WEBVTT

1
00:00:00.000 --> 00:00:07.440
<v Cassidy>The <00:00:00.160>AI <00:00:00.384>land <00:00:00.720>grab <00:00:01.080>has <00:00:01.240>run <00:00:01.392>into <00:00:01.600>a <00:00:01.726>harder <00:00:02.062>question: <00:00:02.773>where <00:00:03.480>are <00:00:03.620>the <00:00:03.720>working <00:00:04.080>machines—and <00:00:05.320>can <00:00:05.460>the <00:00:05.577>models <00:00:05.973>on <00:00:06.112>them <00:00:06.300>survive <00:00:06.800>reality?

2
00:00:09.408 --> 00:00:11.840
<v Cassidy>Here’s <00:00:09.560>how <00:00:09.707>we <00:00:09.820>got <00:00:10.048>here: <00:00:10.464>Microsoft

3
00:00:11.872 --> 00:00:13.360
<v Cassidy>Meta <00:00:12.080>Platforms <00:00:13.074>Oracle

4
00:00:13.429 --> 00:00:21.920
<v Cassidy>Amazon <00:00:13.940>and <00:00:14.062>Alphabet <00:00:14.672>have <00:00:15.152>been <00:00:15.333>stacking <00:00:15.813>up <00:00:16.000>long-term <00:00:16.773>AI <00:00:17.088>data-centre <00:00:17.691>leases <00:00:18.347>as <00:00:18.604>training <00:00:19.000>and <00:00:19.168>inference <00:00:19.669>demand <00:00:20.140>outgrows <00:00:20.667>their <00:00:20.853>owned <00:00:21.191>capacity.

5
00:00:22.060 --> 00:00:29.440
<v Cassidy>Reuters <00:00:23.080>put <00:00:23.240>known <00:00:23.493>commitments <00:00:24.107>at <00:00:24.213>about <00:00:24.560>$1.16 <00:00:25.973>trillion <00:00:27.133>after <00:00:27.648>Meta’s <00:00:27.956>additional <00:00:28.400>July <00:00:28.800>leases.

6
00:00:29.660 --> 00:00:35.918
<v Cassidy>The <00:00:30.364>question <00:00:30.820>now <00:00:31.200>is <00:00:31.400>how <00:00:31.560>those <00:00:31.820>obligations <00:00:32.537>become <00:00:32.949>usable <00:00:33.440>compute <00:00:34.060>and <00:00:34.310>revenue.

7
00:00:35.918 --> 00:00:38.800
<v Bill>AI <00:00:36.560>Daily <00:00:37.027>Briefing—today

8
00:00:38.880 --> 00:00:42.960
<v Bill>one <00:00:39.190>compute <00:00:39.760>deal <00:00:40.160>goes <00:00:40.528>from <00:00:40.800>promise <00:00:41.467>to <00:00:41.751>delivery

9
00:00:43.013 --> 00:00:48.880
<v Bill>while <00:00:43.350>science <00:00:43.893>researchers <00:00:44.533>start <00:00:44.869>stress-testing <00:00:45.720>the <00:00:45.931>models <00:00:46.587>we <00:00:46.848>keep <00:00:47.090>putting <00:00:47.493>on <00:00:47.660>all <00:00:47.888>that <00:00:48.160>hardware.

10
00:00:49.180 --> 00:00:51.200
<v Bill>Let’s <00:00:50.267>start <00:00:50.427>in <00:00:50.587>Texas.

11
00:00:51.824 --> 00:00:53.904
<v Cassidy>Here's <00:00:52.018>Bailey <00:00:52.272>Pemberton <00:00:52.837>at <00:00:52.990>Simply <00:00:53.312>Wall <00:00:53.557>St:

12
00:00:54.511 --> 00:01:06.191
<v Quote>Microsoft <00:00:55.471>(NasdaqGS:MSFT) <00:00:58.551>has <00:00:58.724>accepted <00:00:59.191>the <00:00:59.364>first <00:00:59.871>50MW <00:01:00.858>of <00:01:01.044>AI <00:01:01.378>cloud <00:01:01.782>capacity <00:01:02.399>from <00:01:02.751>IREN <00:01:03.284>at <00:01:03.571>the <00:01:03.684>Childress <00:01:04.180>campus <00:01:04.644>under <00:01:04.951>a <00:01:05.055>multibillion <00:01:05.917>dollar

13
00:01:06.262 --> 00:01:06.911
<v Quote>contract.

14
00:01:07.111 --> 00:01:15.391
<v Quote>The <00:01:07.762>deployment <00:01:08.324>is <00:01:08.484>designed <00:01:08.931>for <00:01:09.060>hyperscale <00:01:09.924>AI <00:01:10.255>workloads <00:01:10.931>and <00:01:11.298>forms <00:01:11.679>part <00:01:11.898>of <00:01:12.007>Microsoft's <00:01:12.724>buildout <00:01:13.204>of <00:01:13.359>next <00:01:13.602>generation <00:01:14.191>cloud <00:01:14.644>infrastructure.

