1
00:00:03,120 --> 00:00:16,360
Speaker 1: Bloomberg Audio Studios, Podcasts, Radio News.

2
00:00:21,040 --> 00:00:25,040
Speaker 2: Hello and welcome to another episode of the Odd Lots Podcast.

3
00:00:25,079 --> 00:00:27,440
Speaker 3: I'm Joe Wisenthal and I'm Tracy Alloway.

4
00:00:27,880 --> 00:00:31,560
Speaker 2: Tracy, Obviously, there's a lot of there's an infinite number

5
00:00:31,680 --> 00:00:35,080
Speaker 2: of angles to artificial intelligence and what's going on in

6
00:00:35,120 --> 00:00:38,200
Speaker 2: the industry. But one thing that I find like striking

7
00:00:38,280 --> 00:00:41,880
Speaker 2: about this that feels very different from other stories in tech,

8
00:00:42,440 --> 00:00:45,280
Speaker 2: and I don't really know what it's about, is why

9
00:00:45,400 --> 00:00:49,240
Speaker 2: you hear it talked about in sort of geopolitical strategic terms,

10
00:00:49,640 --> 00:00:53,040
Speaker 2: and why countries, whether it's the US, China, but also

11
00:00:53,080 --> 00:00:56,240
Speaker 2: Saudi Arabia, Europe, really feel like they need to have

12
00:00:56,360 --> 00:01:00,400
Speaker 2: like their own domestic model or champion. It's almost like

13
00:01:00,440 --> 00:01:03,800
Speaker 2: it's like oil or something where countries have at least

14
00:01:03,880 --> 00:01:06,679
Speaker 2: been convinced that it's very important for them to have

15
00:01:06,880 --> 00:01:08,199
Speaker 2: some sort of homegrown AI.

16
00:01:08,480 --> 00:01:09,039
Speaker 4: Yeah, you're right.

17
00:01:09,120 --> 00:01:12,440
Speaker 3: I hadn't really thought about it, but that's absolutely true.

18
00:01:12,480 --> 00:01:15,720
Speaker 3: People talk about it almost as a strategic resource at

19
00:01:15,720 --> 00:01:18,920
Speaker 3: this point and sort of in existential terms, and you

20
00:01:19,000 --> 00:01:22,120
Speaker 3: never saw anyone talking about like the need for a

21
00:01:22,240 --> 00:01:24,840
Speaker 3: national SaaS strategy or something like that.

22
00:01:24,840 --> 00:01:26,840
Speaker 2: That's exactly what I think right, Like you didn't know

23
00:01:26,880 --> 00:01:28,360
Speaker 2: what it's like, Oh, well, we have to have our

24
00:01:28,400 --> 00:01:29,560
Speaker 2: own Slack or we have to.

25
00:01:29,480 --> 00:01:30,640
Speaker 3: Have our own salesforce.

26
00:01:30,680 --> 00:01:33,080
Speaker 2: We have to have our own British salesforce and the

27
00:01:33,160 --> 00:01:37,640
Speaker 2: Saudi Arabia salesforce. But there's something about AI models and like,

28
00:01:37,760 --> 00:01:40,840
Speaker 2: you know, it's sort of about the chips specifically, and

29
00:01:41,080 --> 00:01:43,480
Speaker 2: you know everyone wants the chips, right, but there also

30
00:01:43,520 --> 00:01:45,720
Speaker 2: seems to be something where like countries want to like

31
00:01:45,959 --> 00:01:49,360
Speaker 2: really develop their own models and have their own open

32
00:01:49,400 --> 00:01:50,520
Speaker 2: AI or whatever it is.

33
00:01:50,600 --> 00:01:52,480
Speaker 3: Yeah, I also wonder how much of it has to

34
00:01:52,520 --> 00:01:55,360
Speaker 3: do with the data and the idea that like the

35
00:01:55,440 --> 00:01:58,720
Speaker 3: data is what you're training the model on, and so

36
00:01:59,040 --> 00:02:03,200
Speaker 3: there's like privacy concerns isn't the right word, but you know,

37
00:02:03,760 --> 00:02:07,640
Speaker 3: countries want to have some control over their own data.

38
00:02:07,640 --> 00:02:09,960
Speaker 3: They don't want full transparency. So I wonder how much

39
00:02:10,040 --> 00:02:12,800
Speaker 3: that feeds into it. But you're right, we have gotten

40
00:02:12,919 --> 00:02:16,680
Speaker 3: used to talking about the AI arms race in one

41
00:02:16,760 --> 00:02:19,960
Speaker 3: way or another, and we should dig into why that is,

42
00:02:20,000 --> 00:02:22,720
Speaker 3: like why is it such a competitive piece of technology?

43
00:02:22,919 --> 00:02:25,160
Speaker 2: Totally? And I'd say, you know, there's two other things here,

44
00:02:25,200 --> 00:02:29,560
Speaker 2: which is that obviously the ultimate in everything these days.

45
00:02:29,840 --> 00:02:32,359
Speaker 2: A lot of it is US China and how US

46
00:02:32,520 --> 00:02:36,160
Speaker 2: models and US companies compared to Chinese models and Chinese companies,

47
00:02:36,200 --> 00:02:39,560
Speaker 2: which I don't think many Americans have much visibility on,

48
00:02:39,639 --> 00:02:41,200
Speaker 2: or at least I certainly don't. I don't really know

49
00:02:41,280 --> 00:02:43,760
Speaker 2: what China's AI strategy is. And then there is of

50
00:02:43,800 --> 00:02:47,480
Speaker 2: course existing policy and how that might change under a

51
00:02:47,520 --> 00:02:50,320
Speaker 2: Trump administration and what that might do both to you

52
00:02:50,360 --> 00:02:52,799
Speaker 2: as China relations in general, and I don't know if

53
00:02:52,800 --> 00:02:54,480
Speaker 2: that's going to be, you know, some sort of shift

54
00:02:54,560 --> 00:02:58,320
Speaker 2: in how Trump might think about AI versus a Biden administration.

