48:11.155 --> 00:48:14.645
<00:48:11.155>Like <00:48:11.310>for <00:48:11.466>example, <00:48:11.621>the <00:48:11.777>30 <00:48:11.932>or <00:48:12.088>40 <00:48:12.243>series <00:48:12.399>from <00:48:12.555>Nvidia?

1000
00:48:15.195 --> 00:48:18.525
<00:48:15.195>Yeah. <00:48:15.395>Um, <00:48:15.595>they're, <00:48:15.795>they're <00:48:15.995>smaller <00:48:16.195>LLMs, <00:48:16.395>uh, <00:48:16.595>which

1001
00:48:18.525 --> 00:48:22.485
<00:48:18.525>for <00:48:18.680>example, <00:48:18.836>uh, <00:48:18.991>have <00:48:19.147>been <00:48:19.302>trained <00:48:19.458>from <00:48:19.613>larger <00:48:19.769>lms, <00:48:19.925>um,

1002
00:48:22.865 --> 00:48:25.925
<00:48:22.865>you <00:48:23.064>know, <00:48:23.265>popular <00:48:23.464>architecture <00:48:23.665>like <00:48:23.865>LAMA <00:48:24.064>from <00:48:24.265>Facebook,

1003
00:48:26.305 --> 00:48:28.405
<00:48:26.305>it <00:48:26.538>has <00:48:26.771>something <00:48:27.004>like <00:48:27.238>70 <00:48:27.471>billion <00:48:27.705>parameters,

1004
00:48:28.465 --> 00:48:31.525
<00:48:28.465>but <00:48:28.640>they <00:48:28.815>have <00:48:28.990>recently, <00:48:29.165>uh, <00:48:29.340>uh, <00:48:29.515>published <00:48:29.690>a <00:48:29.865>model

1005
00:48:31.525 --> 00:48:33.245
<00:48:31.525>with <00:48:31.991>7 <00:48:32.458>billion <00:48:32.925>parameters.

1006
00:48:34.145 --> 00:48:36.445
<00:48:34.145>Um, <00:48:34.495>they <00:48:34.845>are <00:48:35.195>mistral <00:48:35.545>model,

1007
00:48:36.445 --> 00:48:38.325
<00:48:36.445>which <00:48:36.725>also <00:48:37.005>has <00:48:37.285>7 <00:48:37.565>billion <00:48:37.845>parameters.

1008
00:48:38.385 --> 00:48:40.045
<00:48:38.385>So <00:48:38.618>people <00:48:38.851>say <00:48:39.085>good <00:48:39.318>things <00:48:39.551>about <00:48:39.785>it.

1009
00:48:40.145 --> 00:48:42.605
<00:48:40.145>So <00:48:40.425>yeah, <00:48:40.705>uh, <00:48:40.985>check, <00:48:41.265>check, <00:48:41.545>check.

1010
00:48:42.605 --> 00:48:47.325
<00:48:42.605>What <00:48:42.838>is <00:48:43.071>the, <00:48:43.305>um, <00:48:43.538>legal, <00:48:43.771>uh, <00:48:44.005>well,

1011
00:48:47.325 --> 00:48:49.685
<00:48:47.325>what <00:48:47.464>are <00:48:47.605>the <00:48:47.745>legal <00:48:47.884>rights <00:48:48.024>of <00:48:48.165>using <00:48:48.305>this <00:48:48.444>models, <00:48:48.585>of <00:48:48.725>course,

1012
00:48:49.865 --> 00:48:51.925
<00:48:49.865>but <00:48:50.020>I <00:48:50.176>think <00:48:50.331>for <00:48:50.487>Lama, <00:48:50.642>it's, <00:48:50.798>it's <00:48:50.953>not <00:48:51.109>a <00:48:51.265>problem.

1013
00:48:52.265 --> 00:48:55.645
<00:48:52.265>And, <00:48:52.440>um, <00:48:52.615>even <00:48:52.790>I, <00:48:52.964>I <00:48:53.140>even <00:48:53.315>saw <00:48:53.490>people, <00:48:53.665>uh,

1014
00:48:55.755 --> 00:48:57.685
<00:48:55.755>deploying <00:48:56.221>LAMA <00:48:56.688>on <00:48:57.155>Jetson

1015
00:48:57.745 --> 00:49:00.605
<00:48:57.745>or <00:48:57.920>device <00:48:58.095>Jetson <00:48:58.270>is <00:48:58.444>our <00:48:58.620>small <00:48:58.795>embedded, <00:48:58.970>uh, <00:48:59.145>computer,

1016
00:49:00.605 --> 00:49:04.765
<00:49:00.605>which <00:49:00.745>is, <00:49:00.885>uh, <00:49:01.025>which <00:49:01.165>is <00:49:01.305>even <00:49:01.445>less <00:49:01.585>well <00:49:01.725>now, <00:49:01.865>I <00:49:02.005>dunno

1017
00:49:04.765 --> 00:49:06.765
<00:49:04.765>how <00:49:04.964>it <00:49:05.165>compares <00:49:05.365>with <00:49:05.565>30 <00:49:05.765>and <00:49:05.964>40 <00:49:06.165>series.

1018
00:49:06.905 --> 00:49:10.245
<00:49:06.905>But, <00:49:07.080>uh, <00:49:07.255>it's, <00:49:07.430>uh, <00:49:07.605>it's <00:49:07.780>a <00:49:07.955>small, <00:49:08.130>small <00:49:08.305>device

1019
00:49:10.425 --> 00:49:13.525
<00:49:10.425>and <00:49:10.580>you <00:49:10.736>can <00:49:10.891>already <00:49:11.047>deploy <00:49:11.202>generative <00:49:11.358>AI <00:49:11.513>devices <00:49:11.669>like <00:49:11.825>that.