15
00:01:15.802 --> 00:01:23.642
<v Bill>On <00:01:15.942>Big <00:01:16.170>Tech’s <00:01:16.655>AI <00:01:17.002>leases: <00:01:17.690>Microsoft <00:01:18.902>has <00:01:19.135>taken <00:01:19.462>its <00:01:19.709>first <00:01:20.202>50 <00:01:20.582>megawatts <00:01:21.370>from <00:01:21.762>IREN’s <00:01:22.442>Childress <00:01:23.082>campus.

16
00:01:25.855 --> 00:01:34.348
<v Bill>Accepted <00:01:26.042>capacity. <00:01:26.831>Actual <00:01:27.784>hardware <00:01:28.322>and <00:01:28.522>power <00:01:28.982>someone <00:01:29.442>has <00:01:29.619>handed <00:01:30.026>over—not <00:01:31.052>another <00:01:31.482>artist’s <00:01:32.186>rendering <00:01:32.789>of <00:01:32.882>a <00:01:32.991>future <00:01:33.482>data <00:01:33.882>hall.

17
00:01:34.348 --> 00:01:36.362
<v Cassidy>And <00:01:34.842>that <00:01:35.082>distinction <00:01:35.722>matters.

18
00:01:36.502 --> 00:01:45.322
<v Cassidy>The <00:01:36.984>contract <00:01:37.535>is <00:01:37.690>multibillion-dollar, <00:01:39.262>but <00:01:39.370>accepting <00:01:39.882>50MW <00:01:41.082>is <00:01:41.302>the <00:01:41.535>first <00:01:41.850>delivered <00:01:42.354>milestone <00:01:43.055>in <00:01:43.210>this <00:01:43.466>week’s <00:01:43.850>infrastructure <00:01:44.569>commitments.

19
00:01:45.482 --> 00:01:48.510
<v Cassidy>It <00:01:45.862>belongs <00:01:46.309>in <00:01:46.382>the <00:01:46.499>ledger <00:01:46.895>as <00:01:47.082>real <00:01:47.482>capacity.

20
00:01:48.510 --> 00:01:58.842
<v Bill>Fifty <00:01:49.162>megawatts <00:01:50.015>is <00:01:50.382>serious—big <00:01:51.928>enough <00:01:52.234>that <00:01:52.452>latency, <00:01:53.460>networking, <00:01:54.432>cooling, <00:01:55.042>and <00:01:55.222>failure <00:01:55.731>handling <00:01:56.522>stop <00:01:56.895>being <00:01:57.242>slide-deck <00:01:58.015>abstractions.

21
00:01:59.242 --> 00:02:10.682
<v Bill>IREN’s <00:02:00.133>NVIDIA <00:02:00.789>Exemplar <00:02:01.575>Cloud <00:02:02.135>badge <00:02:02.642>and <00:02:02.826>financing <00:02:03.482>package <00:02:04.154>help <00:02:04.662>explain <00:02:05.142>how <00:02:05.375>it <00:02:05.542>got <00:02:05.775>built; <00:02:06.922>Microsoft <00:02:07.698>accepting <00:02:08.255>it <00:02:08.549>is <00:02:08.942>the <00:02:09.066>part <00:02:09.402>I <00:02:09.530>care <00:02:09.802>about.

22
00:02:10.746 --> 00:02:19.242
<v Cassidy>Microsoft <00:02:11.342>and <00:02:11.442>its <00:02:11.622>Big <00:02:11.850>Tech <00:02:12.135>peers <00:02:12.675>reportedly <00:02:13.450>have <00:02:13.682>about <00:02:13.962>$1.16 <00:02:15.269>trillion <00:02:16.175>in <00:02:16.389>AI <00:02:16.666>data-center <00:02:17.282>lease <00:02:17.529>commitments <00:02:18.069>in <00:02:18.142>the <00:02:18.264>pipeline

23
00:02:19.282 --> 00:02:27.642
<v Cassidy>per <00:02:19.459>Simply <00:02:19.850>Wall <00:02:20.095>St. <00:02:20.722>Against <00:02:21.610>that <00:02:21.848>figure, <00:02:22.362>50MW <00:02:23.375>is <00:02:23.582>one <00:02:23.962>tranche—but <00:02:25.295>it <00:02:25.442>comes <00:02:25.674>with <00:02:25.842>a <00:02:25.926>customer <00:02:26.346>sign-off <00:02:26.869>on <00:02:26.984>delivery.