55
00:02:58,560 --> 00:03:02,360
Speaker 3: It is true that, like certainly the semiconductor portion of

56
00:03:02,480 --> 00:03:07,160
Speaker 3: AI is something that has been competitive for a while now,

57
00:03:07,240 --> 00:03:09,600
Speaker 3: and we've seen that in both the Trump administration and

58
00:03:09,760 --> 00:03:13,840
Speaker 3: continued under Biden. So there is that kind of commonality

59
00:03:14,080 --> 00:03:16,760
Speaker 3: running through it. But I am very interested in getting

60
00:03:16,760 --> 00:03:18,959
Speaker 3: more of a handle on exactly what China is doing

61
00:03:19,000 --> 00:03:22,800
Speaker 3: in this space and what it sees potentially as like

62
00:03:22,960 --> 00:03:24,520
Speaker 3: an actual threat here totally.

63
00:03:24,600 --> 00:03:27,440
Speaker 2: Well, I'm very excited to say that we have two

64
00:03:27,480 --> 00:03:30,920
Speaker 2: perfect guests who really know all the dimensions of this

65
00:03:31,040 --> 00:03:33,560
Speaker 2: topic extremely well. We're going to be speaking with Jordan

66
00:03:33,639 --> 00:03:36,800
Speaker 2: Schneider He is the founder of the China Talk podcast

67
00:03:36,840 --> 00:03:40,040
Speaker 2: and newsletter, one of my favorite sources for understanding what's

68
00:03:40,080 --> 00:03:41,720
Speaker 2: going on in China. And we're also going to be

69
00:03:41,720 --> 00:03:43,640
Speaker 2: speaking with Kevin Chu. He is the author of the

70
00:03:43,680 --> 00:03:48,440
Speaker 2: Interconnected newsletter and also a hedge fund manager at Interconnected Capital.

71
00:03:48,680 --> 00:03:51,600
Speaker 2: Writes fantastic stuff like sort of really digging into some

72
00:03:51,640 --> 00:03:54,640
Speaker 2: of the technical and tech sides of all this, so

73
00:03:54,720 --> 00:03:56,520
Speaker 2: hopefully we will walk out of here with a little

74
00:03:56,560 --> 00:03:59,400
Speaker 2: more understanding. Jordan and Kevin, thank you so much for

75
00:03:59,440 --> 00:04:00,000
Speaker 2: coming on OUTE.

76
00:04:00,720 --> 00:04:02,520
Speaker 5: Thank you so much for having us first time.

77
00:04:02,560 --> 00:04:03,600
Speaker 4: Long time. What a treat.

78
00:04:04,760 --> 00:04:06,840
Speaker 2: Love when people say that, I love hearing it reminds

79
00:04:06,880 --> 00:04:08,840
Speaker 2: me of like listening to like sports talk radio as

80
00:04:08,880 --> 00:04:11,800
Speaker 2: a kid. I'll just start with the question, what is

81
00:04:11,840 --> 00:04:15,040
Speaker 2: it about AI? And is it inherent or is it

82
00:04:15,080 --> 00:04:17,719
Speaker 2: sort of the industry's marketing job. But what is it

83
00:04:17,760 --> 00:04:21,200
Speaker 2: about AI that unlike a tracing mention, you know, different

84
00:04:21,200 --> 00:04:25,719
Speaker 2: from like CRM software, social networking or whatever that countries

85
00:04:25,839 --> 00:04:30,360
Speaker 2: really seem to prize having some aspect of it be homegrown.

86
00:04:30,760 --> 00:04:32,880
Speaker 5: Yeah, I couldn't help a chuckle a little bit when

87
00:04:32,920 --> 00:04:36,039
Speaker 5: you guys are doing your intro about the national SAS

88
00:04:36,200 --> 00:04:38,520
Speaker 5: or what is your national CRM since I am a

89
00:04:38,520 --> 00:04:41,800
Speaker 5: software investor by day. You know, it's really interesting if

90
00:04:41,800 --> 00:04:44,680
Speaker 5: you think about how AI, when we talk about it

91
00:04:44,720 --> 00:04:47,600
Speaker 5: generally has been the formulated you know, we're really talking

92
00:04:47,640 --> 00:04:50,920
Speaker 5: about generative AI, right, large language models in particular. There

93
00:04:50,920 --> 00:04:53,880
Speaker 5: are a bunch of other AI machine learning use cases

94
00:04:53,880 --> 00:04:55,919
Speaker 5: that we're gonna you know, just stick to the side

95
00:04:55,920 --> 00:04:59,200
Speaker 5: for now. There's something very deeply cultural when you start

96
00:04:59,240 --> 00:05:03,280
Speaker 5: to model at the entire Internet's information. When it comes

97
00:05:03,279 --> 00:05:07,000
Speaker 5: to large language model, right, it's essentially a three dimensional

98
00:05:07,080 --> 00:05:10,640
Speaker 5: space where every single token, where every single word has

99
00:05:10,720 --> 00:05:12,960
Speaker 5: a line that's drawn from one word to the other.

100
00:05:13,279 --> 00:05:16,640
Speaker 5: That kind of formulates knowledge or formulates you know, truth

101
00:05:16,680 --> 00:05:19,320
Speaker 5: in a lot of ways. And I would say, so far,

102
00:05:20,200 --> 00:05:23,839
Speaker 5: given all the dominant products SHAT, GPT, clode, Gemini, etc.