1020
00:49:15.755 --> 00:49:18.335
<00:49:15.755>That's <00:49:15.988>nice <00:49:16.221>to <00:49:16.455>hear. <00:49:16.688>Thanks. <00:49:16.921>And <00:49:17.155>the

1021
00:49:18.335 --> 00:49:21.455
<00:49:18.335>Models, <00:49:18.510>like <00:49:18.685>the <00:49:18.860>models <00:49:19.035>like <00:49:19.210>diffusion, <00:49:19.385>they <00:49:19.560>are <00:49:19.735>also,

1022
00:49:22.005 --> 00:49:23.535
<00:49:22.005>they <00:49:22.280>are, <00:49:22.555>they're <00:49:22.831>kind <00:49:23.106>of <00:49:23.382>lightweight.

1023
00:49:23.675 --> 00:49:25.735
<00:49:23.675>Uh, <00:49:23.875>so <00:49:24.075>the, <00:49:24.275>for <00:49:24.475>example, <00:49:24.675>as <00:49:24.875>I <00:49:25.075>said,

1024
00:49:25.805 --> 00:49:27.975
<00:49:25.805>this <00:49:26.085>textual <00:49:26.365>inversion <00:49:26.645>example <00:49:26.924>I've <00:49:27.205>demonstrated,

1025
00:49:28.125 --> 00:49:31.375
<00:49:28.125>I've <00:49:28.280>used <00:49:28.436>a <00:49:28.591>single <00:49:28.747>a <00:49:28.902>6,000 <00:49:29.058>GPU, <00:49:29.213>which <00:49:29.369>is <00:49:29.525>comparable

1026
00:49:31.375 --> 00:49:32.775
<00:49:31.375>to <00:49:31.627>40 <00:49:31.879>series, <00:49:32.131>I <00:49:32.383>would <00:49:32.635>say.

1027
00:49:33.355 --> 00:49:36.575
<00:49:33.355>And, <00:49:33.482>uh, <00:49:33.609>it's <00:49:33.736>already <00:49:33.864>enough <00:49:33.991>to <00:49:34.118>use <00:49:34.245>it <00:49:34.373>to <00:49:34.500>train <00:49:34.627>this <00:49:34.755>models

1028
00:49:36.755 --> 00:49:37.855
<00:49:36.755>to <00:49:37.002>fine <00:49:37.250>tune <00:49:37.497>this <00:49:37.745>models.

1029
00:49:39.765 --> 00:49:41.415
<00:49:39.765>Perfect. <00:49:40.231>Thanks <00:49:40.698>for <00:49:41.165>that.

1030
00:49:41.835 --> 00:49:45.135
<00:49:41.835>Um, <00:49:42.035>next <00:49:42.235>question, <00:49:42.435>what <00:49:42.635>about <00:49:42.835>the <00:49:43.035>copyright <00:49:43.235>issue?

1031
00:49:45.315 --> 00:49:49.135
<00:49:45.315>Is <00:49:45.470>there <00:49:45.626>any <00:49:45.781>clear <00:49:45.937>rule <00:49:46.092>agreed <00:49:46.248>upon <00:49:46.403>regarding <00:49:46.559>this <00:49:46.715>aspect

1032
00:49:49.155 --> 00:49:50.975
<00:49:49.155>of <00:49:49.621>using <00:49:50.088>AI <00:49:50.555>technology?

1033
00:49:51.915 --> 00:49:54.735
<00:49:51.915>Um, <00:49:52.070>yeah. <00:49:52.226>Uh, <00:49:52.381>I've <00:49:52.537>already <00:49:52.692>also <00:49:52.848>mentioned, <00:49:53.003>uh, <00:49:53.159>the <00:49:53.315>topic

1034
00:49:54.755 --> 00:49:58.815
<00:49:54.755>of <00:49:54.955>attributions <00:49:55.155>and, <00:49:55.355>um, <00:49:55.555>there's <00:49:55.755>no <00:49:55.955>written <00:49:56.155>rule.

1035
00:49:59.075 --> 00:50:00.415
<00:49:59.075>Uh, <00:49:59.376>but <00:49:59.678>of <00:49:59.979>course, <00:50:00.281>uh,

1036
00:50:00.475 --> 00:50:03.415
<00:50:00.475>you <00:50:00.602>can <00:50:00.729>get <00:50:00.856>into <00:50:00.984>legal <00:50:01.111>trouble <00:50:01.238>if <00:50:01.365>you <00:50:01.493>do <00:50:01.620>not <00:50:01.747>do, <00:50:01.875>uh,

1037
00:50:03.665 --> 00:50:04.695
<00:50:03.665>right <00:50:04.592>attribution.

1038
00:50:05.155 --> 00:50:08.495
<00:50:05.155>The <00:50:05.295>way <00:50:05.435>we <00:50:05.575>are <00:50:05.715>solving <00:50:05.855>that, <00:50:05.995>for <00:50:06.135>example, <00:50:06.275>uh, <00:50:06.415>with <00:50:06.555>our

1039
00:50:09.115 --> 00:50:11.695
<00:50:09.115>Nvidia <00:50:09.395>technologies, <00:50:09.674>um, <00:50:09.955>we <00:50:10.234>are <00:50:10.515>partnering

1040
00:50:11.695 --> 00:50:14.575
<00:50:11.695>with <00:50:11.895>stock <00:50:12.095>companies <00:50:12.295>and <00:50:12.495>we <00:50:12.695>are <00:50:12.895>training <00:50:13.095>models

1041
00:50:14.995 --> 00:50:19.295
<00:50:14.995>for <00:50:15.170>this <00:50:15.345>stock <00:50:15.520>companies, <00:50:15.694>uh, <00:50:15.870>like <00:50:16.045>Getty <00:50:16.220>and <00:50:16.395>Shutterstock.