24
00:02:28.405 --> 00:02:30.005
<v Cassidy>Here's <00:02:28.645>Saf <00:02:28.938>Malik <00:02:29.232>at <00:02:29.356>Capacity:

25
00:02:30.323 --> 00:02:32.563
<v Quote>The <00:02:30.454>deployment <00:02:30.995>will <00:02:31.173>support <00:02:31.603>AI <00:02:31.816>training

26
00:02:32.611 --> 00:02:37.523
<v Quote>inference <00:02:33.336>agentic <00:02:33.896>AI <00:02:34.283>and <00:02:34.687>enterprise <00:02:35.230>AI <00:02:35.507>workloads <00:02:36.023>for <00:02:36.163>Nebius <00:02:36.619>customers

27
00:02:37.563 --> 00:02:38.963
<v Quote>including <00:02:37.983>enterprises

28
00:02:39.023 --> 00:02:42.963
<v Quote>researchers <00:02:40.056>startups <00:02:40.583>and <00:02:40.929>public <00:02:41.260>sector <00:02:41.660>organisations.

29
00:02:43.016 --> 00:02:46.403
<v Quote>It <00:02:43.190>forms <00:02:43.555>part <00:02:43.736>of <00:02:43.903>Nebius’ <00:02:44.483>wider <00:02:45.043>UK <00:02:45.427>expansion

30
00:02:46.430 --> 00:02:54.243
<v Quote>which <00:02:46.616>included <00:02:47.083>a <00:02:47.167>commitment <00:02:47.656>in <00:02:47.843>June <00:02:48.216>to <00:02:48.363>build <00:02:48.623>out <00:02:48.854>approximately <00:02:49.643>£1.7 <00:02:50.593>billion <00:02:51.390>of <00:02:51.514>capacity <00:02:52.140>across <00:02:52.563>four <00:02:53.016>UK <00:02:53.376>sites

31
00:02:54.283 --> 00:02:56.243
<v Quote>among <00:02:54.515>them <00:02:54.683>a <00:02:54.781>deployment <00:02:55.379>with <00:02:55.523>Kao <00:02:55.843>Data.

32
00:02:56.473 --> 00:03:04.153
<v Cassidy>Nebius <00:02:57.006>is <00:02:57.153>now <00:02:57.326>on <00:02:57.481>both <00:02:57.740>sides <00:02:58.100>of <00:02:58.193>the <00:02:58.366>capacity <00:02:58.862>market: <00:02:59.873>selling <00:03:00.260>AI <00:03:00.553>cloud <00:03:00.980>to <00:03:01.089>customers <00:03:01.620>while <00:03:01.803>leasing <00:03:02.222>Nvidia-powered <00:03:02.944>capacity <00:03:03.545>from <00:03:03.713>Vantage <00:03:04.100>at

33
00:03:04.223 --> 00:03:09.296
<v Cassidy>Newport. <00:03:05.033>That <00:03:05.713>tells <00:03:05.966>us <00:03:06.113>a <00:03:06.213>lot <00:03:06.441>more <00:03:06.665>than <00:03:06.833>the <00:03:07.020>usual <00:03:07.433>neocloud <00:03:07.966>press <00:03:08.293>release.

34
00:03:09.296 --> 00:03:11.673
<v Bill>And <00:03:09.726>it’s <00:03:09.953>a <00:03:10.105>long-dated <00:03:11.133>bet.

35
00:03:11.833 --> 00:03:18.873
<v Bill>Nebius <00:03:12.593>has <00:03:12.761>committed <00:03:13.253>roughly <00:03:13.833>£1.7 <00:03:15.333>billion <00:03:16.382>across <00:03:16.873>four <00:03:17.353>UK <00:03:17.766>sites

36
00:03:18.900 --> 00:03:24.199
<v Bill>while <00:03:19.093>its <00:03:19.313>own <00:03:19.497>customers <00:03:20.166>still <00:03:20.681>have <00:03:20.900>to <00:03:21.017>show <00:03:21.326>up <00:03:21.473>and <00:03:21.657>keep <00:03:21.970>buying <00:03:22.366>training <00:03:22.833>and <00:03:23.009>inference.

37
00:03:24.199 --> 00:03:29.353
<v Cassidy>Newport <00:03:25.086>is <00:03:25.233>the <00:03:25.406>first <00:03:25.709>commercial <00:03:26.124>commitment <00:03:26.690>inside <00:03:27.053>the <00:03:27.166>South <00:03:27.486>Wales <00:03:27.860>AI <00:03:28.130>Growth <00:03:28.553>Zone

38
00:03:29.393 --> 00:03:33.193
<v Cassidy>and <00:03:29.553>Vantage <00:03:30.046>is <00:03:30.197>projecting <00:03:30.681>more <00:03:30.905>than <00:03:31.133>one <00:03:31.493>gigawatt <00:03:32.050>across <00:03:32.453>Newport

39
00:03:33.241 --> 00:03:35.033
<v Cassidy>Bridgend <00:03:34.093>and <00:03:34.213>Bro <00:03:34.473>Tathan.

40
00:03:36.044 --> 00:03:39.332
<v Cassidy>The <00:03:36.207>UK <00:03:36.514>has <00:03:36.687>moved <00:03:37.018>past <00:03:37.294>the <00:03:37.405>ceremonial <00:03:37.971>shovel <00:03:38.367>phase <00:03:38.698>here.