103
00:05:24,120 --> 00:05:27,200
Speaker 5: They've been modeling Internet in general, which has a lot

104
00:05:27,240 --> 00:05:29,479
Speaker 5: of content from the American Internet if you want to

105
00:05:29,480 --> 00:05:31,919
Speaker 5: think about that, or you know, Western European Internet, and

106
00:05:31,960 --> 00:05:35,159
Speaker 5: there are other versions of the Internet in Japan and

107
00:05:35,279 --> 00:05:38,760
Speaker 5: China in particular, where the ecosystem is such a wall

108
00:05:38,839 --> 00:05:42,960
Speaker 5: garden that the way knowledge and culture and thoughts and

109
00:05:43,000 --> 00:05:46,960
Speaker 5: the you know, history is being expressed is very different,

110
00:05:47,240 --> 00:05:50,760
Speaker 5: and I think that is what triggered a lot of

111
00:05:50,800 --> 00:05:54,000
Speaker 5: country to come to an understanding that this AI thing

112
00:05:54,240 --> 00:05:58,960
Speaker 5: isn't just a blown up spreadsheet with good UI that

113
00:05:59,040 --> 00:06:01,359
Speaker 5: you can track, you know, pipeline or you know, have

114
00:06:01,480 --> 00:06:04,320
Speaker 5: a bunch of messages going around to help your workplace

115
00:06:04,360 --> 00:06:07,560
Speaker 5: go a little bit faster. It is a codification of

116
00:06:08,000 --> 00:06:09,200
Speaker 5: your people.

117
00:06:09,040 --> 00:06:10,920
Speaker 2: And collective truth and culture.

118
00:06:11,040 --> 00:06:14,400
Speaker 5: And I will say collectively, your public discourse, maybe even

119
00:06:14,440 --> 00:06:17,440
Speaker 5: an element of your private discourse that's been published on

120
00:06:17,480 --> 00:06:21,360
Speaker 5: the internet is being codified in a particular large language model.

121
00:06:21,440 --> 00:06:23,240
Speaker 4: Just to add to that and bring up the US

122
00:06:23,360 --> 00:06:25,680
Speaker 4: China context. So you know, if we take as a

123
00:06:25,720 --> 00:06:28,960
Speaker 4: premise that the US and China are in a strategic competition,

124
00:06:29,360 --> 00:06:31,960
Speaker 4: what nations do when they try to when they compete

125
00:06:32,040 --> 00:06:35,000
Speaker 4: is they try to, you know, maximize national power. What

126
00:06:35,240 --> 00:06:38,359
Speaker 4: is national power composed of how big your economy is

127
00:06:38,400 --> 00:06:41,680
Speaker 4: and how powerful your military is. Now, you know, there's

128
00:06:41,720 --> 00:06:44,039
Speaker 4: still a lot of debate to be about just how

129
00:06:44,080 --> 00:06:47,280
Speaker 4: important AI might be to the future of militaries and

130
00:06:47,360 --> 00:06:50,880
Speaker 4: the future of productivity growth, But it is the technology

131
00:06:50,920 --> 00:06:53,960
Speaker 4: today that has the most potential to really change long

132
00:06:54,040 --> 00:06:57,480
Speaker 4: term military and economic trajectories. And I think we've seen

133
00:06:57,560 --> 00:07:00,520
Speaker 4: over the past few years the realization, maybe starting with

134
00:07:00,560 --> 00:07:04,960
Speaker 4: the October seventh, twenty twenty two export controls, that having

135
00:07:04,960 --> 00:07:08,440
Speaker 4: a completely open exchange between the US and China is

136
00:07:08,520 --> 00:07:12,720
Speaker 4: too much of a risk for American policy makers to stomach.

137
00:07:12,760 --> 00:07:16,880
Speaker 4: And sort of in response to that, simultaneously and particularly

138
00:07:16,920 --> 00:07:20,480
Speaker 4: in response to the October seventh export controls, you've seen

139
00:07:20,560 --> 00:07:24,600
Speaker 4: trying to really aggressively start to invest and support both

140
00:07:24,640 --> 00:07:27,800
Speaker 4: their semiconductor as well as broader AI ecosystems.

141
00:07:28,120 --> 00:07:30,360
Speaker 3: Wait, so I want to press on this point that

142
00:07:30,600 --> 00:07:34,920
Speaker 3: Kevin also brought up. But the idea of like AI

143
00:07:35,440 --> 00:07:40,280
Speaker 3: as a means of like cementing cultural and informational control.

144
00:07:40,800 --> 00:07:44,600
Speaker 3: Did anyone play around with the what was it called

145
00:07:44,640 --> 00:07:45,320
Speaker 3: like the study?

146
00:07:45,520 --> 00:07:47,120
Speaker 4: She thought app?

147
00:07:48,840 --> 00:07:51,320
Speaker 3: Yeah, it was like the open AI app that was

148
00:07:51,400 --> 00:07:53,880
Speaker 3: trained on. Xi Jinping thought, here's.

149
00:07:53,640 --> 00:07:56,120
Speaker 4: The funny thing is that app is like actually kind

150
00:07:56,120 --> 00:07:59,760
Speaker 4: of vaporware. We tried to find it on China Talk

151
00:08:00,080 --> 00:08:02,760
Speaker 4: you guys should all subscribe to and couldn't. And I

152
00:08:02,760 --> 00:08:06,120
Speaker 4: think this is one of the really interesting pieces of

153
00:08:06,200 --> 00:08:09,640
Speaker 4: the Chinese government and how they're specifically playing on the

154
00:08:09,720 --> 00:08:12,320
Speaker 4: sort of AI model side of things. As I've sort

155
00:08:12,320 --> 00:08:15,400
Speaker 4: of if you look at a lot of the Chinese

156
00:08:15,440 --> 00:08:19,080
Speaker 4: government sponsored projects so far, particularly on the model generation

157
00:08:19,360 --> 00:08:22,240
Speaker 4: or like national data set side, and you try to