1042
00:50:20.035 --> 00:50:24.695
<00:50:20.035>And, <00:50:20.162>uh, <00:50:20.289>then <00:50:20.416>we <00:50:20.544>know <00:50:20.671>like <00:50:20.798>which <00:50:20.925>data <00:50:21.053>our <00:50:21.180>models <00:50:21.307>have <00:50:21.435>been

1043
00:50:24.695 --> 00:50:28.245
<00:50:24.695>trained <00:50:24.850>on, <00:50:25.006>and <00:50:25.161>then <00:50:25.317>we <00:50:25.472>are <00:50:25.628>able <00:50:25.783>to <00:50:25.939>pay <00:50:26.095>attributions.

1044
00:50:28.425 --> 00:50:31.285
<00:50:28.425>Uh, <00:50:28.552>if <00:50:28.679>this <00:50:28.806>model <00:50:28.934>is <00:50:29.061>used, <00:50:29.188>we <00:50:29.315>are <00:50:29.443>able <00:50:29.570>to <00:50:29.697>pay <00:50:29.825>attributions

1045
00:50:31.505 --> 00:50:36.165
<00:50:31.505>for <00:50:31.680>the, <00:50:31.855>the, <00:50:32.030>the <00:50:32.205>creators <00:50:32.380>of <00:50:32.555>the <00:50:32.730>initial <00:50:32.905>content.

1046
00:50:37.685 --> 00:50:39.965
<00:50:37.685>I <00:50:37.825>think <00:50:37.965>that's, <00:50:38.105>that's <00:50:38.245>the <00:50:38.384>right, <00:50:38.525>the <00:50:38.665>only <00:50:38.805>right <00:50:38.945>way. <00:50:39.085>Um,

1047
00:50:41.815 --> 00:50:43.425
<00:50:41.815>Yeah, <00:50:42.095>seems <00:50:42.375>like <00:50:42.655>the <00:50:42.935>right <00:50:43.215>process.

1048
00:50:44.925 --> 00:50:49.705
<00:50:44.925>We, <00:50:45.100>thanks <00:50:45.275>for <00:50:45.450>answering <00:50:45.625>also <00:50:45.800>that <00:50:45.975>we <00:50:46.150>got, <00:50:46.325>uh,

1049
00:50:49.865 --> 00:50:51.825
<00:50:49.865>a <00:50:50.064>follow-up <00:50:50.265>question <00:50:50.464>to <00:50:50.665>one <00:50:50.865>of <00:50:51.064>the <00:50:51.265>last

1050
00:50:51.825 --> 00:50:54.025
<00:50:51.825>regarding <00:50:52.291>the <00:50:52.758>programming <00:50:53.225>language.

1051
00:50:54.205 --> 00:50:56.545
<00:50:54.205>Mm-Hmm. <00:50:54.404>Because, <00:50:54.605>uh, <00:50:54.805>somebody <00:50:55.005>in <00:50:55.205>the <00:50:55.404>audience <00:50:55.605>wants

1052
00:50:56.545 --> 00:51:00.025
<00:50:56.545>to <00:50:56.720>know <00:50:56.895>why <00:50:57.070>is <00:50:57.245>Python <00:50:57.420>still <00:50:57.595>most <00:50:57.770>popular <00:50:57.945>language

1053
00:51:00.045 --> 00:51:03.585
<00:51:00.045>for <00:51:00.200>ai <00:51:00.356>when <00:51:00.511>we <00:51:00.667>have <00:51:00.822>some <00:51:00.978>other <00:51:01.133>languages <00:51:01.289>like <00:51:01.445>Julia,

1054
00:51:04.575 --> 00:51:07.425
<00:51:04.575>Some <00:51:04.774>people <00:51:04.975>use <00:51:05.174>Julia <00:51:05.375>for <00:51:05.575>AI <00:51:05.774>machine <00:51:05.975>learning.

1055
00:51:07.425 --> 00:51:11.265
<00:51:07.425>However, <00:51:07.565>I <00:51:07.705>think <00:51:07.845>the <00:51:07.985>power <00:51:08.125>of <00:51:08.265>Python <00:51:08.405>is <00:51:08.545>in <00:51:08.685>the <00:51:08.825>community.

1056
00:51:11.645 --> 00:51:13.665
<00:51:11.645>Um, <00:51:12.111>and <00:51:12.578>uh, <00:51:13.045>the,

1057
00:51:13.685 --> 00:51:16.105
<00:51:13.685>the <00:51:13.884>whole <00:51:14.085>open <00:51:14.285>source <00:51:14.485>community <00:51:14.685>is <00:51:14.884>about <00:51:15.085>Python,

1058
00:51:16.325 --> 00:51:18.305
<00:51:16.325>and <00:51:16.524>you <00:51:16.725>can <00:51:16.924>get <00:51:17.125>good <00:51:17.325>examples <00:51:17.524>in <00:51:17.725>Python

1059
00:51:18.565 --> 00:51:21.065
<00:51:18.565>and <00:51:18.720>also <00:51:18.876>like <00:51:19.031>popular <00:51:19.187>frameworks, <00:51:19.342>like <00:51:19.498>tens <00:51:19.653>of <00:51:19.809>low <00:51:19.965>Python,

1060
00:51:21.335 --> 00:51:23.105
<00:51:21.335>they <00:51:21.685>started <00:51:22.035>like <00:51:22.385>with <00:51:22.735>Python.

1061
00:51:23.405 --> 00:51:26.825
<00:51:23.405>And, <00:51:23.560>uh, <00:51:23.716>that's, <00:51:23.871>that's <00:51:24.027>the <00:51:24.182>main <00:51:24.338>reason, <00:51:24.493>uh, <00:51:24.649>the <00:51:24.805>community.