41
00:03:39.332 --> 00:03:46.074
<v Bill>Childress <00:03:40.522>gave <00:03:40.767>us <00:03:40.954>a <00:03:41.274>50-megawatt <00:03:42.387>acceptance <00:03:43.327>as <00:03:43.807>an <00:03:43.994>actual <00:03:44.527>delivery <00:03:45.023>marker.

42
00:03:46.154 --> 00:03:52.954
<v Bill>Newport <00:03:46.794>is <00:03:47.007>still <00:03:47.354>a <00:03:47.496>capacity <00:03:48.059>commitment—but <00:03:49.407>at <00:03:49.554>least <00:03:49.814>Nebius <00:03:50.394>is <00:03:50.594>a <00:03:50.691>tenant <00:03:51.082>with <00:03:51.314>a <00:03:51.434>real <00:03:51.845>campus

43
00:03:53.014 --> 00:03:56.634
<v Bill>not <00:03:53.224>another <00:03:53.504>company <00:03:54.054>admiring <00:03:54.623>future <00:03:55.101>racks <00:03:55.466>from <00:03:55.634>a <00:03:55.741>slide <00:03:56.154>deck.

44
00:03:57.192 --> 00:03:58.592
<v Cassidy>This <00:03:57.312>one's <00:03:57.528>from <00:03:57.649>Nature <00:03:58.062>Methods:

45
00:03:59.117 --> 00:04:06.637
<v Quote>Here <00:03:59.330>we <00:03:59.477>demonstrate <00:03:59.977>how <00:04:00.210>to <00:04:00.357>provide <00:04:00.733>task-specific <00:04:01.644>information <00:04:02.237>without <00:04:02.854>losing <00:04:03.237>the <00:04:03.397>general <00:04:03.725>knowledge <00:04:04.207>learned <00:04:04.580>during <00:04:04.777>pretraining <00:04:05.704>by <00:04:05.970>using <00:04:06.294>direct

46
00:04:06.688 --> 00:04:14.397
<v Quote>preference <00:04:07.259>optimization <00:04:08.130>to <00:04:08.464>align <00:04:08.877>a <00:04:08.997>structure-conditioned <00:04:09.887>protein <00:04:10.281>language <00:04:10.690>model <00:04:11.090>to <00:04:11.426>preferentially <00:04:12.210>generate <00:04:12.706>stable <00:04:13.177>protein <00:04:13.669>sequences.

47
00:04:14.677 --> 00:04:17.917
<v Quote>Our <00:04:15.567>aligned <00:04:15.984>model <00:04:16.547>ProteinDPO

48
00:04:17.961 --> 00:04:25.837
<v Quote>achieves <00:04:18.373>stability <00:04:18.921>prediction <00:04:19.410>competitive <00:04:20.050>to <00:04:20.221>task-specific <00:04:21.014>models <00:04:21.497>and <00:04:21.979>consistently <00:04:22.777>outperforms <00:04:23.537>unsupervised <00:04:24.437>and <00:04:24.621>fine-tuned <00:04:25.250>versions <00:04:25.704>of <00:04:25.777>the

49
00:04:25.890 --> 00:04:26.317
<v Quote>model.

50
00:04:26.681 --> 00:04:32.361
<v Bill>After <00:04:26.981>all <00:04:27.209>that <00:04:27.481>capacity <00:04:28.217>talk, <00:04:29.113>here’s <00:04:29.361>something <00:04:29.801>I <00:04:29.925>actually <00:04:30.345>want <00:04:30.651>running <00:04:31.054>on <00:04:31.214>those <00:04:31.574>racks.

51
00:04:32.481 --> 00:04:42.201
<v Bill>ProteinDPO <00:04:33.934>takes <00:04:34.281>direct <00:04:34.826>preference <00:04:35.573>optimization—the <00:04:37.321>chat-model <00:04:38.097>alignment <00:04:38.668>trick—and <00:04:39.721>aims <00:04:40.068>it <00:04:40.201>at <00:04:40.431>protein <00:04:40.993>stability

52
00:04:42.214 --> 00:04:44.921
<v Bill>where <00:04:42.321>the <00:04:42.510>target <00:04:42.974>is <00:04:43.130>experimental <00:04:43.801>fitness

53
00:04:44.981 --> 00:04:47.426
<v Bill>not <00:04:45.201>a <00:04:45.308>crowd-sourced <00:04:46.270>thumbs-up.