158
00:08:22,240 --> 00:08:24,120
Speaker 4: get to ground truth on them, there's really not a

159
00:08:24,120 --> 00:08:26,840
Speaker 4: lot of there there. But that's not to say that

160
00:08:26,880 --> 00:08:31,680
Speaker 4: the Chinese sort of VC backed model ecosystem isn't the

161
00:08:31,720 --> 00:08:35,640
Speaker 4: real deal. We've seen companies raise multiple billions of dollars

162
00:08:35,679 --> 00:08:38,200
Speaker 4: and be able to publish models, which are you know,

163
00:08:38,280 --> 00:08:40,960
Speaker 4: maybe six months behind the leading edge when you when

164
00:08:41,040 --> 00:08:43,720
Speaker 4: you're looking at open AI or anthropic and the sort

165
00:08:43,720 --> 00:08:46,880
Speaker 4: of delta between the government back stuff that you know,

166
00:08:46,960 --> 00:08:49,000
Speaker 4: can't pay for the best engineers or what have you

167
00:08:49,320 --> 00:08:53,680
Speaker 4: and these you know, DRIPUAI, Moonshot, Ernie it Baydu. Those

168
00:08:53,720 --> 00:08:56,079
Speaker 4: folks are really running just as hard as what you're

169
00:08:56,080 --> 00:08:56,800
Speaker 4: seeing in the West.

170
00:08:57,760 --> 00:09:00,720
Speaker 3: Let me ask a related question, but in terms of

171
00:09:01,360 --> 00:09:04,319
Speaker 3: Chinese AI models, what do we know about the data

172
00:09:04,320 --> 00:09:06,880
Speaker 3: sets that they're being trained on and are there any

173
00:09:06,920 --> 00:09:10,280
Speaker 3: limitations that are introduced by the fact that you know,

174
00:09:10,400 --> 00:09:14,560
Speaker 3: they are being developed within what is ostensibly all walled,

175
00:09:14,800 --> 00:09:17,679
Speaker 3: you know, the Great Firewall of China Internet. Do they

176
00:09:17,679 --> 00:09:21,120
Speaker 3: still have access to X China data sets that they

177
00:09:21,120 --> 00:09:23,400
Speaker 3: can pull in even if those aren't being sort of

178
00:09:23,679 --> 00:09:26,320
Speaker 3: publicly dispensed to normal people.

179
00:09:26,679 --> 00:09:29,480
Speaker 5: So I think they definitely do. The one thing that

180
00:09:29,520 --> 00:09:32,320
Speaker 5: all these Chinese tech companies, the model makers in particular,

181
00:09:32,320 --> 00:09:34,240
Speaker 5: are very aware of, and I would say any tech

182
00:09:34,240 --> 00:09:36,240
Speaker 5: company in China has to be aware of, is the

183
00:09:36,360 --> 00:09:39,480
Speaker 5: redline that cannot cross when it comes to certain things

184
00:09:39,520 --> 00:09:42,120
Speaker 5: that their product may generate, right, And I think this

185
00:09:42,240 --> 00:09:44,439
Speaker 5: is one of the reasons why building a general of

186
00:09:44,440 --> 00:09:47,760
Speaker 5: AI application or company in China is so hard. Yes,

187
00:09:47,800 --> 00:09:50,720
Speaker 5: they do have access to you know, Wikipedia or the

188
00:09:50,760 --> 00:09:54,559
Speaker 5: Common Crawl, any of the kind of the standard table

189
00:09:54,679 --> 00:09:57,560
Speaker 5: stake data sets that you use to train all the

190
00:09:57,640 --> 00:10:00,840
Speaker 5: large language models so far, and then they can inject

191
00:10:00,960 --> 00:10:04,880
Speaker 5: some of the data that is available inside the Chinese

192
00:10:04,920 --> 00:10:07,880
Speaker 5: Internet kind of garden a wall garden, if you will.

193
00:10:08,160 --> 00:10:11,080
Speaker 5: But there are two interesting things about the Chinese Internet right.

194
00:10:11,280 --> 00:10:14,040
Speaker 5: One is I'm assuming everybody who's listening to this is

195
00:10:14,080 --> 00:10:17,400
Speaker 5: aware of the level of censorship that's in public discourse

196
00:10:17,600 --> 00:10:21,319
Speaker 5: on the Chinese Internet. That does not stop Chinese nedicence

197
00:10:21,320 --> 00:10:24,560
Speaker 5: from wanting to express themselves in a certain way. So

198
00:10:24,720 --> 00:10:27,520
Speaker 5: if you are in a ficianadel of the Chinese language.

199
00:10:27,600 --> 00:10:30,800
Speaker 5: You will see a lot of contortions of linguistic kind

200
00:10:30,800 --> 00:10:34,560
Speaker 5: of acrobats being used in all kinds of ways to

201
00:10:34,600 --> 00:10:38,360
Speaker 5: publish public in China that you can convey certain meanings,

202
00:10:38,400 --> 00:10:41,080
Speaker 5: but when it comes to modeling the language actually becomes

203
00:10:41,240 --> 00:10:44,439
Speaker 5: very difficult to model the accuracy of the language. So

204
00:10:44,480 --> 00:10:47,800
Speaker 5: that's number one and number two because of again the

205
00:10:47,840 --> 00:10:50,760
Speaker 5: red line of censorship moving at all times, so it's

206
00:10:50,800 --> 00:10:54,080
Speaker 5: not like a static target. A lot of contents actually

207
00:10:54,080 --> 00:10:57,480
Speaker 5: get deleted on the Chinese Internet. Things just disappear, you know,

208
00:10:57,480 --> 00:10:59,760
Speaker 5: if you don't archive of a certain article, when you

209
00:11:00,240 --> 00:11:02,480
Speaker 5: read it two days later, it might be gone for

210
00:11:03,080 --> 00:11:04,920
Speaker 5: you know, reasons that a lot of people can't even

211
00:11:05,000 --> 00:11:07,160
Speaker 5: decipher even if you think about this all the time,

212
00:11:07,440 --> 00:11:10,760
Speaker 5: and that reduces the quantity. And I would say the

213
00:11:10,840 --> 00:11:14,559
Speaker 5: quality of training data that exists on their Chinese Internet

214
00:11:14,600 --> 00:11:18,120
Speaker 5: in particular, that I think is almost built in disadvantage

215
00:11:18,720 --> 00:11:21,520
Speaker 5: to a lot of Chinese model makers. They're building their

216
00:11:21,559 --> 00:11:24,880
Speaker 5: own lom you know, despite all odds there of course,

217
00:11:24,920 --> 00:11:27,320
Speaker 5: still continuing their effort to do so.

