WEBVTT

1
00:00:00.070 --> 00:00:07.040
<v Cassidy>A <00:00:00.147>voice <00:00:00.613>AI <00:00:00.950>startup <00:00:01.480>pulls <00:00:01.720>a <00:00:01.800>hundred <00:00:02.130>million <00:00:02.533>in <00:00:02.704>seed <00:00:03.027>money — <00:00:04.020>and <00:00:04.100>the <00:00:04.192>name <00:00:04.352>that <00:00:04.533>jumps <00:00:04.840>out <00:00:04.987>on <00:00:05.060>the <00:00:05.180>cap <00:00:05.427>table <00:00:05.840>isn't <00:00:06.280>a <00:00:06.533>VC.

2
00:00:07.253 --> 00:00:11.043
<v Cassidy>It's <00:00:07.840>Nvidia. <00:00:09.582>This <00:00:09.656>is <00:00:09.870>AI <00:00:10.136>Daily <00:00:10.456>Briefing.

3
00:00:11.190 --> 00:00:17.603
<v Cassidy>Today <00:00:11.950>we <00:00:12.056>pivot <00:00:12.383>off <00:00:12.583>the <00:00:12.703>megawatts <00:00:13.423>and <00:00:13.651>into <00:00:13.863>the <00:00:13.983>science <00:00:14.456>layer — <00:00:15.403>a <00:00:15.491>chipmaker <00:00:15.969>buying <00:00:16.275>into <00:00:16.443>a <00:00:16.536>voice <00:00:16.870>round

4
00:00:17.667 --> 00:00:22.403
<v Cassidy>plus <00:00:17.983>two <00:00:18.209>Nature <00:00:18.552>papers <00:00:18.995>that <00:00:19.193>quietly <00:00:19.729>matter <00:00:20.067>more <00:00:20.355>than <00:00:20.503>the <00:00:20.613>funding <00:00:21.034>headline.

5
00:00:22.423 --> 00:00:25.763
<v Bill>And <00:00:22.523>I <00:00:22.595>want <00:00:22.750>to <00:00:22.851>know <00:00:23.091>what <00:00:23.386>Nvidia's <00:00:24.225>actually <00:00:24.803>buying <00:00:25.363>here.

6
00:00:25.923 --> 00:00:27.523
<v Bill>Let's <00:00:26.523>start <00:00:26.755>with <00:00:26.936>Gradium.

7
00:00:27.649 --> 00:00:32.483
<v Bill>Follow <00:00:28.443>the <00:00:28.675>show <00:00:29.163>and <00:00:29.303>the <00:00:29.427>next <00:00:29.736>briefing <00:00:30.243>lands <00:00:30.670>in <00:00:30.755>your <00:00:30.979>feed <00:00:31.443>on <00:00:31.663>its <00:00:31.943>own.

8
00:00:33.245 --> 00:00:34.285
<v Cassidy>From <00:00:33.455>Ventureburn:

9
00:00:34.692 --> 00:00:40.932
<v Quote>Paris-based <00:00:35.399>voice <00:00:35.839>AI <00:00:36.102>startup <00:00:36.532>Gradium <00:00:37.092>has <00:00:37.372>announced <00:00:37.772>a <00:00:37.892>major <00:00:38.372>$100 <00:00:39.102>million <00:00:39.812>seed <00:00:40.172>funding <00:00:40.505>round.

10
00:00:41.072 --> 00:00:48.292
<v Quote>The <00:00:41.542>company <00:00:41.945>recently <00:00:42.425>added <00:00:42.772>a <00:00:42.892>$30 <00:00:43.252>million <00:00:43.868>extension <00:00:44.505>to <00:00:44.812>its <00:00:44.982>initial <00:00:45.412>$70 <00:00:45.892>million <00:00:46.516>tranche <00:00:46.932>raised <00:00:47.476>last <00:00:47.776>December.

11
00:00:48.420 --> 00:00:51.812
<v Quote>Tech <00:00:48.999>giant <00:00:49.401>Nvidia <00:00:49.858>joined <00:00:50.152>the <00:00:50.265>round <00:00:50.559>as <00:00:50.652>the <00:00:50.808>headline <00:00:51.154>investor.

12
00:00:52.127 --> 00:00:59.647
<v Bill>So <00:00:52.340>Gradium — <00:00:53.327>Paris, <00:00:53.914>voice <00:00:54.714>AI, <00:00:55.620>ultra-low <00:00:56.187>latency <00:00:56.670>transcription — <00:00:57.887>closes <00:00:58.447>a <00:00:58.577>hundred <00:00:59.007>million.

