WEBVTT

1
00:00:00.080 --> 00:00:05.080
<v Speaker 1>AI daily briefing, I med sharp, thanks for joining me today.

2
00:00:05.360 --> 00:00:08.880
<v Speaker 1>Can labs self regulate or do they need refs? The

3
00:00:09.000 --> 00:00:12.919
<v Speaker 1>leaders of Anthropic, open Ai, Meta and deep Mind just

4
00:00:12.960 --> 00:00:15.279
<v Speaker 1>did something that would have seemed unthinkable a year ago.

5
00:00:15.720 --> 00:00:18.839
<v Speaker 1>They signed a joint public statement asking governments to step

6
00:00:18.839 --> 00:00:22.519
<v Speaker 1>in and slow them down. The statement, called Pacing the Frontier,

7
00:00:22.879 --> 00:00:26.760
<v Speaker 1>calls for international government backed tools to control recursive self

8
00:00:26.760 --> 00:00:30.960
<v Speaker 1>improvement risks. The core argument is straightforward. No single lab

9
00:00:31.000 --> 00:00:34.719
<v Speaker 1>can afford to unilaterally slow down when competitors won't, so

10
00:00:34.759 --> 00:00:38.119
<v Speaker 1>they're asking for a referee. That shift from internal safety

11
00:00:38.119 --> 00:00:41.920
<v Speaker 1>teams to public coordination requests is the clearest signal yet

12
00:00:42.119 --> 00:00:46.000
<v Speaker 1>that the industry believes market forces alone can't manage what's coming.

13
00:00:46.479 --> 00:00:49.280
<v Speaker 1>The critical risk they're flagging is the point where AI

14
00:00:49.439 --> 00:00:53.399
<v Speaker 1>systems begin designing their own successors faster than humans can

15
00:00:53.479 --> 00:00:58.359
<v Speaker 1>verify whether those successors are aligned. That threshold doesn't have

16
00:00:58.399 --> 00:01:02.560
<v Speaker 1>a consensus definition. Open ai has already invoked safety pause

17
00:01:02.600 --> 00:01:06.480
<v Speaker 1>protocols on at least one pre release model. What triggers

18
00:01:06.519 --> 00:01:10.359
<v Speaker 1>those pauses isn't publicly specified, which makes the whole framework

19
00:01:10.439 --> 00:01:14.120
<v Speaker 1>hard to evaluate from the outside. The important distinction here

20
00:01:14.200 --> 00:01:17.680
<v Speaker 1>is between a genuine coordination problem and a credibility move.

21
00:01:18.239 --> 00:01:22.239
<v Speaker 1>These are competing companies publicly agreeing they need external control.

22
00:01:22.920 --> 00:01:25.719
<v Speaker 1>That's either a meaningful safety signal or a way to

23
00:01:25.799 --> 00:01:30.560
<v Speaker 1>slow rivals using regulatory friction, probably some of both. While

24
00:01:30.599 --> 00:01:34.200
<v Speaker 1>the labs are asking for oversight, Microsoft is quietly repositioning

25
00:01:34.280 --> 00:01:36.920
<v Speaker 1>away from them. Sati and the dell Are told Wall

26
00:01:36.920 --> 00:01:40.200
<v Speaker 1>Street this week that Microsoft is actively selling enterprises its

27
00:01:40.280 --> 00:01:44.599
<v Speaker 1>own homegrown MAI models and multi agent systems as alternatives

28
00:01:44.640 --> 00:01:48.319
<v Speaker 1>to relying on OpenAI and anthropic. He cited the hugging

29
00:01:48.319 --> 00:01:51.120
<v Speaker 1>face incident as evidence of the risk of lab dependency.

30
00:01:51.719 --> 00:01:56.560
<v Speaker 1>The messaging is explicit, Microsoft controlled infrastructure is safer than

31
00:01:56.640 --> 00:02:00.000
<v Speaker 1>betting on any single frontier lab. The signal here is

32
00:02:00.280 --> 00:02:04.159
<v Speaker 1>a shift in Microsoft's public posture. Previously, the company presented

33
00:02:04.200 --> 00:02:07.879
<v Speaker 1>itself as a neutral cloud provider. That framing is gone.

34
00:02:08.080 --> 00:02:12.360
<v Speaker 1>Nadella is now making a competitive argument directly to enterprise customers.

35
00:02:13.080 --> 00:02:16.199
<v Speaker 1>Whether that strains the OpenAI partnership is an open question,

36
00:02:16.719 --> 00:02:20.360
<v Speaker 1>but the incentives are clearly pulling in different directions. Here's

37
00:02:20.400 --> 00:02:23.439
<v Speaker 1>the thing. Separate from the lab dynamics. The White House

38
00:02:23.479 --> 00:02:27.319
<v Speaker 1>has formalized a defense AI strategy that deserves attention. The

39
00:02:27.400 --> 00:02:30.840
<v Speaker 1>Genesis Mission directs the Pentagon to feed decades of classified

40
00:02:30.840 --> 00:02:34.800
<v Speaker 1>military research data into the Department of Energy's American Science