218
00:11:27,520 --> 00:11:29,800
Speaker 2: One thing I like about some of the AI conversations

219
00:11:29,920 --> 00:11:32,880
Speaker 2: is they get into the realm of linguistic and linguistic theory.

220
00:11:32,960 --> 00:11:34,920
Speaker 2: Can I just press you on that point about the

221
00:11:35,000 --> 00:11:38,640
Speaker 2: nature of the language itself and the effect that that

222
00:11:38,920 --> 00:11:41,600
Speaker 2: has on building a high quality model.

223
00:11:41,800 --> 00:11:44,360
Speaker 5: Yeah, so, I would say two things that's interesting about

224
00:11:44,400 --> 00:11:47,800
Speaker 5: their Chinese language in particular is that one, there are

225
00:11:47,800 --> 00:11:50,040
Speaker 5: a lot of homonyms. There are a lot of sounds.

226
00:11:50,080 --> 00:11:52,320
Speaker 5: There are a lot of characters that sound almost exactly

227
00:11:52,360 --> 00:11:55,360
Speaker 5: the same. That means, you know, nothing similar.

228
00:11:55,640 --> 00:11:58,360
Speaker 3: The famous example is the whole like It's either a

229
00:11:58,400 --> 00:12:02,040
Speaker 3: poem or a very long sentence with like ma like force.

230
00:12:02,240 --> 00:12:05,720
Speaker 3: And all these ma can be depending on the intonation,

231
00:12:05,800 --> 00:12:08,040
Speaker 3: can like mean a bunch of different things. You can

232
00:12:08,080 --> 00:12:12,240
Speaker 3: construct a whole poem or sentence out of that single syllable.

233
00:12:11,960 --> 00:12:15,040
Speaker 5: That's right like ma ma ma right the four tones

234
00:12:15,080 --> 00:12:17,720
Speaker 5: of the standard Chinese Mandarin language. And if you just

235
00:12:17,760 --> 00:12:21,760
Speaker 5: put that into the tech context, you know, Jack Ma Mayun,

236
00:12:21,800 --> 00:12:24,679
Speaker 5: the founder of Ali Baba, has been either a celebrated

237
00:12:24,800 --> 00:12:27,880
Speaker 5: entrepreneur or a villain, depending on how you think about

238
00:12:27,880 --> 00:12:31,400
Speaker 5: his current public stature. Right. That makes the modeling of

239
00:12:31,480 --> 00:12:34,880
Speaker 5: any let's just say, publicly available information about Jack Ma

240
00:12:35,280 --> 00:12:39,480
Speaker 5: pretty difficult when you're also building a generative application on

241
00:12:39,559 --> 00:12:41,000
Speaker 5: top of it. And there are a lot of these

242
00:12:41,080 --> 00:12:45,600
Speaker 5: examples in the Chinese linguistic case. That makes not just

243
00:12:45,640 --> 00:12:48,719
Speaker 5: the modeling difficult on a technical sense, but also what

244
00:12:48,800 --> 00:12:51,320
Speaker 5: can or you cannot generate based on the red lines

245
00:12:51,320 --> 00:12:52,760
Speaker 5: that you may or may not have to cross.

246
00:12:52,960 --> 00:12:55,880
Speaker 4: Yeah, So, just to play Devil's advocate here for a second,

247
00:12:56,240 --> 00:12:59,840
Speaker 4: GPT four was trained on zero point two percent Mandarin

248
00:13:00,160 --> 00:13:03,079
Speaker 4: and is ninety nine percent as smart in Chinese as

249
00:13:03,120 --> 00:13:05,920
Speaker 4: it is in English. And when you look at what

250
00:13:06,320 --> 00:13:09,400
Speaker 4: some folks out of the leading Chinese AI labs are saying,

251
00:13:09,679 --> 00:13:12,840
Speaker 4: like they're perfectly happy getting the vast majority of their

252
00:13:12,880 --> 00:13:16,480
Speaker 4: training data from English and sort of pourting it over

253
00:13:16,559 --> 00:13:19,000
Speaker 4: to Chinese. So there is a little challenge there. But

254
00:13:19,040 --> 00:13:21,280
Speaker 4: I also think it's it's the other thing on this

255
00:13:21,400 --> 00:13:24,280
Speaker 4: sort of are the model is going to be lobotomized

256
00:13:24,360 --> 00:13:27,320
Speaker 4: because they will be able to be too politically sensitive.

257
00:13:27,320 --> 00:13:29,440
Speaker 4: I mean, look, I feel like this is basically a

258
00:13:29,480 --> 00:13:33,720
Speaker 4: solved problem. If you try to make Chatgypt say something

259
00:13:33,840 --> 00:13:38,199
Speaker 4: racist or sexist, it basically won't. And if Chinese internet

260
00:13:38,200 --> 00:13:41,360
Speaker 4: companies have data sets on anything, it's all the posts

261
00:13:41,440 --> 00:13:44,080
Speaker 4: that they've censored over the past twenty years, you know.

