WEBVTT

1
00:00:00.080 --> 00:00:04.960
<v Speaker 1>AI Daily Briefing imed sharp thanks for joining me today

2
00:00:05.120 --> 00:00:07.719
<v Speaker 1>the K three shock. While US chip dominance no longer

3
00:00:07.719 --> 00:00:11.839
<v Speaker 1>guarantees AI lead moonshot AI just dropped a model that's

4
00:00:11.919 --> 00:00:16.320
<v Speaker 1>rattling markets, rewriting timelines, and forcing a hard question. If

5
00:00:16.440 --> 00:00:19.359
<v Speaker 1>China can match US frontier AI at half the price,

6
00:00:19.760 --> 00:00:23.640
<v Speaker 1>what exactly are the chip export controls protecting Kimmi. K

7
00:00:23.800 --> 00:00:27.320
<v Speaker 1>three landed on July seventeenth. It's a two point eight

8
00:00:27.359 --> 00:00:30.440
<v Speaker 1>trillion parameter model from a three year old Chinese startup,

9
00:00:30.760 --> 00:00:33.920
<v Speaker 1>and early rankings put it competitive with or a head

10
00:00:33.960 --> 00:00:38.200
<v Speaker 1>of clawed Opus and GPT five point six Soul. The

11
00:00:38.280 --> 00:00:41.880
<v Speaker 1>price gap is the sharper signal. K three runs at

12
00:00:41.880 --> 00:00:46.520
<v Speaker 1>fifteen dollars per million tokens. It's US competitors charge fifty.

13
00:00:47.159 --> 00:00:51.880
<v Speaker 1>Markets didn't wait for independent verification. TSMC fell seven percent,

14
00:00:52.439 --> 00:00:56.600
<v Speaker 1>Soft Bank dropped nine, Invidiouslid one point two percent, Meta

15
00:00:56.880 --> 00:00:59.679
<v Speaker 1>two point four The Nasdaq one hundred was down one

16
00:00:59.679 --> 00:01:03.159
<v Speaker 1>percent on the day. That's a significant repricing for a

17
00:01:03.200 --> 00:01:06.760
<v Speaker 1>single model announcement from a company most Western investors hadn't

18
00:01:06.799 --> 00:01:10.319
<v Speaker 1>heard of forty eight hours earlier. The important distinction is

19
00:01:10.319 --> 00:01:12.879
<v Speaker 1>that this wasn't paddic about one model. It was a

20
00:01:12.920 --> 00:01:17.120
<v Speaker 1>repricing of the underlying assumption that US chip dominance translates

21
00:01:17.159 --> 00:01:21.519
<v Speaker 1>reliably into AI capability. Dominance. K three puts that logic

22
00:01:21.640 --> 00:01:25.680
<v Speaker 1>under serious pressure. The more consequential move comes later. On

23
00:01:25.799 --> 00:01:29.120
<v Speaker 1>July twenty seventh, Kimmi K three becomes the world's first

24
00:01:29.159 --> 00:01:33.760
<v Speaker 1>freely downloadable three trillion parameter model. Developers worldwide will be

25
00:01:33.799 --> 00:01:37.079
<v Speaker 1>able to modify it, run it locally, deploy it without

26
00:01:37.120 --> 00:01:40.680
<v Speaker 1>inference fees. That's not just a product launch. That's a

27
00:01:40.719 --> 00:01:43.640
<v Speaker 1>structural challenge to the closed model rivenue model that Anthropic

28
00:01:43.680 --> 00:01:46.439
<v Speaker 1>and open AI are built on. Chinese labs have been

29
00:01:46.439 --> 00:01:50.359
<v Speaker 1>here before Deep Seek did this. The pattern is consistent,

30
00:01:50.599 --> 00:01:54.400
<v Speaker 1>release capable models for free, commoditize the frontier, and let

31
00:01:54.439 --> 00:01:57.760
<v Speaker 1>the pricing math do the damage. Deep Seek currently charges

32
00:01:57.799 --> 00:02:01.719
<v Speaker 1>eighty seven cents per million tokens. K three charges fifteen dollars.

33
00:02:02.040 --> 00:02:05.159
<v Speaker 1>Open ai charges fifty dollars. The gap at each level

34
00:02:05.280 --> 00:02:09.400
<v Speaker 1>is widening, not closing. Here's the thing. The developer adoption

35
00:02:09.520 --> 00:02:12.879
<v Speaker 1>data reinforces this. All five of the most used models

36
00:02:12.919 --> 00:02:17.280
<v Speaker 1>on open router right now are Chinese owned ten Cent Shalmie,

37
00:02:17.560 --> 00:02:21.120
<v Speaker 1>deep Seek, Mini Max, and z dot ai. That's a

38
00:02:21.120 --> 00:02:24.240
<v Speaker 1>clean sweep. A year ago, that list would have looked

39
00:02:24.319 --> 00:02:27.639
<v Speaker 1>very different. The signal here is that developer preference has

40
00:02:27.719 --> 00:02:32.199
<v Speaker 1>already shifted before K three even fully releases. Within China's

41
00:02:32.240 --> 00:02:35.759
<v Speaker 1>own AI market, K three's arrival triggered sharp the kinds

42
00:02:36.439 --> 00:02:40.639
<v Speaker 1>z dot ai fell twenty seven percent minimax drops sixteen.