54
00:04:47.426 --> 00:04:55.321
<v Cassidy>Nature <00:04:48.181>Methods <00:04:48.661>reports <00:04:49.113>that <00:04:49.291>roughly <00:04:49.881>80% <00:04:50.814>of <00:04:50.961>its <00:04:51.121>hemagglutinin <00:04:51.791>designs <00:04:52.351>matched <00:04:52.894>or <00:04:53.072>improved <00:04:53.694>on <00:04:53.847>native <00:04:54.177>stability

55
00:04:55.353 --> 00:05:01.081
<v Cassidy>with <00:04:55.548>gains <00:04:55.934>up <00:04:56.121>to <00:04:56.361>32 <00:04:56.801>degrees <00:04:57.191>Celsius <00:04:58.041>against <00:04:58.663>recently <00:04:59.231>emerged <00:04:59.769>mammalian <00:05:00.371>strains.

56
00:05:01.945 --> 00:05:07.567
<v Cassidy>That’s <00:05:02.190>tied <00:05:02.473>to <00:05:02.566>a <00:05:02.666>wet-lab <00:05:03.148>property, <00:05:04.073>so <00:05:04.179>it <00:05:04.326>carries <00:05:04.646>a <00:05:04.786>lot <00:05:05.118>more <00:05:05.303>weight <00:05:05.678>than <00:05:05.846>another <00:05:06.179>model <00:05:06.506>leaderboard.

57
00:05:07.567 --> 00:05:14.686
<v Bill>The <00:05:08.126>paper <00:05:08.446>calls <00:05:08.894>this <00:05:09.166>an <00:05:09.398>alignment <00:05:10.146>gap: <00:05:11.086>pretraining <00:05:11.795>learns <00:05:12.299>broad <00:05:12.846>biological <00:05:13.708>patterns

58
00:05:14.718 --> 00:05:18.446
<v Bill>then <00:05:14.899>falls <00:05:15.246>apart <00:05:15.694>when <00:05:15.906>you <00:05:16.046>ask <00:05:16.326>for <00:05:16.526>one <00:05:16.846>very <00:05:17.237>specific <00:05:17.876>outcome.

59
00:05:19.456 --> 00:05:28.526
<v Bill>Familiar. <00:05:20.206>It’s <00:05:20.786>the <00:05:20.906>protein-design <00:05:21.856>version <00:05:22.259>of <00:05:22.393>an <00:05:22.526>agent <00:05:22.958>that <00:05:23.139>looks <00:05:23.462>brilliant <00:05:24.006>through <00:05:24.366>step <00:05:24.806>six <00:05:25.326>and <00:05:25.854>then <00:05:26.126>quietly <00:05:26.755>drives <00:05:27.214>into <00:05:27.446>a <00:05:27.566>ditch.

60
00:05:28.526 --> 00:05:33.087
<v Cassidy>And <00:05:29.034>it’s <00:05:29.207>a <00:05:29.304>useful <00:05:29.709>reminder <00:05:30.255>that <00:05:30.507>control <00:05:30.954>of <00:05:31.027>the <00:05:31.167>fine-tuning <00:05:31.874>stack <00:05:32.297>matters.

61
00:05:33.407 --> 00:05:41.007
<v Cassidy>A <00:05:33.791>base <00:05:34.100>model <00:05:34.427>has <00:05:34.655>radically <00:05:35.119>different <00:05:35.500>value <00:05:36.020>if <00:05:36.167>you <00:05:36.307>can <00:05:36.434>align <00:05:36.820>it <00:05:36.980>to <00:05:37.123>measured <00:05:37.487>biophysics <00:05:38.387>without <00:05:39.127>scrubbing <00:05:39.527>out <00:05:39.667>the <00:05:39.767>general <00:05:40.087>knowledge <00:05:40.500>it <00:05:40.657>started

62
00:05:41.087 --> 00:05:41.407
<v Cassidy>with.

63
00:05:42.220 --> 00:05:50.980
<v Cassidy>From <00:05:42.380>Liqin <00:05:42.820>Tan, <00:05:43.580>Xiean <00:05:43.932>Wang, <00:05:44.620>Yuexin <00:05:45.300>Zou, <00:05:45.900>Pin <00:05:46.476>Chen, <00:05:47.180>Qingsong <00:05:47.900>Zou <00:05:48.513>at <00:05:49.073>npj <00:05:49.546>Computational <00:05:50.220>Materials:

64
00:05:51.386 --> 00:06:01.906
<v Quote>Reliable <00:05:51.919>uncertainty <00:05:52.559>quantification <00:05:53.626>(UQ) <00:05:54.526>for <00:05:54.653>graph <00:05:55.020>neural <00:05:55.297>networks <00:05:55.986>(GNNs) <00:05:57.106>under <00:05:57.366>out-of-distribution <00:05:58.586>(OOD) <00:05:59.489>shifts <00:06:00.186>remains <00:06:00.801>insufficiently

65
00:06:01.937 --> 00:06:03.706
<v Quote>characterized <00:06:02.373>in <00:06:02.482>materials <00:06:03.042>discovery.