262
00:13:44,679 --> 00:13:47,760
Speaker 4: Having also tried to do this exercise of getting Chinese

263
00:13:47,800 --> 00:13:50,439
Speaker 4: models to say politically racy stuff. It is a pretty

264
00:13:50,480 --> 00:13:53,920
Speaker 4: hard slog. So you know, on the plus side, I think,

265
00:13:53,960 --> 00:13:56,400
Speaker 4: you know, not all data that's going to matter for

266
00:13:56,440 --> 00:13:59,320
Speaker 4: the future of AI and it's economic and military impact

267
00:13:59,400 --> 00:14:02,480
Speaker 4: is going to be linguistic and text based. And you know,

268
00:14:02,600 --> 00:14:05,960
Speaker 4: being the world's manufacturing center, putting cameras into all those

269
00:14:06,000 --> 00:14:09,400
Speaker 4: factories may help you train a sort of like AI

270
00:14:10,400 --> 00:14:13,200
Speaker 4: manufacturing robot a lot better. So there are real trade

271
00:14:13,240 --> 00:14:15,959
Speaker 4: offs here, and it's not just like China is sitting

272
00:14:15,960 --> 00:14:28,680
Speaker 4: behind the eight bombs.

273
00:14:32,560 --> 00:14:34,760
Speaker 3: I just realized how old I am because I remember

274
00:14:34,760 --> 00:14:39,000
Speaker 3: when China's censors like actually physically used to cross things

275
00:14:39,080 --> 00:14:42,480
Speaker 3: out in newspapers that were like important in Beijing. This

276
00:14:42,480 --> 00:14:44,480
Speaker 3: would have been around like two thousand and four, two

277
00:14:44,560 --> 00:14:46,520
Speaker 3: thousand and five. It used to be done by hand.

278
00:14:46,880 --> 00:14:49,000
Speaker 3: One thing I wanted to ask, since Kevin brought up

279
00:14:49,000 --> 00:14:53,080
Speaker 3: the sort of like vagaries of Chinese politics where one

280
00:14:53,160 --> 00:14:56,400
Speaker 3: day something is really in and the next day something

281
00:14:56,560 --> 00:14:59,920
Speaker 3: is very much out, we have seen that bleed into

282
00:15:00,840 --> 00:15:06,520
Speaker 3: financial markets and specifically you know, funding for technology, primarily

283
00:15:06,600 --> 00:15:10,560
Speaker 3: internet technology SaaS in recent years, but also into housing

284
00:15:10,600 --> 00:15:14,720
Speaker 3: and things like that with the three redlines policy. So

285
00:15:14,720 --> 00:15:18,520
Speaker 3: I'm curious, like, what does the funding environment actually look

286
00:15:18,680 --> 00:15:21,440
Speaker 3: like for AI at the moment? Is there a lot

287
00:15:21,480 --> 00:15:24,440
Speaker 3: of like government support for developing this technology?

288
00:15:24,480 --> 00:15:26,480
Speaker 2: And just while you're answering that, like who are some

289
00:15:26,520 --> 00:15:28,560
Speaker 2: of the companies, Yeah, you shouldn't, like who are the

290
00:15:28,600 --> 00:15:30,800
Speaker 2: open AI and anthropic of China that we should know

291
00:15:30,800 --> 00:15:31,120
Speaker 2: about it?

292
00:15:31,160 --> 00:15:31,320
Speaker 5: Yeah?

293
00:15:31,320 --> 00:15:33,120
Speaker 3: And how are they raising money? Like is most of

294
00:15:33,120 --> 00:15:37,480
Speaker 3: it coming from vcs, vcs or public vironments?

295
00:15:37,560 --> 00:15:41,560
Speaker 5: Yeah? So when it comes to the AI model makers, right,

296
00:15:41,600 --> 00:15:44,320
Speaker 5: I'll put that in one category of new startups, which

297
00:15:44,360 --> 00:15:47,480
Speaker 5: actually kind of came up around the same time as

298
00:15:47,600 --> 00:15:49,600
Speaker 5: all the other ones that we already know here in

299
00:15:49,640 --> 00:15:53,800
Speaker 5: the US as well. You have Dripai, which is, you know,

300
00:15:54,000 --> 00:15:57,840
Speaker 5: purported to be one of the company that's closest to

301
00:15:57,920 --> 00:16:01,080
Speaker 5: open ai. You know, they're avow to all so develop AGI.

302
00:16:01,760 --> 00:16:05,240
Speaker 5: They have strong academic roots from Qinghuai University. You have

303
00:16:05,360 --> 00:16:09,440
Speaker 5: Moonsha that I think Jordan mentioned. You have zero one

304
00:16:09,520 --> 00:16:12,480
Speaker 5: dot Ai, which is a lee hai Fu's a new

305
00:16:12,600 --> 00:16:15,560
Speaker 5: venture building models as well. And I would say they're

306
00:16:15,560 --> 00:16:20,680
Speaker 5: probably six or seven Chinese unicorns that currently exist that

307
00:16:20,960 --> 00:16:25,480
Speaker 5: receive funding mostly from your Chinese standard vcs. These are

308
00:16:25,560 --> 00:16:27,640
Speaker 5: vcs that are just like the Sequoias and the and

309
00:16:27,800 --> 00:16:30,400
Speaker 5: reasons of the world. Right, you have Jung Fund, you

310
00:16:30,520 --> 00:16:33,200
Speaker 5: have a Hillhouse Capital, you have home Shin who was

311
00:16:33,280 --> 00:16:37,160
Speaker 5: formerly a Sequoia capital China, and a few others pouring

312
00:16:37,240 --> 00:16:41,480
Speaker 5: money into this particular lane because one thing I think

313
00:16:41,960 --> 00:16:43,680
Speaker 5: we should back up is that this is one of