1062
00:51:27.265 --> 00:51:28.945
<00:51:27.265>I <00:51:27.464>mean, <00:51:27.665>when <00:51:27.865>I <00:51:28.065>was <00:51:28.265>doing <00:51:28.464>my <00:51:28.665>PhD

1063
00:51:29.045 --> 00:51:32.105
<00:51:29.045>and <00:51:29.185>masters, <00:51:29.325>I <00:51:29.465>was <00:51:29.605>using <00:51:29.745>matlab, <00:51:29.885>uh, <00:51:30.025>which <00:51:30.165>was <00:51:30.305>also <00:51:30.445>useful.

1064
00:51:32.305 --> 00:51:34.105
<00:51:32.305>I <00:51:32.480>could <00:51:32.654>do <00:51:32.830>machine <00:51:33.004>learning <00:51:33.180>with <00:51:33.355>MATLAB <00:51:33.529>as <00:51:33.705>well,

1065
00:51:34.725 --> 00:51:36.425
<00:51:34.725>but <00:51:35.005>it's <00:51:35.285>just <00:51:35.565>a <00:51:35.845>community <00:51:36.125>thing.

1066
00:51:37.255 --> 00:51:38.625
<00:51:37.255>Yeah, <00:51:37.666>you <00:51:38.077>get <00:51:38.488>lots

1067
00:51:38.625 --> 00:51:41.265
<00:51:38.625>of <00:51:38.825>help <00:51:39.025>when <00:51:39.225>you're <00:51:39.425>using <00:51:39.625>a <00:51:39.825>popular <00:51:40.025>language.

1068
00:51:41.365 --> 00:51:43.385
<00:51:41.365>So <00:51:41.645>yeah, <00:51:41.924>start <00:51:42.205>with <00:51:42.484>that. <00:51:42.765>Yes,

1069
00:51:43.845 --> 00:51:44.845
<00:51:43.845>Yes, <00:51:44.745>yes.

1070
00:51:45.365 --> 00:51:49.385
<00:51:45.365>All <00:51:45.520>right. <00:51:45.676>And <00:51:45.831>we <00:51:45.987>are <00:51:46.142>getting <00:51:46.298>to <00:51:46.453>our <00:51:46.609>next <00:51:46.765>question.

1071
00:51:52.005 --> 00:51:55.305
<00:51:52.005>Do <00:51:52.180>you <00:51:52.355>think <00:51:52.530>it's <00:51:52.705>beneficial <00:51:52.880>to <00:51:53.055>take <00:51:53.230>courses <00:51:53.405>such

1072
00:51:55.305 --> 00:51:58.745
<00:51:55.305>as <00:51:55.504>those <00:51:55.705>on <00:51:55.904>Coursera <00:51:56.105>to <00:51:56.305>obtain <00:51:56.504>obtain <00:51:56.705>certificates?

1073
00:51:59.295 --> 00:52:01.745
<00:51:59.295>Does <00:51:59.495>it <00:51:59.695>add <00:51:59.895>any <00:52:00.095>value <00:52:00.295>to <00:52:00.495>the <00:52:00.695>curriculum?

1074
00:52:02.685 --> 00:52:06.025
<00:52:02.685>Um, <00:52:02.965>my <00:52:03.245>personal <00:52:03.525>opinion <00:52:03.805>is <00:52:04.085>that

1075
00:52:06.805 --> 00:52:08.025
<00:52:06.805>you, <00:52:07.079>you <00:52:07.354>learn <00:52:07.628>for <00:52:07.903>yourself.

1076
00:52:08.045 --> 00:52:09.785
<00:52:08.045>You <00:52:08.245>don't <00:52:08.445>learn <00:52:08.645>for <00:52:08.845>the <00:52:09.045>piece <00:52:09.245>of <00:52:09.445>paper

1077
00:52:10.125 --> 00:52:12.065
<00:52:10.125>and, <00:52:10.325>uh, <00:52:10.525>it <00:52:10.725>doesn't <00:52:10.925>matter <00:52:11.125>where <00:52:11.325>you <00:52:11.525>learn

1078
00:52:12.165 --> 00:52:14.425
<00:52:12.165>or <00:52:12.365>whether <00:52:12.565>you <00:52:12.765>obtain <00:52:12.965>a <00:52:13.165>certificate <00:52:13.365>or <00:52:13.565>not.

1079
00:52:14.605 --> 00:52:15.825
<00:52:14.605>Um, <00:52:14.971>I <00:52:15.337>wouldn't <00:52:15.703>put

1080
00:52:15.825 --> 00:52:18.185
<00:52:15.825>that <00:52:16.105>much <00:52:16.384>value <00:52:16.665>into <00:52:16.944>certificates, <00:52:17.225>honestly.

1081
00:52:21.895 --> 00:52:23.515
<00:52:21.895>Thanks. <00:52:22.595>Another <00:52:23.295>question.

1082
00:52:24.215 --> 00:52:28.075
<00:52:24.215>Can <00:52:24.355>you <00:52:24.495>please <00:52:24.635>shed <00:52:24.775>some <00:52:24.915>light <00:52:25.055>in <00:52:25.195>recent <00:52:25.335>developments <00:52:25.475>in <00:52:25.615>IOT

1083
00:52:28.075 --> 00:52:30.315
<00:52:28.075>devices <00:52:28.424>or <00:52:28.774>Nvidia <00:52:29.125>Jetson <00:52:29.475>family?

1084
00:52:31.785 --> 00:52:34.045
<00:52:31.785>Um, <00:52:32.251>a <00:52:32.718>good <00:52:33.185>question.