13
00:00:59.807 --> 00:01:03.887
<v Bill>But <00:01:00.319>look <00:01:00.474>at <00:01:00.567>how <00:01:00.719>they <00:01:00.907>got <00:01:01.140>there: <00:01:02.037>seventy <00:01:02.607>in <00:01:02.829>December

14
00:01:03.935 --> 00:01:06.127
<v Bill>then <00:01:04.207>a <00:01:04.344>thirty-million <00:01:05.071>extension

15
00:01:06.167 --> 00:01:10.927
<v Bill>and <00:01:06.307>the <00:01:06.431>extension <00:01:07.114>is <00:01:07.367>where <00:01:07.658>Nvidia <00:01:08.260>walks <00:01:08.634>in <00:01:08.927>as <00:01:09.149>headline <00:01:09.709>investor.

16
00:01:10.947 --> 00:01:13.007
<v Cassidy>And <00:01:11.055>that's <00:01:11.347>the <00:01:11.439>part <00:01:11.607>I <00:01:11.695>keep <00:01:11.947>circling.

17
00:01:13.076 --> 00:01:15.567
<v Cassidy>Nvidia <00:01:13.527>isn't <00:01:13.767>writing <00:01:14.063>that <00:01:14.180>check <00:01:14.487>for <00:01:14.627>the <00:01:14.777>returns.

18
00:01:15.679 --> 00:01:23.542
<v Cassidy>When <00:01:16.147>the <00:01:16.239>chip <00:01:16.417>company <00:01:16.820>shows <00:01:17.114>up <00:01:17.220>at <00:01:17.347>the <00:01:17.457>latency <00:01:17.860>layer <00:01:18.154>of <00:01:18.247>a <00:01:18.351>multilingual <00:01:18.980>voice <00:01:19.314>model, <00:01:20.143>they're <00:01:20.321>buying <00:01:20.607>access <00:01:21.140>to <00:01:21.287>how <00:01:21.471>inference <00:01:22.109>actually <00:01:22.655>gets <00:01:22.947>run.

19
00:01:23.542 --> 00:01:28.047
<v Bill>Right — <00:01:24.660>so <00:01:24.820>which <00:01:25.140>is <00:01:25.300>it: <00:01:25.940>customer, <00:01:26.657>partner, <00:01:27.300>or <00:01:27.500>hedge?

20
00:01:28.407 --> 00:01:33.007
<v Bill>A <00:01:28.834>voice <00:01:29.227>startup <00:01:29.700>at <00:01:29.855>this <00:01:30.114>scale <00:01:30.500>is <00:01:30.660>going <00:01:30.954>to <00:01:31.034>be <00:01:31.127>a <00:01:31.247>serious <00:01:31.987>GPU <00:01:32.594>buyer.

21
00:01:33.135 --> 00:01:37.274
<v Bill>That's <00:01:34.058>Nvidia <00:01:34.417>funding <00:01:34.827>its <00:01:35.067>own <00:01:35.316>future <00:01:35.780>purchase <00:01:36.260>order.

22
00:01:37.291 --> 00:01:42.207
<v Cassidy>We <00:01:37.367>saw <00:01:37.527>a <00:01:37.607>version <00:01:37.914>of <00:01:37.999>this <00:01:38.191>last <00:01:38.495>week <00:01:38.719>with <00:01:38.887>Cerebras <00:01:39.547>subletting <00:01:40.020>European <00:01:40.518>capacity <00:01:41.117>through <00:01:41.407>a <00:01:41.519>chip <00:01:41.707>company.

23
00:01:42.383 --> 00:01:49.247
<v Cassidy>That's <00:01:43.234>twice <00:01:43.594>in <00:01:43.687>a <00:01:43.767>few <00:01:43.951>days <00:01:44.276>Nvidia <00:01:44.727>has <00:01:44.881>gotten <00:01:45.135>into <00:01:45.383>inference <00:01:45.907>through <00:01:46.087>a <00:01:46.183>financial <00:01:46.738>instrument, <00:01:47.687>not <00:01:47.847>a <00:01:47.940>plain <00:01:48.257>silicon <00:01:48.703>sale.

24
00:01:49.427 --> 00:01:53.047
<v Cassidy>The <00:01:50.017>pattern's <00:01:50.477>getting <00:01:50.751>hard <00:01:51.034>to <00:01:51.135>call <00:01:51.367>a <00:01:51.480>coincidence.