66
00:06:03.857 --> 00:06:12.826
<v Quote>Existing <00:06:05.124>benchmarks <00:06:05.759>based <00:06:06.079>on <00:06:06.243>random <00:06:06.666>splits <00:06:07.206>can <00:06:07.466>overestimate <00:06:08.239>model <00:06:08.579>reliability <00:06:09.519>by <00:06:09.839>underrepresenting <00:06:10.717>structural <00:06:11.288>extrapolation <00:06:12.084>challenges.

67
00:06:13.018 --> 00:06:16.186
<v Quote>Here <00:06:13.839>we <00:06:14.002>introduce <00:06:14.666>MatUQ

68
00:06:16.266 --> 00:06:24.266
<v Quote>a <00:06:16.394>benchmark <00:06:16.879>built <00:06:17.199>on <00:06:17.354>structure-aware <00:06:18.266>Smooth <00:06:18.826>Overlap <00:06:19.413>of <00:06:19.535>Atomic <00:06:20.002>Positions <00:06:20.573>Leave-One-Cluster-Out <00:06:22.226>(SOAP-LOCO) <00:06:23.498>splitting

69
00:06:24.310 --> 00:06:31.386
<v Quote>together <00:06:24.698>with <00:06:24.866>a <00:06:24.950>training <00:06:25.368>protocol <00:06:25.898>that <00:06:26.088>combines <00:06:26.650>Deep <00:06:27.026>Evidential <00:06:27.531>Regression <00:06:28.306>(DER) <00:06:29.418>with <00:06:29.616>dropout <00:06:30.170>regularization

70
00:06:31.406 --> 00:06:35.706
<v Quote>for <00:06:31.524>evaluating <00:06:32.306>GNN <00:06:32.813>reliability <00:06:33.599>under <00:06:33.917>structural <00:06:34.481>distribution <00:06:35.203>shifts.

71
00:06:36.155 --> 00:06:47.115
<v Bill>MatUQ <00:06:37.083>asks <00:06:37.635>the <00:06:37.808>question <00:06:38.307>benchmark <00:06:39.046>charts <00:06:39.675>usually <00:06:40.342>dodge: <00:06:41.227>when <00:06:41.715>a <00:06:41.811>materials <00:06:42.475>GNN <00:06:43.051>sees <00:06:43.515>a <00:06:43.651>structure <00:06:44.265>outside <00:06:45.075>its <00:06:45.248>training <00:06:45.724>neighborhood

72
00:06:47.147 --> 00:06:48.875
<v Bill>does <00:06:47.328>it <00:06:47.483>know <00:06:47.702>it <00:06:47.795>might <00:06:48.048>be <00:06:48.248>wrong?

73
00:06:49.022 --> 00:06:58.475
<v Bill>Their <00:06:49.835>SOAP-LOCO <00:06:50.646>splits <00:06:51.307>make <00:06:51.795>for <00:06:51.955>a <00:06:52.043>much <00:06:52.347>nastier <00:06:53.051>test <00:06:53.499>than <00:06:53.835>shuffling <00:06:54.355>the <00:06:54.539>same <00:06:54.875>dataset <00:06:55.555>and <00:06:55.965>calling <00:06:56.368>it <00:06:56.555>generalization.

74
00:06:58.491 --> 00:06:59.995
<v Cassidy>They <00:06:58.612>tested <00:06:59.015>six <00:06:59.266>datasets

75
00:07:00.041 --> 00:07:01.115
<v Cassidy>twelve <00:07:00.366>architectures

76
00:07:01.155 --> 00:07:03.115
<v Cassidy>and <00:07:01.328>eight <00:07:01.642>uncertainty <00:07:02.215>methods

77
00:07:03.155 --> 00:07:07.755
<v Cassidy>and <00:07:03.337>accuracy <00:07:03.888>split <00:07:04.203>from <00:07:04.448>uncertainty <00:07:05.105>quality <00:07:05.648>under <00:07:05.968>those <00:07:06.295>out-of-distribution <00:07:07.208>tests.

78
00:07:08.022 --> 00:07:14.962
<v Cassidy>So <00:07:08.595>the <00:07:08.755>model <00:07:08.987>with <00:07:09.135>the <00:07:09.243>prettiest <00:07:09.711>prediction <00:07:10.142>score <00:07:10.535>may <00:07:10.779>also <00:07:11.088>be <00:07:11.235>the <00:07:11.415>one <00:07:11.643>most <00:07:11.895>confidently <00:07:12.601>hallucinating <00:07:13.315>a <00:07:13.399>material <00:07:13.817>property.

79
00:07:14.962 --> 00:07:19.275
<v Bill>Exactly—my <00:07:16.315>step-seven <00:07:17.175>problem <00:07:17.675>in <00:07:17.915>lab-coat <00:07:18.635>form.

