WEBVTT

1
00:00:00.040 --> 00:00:04.200
<v Speaker 1>AI Daily Briefing, Amed Sharp, thanks for joining me today.

2
00:00:04.480 --> 00:00:08.119
<v Speaker 1>Who gets to define AI safety? Hassabis versus the market

3
00:00:09.080 --> 00:00:12.199
<v Speaker 1>demis Hassebis wants the US government to approve AI models

4
00:00:12.199 --> 00:00:15.119
<v Speaker 1>before they ship. That's the sharpest version of what the

5
00:00:15.119 --> 00:00:18.679
<v Speaker 1>Google DeepMind CEO proposed this week, calling for a federally

6
00:00:18.719 --> 00:00:22.879
<v Speaker 1>overseen public private standards body modeled on FINGRA, the Financial

7
00:00:22.879 --> 00:00:27.600
<v Speaker 1>industry regulator. The proposal includes mandatory thirty day pre release testing,

8
00:00:27.960 --> 00:00:31.800
<v Speaker 1>federal funding, and a pathway from voluntary to mandatory compliance.

9
00:00:32.359 --> 00:00:35.719
<v Speaker 1>It's specific enough to be taken seriously and ambitious enough

10
00:00:35.840 --> 00:00:39.719
<v Speaker 1>to face serious resistance. The important distinction here is who

11
00:00:39.759 --> 00:00:46.159
<v Speaker 1>this proposal serves. Hassebis frames it around AGI risk, citing cybersecurity, biological,

12
00:00:46.320 --> 00:00:50.479
<v Speaker 1>and nuclear level threats from frontier models. That framing is genuine,

13
00:00:51.079 --> 00:00:53.600
<v Speaker 1>But the company's best position to pass a thirty day

14
00:00:53.640 --> 00:00:58.119
<v Speaker 1>federal review are the ones that already have safety teams, compliance, infrastructure,

15
00:00:58.320 --> 00:01:03.119
<v Speaker 1>and regulatory relationships. That's Google deep Mind. The proposal doesn't

16
00:01:03.159 --> 00:01:06.359
<v Speaker 1>just define AI safety, it defines who gets to be

17
00:01:06.439 --> 00:01:10.719
<v Speaker 1>credible at it. The market's counter argument is speed. Any

18
00:01:10.760 --> 00:01:14.359
<v Speaker 1>mandatory review window creates a gap that competitives, including state

19
00:01:14.400 --> 00:01:19.159
<v Speaker 1>backed labs won't observe. Hassabis acknowledges this. His model asks

20
00:01:19.200 --> 00:01:22.319
<v Speaker 1>industry to fund the body and phase in requirements gradually.

21
00:01:22.599 --> 00:01:25.920
<v Speaker 1>What's still unresolved is whether voluntary compliance can hold long

22
00:01:26.040 --> 00:01:29.079
<v Speaker 1>enough to become mandatory, and whether a federal body has

23
00:01:29.079 --> 00:01:32.480
<v Speaker 1>the technical capacity to evaluate frontier models under a thirty

24
00:01:32.519 --> 00:01:37.079
<v Speaker 1>day clock. Those aren't minor details, they're the mechanism. Here's

25
00:01:37.079 --> 00:01:40.920
<v Speaker 1>what makes this moment distinct. Governance is no longer peripheral

26
00:01:40.920 --> 00:01:44.920
<v Speaker 1>to AI strategy. The companies willing to embrace oversight are

27
00:01:45.040 --> 00:01:48.400
<v Speaker 1>using it as a signal of legitimacy, and legitimacy is

28
00:01:48.439 --> 00:01:54.200
<v Speaker 1>converting into enterprise contracts, government partnerships, and regulatory goodwill safety

29
00:01:54.280 --> 00:01:59.239
<v Speaker 1>frameworks are becoming competitive assets. Apple cleared China's cyberspace administration

30
00:01:59.319 --> 00:02:02.599
<v Speaker 1>after twenty two months. Apple Intelligence is now approved in

31
00:02:02.599 --> 00:02:06.079
<v Speaker 1>the world's largest smartphone market, but the terms matter. It

32
00:02:06.159 --> 00:02:09.520
<v Speaker 1>runs on Ali Barber's Qan and Baydu's models, not Apple's

33
00:02:09.560 --> 00:02:12.879
<v Speaker 1>own stack. The cloud dependent features are still being negotiated,

34
00:02:13.280 --> 00:02:15.520
<v Speaker 1>and both Ali Barba and Baidu are on the U

35
00:02:15.560 --> 00:02:18.759
<v Speaker 1>S Defense Department Section one two six to zero H list,

36
00:02:19.120 --> 00:02:22.960
<v Speaker 1>which flags firms with military links that creates political friction.