25
00:01:53.057 --> 00:01:56.527
<v Bill>And <00:01:53.127>a <00:01:53.207>hundred-million <00:01:54.095>seed. <00:01:54.959>Seed.

26
00:01:56.623 --> 00:02:06.687
<v Bill>That <00:01:57.064>number <00:01:57.460>would've <00:01:57.864>raised <00:01:58.234>an <00:01:58.367>eyebrow <00:01:59.007>in <00:01:59.231>late-2021 <00:02:00.487>SaaS, <00:02:01.527>and <00:02:01.700>it <00:02:01.830>should <00:02:02.020>raise <00:02:02.347>one <00:02:02.687>now — <00:02:03.594>it <00:02:03.714>reads <00:02:04.079>like <00:02:04.247>a <00:02:04.359>strategic <00:02:05.076>anchor <00:02:05.535>with <00:02:05.767>a <00:02:05.887>seed <00:02:06.260>label.

27
00:02:07.332 --> 00:02:08.492
<v Cassidy>From <00:02:07.549>Nature <00:02:07.954>Medicine:

28
00:02:08.929 --> 00:02:10.849
<v Quote>We <00:02:09.129>introduce <00:02:09.556>NeuroVFM

29
00:02:10.889 --> 00:02:19.649
<v Quote>a <00:02:10.975>visual <00:02:11.300>foundation <00:02:11.862>model <00:02:12.169>trained <00:02:12.502>on <00:02:12.769>5.24 <00:02:13.859>million <00:02:14.253>clinical <00:02:14.769>MRI <00:02:15.349>and <00:02:15.809>CT <00:02:16.189>volumes <00:02:16.702>using <00:02:17.209>a <00:02:17.311>scalable <00:02:17.876>volumetric <00:02:18.493>predictive <00:02:18.978>architecture.

30
00:02:19.862 --> 00:02:25.329
<v Quote>NeuroVFM <00:02:21.535>learns <00:02:21.855>comprehensive <00:02:22.504>representations <00:02:23.356>of <00:02:23.449>brain <00:02:23.719>anatomy <00:02:24.229>and <00:02:24.353>pathology

31
00:02:25.369 --> 00:02:29.089
<v Quote>achieving <00:02:25.782>state-of-the-art <00:02:26.416>performance <00:02:26.975>across <00:02:27.302>multiple <00:02:27.773>clinical <00:02:28.222>tasks

32
00:02:29.129 --> 00:02:32.049
<v Quote>including <00:02:29.537>radiologic <00:02:30.193>diagnosis <00:02:30.829>and <00:02:31.055>report <00:02:31.387>generation.

33
00:02:32.302 --> 00:02:36.302
<v Cassidy>So <00:02:32.510>here's <00:02:32.802>the <00:02:32.922>one <00:02:33.129>in <00:02:33.235>today's <00:02:33.662>rundown <00:02:34.422>I <00:02:34.658>actually <00:02:35.002>can't <00:02:35.326>stop <00:02:35.618>thinking <00:02:35.955>about.

34
00:02:36.502 --> 00:02:45.102
<v Cassidy>NeuroVFM — <00:02:38.662>trained <00:02:39.022>on <00:02:39.183>5.24 <00:02:40.192>million <00:02:40.586>clinical <00:02:41.102>MRI <00:02:41.622>and <00:02:41.875>CT <00:02:42.222>volumes, <00:02:43.013>straight <00:02:43.622>out <00:02:43.795>of <00:02:43.942>routine <00:02:44.266>hospital <00:02:44.702>care.

35
00:02:45.302 --> 00:02:53.422
<v Cassidy>And <00:02:45.922>the <00:02:46.028>reason <00:02:46.334>that <00:02:46.532>matters: <00:02:47.315>brain <00:02:48.049>scans <00:02:48.422>are <00:02:48.590>basically <00:02:49.091>absent <00:02:49.534>from <00:02:49.682>the <00:02:49.799>public <00:02:50.222>web, <00:02:50.622>because <00:02:51.202>you <00:02:51.302>can <00:02:51.469>reconstruct <00:02:52.022>a <00:02:52.126>face <00:02:52.414>from <00:02:52.569>an <00:02:52.802>MRI.