80
00:07:19.395 --> 00:07:25.915
<v Bill>The <00:07:19.808>paper <00:07:20.168>finds <00:07:20.667>that <00:07:20.815>the <00:07:20.935>uncertainty <00:07:21.652>winner <00:07:22.075>on <00:07:22.295>one <00:07:22.537>property <00:07:23.155>often <00:07:23.808>doesn’t <00:07:24.208>transfer <00:07:24.715>to <00:07:24.925>another

81
00:07:25.955 --> 00:07:28.155
<v Bill>and <00:07:26.107>when <00:07:26.279>training <00:07:26.699>data <00:07:27.019>gets <00:07:27.355>thin

82
00:07:28.235 --> 00:07:33.069
<v Bill>plain <00:07:28.715>ensemble <00:07:29.417>variance <00:07:30.008>beats <00:07:30.495>the <00:07:30.635>clever <00:07:31.195>evidential <00:07:31.925>hybrids.

83
00:07:33.069 --> 00:07:37.836
<v Cassidy>We <00:07:33.388>just <00:07:33.556>covered <00:07:33.896>ProteinDPO <00:07:34.849>using <00:07:35.165>experimental <00:07:35.816>fitness <00:07:36.289>to <00:07:36.436>steer <00:07:36.686>protein <00:07:37.116>models.

84
00:07:38.016 --> 00:07:43.436
<v Cassidy>MatUQ <00:07:38.938>supplies <00:07:39.456>the <00:07:39.556>other <00:07:39.836>half: <00:07:40.350>before <00:07:41.076>sending <00:07:41.396>a <00:07:41.484>materials <00:07:41.964>candidate <00:07:42.423>to <00:07:42.496>the <00:07:42.676>lab

85
00:07:43.484 --> 00:07:49.356
<v Cassidy>make <00:07:43.696>the <00:07:43.823>model <00:07:44.204>show <00:07:44.436>its <00:07:44.600>confidence <00:07:45.116>work—and <00:07:46.042>stress-test <00:07:46.652>that <00:07:46.774>confidence <00:07:47.369>where <00:07:47.456>the <00:07:47.560>training <00:07:47.965>distribution <00:07:48.668>ends.

86
00:07:50.126 --> 00:07:52.286
<v Cassidy>Here's <00:07:50.499>npj <00:07:50.972>Computational <00:07:51.630>Materials:

87
00:07:52.781 --> 00:08:01.661
<v Quote>We <00:07:52.991>present <00:07:53.474>OMOL-1k-MD, <00:07:55.661>a <00:07:55.781>new <00:07:55.961>dataset <00:07:56.421>for <00:07:56.561>the <00:07:56.677>benchmark <00:07:57.201>and <00:07:57.314>training <00:07:57.768>of <00:07:57.885>universal <00:07:58.521>machine-learning <00:07:59.314>interatomic <00:08:00.039>potentials <00:08:00.781>(uMLIPs).

88
00:08:01.954 --> 00:08:12.381
<v Quote>It <00:08:02.603>contains <00:08:03.141>three <00:08:03.408>independent <00:08:04.248>ab <00:08:04.525>initio <00:08:04.989>molecular <00:08:05.501>dynamics <00:08:06.261>(AIMD) <00:08:07.323>trajectories <00:08:08.088>of <00:08:08.301>10 <00:08:08.674>ps <00:08:09.005>each <00:08:09.321>for <00:08:09.581>1000 <00:08:10.361>arbitrarily <00:08:11.078>chosen <00:08:11.481>neutral <00:08:11.970>closed

89
00:08:12.421 --> 00:08:15.901
<v Quote>shell <00:08:12.669>molecular <00:08:13.161>systems <00:08:13.629>from <00:08:13.841>the <00:08:13.981>OMOL25 <00:08:15.011>dataset

90
00:08:15.959 --> 00:08:19.501
<v Quote>calculated <00:08:16.594>at <00:08:16.721>the <00:08:16.941>PBE <00:08:17.474>level <00:08:17.768>of <00:08:17.878>theory <00:08:18.301>at <00:08:18.541>300 <00:08:19.221>K.

91
00:08:19.940 --> 00:08:30.260
<v Bill>Thirty <00:08:20.240>million <00:08:20.724>room-temperature <00:08:21.598>structures <00:08:22.306>across <00:08:23.220>a <00:08:23.478>thousand <00:08:24.364>molecules, <00:08:25.880>and <00:08:25.956>they’re <00:08:26.160>testing <00:08:26.640>whether <00:08:27.100>the <00:08:27.291>dynamics <00:08:27.956>hold <00:08:28.460>for <00:08:28.900>10 <00:08:29.320>picoseconds.

92
00:08:30.452 --> 00:08:41.140
<v Bill>Good. <00:08:31.660>A <00:08:31.916>potential <00:08:32.468>that <00:08:32.740>nails <00:08:33.180>the <00:08:33.350>resting <00:08:33.780>pose <00:08:34.220>but <00:08:34.660>mangles <00:08:35.240>the <00:08:35.373>vibrations <00:08:36.313>is <00:08:36.820>useless <00:08:37.480>the <00:08:37.826>minute <00:08:38.132>your <00:08:38.325>simulation <00:08:39.051>actually <00:08:39.727>moves.