37
00:02:23.080 --> 00:02:26.039
<v Speaker 1>Apple hasn't fully resolved. The signal for the rest of

38
00:02:26.039 --> 00:02:29.520
<v Speaker 1>the industry is clear. Market access in China now requires

39
00:02:29.520 --> 00:02:33.759
<v Speaker 1>handing technical control to domestic partners. Apple's playbook is already

40
00:02:33.800 --> 00:02:37.759
<v Speaker 1>being studied by Western firms watching from outside. Operate through

41
00:02:37.800 --> 00:02:42.919
<v Speaker 1>local models, or don't operate. Here's the thing. Moonshot, the

42
00:02:43.039 --> 00:02:47.120
<v Speaker 1>Chinese AI startup, launched Kimi K three this week. It's

43
00:02:47.120 --> 00:02:50.400
<v Speaker 1>an open weight model with two point eight trillion parameters

44
00:02:50.680 --> 00:02:54.159
<v Speaker 1>and a one million token context window, and it's performing

45
00:02:54.199 --> 00:02:58.599
<v Speaker 1>competitively with anthropics fable on benchmarks. That's a direct challenge

46
00:02:58.599 --> 00:03:01.439
<v Speaker 1>to the assumption that Chinese lanfe aps trail Western frontier

47
00:03:01.439 --> 00:03:04.520
<v Speaker 1>models by six months or more. The gap is narrowing,

48
00:03:04.800 --> 00:03:07.759
<v Speaker 1>and Chemik three is free to deploy the cost and

49
00:03:07.800 --> 00:03:12.400
<v Speaker 1>capability calculus for enterprise buyers just shifted. Moonshot is also

50
00:03:12.520 --> 00:03:15.560
<v Speaker 1>raising two billion dollars at a thirty billion dollar valuation

51
00:03:15.919 --> 00:03:19.919
<v Speaker 1>while releasing that model openly. The unit economics are transparent,

52
00:03:20.000 --> 00:03:24.639
<v Speaker 1>yet that tension between open weight strategy and valuation sustainability

53
00:03:24.919 --> 00:03:28.159
<v Speaker 1>is worth watching. Fireworks Ai closed a one point five

54
00:03:28.199 --> 00:03:31.520
<v Speaker 1>billion dollar Series D at a seventeen point five billion

55
00:03:31.560 --> 00:03:36.000
<v Speaker 1>dollar valuation The company processes forty trillion tokens daily and

56
00:03:36.120 --> 00:03:39.479
<v Speaker 1>is running over one billion dollars in annualized revenue. The

57
00:03:39.520 --> 00:03:43.919
<v Speaker 1>infrastructure layer is consolidating fast, and specialized platforms are proving

58
00:03:43.960 --> 00:03:48.439
<v Speaker 1>more durable than early observers expected, and Thropic launched clawd reflect,

59
00:03:48.719 --> 00:03:52.919
<v Speaker 1>an analytics dashboard that visualizes user AI patterns and suggests

60
00:03:52.960 --> 00:03:57.479
<v Speaker 1>deeper clawed integration. The framing is productivity and mindful usage.

61
00:03:57.800 --> 00:04:02.319
<v Speaker 1>The function is retention and competitive locking. Those aren't mutually exclusive,

62
00:04:02.639 --> 00:04:06.560
<v Speaker 1>but they're worth distinguishing. In Japan, in Vidia's Jensen Huhan

63
00:04:06.599 --> 00:04:10.000
<v Speaker 1>announced the Notre partnership with Sony and Honda, anchored by

64
00:04:10.000 --> 00:04:14.919
<v Speaker 1>a twenty seven five hundred GPU national AI superfactory. The

65
00:04:14.960 --> 00:04:18.959
<v Speaker 1>focus is physical AI and robotics. In Vidia also resumed

66
00:04:18.959 --> 00:04:22.720
<v Speaker 1>a thirty year Sega Gaming partnership. Japan is positioning itself

67
00:04:22.759 --> 00:04:25.839
<v Speaker 1>as a hardware AI manufacturing hub, and this is the

68
00:04:25.879 --> 00:04:29.279
<v Speaker 1>infrastructure bet that underpins it. The thread across all of

69
00:04:29.319 --> 00:04:33.199
<v Speaker 1>this is governance's leverage. HASSEBIS wants to define the standard.

70
00:04:33.600 --> 00:04:37.920
<v Speaker 1>China already controls the gate. Anthropic is shaping behavior. Japan

71
00:04:38.000 --> 00:04:41.000
<v Speaker 1>is building the infrastructure. The question of who defines AI

72
00:04:41.079 --> 00:04:45.199
<v Speaker 1>safety isn't just philosophical. It's determining who operates, at what

73
00:04:45.360 --> 00:04:49.319
<v Speaker 1>scale and in which markets. Watch whether the Hasabis proposal

74
00:04:49.360 --> 00:04:53.000
<v Speaker 1>gains any congressional traction, and watch whether Apple's China approval

75
00:04:53.000 --> 00:04:56.439
<v Speaker 1>survives its Pentagon friction. Those two data points will tell

76
00:04:56.519 --> 00:04:59.000
<v Speaker 1>us a lot about where the governance frame actually holds.

77
00:04:59.439 --> 00:05:03.560
<v Speaker 1>Thanksfully listening. This podcast was built using AI technology, A

78
00:05:03.680 --> 00:05:04.600
<v Speaker 1>Yes We production