314
00:16:43,720 --> 00:16:47,120
Speaker 5: the very few lanes that vcs in China can actually

315
00:16:47,120 --> 00:16:50,040
Speaker 5: put a lot of money into and expect the kind

316
00:16:50,080 --> 00:16:53,840
Speaker 5: of return that you would expect from a pure financial

317
00:16:53,920 --> 00:16:56,720
Speaker 5: VC because a lot of other parts of the Chinese

318
00:16:56,760 --> 00:16:59,960
Speaker 5: tech ecosystem has been battered down quite a bit because

319
00:17:00,160 --> 00:17:03,160
Speaker 5: of all the differences in the policy crackdown that happened

320
00:17:03,160 --> 00:17:05,080
Speaker 5: in the last few years or so. So those are

321
00:17:05,600 --> 00:17:08,399
Speaker 5: some of the big model maker players, and a lot

322
00:17:08,440 --> 00:17:12,159
Speaker 5: of them are attracting large corporate venture investments from Alibaba

323
00:17:12,200 --> 00:17:15,200
Speaker 5: in particular, but also Huaiwi, ten Cent, all the big

324
00:17:15,280 --> 00:17:18,600
Speaker 5: dogs pouring into it. Very similar again to what we

325
00:17:18,680 --> 00:17:21,920
Speaker 5: have here in the US, where Microsoft, Google and Amazon

326
00:17:22,040 --> 00:17:24,520
Speaker 5: all have like their player on the field, if you will,

327
00:17:24,680 --> 00:17:27,119
Speaker 5: when it comes to startups, and even some of the

328
00:17:27,160 --> 00:17:30,560
Speaker 5: deal structures are similar I think one of Ali Baba's

329
00:17:30,640 --> 00:17:33,840
Speaker 5: deal is that half of the investment they're making has

330
00:17:33,880 --> 00:17:37,720
Speaker 5: to go into cloud credits to use the GPU service

331
00:17:37,760 --> 00:17:40,480
Speaker 5: that are in Alibaba Cloud to be able to train

332
00:17:40,640 --> 00:17:43,359
Speaker 5: their model as a way to entice, but also to

333
00:17:43,520 --> 00:17:46,560
Speaker 5: tie them into Ali Baba's on ecosystem. So actually they're

334
00:17:46,560 --> 00:17:49,320
Speaker 5: probably more similarities, I would say, than differences when it

335
00:17:49,359 --> 00:17:52,480
Speaker 5: comes to the AI model makers funded by vcs, But

336
00:17:52,600 --> 00:17:54,800
Speaker 5: that's about a different as similar as it gets.

337
00:17:55,000 --> 00:17:58,399
Speaker 2: Jordan, you mentioned the robots and the idea that another

338
00:17:58,520 --> 00:18:01,560
Speaker 2: form of data is just all the visual information that's

339
00:18:01,560 --> 00:18:04,640
Speaker 2: inside a factory, and we did an episode on sort

340
00:18:04,640 --> 00:18:08,119
Speaker 2: of the connection with AI and robotics with Josh Wolf

341
00:18:08,160 --> 00:18:10,840
Speaker 2: and we talked about some of these ideas. But you know,

342
00:18:10,880 --> 00:18:12,680
Speaker 2: you mentioned you know you're talking about okay, they also

343
00:18:12,680 --> 00:18:15,679
Speaker 2: want to achieve AGI. Are there distinct aspects though of

344
00:18:15,720 --> 00:18:18,399
Speaker 2: the strategy that is like oh no, like okay, we

345
00:18:18,600 --> 00:18:21,639
Speaker 2: are thinking about AI in the chatbot sense, So that

346
00:18:21,800 --> 00:18:24,840
Speaker 2: is the way most people in the US think about

347
00:18:24,920 --> 00:18:28,600
Speaker 2: what these AI companies have delivered. Is it the same

348
00:18:28,840 --> 00:18:33,080
Speaker 2: with these Chinese powerhouses or are they thinking about different

349
00:18:33,160 --> 00:18:36,320
Speaker 2: forms of expression maybe in a more industrial manner. Et

350
00:18:36,320 --> 00:18:39,680
Speaker 2: cetera than the sort of chatbot idiom.

351
00:18:39,680 --> 00:18:41,840
Speaker 5: So Joe one thing that this is more of a

352
00:18:41,880 --> 00:18:45,280
Speaker 5: recent development. So I think just the last week, Shanghai

353
00:18:45,320 --> 00:18:50,840
Speaker 5: hosted its annual World AI conference, right, and during this

354
00:18:51,000 --> 00:18:53,760
Speaker 5: conference is a big show and there I believe twenty

355
00:18:53,840 --> 00:18:57,720
Speaker 5: five different brands of humanoid robots that were on display

356
00:18:58,200 --> 00:19:01,280
Speaker 5: in the expo you know, hall of this center. So

357
00:19:01,359 --> 00:19:04,320
Speaker 5: that illustrates I think to you that I think the

358
00:19:04,400 --> 00:19:06,399
Speaker 5: direction I think Tracy asked about this, you know, where's

359
00:19:06,440 --> 00:19:07,240
Speaker 5: the funding going to?

360
00:19:07,440 --> 00:19:07,600
Speaker 3: Right?

361
00:19:07,800 --> 00:19:10,720
Speaker 5: I think if the national government were to have a say,

362
00:19:10,960 --> 00:19:14,159
Speaker 5: a lot of the funding ought to be go into

363
00:19:14,359 --> 00:19:17,520
Speaker 5: AI applications that are what I call heart tech things

364
00:19:17,520 --> 00:19:20,040
Speaker 5: that you can actually touch, things that actually do certain

365
00:19:20,080 --> 00:19:22,840
Speaker 5: things in the real world, whether it's in a car factory,

366
00:19:23,119 --> 00:19:27,320
Speaker 5: in a manufacturing shop, or even actual robots that will

367
00:19:27,320 --> 00:19:29,919
Speaker 5: be in your home. And I think that is the

368
00:19:29,960 --> 00:19:32,639
Speaker 5: direction that a lot of Chinese companies that are not

369
00:19:33,200 --> 00:19:35,760
Speaker 5: kind of in the model building world are going into.