93
00:08:41.180 --> 00:08:52.100
<v Cassidy>MACE <00:08:41.567>MP-0 <00:08:42.233>leads <00:08:42.540>the <00:08:42.767>ML <00:08:43.060>pack, <00:08:43.816>followed <00:08:44.153>by <00:08:44.327>SevenNet <00:08:45.020>and <00:08:45.240>Orb <00:08:45.620>OMat—but <00:08:47.033>GFN2-xTB, <00:08:49.000>the <00:08:49.124>semiempirical <00:08:49.929>method, <00:08:50.527>beats <00:08:51.127>every <00:08:51.500>one <00:08:51.673>of <00:08:51.828>them.

94
00:08:52.260 --> 00:08:57.251
<v Cassidy>That <00:08:52.950>ranking <00:08:53.340>tells <00:08:53.600>you <00:08:53.840>far <00:08:54.084>more <00:08:54.372>than <00:08:54.540>another <00:08:54.900>glossy <00:08:55.444>universal-model <00:08:56.353>claim.

95
00:08:57.251 --> 00:09:04.580
<v Bill>And <00:08:57.753>it <00:08:57.917>echoes <00:08:58.440>MatUQ <00:08:59.732>from <00:08:59.930>earlier: <00:09:00.807>clean <00:09:01.604>benchmark <00:09:02.238>conditions <00:09:03.007>flatter <00:09:03.791>models.

96
00:09:04.708 --> 00:09:09.860
<v Bill>Here, <00:09:05.560>the <00:09:05.684>miss <00:09:06.020>shows <00:09:06.500>up <00:09:06.724>away <00:09:07.028>from <00:09:07.273>local <00:09:07.600>minima, <00:09:08.447>in <00:09:08.564>molecular <00:09:09.220>motion.

97
00:09:10.000 --> 00:09:20.100
<v Bill>The <00:09:10.480>authors’ <00:09:11.428>diagnosis—more <00:09:12.760>off-equilibrium <00:09:13.904>training <00:09:14.340>data—will <00:09:15.513>sound <00:09:15.836>painfully <00:09:16.393>familiar <00:09:16.900>to <00:09:17.117>anyone <00:09:17.500>who’s <00:09:17.820>watched <00:09:18.140>a <00:09:18.260>system <00:09:18.836>fail <00:09:19.289>around <00:09:19.780>step

98
00:09:20.193 --> 00:09:20.820
<v Bill>seven.

99
00:09:21.541 --> 00:09:27.141
<v Cassidy>If <00:09:21.641>you’re <00:09:21.825>enjoying <00:09:22.221>the <00:09:22.474>AI <00:09:22.794>Daily <00:09:23.114>Briefing, <00:09:23.653>take <00:09:23.821>a <00:09:23.907>moment <00:09:24.234>to <00:09:24.413>subscribe <00:09:25.114>or <00:09:25.261>leave <00:09:25.501>a <00:09:25.598>review <00:09:25.985>wherever <00:09:26.361>you’re <00:09:26.549>listening.

100
00:09:27.251 --> 00:09:33.541
<v Cassidy>Reviews <00:09:28.069>help <00:09:28.301>other <00:09:28.547>people <00:09:28.869>find <00:09:29.081>the <00:09:29.221>show, <00:09:29.981>and <00:09:30.133>your <00:09:30.311>support <00:09:30.701>helps <00:09:30.954>us <00:09:31.093>keep <00:09:31.265>bringing <00:09:31.661>you <00:09:31.801>the <00:09:31.961>day’s <00:09:32.501>AI <00:09:32.949>news.

101
00:09:34.281 --> 00:09:37.001
<v Cassidy>Links <00:09:34.548>to <00:09:34.641>every <00:09:34.894>story <00:09:35.214>we <00:09:35.361>covered <00:09:35.681>today <00:09:35.921>are <00:09:36.068>in <00:09:36.141>the <00:09:36.249>show <00:09:36.521>notes.

102
00:09:37.188 --> 00:09:41.001
<v Cassidy>If <00:09:37.601>one <00:09:37.767>caught <00:09:38.057>your <00:09:38.169>attention, <00:09:38.729>take <00:09:39.001>a <00:09:39.104>moment <00:09:39.294>to <00:09:39.435>follow <00:09:39.681>the <00:09:39.865>link <00:09:40.161>and <00:09:40.329>read <00:09:40.601>more.

103
00:09:41.113 --> 00:09:43.481
<v Cassidy>That’s <00:09:41.774>AI <00:09:42.094>Daily <00:09:42.414>Briefing <00:09:42.801>for <00:09:43.001>today.

104
00:09:43.625 --> 00:09:45.721
<v Cassidy>This <00:09:44.254>is <00:09:44.441>a <00:09:44.561>Lantern <00:09:44.941>Podcast.