1085
00:52:34.085 --> 00:52:35.285
<00:52:34.085>I, <00:52:34.301>I <00:52:34.517>don't <00:52:34.732>think <00:52:34.949>I'm <00:52:35.165>supposed

1086
00:52:35.285 --> 00:52:37.485
<00:52:35.285>to <00:52:35.518>talk <00:52:35.751>about <00:52:35.984>the <00:52:36.218>roadmap <00:52:36.451>about <00:52:36.685>products,

1087
00:52:37.825 --> 00:52:40.085
<00:52:37.825>but <00:52:38.058>yeah, <00:52:38.291>we <00:52:38.524>have <00:52:38.758>a <00:52:38.991>Jetson <00:52:39.225>family.

1088
00:52:40.355 --> 00:52:44.165
<00:52:40.355>It's <00:52:40.495>a, <00:52:40.635>it's <00:52:40.775>a <00:52:40.915>good <00:52:41.055>family <00:52:41.195>of <00:52:41.335>embedded <00:52:41.475>devices <00:52:41.615>which, <00:52:41.755>uh,

1089
00:52:44.385 --> 00:52:47.565
<00:52:44.385>are <00:52:44.618>used <00:52:44.851>in <00:52:45.085>robotics <00:52:45.318>in, <00:52:45.551>uh, <00:52:45.785>surveillance.

1090
00:52:47.825 --> 00:52:49.485
<00:52:47.825>Uh, <00:52:48.058>this, <00:52:48.291>these <00:52:48.524>device <00:52:48.758>is <00:52:48.991>a <00:52:49.225>good

1091
00:52:49.485 --> 00:52:52.445
<00:52:49.485>to <00:52:49.612>use <00:52:49.739>in <00:52:49.866>the <00:52:49.994>field <00:52:50.121>on <00:52:50.248>the <00:52:50.375>edge <00:52:50.503>if <00:52:50.630>you <00:52:50.757>don't <00:52:50.885>necessarily

1092
00:52:52.515 --> 00:52:54.005
<00:52:52.515>want <00:52:52.706>to <00:52:52.898>send <00:52:53.089>your <00:52:53.281>data <00:52:53.472>to <00:52:53.664>the <00:52:53.856>cloud

1093
00:52:54.105 --> 00:52:56.165
<00:52:54.105>or <00:52:54.305>to <00:52:54.505>the <00:52:54.705>internet, <00:52:54.905>it's <00:52:55.105>a <00:52:55.305>great <00:52:55.505>devices.

1094
00:52:56.585 --> 00:52:57.845
<00:52:56.585>So <00:52:56.962>yeah, <00:52:57.341>uh, <00:52:57.719>go

1095
00:52:57.885 --> 00:53:00.525
<00:52:57.885>and <00:52:58.085>check <00:52:58.285>our <00:52:58.485>website, <00:52:58.685>uh, <00:52:58.885>for <00:52:59.085>Jetson <00:52:59.285>products.

1096
00:53:00.545 --> 00:53:01.845
<00:53:00.545>Uh, <00:53:00.779>maybe <00:53:01.013>it's <00:53:01.247>something <00:53:01.481>for <00:53:01.714>you.

1097
00:53:04.035 --> 00:53:07.165
<00:53:04.035>Perfect. <00:53:04.501>Alright, <00:53:04.968>another <00:53:05.435>question.

1098
00:53:07.745 --> 00:53:11.125
<00:53:07.745>Can <00:53:08.025>multimodal <00:53:08.305>generative <00:53:08.585>AI <00:53:08.865>work <00:53:09.145>together

1099
00:53:11.195 --> 00:53:14.205
<00:53:11.195>with <00:53:11.475>computer <00:53:11.755>vision, <00:53:12.035>natural <00:53:12.315>language <00:53:12.595>processing

1100
00:53:14.305 --> 00:53:17.525
<00:53:14.305>or <00:53:14.538>audio <00:53:14.771>processing <00:53:15.004>to <00:53:15.238>create <00:53:15.471>better <00:53:15.705>outcomes?

1101
00:53:18.445 --> 00:53:19.565
<00:53:18.445>Absolutely. <00:53:18.949>That's <00:53:19.453>exactly

1102
00:53:19.595 --> 00:53:21.525
<00:53:19.595>what <00:53:19.944>multimodal <00:53:20.294>generative <00:53:20.645>AI <00:53:20.995>does.

1103
00:53:21.695 --> 00:53:24.845
<00:53:21.695>Multimodal <00:53:21.975>multimodality <00:53:22.255>brings <00:53:22.535>all <00:53:22.815>these <00:53:23.095>modalities

1104
00:53:25.205 --> 00:53:27.005
<00:53:25.205>together <00:53:25.671>and, <00:53:26.138>uh, <00:53:26.605>basically

1105
00:53:27.065 --> 00:53:29.845
<00:53:27.065>by <00:53:27.298>including <00:53:27.531>some <00:53:27.765>modalities <00:53:27.998>you <00:53:28.231>enhance, <00:53:28.465>uh,

1106
00:53:30.305 --> 00:53:32.685
<00:53:30.305>you <00:53:30.480>enhance <00:53:30.654>the <00:53:30.830>use <00:53:31.004>of, <00:53:31.180>uh, <00:53:31.355>the <00:53:31.529>other <00:53:31.705>modalities.

1107
00:53:32.865 --> 00:53:35.845
<00:53:32.865>And, <00:53:33.020>um, <00:53:33.176>I <00:53:33.331>really <00:53:33.487>liked <00:53:33.642>some <00:53:33.798>research <00:53:33.953>paper <00:53:34.109>by <00:53:34.265>published

1108
00:53:35.845 --> 00:53:37.085
<00:53:35.845>by <00:53:36.124>meta <00:53:36.403>some <00:53:36.682>time <00:53:36.961>ago.