36
00:02:54.269 --> 00:02:59.262
<v Cassidy>Which <00:02:54.355>means <00:02:54.642>the <00:02:54.755>frontier <00:02:55.239>models <00:02:55.635>everyone <00:02:56.113>benchmarks <00:02:56.675>to <00:02:56.835>death <00:02:57.169>never <00:02:57.542>had <00:02:57.710>this <00:02:57.966>data <00:02:58.262>and <00:02:58.442>can't <00:02:58.762>buy <00:02:59.022>it.

37
00:02:59.454 --> 00:03:03.891
<v Cassidy>Here, <00:03:00.322>the <00:03:00.435>moat <00:03:00.662>comes <00:03:00.894>from <00:03:01.073>permission: <00:03:01.842>who's <00:03:02.502>allowed <00:03:02.809>to <00:03:02.902>touch <00:03:03.122>the <00:03:03.302>scans.

38
00:03:03.891 --> 00:03:10.462
<v Bill>Yeah, <00:03:04.422>and <00:03:04.606>this <00:03:04.915>is <00:03:05.062>the <00:03:05.249>first <00:03:05.635>clinical-data <00:03:06.542>story <00:03:07.022>all <00:03:07.342>week <00:03:07.726>that <00:03:08.042>isn't <00:03:08.455>about <00:03:08.702>megawatts <00:03:09.502>or <00:03:09.753>leases.

39
00:03:10.682 --> 00:03:18.702
<v Bill>You <00:03:11.382>can <00:03:11.550>pour <00:03:11.822>a <00:03:11.962>gigawatt <00:03:12.526>into <00:03:12.822>a <00:03:12.926>data <00:03:13.239>center <00:03:13.662>and <00:03:13.955>it <00:03:14.102>still <00:03:14.342>doesn't <00:03:14.662>get <00:03:14.842>you <00:03:15.079>access <00:03:15.475>to <00:03:15.662>a <00:03:15.795>hospital's <00:03:16.489>uncurated <00:03:17.475>CT <00:03:17.982>volumes.

40
00:03:19.487 --> 00:03:23.262
<v Bill>The <00:03:19.550>line <00:03:19.774>that <00:03:19.982>got <00:03:20.302>me — <00:03:20.976>paired <00:03:21.214>with <00:03:21.406>open-source <00:03:22.106>language <00:03:22.531>models

41
00:03:23.342 --> 00:03:28.942
<v Bill>it <00:03:23.555>beats <00:03:23.842>the <00:03:23.955>frontier <00:03:24.439>models <00:03:24.862>on <00:03:25.091>report <00:03:25.582>accuracy <00:03:26.362>and <00:03:26.862>cuts <00:03:27.262>hallucinated <00:03:28.142>findings.

42
00:03:29.150 --> 00:03:33.262
<v Bill>That's <00:03:30.102>the <00:03:30.270>part <00:03:30.494>that <00:03:30.693>survives <00:03:31.352>contact <00:03:32.014>with <00:03:32.242>a <00:03:32.382>radiologist.

43
00:03:33.482 --> 00:03:38.782
<v Bill>Fewer <00:03:34.364>critical <00:03:34.953>errors <00:03:35.555>is <00:03:35.682>the <00:03:35.806>only <00:03:36.142>spec <00:03:36.510>that <00:03:36.752>matters <00:03:37.155>in <00:03:37.310>that <00:03:37.582>room.

44
00:03:38.802 --> 00:03:46.142
<v Cassidy>And <00:03:38.942>notice, <00:03:39.822>the <00:03:40.111>architecture <00:03:40.759>solves <00:03:41.142>the <00:03:41.302>thing <00:03:41.529>we've <00:03:41.694>been <00:03:41.875>stuck <00:03:42.195>on <00:03:42.362>all <00:03:42.622>week: <00:03:43.702>who <00:03:43.893>controls <00:03:44.526>fine-tuning <00:03:45.302>and <00:03:45.486>inference.

45
00:03:46.462 --> 00:03:52.222
<v Cassidy>In <00:03:47.125>health <00:03:47.342>systems, <00:03:48.355>federated <00:03:48.826>learning <00:03:49.222>has <00:03:49.382>a <00:03:49.466>clinical-grade <00:03:50.199>answer <00:03:50.609>baked <00:03:51.022>in <00:03:51.209>by <00:03:51.342>regulation.

46
00:03:52.402 --> 00:03:55.535
<v Cassidy>The <00:03:53.006>data <00:03:53.310>literally <00:03:53.822>can't <00:03:54.169>leave <00:03:54.342>the <00:03:54.506>building.