43
00:02:41.240 --> 00:02:45.719
<v Speaker 1>Competitive displacement within the Chinese ecosystem is apparently happening faster

44
00:02:45.800 --> 00:02:49.280
<v Speaker 1>than anyone expected. The geopolitical framing around all of this

45
00:02:49.639 --> 00:02:53.919
<v Speaker 1>was deliberate. At Shanghai's World AI Conference this week, President

46
00:02:53.960 --> 00:02:58.159
<v Speaker 1>She publicly endorsed open source AI, called for international corporation,

47
00:02:58.479 --> 00:03:02.039
<v Speaker 1>and criticized what he described as US overreach in applying

48
00:03:02.120 --> 00:03:07.479
<v Speaker 1>national security justifications to technology controls. The timing wasn't coincidental.

49
00:03:08.159 --> 00:03:13.759
<v Speaker 1>Alongside that speech, twenty nine countries, including Russia, Pakistan, Kazakhstan

50
00:03:13.879 --> 00:03:18.520
<v Speaker 1>and Laos signed an agreement establishing WAYCO, a new intergominental

51
00:03:18.560 --> 00:03:22.680
<v Speaker 1>AI cooperation body, with the UN Secretary General present. Whether

52
00:03:22.719 --> 00:03:26.280
<v Speaker 1>WAYCO delivers substance is an open question, but China is

53
00:03:26.319 --> 00:03:29.240
<v Speaker 1>now formally positioning itself as the rules maker for AI

54
00:03:29.319 --> 00:03:32.560
<v Speaker 1>governance in the developing world, and the US was notally

55
00:03:32.599 --> 00:03:36.479
<v Speaker 1>absent from the room. Consider this. Here's the harder implication.

56
00:03:36.560 --> 00:03:40.199
<v Speaker 1>For Washington. US export controls were designed to slow China's

57
00:03:40.199 --> 00:03:44.240
<v Speaker 1>AI progress by restricting chip access. The evidence now suggests

58
00:03:44.240 --> 00:03:48.960
<v Speaker 1>those restrictions accelerated a different outcome. Chinese lads optimized aggressively

59
00:03:49.000 --> 00:03:52.960
<v Speaker 1>for performance per dollar and per GPU, producing architectures that

60
00:03:53.000 --> 00:03:57.120
<v Speaker 1>compete despite lower compute access. Mai Twan trained a one

61
00:03:57.159 --> 00:04:01.120
<v Speaker 1>point six tree in parameter model entirely on domestic Chinese processes.

62
00:04:01.560 --> 00:04:05.000
<v Speaker 1>Quarley's Atlas nine to five zero was showcased at Shanghai

63
00:04:05.080 --> 00:04:09.199
<v Speaker 1>as a credible hardware alternative the bottle Maker's narrowing not gone,

64
00:04:09.439 --> 00:04:13.280
<v Speaker 1>but narrowing. A few months ago, senior voices in USAI

65
00:04:13.439 --> 00:04:17.040
<v Speaker 1>expected China to reach capability parity somewhere around early twenty

66
00:04:17.079 --> 00:04:20.639
<v Speaker 1>twenty seven. K three's arrival collapsed that timeline by at

67
00:04:20.680 --> 00:04:24.480
<v Speaker 1>least six months, possibly more. The honest answer right now

68
00:04:24.560 --> 00:04:27.079
<v Speaker 1>is that we don't know exactly how wide the remaining

69
00:04:27.120 --> 00:04:30.959
<v Speaker 1>gap is because K three's full capabilities haven't been independently

70
00:04:31.079 --> 00:04:35.399
<v Speaker 1>verified yet. July twenty seventh changes that the real test

71
00:04:35.519 --> 00:04:39.040
<v Speaker 1>isn't the benchmark score, it's whether the open source release

72
00:04:39.079 --> 00:04:42.600
<v Speaker 1>performs as claimed under real world conditions. If it does,

73
00:04:42.879 --> 00:04:46.240
<v Speaker 1>the competitive map looks materially different than it did last week.

74
00:04:46.600 --> 00:04:50.000
<v Speaker 1>That's what to watch. Thanks for listening. This podcast was

75
00:04:50.040 --> 00:04:53.439
<v Speaker 1>built using AI technology, a Yes We production