1109
00:53:37.515 --> 00:53:40.805
<00:53:37.515>They <00:53:37.654>were <00:53:37.795>able <00:53:37.935>to <00:53:38.075>show <00:53:38.214>that <00:53:38.355>when <00:53:38.495>you <00:53:38.634>train <00:53:38.775>something <00:53:38.915>like

1110
00:53:41.875 --> 00:53:46.125
<00:53:41.875>between <00:53:42.050>modalities, <00:53:42.225>like <00:53:42.400>image <00:53:42.575>to <00:53:42.750>text, <00:53:42.925>audio <00:53:43.100>to <00:53:43.275>text,

1111
00:53:46.435 --> 00:53:49.045
<00:53:46.435>then <00:53:46.668>you <00:53:46.901>also <00:53:47.134>get <00:53:47.368>automatically <00:53:47.601>this <00:53:47.835>breach

1112
00:53:49.045 --> 00:53:50.525
<00:53:49.045>between <00:53:49.489>image <00:53:49.933>to <00:53:50.377>audio.

1113
00:53:50.635 --> 00:53:52.245
<00:53:50.635>That <00:53:50.915>was <00:53:51.195>really <00:53:51.475>fascinating <00:53:51.755>to <00:53:52.035>see.

1114
00:53:52.505 --> 00:53:54.445
<00:53:52.505>And <00:53:52.738>they've <00:53:52.971>demonstrated <00:53:53.205>that <00:53:53.438>it <00:53:53.671>actually <00:53:53.905>works.

1115
00:53:57.155 --> 00:53:59.565
<00:53:57.155>Perfect. <00:53:57.435>Thanks <00:53:57.715>for <00:53:57.995>also <00:53:58.275>explaining <00:53:58.555>this.

1116
00:54:00.115 --> 00:54:05.085
<00:54:00.115>Another <00:54:00.348>question, <00:54:00.581>um, <00:54:00.814>regarding <00:54:01.048>the <00:54:01.281>legal <00:54:01.515>stuff.

1117
00:54:05.105 --> 00:54:07.205
<00:54:05.105>We <00:54:05.280>already <00:54:05.455>talked <00:54:05.630>a <00:54:05.805>bit <00:54:05.980>about, <00:54:06.155>do <00:54:06.330>you <00:54:06.505>think

1118
00:54:07.205 --> 00:54:08.485
<00:54:07.205>that <00:54:07.589>countries <00:54:07.973>are <00:54:08.357>missing

1119
00:54:08.505 --> 00:54:11.445
<00:54:08.505>to <00:54:08.705>address <00:54:08.905>AI <00:54:09.105>models <00:54:09.305>in <00:54:09.505>laws <00:54:09.705>and <00:54:09.905>regulations?

1120
00:54:12.085 --> 00:54:13.765
<00:54:12.085>I <00:54:12.435>guess <00:54:12.785>we <00:54:13.135>postponed <00:54:13.485>this,

1121
00:54:13.985 --> 00:54:16.485
<00:54:13.985>but <00:54:14.218>the <00:54:14.451>models <00:54:14.685>are <00:54:14.918>growing <00:54:15.151>super <00:54:15.385>rapid.

1122
00:54:18.225 --> 00:54:23.165
<00:54:18.225>Um, <00:54:18.424>yeah, <00:54:18.625>it's, <00:54:18.825>uh, <00:54:19.025>I <00:54:19.225>think <00:54:19.424>the <00:54:19.625>countries,

1123
00:54:23.165 --> 00:54:26.365
<00:54:23.165>of <00:54:23.305>course, <00:54:23.445>uh, <00:54:23.585>there, <00:54:23.725>there <00:54:23.865>are <00:54:24.005>many <00:54:24.145>debates <00:54:24.285>about <00:54:24.425>this, <00:54:24.565>uh,

1124
00:54:26.495 --> 00:54:29.485
<00:54:26.495>about, <00:54:26.650>uh, <00:54:26.806>how <00:54:26.961>to <00:54:27.117>use <00:54:27.272>AI <00:54:27.428>in <00:54:27.583>the <00:54:27.739>proper <00:54:27.895>way.

1125
00:54:29.825 --> 00:54:33.645
<00:54:29.825>Um, <00:54:30.058>it's, <00:54:30.291>it's, <00:54:30.524>it's <00:54:30.758>a <00:54:30.991>tricky <00:54:31.225>question

1126
00:54:33.645 --> 00:54:36.005
<00:54:33.645>because <00:54:33.820>we, <00:54:33.995>we, <00:54:34.170>we <00:54:34.345>also <00:54:34.520>cannot <00:54:34.695>stop <00:54:34.870>the <00:54:35.045>progress

1127
00:54:36.345 --> 00:54:39.405
<00:54:36.345>and, <00:54:36.500>uh, <00:54:36.656>the <00:54:36.811>progress <00:54:36.967>is <00:54:37.122>not <00:54:37.278>necessarily <00:54:37.433>a <00:54:37.589>bad <00:54:37.745>thing.

1128
00:54:39.565 --> 00:54:41.525
<00:54:39.565>I <00:54:39.765>think <00:54:39.965>the <00:54:40.165>bad <00:54:40.365>thing <00:54:40.565>is, <00:54:40.765>uh, <00:54:40.965>to

1129
00:54:42.465 --> 00:54:47.365
<00:54:42.465>use <00:54:42.640>this <00:54:42.815>well <00:54:42.990>technology <00:54:43.165>in <00:54:43.340>the, <00:54:43.515>uh, <00:54:43.690>unethical <00:54:43.865>way

1130
00:54:47.585 --> 00:54:50.805
<00:54:47.585>and, <00:54:47.760>um, <00:54:47.935>I <00:54:48.110>believe, <00:54:48.285>yeah, <00:54:48.460>some, <00:54:48.635>some <00:54:48.810>regulation, <00:54:48.985>uh,

1131
00:54:50.805 --> 00:54:52.365
<00:54:50.805>would <00:54:51.038>be <00:54:51.271>of <00:54:51.504>course <00:54:51.738>helpful <00:54:51.971>for <00:54:52.205>that.