47
00:03:55.535 --> 00:04:02.222
<v Bill>Right, <00:03:56.282>and <00:03:56.506>remember <00:03:56.894>that <00:03:57.072>startup <00:03:57.502>study — <00:03:58.562>the <00:03:58.682>one <00:03:58.919>saying <00:03:59.369>reorganizing <00:04:00.286>work <00:04:00.606>beat <00:04:00.953>access <00:04:01.555>to <00:04:01.682>the <00:04:01.822>tool?

48
00:04:03.446 --> 00:04:04.470
<v Bill>Here's <00:04:03.560>the <00:04:03.630>counter-case.

49
00:04:04.653 --> 00:04:07.350
<v Bill>Nobody <00:04:05.857>reorganized <00:04:06.670>a <00:04:06.772>hospital.

50
00:04:07.570 --> 00:04:14.790
<v Bill>The <00:04:08.283>model <00:04:08.610>can <00:04:08.870>do <00:04:09.078>something <00:04:09.542>that <00:04:09.734>wasn't <00:04:10.141>possible <00:04:10.790>before <00:04:11.360>because <00:04:12.030>the <00:04:12.214>data-access <00:04:13.098>architecture <00:04:14.010>changed.

51
00:04:15.030 --> 00:04:19.750
<v Bill>You <00:04:15.790>get <00:04:15.950>a <00:04:16.054>real <00:04:16.375>capability <00:04:17.110>jump <00:04:17.478>here, <00:04:18.190>not <00:04:18.358>just <00:04:18.590>a <00:04:18.699>faster <00:04:19.190>horse.

52
00:04:20.907 --> 00:04:23.227
<v Cassidy>Here's <00:04:21.107>Min-A <00:04:21.467>Kang <00:04:21.840>at <00:04:21.993>Nature <00:04:22.331>Communications:

53
00:04:23.645 --> 00:04:32.365
<v Quote>Arrays <00:04:24.178>of <00:04:24.312>these <00:04:24.525>cerebellum-inspired <00:04:25.685>memtransistors <00:04:26.755>exploit <00:04:27.285>the <00:04:27.458>evolving <00:04:27.938>interplay <00:04:28.425>between <00:04:28.912>excitatory <00:04:29.765>and <00:04:29.938>inhibitory <00:04:30.621>responses <00:04:31.592>to <00:04:31.865>emulate <00:04:32.305>the

54
00:04:32.432 --> 00:04:35.645
<v Quote>emergent <00:04:32.987>synaptic <00:04:33.540>differentiation <00:04:34.418>of <00:04:34.545>the <00:04:34.669>cerebellum

55
00:04:35.698 --> 00:04:38.605
<v Quote>enabling <00:04:36.192>rapid <00:04:36.578>identification <00:04:37.405>of <00:04:37.632>novel <00:04:38.045>events.

56
00:04:38.781 --> 00:04:41.725
<v Quote>When <00:04:39.535>applied <00:04:39.938>to <00:04:40.105>electrocardiogram <00:04:41.229>data

57
00:04:41.785 --> 00:04:48.765
<v Quote>arrhythmias <00:04:42.385>are <00:04:42.498>detected <00:04:42.971>within <00:04:43.285>a <00:04:43.371>single <00:04:43.693>heartbeat <00:04:44.173>with <00:04:44.445>10,000-fold <00:04:45.392>fewer <00:04:45.790>operations <00:04:46.493>than <00:04:46.738>existing <00:04:47.245>silicon-based <00:04:48.103>approaches.

58
00:04:49.054 --> 00:04:54.734
<v Bill>Okay <00:04:49.806>here's <00:04:50.054>the <00:04:50.208>number <00:04:50.446>that <00:04:50.634>stopped <00:04:51.107>me: <00:04:51.534>arrhythmia <00:04:52.390>detection <00:04:52.940>within <00:04:53.254>a <00:04:53.351>single <00:04:53.758>heartbeat

59
00:04:54.814 --> 00:04:59.614
<v Bill>at <00:04:55.014>ten <00:04:55.187>thousand <00:04:55.707>times <00:04:56.267>fewer <00:04:56.807>operations <00:04:57.646>than <00:04:58.114>the <00:04:58.234>silicon <00:04:58.725>approach.

60
00:04:59.999 --> 00:05:05.116
<v Bill>You <00:05:00.059>don't <00:05:00.323>hand-wave <00:05:00.851>that <00:05:01.059>away <00:05:01.432>as <00:05:01.579>a <00:05:01.681>rounding <00:05:02.232>error; <00:05:03.032>it's <00:05:03.179>a <00:05:03.259>different <00:05:03.686>order <00:05:04.046>of <00:05:04.179>magnitude.