41
00:02:34.840 --> 00:02:39.120
<v Speaker 1>and Security Platform. The goal is automated AI driven scientific

42
00:02:39.159 --> 00:02:43.360
<v Speaker 1>breakthroughs for defense innovation. This isn't a pilot program. It's

43
00:02:43.400 --> 00:02:47.319
<v Speaker 1>a centralized consolidation of the defense establishment's most valuable data

44
00:02:47.400 --> 00:02:51.439
<v Speaker 1>assets into a single AI powered discovery platform. What that

45
00:02:51.479 --> 00:02:54.400
<v Speaker 1>means in practice is a significant acceleration of the fence

46
00:02:54.439 --> 00:02:57.719
<v Speaker 1>sector AI capability built on data sets that have no

47
00:02:57.840 --> 00:03:03.080
<v Speaker 1>civilian equivalent. The transparenturrency questions around that strategy remain largely unaddressed.

48
00:03:03.520 --> 00:03:06.919
<v Speaker 1>The regulatory picture in Europe is sharpening, and the numbers

49
00:03:06.960 --> 00:03:11.080
<v Speaker 1>are now concrete enough to matter. Consider this. The EUAI

50
00:03:11.159 --> 00:03:14.479
<v Speaker 1>Act imposes fines up to thirty five million euros or

51
00:03:14.560 --> 00:03:18.199
<v Speaker 1>seven percent of global revenue. A single high risk system

52
00:03:18.240 --> 00:03:21.719
<v Speaker 1>costs roughly fifty thousand euros to bring into compliance initially

53
00:03:22.039 --> 00:03:26.599
<v Speaker 1>plus twenty nine thousand annually. The Act applies extra territorially,

54
00:03:27.120 --> 00:03:30.639
<v Speaker 1>meaning any company whose AI outputs reach EU uses is

55
00:03:30.719 --> 00:03:35.000
<v Speaker 1>subject to it regardless of where they're based. Meanwhile, US

56
00:03:35.120 --> 00:03:38.120
<v Speaker 1>private AI investment hit two hundred and eighty five point

57
00:03:38.240 --> 00:03:41.479
<v Speaker 1>nine billion dollars in twenty twenty five. That's a one

58
00:03:41.520 --> 00:03:44.759
<v Speaker 1>hundred and sixty percent year on year growth. Europe grew

59
00:03:44.840 --> 00:03:48.840
<v Speaker 1>seven point two percent. Thirty percent of European AI unicorns

60
00:03:48.879 --> 00:03:52.319
<v Speaker 1>have already relocated to the United States. The trade off

61
00:03:52.360 --> 00:03:55.599
<v Speaker 1>is real. Europe is building a governance framework while the

62
00:03:55.719 --> 00:03:59.919
<v Speaker 1>US and China are scaling infrastructure and capital. Compliance costs

63
00:04:00.240 --> 00:04:04.599
<v Speaker 1>without parallel investment don't just slow innovation, they redirect it

64
00:04:04.680 --> 00:04:07.840
<v Speaker 1>somewhere else. Three things are worth tracking closely. From here,

65
00:04:08.240 --> 00:04:11.080
<v Speaker 1>you see the pacing. The frontier statement means to produce

66
00:04:11.080 --> 00:04:15.240
<v Speaker 1>something concrete to be meaningful. Vague coordination requests without enforcement

67
00:04:15.240 --> 00:04:19.560
<v Speaker 1>mechanisms are aspirational, not operational. The project watch where the

68
00:04:19.600 --> 00:04:22.800
<v Speaker 1>governments engage with the verification problem, because that's where the

69
00:04:22.839 --> 00:04:27.040
<v Speaker 1>credibility of the whole initiative lives. On Microsoft watch enterprise

70
00:04:27.120 --> 00:04:30.879
<v Speaker 1>procurement patterns over the next two quarters. If customers start

71
00:04:30.959 --> 00:04:35.879
<v Speaker 1>actively substituting MAI models for open AI API access at scale,

72
00:04:36.240 --> 00:04:40.519
<v Speaker 1>the partnership dynamic changes materially and on Europe. The question

73
00:04:40.680 --> 00:04:43.240
<v Speaker 1>isn't whether the AI Act is good policy in isolation.

74
00:04:43.879 --> 00:04:47.519
<v Speaker 1>It's whether the compliance cost curve intersects with investment trends

75
00:04:47.680 --> 00:04:51.279
<v Speaker 1>in a way that proves irreversible for European AI competitiveness.

76
00:04:52.040 --> 00:04:55.519
<v Speaker 1>The labs can't self regulate because no single one will

77
00:04:55.519 --> 00:04:58.800
<v Speaker 1>blink first. That's the honest read of this week. The

78
00:04:58.839 --> 00:05:01.360
<v Speaker 1>refs have been called on Toto the field, whether they

79
00:05:01.360 --> 00:05:05.560
<v Speaker 1>have the tools to actually officiate is still very much unresolved.

80
00:05:06.160 --> 00:05:09.920
<v Speaker 1>Thanks for listening. This podcast was built using AI technology

81
00:05:10.360 --> 00:05:11.439
<v Speaker 1>A Yes We production