1132
00:54:53.855 --> 00:54:57.485
<00:54:53.855>Maybe <00:54:53.982>it <00:54:54.109>reminds <00:54:54.236>a <00:54:54.364>bit <00:54:54.491>to <00:54:54.618>the <00:54:54.745>early <00:54:54.873>days <00:54:55.000>of <00:54:55.127>the <00:54:55.255>internet

1133
00:54:57.485 --> 00:54:59.725
<00:54:57.485>where <00:54:57.765>technology <00:54:58.045>was <00:54:58.325>moving <00:54:58.605>way <00:54:58.885>faster

1134
00:55:00.235 --> 00:55:02.765
<00:55:00.235>than <00:55:00.468>the <00:55:00.701>legal <00:55:00.935>regulations <00:55:01.168>and <00:55:01.401>yeah, <00:55:01.635>it's,

1135
00:55:02.985 --> 00:55:04.365
<00:55:02.985>Yes, <00:55:03.399>maybe <00:55:03.813>the <00:55:04.227>same.

1136
00:55:05.145 --> 00:55:06.525
<00:55:05.145>You <00:55:05.352>cannot <00:55:05.559>stop, <00:55:05.766>you <00:55:05.973>cannot <00:55:06.180>stop <00:55:06.387>the

1137
00:55:06.725 --> 00:55:07.925
<00:55:06.725>progress <00:55:07.085>and <00:55:07.445>you <00:55:07.805>shouldn't.

1138
00:55:08.075 --> 00:55:09.725
<00:55:08.075>Yeah, <00:55:08.355>that, <00:55:08.634>that's <00:55:08.915>my <00:55:09.194>opinion. <00:55:09.475>Of

1139
00:55:09.725 --> 00:55:10.725
<00:55:09.725>Course <00:55:09.950>not. <00:55:10.174>Of <00:55:10.400>course <00:55:10.625>not.

1140
00:55:10.725 --> 00:55:13.005
<00:55:10.725>Alright, <00:55:11.424>thank <00:55:12.125>you.

1141
00:55:15.085 --> 00:55:17.215
<00:55:15.085>Next <00:55:15.551>question. <00:55:16.018>What <00:55:16.485>experiments

1142
00:55:17.245 --> 00:55:20.695
<00:55:17.245>with <00:55:17.400>data <00:55:17.556>sets <00:55:17.711>you <00:55:17.867>recommend <00:55:18.022>adding <00:55:18.178>to <00:55:18.333>my <00:55:18.489>GitHub <00:55:18.645>portfolio?

1143
00:55:21.195 --> 00:55:24.495
<00:55:21.195>For <00:55:21.350>example, <00:55:21.506>I <00:55:21.661>have <00:55:21.817>common <00:55:21.972>data <00:55:22.128>sets <00:55:22.283>like <00:55:22.439>Boston <00:55:22.595>houses,

1144
00:55:24.915 --> 00:55:26.775
<00:55:24.915>but <00:55:25.148>is <00:55:25.381>a <00:55:25.615>very <00:55:25.848>easy <00:55:26.081>data <00:55:26.315>set.

1145
00:55:27.005 --> 00:55:28.575
<00:55:27.005>Does <00:55:27.238>it <00:55:27.471>make <00:55:27.705>sense <00:55:27.938>to <00:55:28.171>add <00:55:28.405>it?

1146
00:55:29.975 --> 00:55:31.895
<00:55:29.975>I <00:55:30.208>think <00:55:30.441>what <00:55:30.674>makes <00:55:30.908>sense <00:55:31.141>to <00:55:31.375>add

1147
00:55:31.895 --> 00:55:35.295
<00:55:31.895>to <00:55:32.128>your <00:55:32.361>GitHub <00:55:32.595>portfolio <00:55:32.828>is, <00:55:33.061>uh, <00:55:33.295>creation

1148
00:55:35.295 --> 00:55:36.535
<00:55:35.295>of <00:55:35.667>some <00:55:36.038>unique <00:55:36.411>value.

1149
00:55:36.895 --> 00:55:39.135
<00:55:36.895>I <00:55:37.128>mean, <00:55:37.361>you, <00:55:37.595>everyone <00:55:37.828>can <00:55:38.061>do <00:55:38.295>some

1150
00:55:39.135 --> 00:55:40.415
<00:55:39.135>experiments <00:55:39.519>and <00:55:39.903>then <00:55:40.287>publish.

1151
00:55:40.595 --> 00:55:42.215
<00:55:40.595>It <00:55:40.794>doesn't, <00:55:40.995>when <00:55:41.194>it <00:55:41.395>doesn't <00:55:41.595>bring <00:55:41.794>the <00:55:41.995>value

1152
00:55:42.235 --> 00:55:43.695
<00:55:42.235>to <00:55:42.497>the <00:55:42.760>community, <00:55:43.023>don't <00:55:43.286>do <00:55:43.549>that.