61
00:05:05.116 --> 00:05:10.899
<v Cassidy>Cerebellum-inspired <00:05:07.032>MoS2 <00:05:07.859>memtransistors, <00:05:09.139>out <00:05:09.432>of <00:05:09.585>Nature <00:05:09.923>Communications.

62
00:05:11.099 --> 00:05:19.740
<v Cassidy>The <00:05:11.726>whole <00:05:11.939>premise <00:05:12.526>is <00:05:12.611>that <00:05:12.819>data-center <00:05:13.449>silicon <00:05:13.992>is <00:05:14.179>drowning <00:05:14.846>in <00:05:14.959>its <00:05:15.179>own <00:05:15.352>power <00:05:15.683>draw, <00:05:16.712>so <00:05:16.851>they <00:05:17.011>went <00:05:17.179>looking <00:05:17.486>at <00:05:17.609>biology <00:05:18.139>for <00:05:18.299>a <00:05:18.389>cheaper <00:05:18.811>substrate.

63
00:05:19.740 --> 00:05:25.939
<v Bill>And <00:05:20.099>the <00:05:20.259>target <00:05:20.659>is <00:05:21.219>edge — <00:05:22.477>healthcare, <00:05:23.388>robotics, <00:05:24.477>autonomous <00:05:25.139>vehicles.

64
00:05:26.710 --> 00:05:34.180
<v Bill>The <00:05:26.807>exact <00:05:27.209>places <00:05:27.660>where <00:05:27.900>you <00:05:28.080>don't <00:05:28.400>get <00:05:28.607>to <00:05:28.713>phone <00:05:29.060>home <00:05:29.540>to <00:05:29.940>a <00:05:30.100>401-megawatt <00:05:31.620>campus <00:05:32.193>every <00:05:32.740>time <00:05:33.080>you <00:05:33.188>need <00:05:33.407>to <00:05:33.540>think.

65
00:05:34.180 --> 00:05:36.820
<v Cassidy>That's <00:05:34.840>the <00:05:34.962>contrast <00:05:35.500>I <00:05:35.588>keep <00:05:35.810>sitting <00:05:36.052>with <00:05:36.228>this <00:05:36.484>week.

66
00:05:37.060 --> 00:05:41.620
<v Cassidy>We've <00:05:37.673>spent <00:05:37.980>three <00:05:38.260>days <00:05:38.607>on <00:05:38.720>gigawatts <00:05:39.340>and <00:05:39.510>balance <00:05:39.940>sheets — <00:05:40.776>building <00:05:41.129>bigger.

67
00:05:42.548 --> 00:05:48.588
<v Cassidy>This <00:05:42.647>paper <00:05:43.033>is <00:05:43.210>quietly <00:05:43.769>asking <00:05:44.210>whether <00:05:44.489>bigger <00:05:44.980>is <00:05:45.204>even <00:05:45.480>the <00:05:45.580>right <00:05:45.852>direction <00:05:46.628>once <00:05:47.160>you're <00:05:47.360>out <00:05:47.567>at <00:05:47.660>the <00:05:47.876>edge.

68
00:05:48.588 --> 00:05:58.420
<v Bill>Right — <00:05:49.833>if <00:05:49.988>edge <00:05:50.228>inference <00:05:50.724>gets <00:05:51.028>this <00:05:51.287>cheap <00:05:51.647>on <00:05:51.740>a <00:05:51.847>novel <00:05:52.244>substrate, <00:05:53.500>the <00:05:53.673>whole <00:05:54.052>cost-per-operation <00:05:55.540>math <00:05:55.951>shifts <00:05:56.557>before <00:05:56.940>you <00:05:57.124>ever <00:05:57.444>rent <00:05:57.780>a <00:05:57.940>rack.

69
00:05:58.628 --> 00:06:02.900
<v Bill>That's <00:05:59.580>the <00:05:59.764>part <00:06:00.060>the <00:06:00.228>parameter-count <00:06:01.060>crowd <00:06:01.527>never <00:06:01.940>prices <00:06:02.527>in.

70
00:06:03.552 --> 00:06:05.432
<v Cassidy>This <00:06:03.752>one's <00:06:03.984>from <00:06:04.158>Nature <00:06:04.496>Communications:

71
00:06:05.877 --> 00:06:09.477
<v Quote>Here <00:06:06.090>we <00:06:06.245>show <00:06:06.469>that <00:06:06.650>chemical <00:06:07.121>language <00:06:07.534>models <00:06:07.917>can <00:06:08.085>gain <00:06:08.357>a <00:06:08.494>comprehensive

72
00:06:09.530 --> 00:06:14.277
<v Quote>multi-modal <00:06:10.248>understanding <00:06:10.944>of <00:06:11.085>molecules <00:06:11.917>through <00:06:12.243>heterogeneous <00:06:12.885>molecular <00:06:13.379>encoding

73
00:06:14.304 --> 00:06:16.837
<v Quote>which <00:06:14.481>integrates <00:06:14.997>one-dimensional <00:06:15.781>sequences

74
00:06:16.857 --> 00:06:18.197
<v Quote>two-dimensional <00:06:17.468>topology

75
00:06:18.290 --> 00:06:20.357
<v Quote>three-dimensional <00:06:19.308>geometry

76
00:06:20.377 --> 00:06:22.997
<v Quote>and <00:06:20.488>statistically <00:06:21.217>derived <00:06:21.701>molecular <00:06:22.333>fragments.

77
00:06:23.237 --> 00:06:33.237
<v Bill>Okay <00:06:23.877>so <00:06:24.133>this <00:06:24.557>one's <00:06:24.797>a <00:06:24.908>chemical <00:06:25.539>language <00:06:26.104>model <00:06:26.517>paper — <00:06:27.653>they're <00:06:27.837>basically <00:06:28.254>saying <00:06:28.637>the <00:06:28.864>SMILES-string <00:06:30.010>approach <00:06:30.557>loses <00:06:31.197>the <00:06:31.397>actual <00:06:32.037>3D <00:06:32.664>shape <00:06:33.024>of <00:06:33.117>the

78
00:06:33.308 --> 00:06:39.317
<v Bill>molecule <00:06:34.397>and <00:06:34.533>they're <00:06:34.717>bolting <00:06:35.077>on <00:06:35.317>topology <00:06:36.017>and <00:06:36.277>geometry <00:06:37.077>to <00:06:37.477>close <00:06:37.925>that <00:06:38.237>gap.

79
00:06:39.337 --> 00:06:43.797
<v Cassidy>And <00:06:39.445>they <00:06:39.664>built <00:06:39.877>a <00:06:40.007>million-scale <00:06:40.737>dataset <00:06:41.197>for <00:06:41.384>multi-objective <00:06:42.245>molecular <00:06:42.746>design <00:06:43.184>to <00:06:43.290>train <00:06:43.610>it.

80
00:06:43.957 --> 00:06:50.561
<v Cassidy>That's <00:06:44.697>the <00:06:44.789>part <00:06:44.933>that <00:06:45.047>catches <00:06:45.477>me — <00:06:46.297>the <00:06:46.410>moat <00:06:46.661>here <00:06:46.970>is <00:06:47.117>the <00:06:47.317>data, <00:06:48.181>same <00:06:48.464>as <00:06:48.537>the <00:06:48.664>neuroimaging <00:06:49.370>piece <00:06:49.664>we <00:06:49.765>just <00:06:50.057>hit.

81
00:06:50.561 --> 00:07:01.478
<v Bill>Right, <00:06:51.338>and <00:06:51.438>the <00:06:51.638>honest <00:06:52.198>read: <00:06:52.822>this <00:06:53.478>is <00:06:53.678>a <00:06:53.792>comprehension-and-design <00:06:55.726>benchmark <00:06:56.698>win — <00:06:57.838>quote, <00:06:58.205>consistent <00:06:59.133>improvements <00:06:59.846>over <00:07:00.107>strong <00:07:00.578>baselines.

82
00:07:02.325 --> 00:07:07.975
<v Bill>Which <00:07:02.411>is <00:07:02.538>the <00:07:02.667>phrase <00:07:03.131>every <00:07:03.465>paper <00:07:03.894>uses <00:07:04.318>right <00:07:04.564>before <00:07:04.858>you <00:07:04.982>find <00:07:05.298>out <00:07:05.510>what <00:07:05.878>'strong <00:07:06.518>baselines' <00:07:07.371>meant.