1153
00:55:44.315 --> 00:55:45.815
<00:55:44.315>Uh, <00:55:44.585>if <00:55:44.855>you <00:55:45.125>create <00:55:45.395>something <00:55:45.665>unique,

1154
00:55:46.075 --> 00:55:49.375
<00:55:46.075>if <00:55:46.250>you <00:55:46.424>develop <00:55:46.600>something <00:55:46.774>unique, <00:55:46.950>some <00:55:47.125>unique <00:55:47.299>approach, <00:55:47.475>then

1155
00:55:49.475 --> 00:55:51.855
<00:55:49.475>of <00:55:49.674>course, <00:55:49.875>uh, <00:55:50.075>it's <00:55:50.275>beneficial <00:55:50.475>for <00:55:50.674>the <00:55:50.875>community,

1156
00:55:51.855 --> 00:55:53.295
<00:55:51.855>then <00:55:52.287>please <00:55:52.719>publish <00:55:53.151>that.

1157
00:55:53.435 --> 00:55:58.215
<00:55:53.435>But <00:55:53.590>if <00:55:53.746>it's <00:55:53.901>just <00:55:54.057>a <00:55:54.212>generic <00:55:54.368>exercise, <00:55:54.523>uh, <00:55:54.679>don't <00:55:54.835>please.

1158
00:56:01.235 --> 00:56:04.375
<00:56:01.235>All <00:56:01.375>right, <00:56:01.515>thanks. <00:56:01.655>And <00:56:01.795>we <00:56:01.935>have <00:56:02.075>time <00:56:02.215>for <00:56:02.355>one <00:56:02.495>more <00:56:02.635>question.

1159
00:56:05.585 --> 00:56:06.975
<00:56:05.585>Let's <00:56:05.897>maybe <00:56:06.210>take <00:56:06.523>this <00:56:06.836>one.

1160
00:56:07.325 --> 00:56:11.415
<00:56:07.325>What <00:56:07.500>type <00:56:07.674>of <00:56:07.850>hardware <00:56:08.024>challenges, <00:56:08.200>challenges <00:56:08.375>when <00:56:08.549>it's, <00:56:08.725>uh,

1161
00:56:11.485 --> 00:56:14.455
<00:56:11.485>what <00:56:11.625>type <00:56:11.765>of <00:56:11.905>hardware <00:56:12.045>challenges <00:56:12.185>are <00:56:12.325>we <00:56:12.465>facing <00:56:12.605>when <00:56:12.745>it <00:56:12.885>comes

1162
00:56:14.475 --> 00:56:17.535
<00:56:14.475>to <00:56:14.630>AI <00:56:14.786>and <00:56:14.941>how <00:56:15.097>does <00:56:15.252>Nvidia <00:56:15.408>plan <00:56:15.563>to <00:56:15.719>tackle <00:56:15.875>it?

1163
00:56:18.605 --> 00:56:21.735
<00:56:18.605>Well, <00:56:18.838>NVIDIA <00:56:19.071>is <00:56:19.305>providing <00:56:19.538>hardware <00:56:19.771>for <00:56:20.005>ai.

1164
00:56:22.515 --> 00:56:24.085
<00:56:22.515>That's <00:56:22.795>our <00:56:23.075>plan <00:56:23.355>to <00:56:23.634>tackle <00:56:23.915>it.

1165
00:56:25.635 --> 00:56:27.365
<00:56:25.635>Even <00:56:25.985>specialized. <00:56:26.335>Yeah, <00:56:26.685>in <00:56:27.035>it.

1166
00:56:28.555 --> 00:56:31.165
<00:56:28.555>Alright, <00:56:29.955>perfect.

1167
00:56:31.235 --> 00:56:33.045
<00:56:31.235>Then <00:56:31.410>I <00:56:31.585>would <00:56:31.760>say <00:56:31.935>we <00:56:32.110>close <00:56:32.285>with <00:56:32.460>this <00:56:32.635>question.

1168
00:56:34.095 --> 00:56:35.965
<00:56:34.095>Katya, <00:56:34.270>thank <00:56:34.444>you <00:56:34.620>so <00:56:34.794>much <00:56:34.970>for <00:56:35.145>joining <00:56:35.319>us <00:56:35.495>today

1169
00:56:35.965 --> 00:56:39.365
<00:56:35.965>and <00:56:36.165>giving <00:56:36.365>this <00:56:36.565>talk <00:56:36.765>and <00:56:36.965>answering <00:56:37.165>these <00:56:37.365>multiple

1170
00:56:39.725 --> 00:56:41.165
<00:56:39.725>questions <00:56:39.984>in <00:56:40.243>the <00:56:40.502>q <00:56:40.761>and <00:56:41.020>a.

1171
00:56:41.165 --> 00:56:42.525
<00:56:41.165>It <00:56:41.369>was <00:56:41.573>a <00:56:41.777>pleasure <00:56:41.981>to <00:56:42.185>have <00:56:42.389>you

1172
00:56:43.895 --> 00:56:47.115
<00:56:43.895>and <00:56:44.022>yeah, <00:56:44.149>hopefully <00:56:44.276>see <00:56:44.404>you <00:56:44.531>soon <00:56:44.658>on <00:56:44.785>a <00:56:44.913>lot <00:56:45.040>of <00:56:45.167>lifetime <00:56:45.295>maybe.

1173
00:56:47.725 --> 00:56:50.315
<00:56:47.725>Thank <00:56:47.880>you <00:56:48.036>very <00:56:48.191>much. <00:56:48.347>Was <00:56:48.502>super <00:56:48.658>fun <00:56:48.813>to <00:56:48.969>be <00:56:49.125>here

1174
00:56:50.695 --> 00:56:53.915
<00:56:50.695>and, <00:56:50.822>uh, <00:56:50.949>enjoy <00:56:51.076>the <00:56:51.204>rest <00:56:51.331>of <00:56:51.458>the <00:56:51.585>day <00:56:51.713>and <00:56:51.840>uh, <00:56:51.967>take <00:56:52.095>care.

1175
00:56:54.105 --> 00:56:54.595
Bye-Bye.