83
00:07:07.975 --> 00:07:15.158
<v Cassidy>The <00:07:08.318>chain-of-fragment <00:07:09.134>mechanism <00:07:09.691>is <00:07:09.838>the <00:07:09.991>interesting <00:07:10.418>bit <00:07:10.638>for <00:07:10.891>me — <00:07:11.734>they're <00:07:11.918>guiding <00:07:12.256>generation <00:07:12.868>through <00:07:13.118>a <00:07:13.213>hierarchical <00:07:13.942>blueprint <00:07:14.558>instead <00:07:14.865>of <00:07:14.978>one

84
00:07:15.238 --> 00:07:19.958
<v Cassidy>shot. <00:07:15.918>The <00:07:16.558>structure <00:07:16.985>fills <00:07:17.371>in <00:07:17.510>what <00:07:17.705>plain <00:07:18.082>language <00:07:18.488>strings <00:07:18.934>miss.

85
00:07:20.065 --> 00:07:22.278
<v Bill>It's <00:07:20.451>chain-of-thought <00:07:21.358>for <00:07:21.550>molecules.

86
00:07:22.358 --> 00:07:30.758
<v Bill>Whether <00:07:22.971>it <00:07:23.140>survives <00:07:23.718>contact <00:07:24.310>with <00:07:24.491>an <00:07:24.667>actual <00:07:25.142>synthesis <00:07:25.838>lab <00:07:26.358>is <00:07:26.758>a <00:07:26.905>whole <00:07:27.286>different <00:07:27.785>error <00:07:28.198>rate <00:07:28.582>than <00:07:28.878>a <00:07:28.982>benchmark <00:07:29.638>score

87
00:07:30.778 --> 00:07:36.518
<v Bill>but <00:07:30.878>the <00:07:31.051>fragment-by-fragment <00:07:32.331>approach <00:07:32.838>is <00:07:33.025>at <00:07:33.131>least <00:07:33.489>honest <00:07:34.131>about <00:07:34.651>where <00:07:34.971>these <00:07:35.307>models <00:07:35.798>fall <00:07:36.182>over.

88
00:07:37.798 --> 00:07:40.918
<v Cassidy>If <00:07:37.898>you <00:07:38.004>follow <00:07:38.385>AI <00:07:38.651>Daily <00:07:38.958>Briefing <00:07:39.358>for <00:07:39.498>the <00:07:39.638>policy <00:07:40.175>stakes

89
00:07:40.998 --> 00:07:42.918
<v Cassidy>try <00:07:41.302>Anthropic <00:07:41.940>Pentagon <00:07:42.505>Watch

90
00:07:43.198 --> 00:07:47.078
<v Cassidy>a <00:07:43.531>daily <00:07:43.842>briefing <00:07:44.251>on <00:07:44.406>Anthropic’s <00:07:45.105>fight <00:07:45.270>with <00:07:45.418>the <00:07:45.611>DoD <00:07:46.166>over <00:07:46.438>Claude

91
00:07:47.176 --> 00:07:50.358
<v Cassidy>military <00:07:48.065>AI <00:07:48.418>use <00:07:49.318>autonomous <00:07:49.778>weapons

92
00:07:50.458 --> 00:07:52.918
<v Cassidy>and <00:07:50.865>AI <00:07:51.125>procurement <00:07:51.691>blacklisting.

93
00:07:53.014 --> 00:07:55.318
<v Cassidy>Find <00:07:53.451>it <00:07:53.602>wherever <00:07:53.978>you <00:07:54.072>listen <00:07:54.331>to <00:07:54.518>podcasts.

94
00:07:56.191 --> 00:08:03.071
<v Cassidy>We’ve <00:07:56.391>put <00:07:56.564>links <00:07:56.858>to <00:07:56.955>everything <00:07:57.444>we <00:07:57.591>covered <00:07:57.911>today <00:07:58.138>in <00:07:58.211>the <00:07:58.319>show <00:07:58.578>notes, <00:07:59.364>so <00:07:59.524>if <00:07:59.671>a <00:07:59.764>story <00:08:00.079>made <00:08:00.291>you <00:08:00.399>want <00:08:00.611>the <00:08:00.717>fuller <00:08:01.071>read, <00:08:01.743>that’s <00:08:01.971>the <00:08:02.084>place <00:08:02.378>to <00:08:02.511>start.

95
00:08:03.215 --> 00:08:05.791
<v Cassidy>That’s <00:08:04.111>AI <00:08:04.324>Daily <00:08:04.644>Briefing <00:08:05.031>for <00:08:05.244>today.

96
00:08:05.919 --> 00:08:08.191
<v Cassidy>This <00:08:06.458>is <00:08:06.551>a <00:08:06.641>Lantern <00:08:07.091>Podcast.