370
00:19:36,240 --> 00:19:38,919
Speaker 5: And I think there's actually one very important, kind of

371
00:19:39,080 --> 00:19:42,040
Speaker 5: larger existential crisis that China's dealing with and a lot

372
00:19:42,080 --> 00:19:44,119
Speaker 5: of people don't talk about in the context of AI,

373
00:19:44,480 --> 00:19:46,840
Speaker 5: which is that by the end of this century, the

374
00:19:47,000 --> 00:19:50,640
Speaker 5: UN is projecting that the entire population of China will

375
00:19:50,680 --> 00:19:53,440
Speaker 5: go down to around eight hundred million to about one

376
00:19:53,440 --> 00:19:56,720
Speaker 5: point four billion right now, so roughly have and half

377
00:19:56,760 --> 00:19:59,320
Speaker 5: of that eight hundred million people will be people who

378
00:19:59,359 --> 00:20:04,239
Speaker 5: are sixty years or older. So demography is destiny in

379
00:20:04,280 --> 00:20:06,520
Speaker 5: a lot of ways. And if you're a Chinese policymaker

380
00:20:06,600 --> 00:20:09,160
Speaker 5: steering into this abyss that we're just going to basically

381
00:20:09,200 --> 00:20:11,920
Speaker 5: rent out of people, then you know, having a lot

382
00:20:11,920 --> 00:20:16,200
Speaker 5: of AI robotics is almost a national imperative for China

383
00:20:16,280 --> 00:20:20,000
Speaker 5: to solve its own demographic problem with or without, you know,

384
00:20:20,280 --> 00:20:21,879
Speaker 5: everything else going on around the world.

385
00:20:22,800 --> 00:20:27,360
Speaker 3: Okay, So one of the benefits of having a command

386
00:20:27,520 --> 00:20:31,760
Speaker 3: economy essentially is that Chinese policy makers can come out

387
00:20:31,840 --> 00:20:35,560
Speaker 3: and say, we want everyone to direct their money away

388
00:20:35,680 --> 00:20:40,159
Speaker 3: from I don't know, online internet e commerce platforms and

389
00:20:40,280 --> 00:20:45,080
Speaker 3: into hard technology like semiconductors or developing some new chatbot

390
00:20:45,200 --> 00:20:48,840
Speaker 3: or something like that. But the downside is that you

391
00:20:49,040 --> 00:20:53,479
Speaker 3: can get excess capital to put it politely crowded into

392
00:20:53,600 --> 00:20:56,800
Speaker 3: these things. Is that a concern when it comes to AI,

393
00:20:57,000 --> 00:21:00,639
Speaker 3: or is it that the technology is so strategically important

394
00:21:00,680 --> 00:21:05,399
Speaker 3: and so competitive that it's not really a concern at

395
00:21:05,400 --> 00:21:06,159
Speaker 3: the moment, you know.

396
00:21:06,200 --> 00:21:08,840
Speaker 4: So I think it's an interesting contrast you can look

397
00:21:08,920 --> 00:21:12,240
Speaker 4: at between sort of the AI model makers and the

398
00:21:12,280 --> 00:21:17,080
Speaker 4: semiconductor and the broader semiconductor ecosystem. So what Kevin laid out,

399
00:21:17,119 --> 00:21:19,239
Speaker 4: which is what I see as well, is basically like

400
00:21:19,560 --> 00:21:23,360
Speaker 4: a largely private sector driven thing with the same kind

401
00:21:23,400 --> 00:21:26,680
Speaker 4: of boom and bust or boom and like question mark

402
00:21:26,720 --> 00:21:29,680
Speaker 4: that we're current living living into in the West, happening

403
00:21:29,720 --> 00:21:30,200
Speaker 4: in China.

404
00:21:30,280 --> 00:21:30,480
Speaker 3: Now.

405
00:21:30,720 --> 00:21:34,080
Speaker 4: You know, what's happened in the Chinese semiconductor ecosystem over

406
00:21:34,119 --> 00:21:37,919
Speaker 4: the past decade or so is really different. So you know,

407
00:21:38,119 --> 00:21:40,440
Speaker 4: if you're talking about the US Chips Act, which didn't

408
00:21:40,440 --> 00:21:43,040
Speaker 4: come online until twenty twenty two, and it's going to

409
00:21:43,080 --> 00:21:46,480
Speaker 4: spend roughly seventy five billion dollars, like we've seen that

410
00:21:46,920 --> 00:21:49,920
Speaker 4: level of investment doubled or even trebled over the past

411
00:21:49,920 --> 00:21:52,480
Speaker 4: ten years in the Chinese government, you know, she is

412
00:21:52,600 --> 00:21:57,760
Speaker 4: really semiconductor and hardware and industrial pilled. And I think

413
00:21:57,800 --> 00:22:01,320
Speaker 4: he also from a long time ago saw the writing

414
00:22:01,359 --> 00:22:03,960
Speaker 4: on the wall of these you know, of these export

415
00:22:04,000 --> 00:22:07,920
Speaker 4: restrictions coming on as uh as semi conductors became more

416
00:22:07,960 --> 00:22:10,480
Speaker 4: and more, more and more and more strategic technology.

417
00:22:10,600 --> 00:22:12,639
Speaker 2: She she should ping. Sounds like he could be the

418
00:22:12,680 --> 00:22:17,320
Speaker 2: perfect guest if anyone wants to disseminate that. If anyone's listening,

419
00:22:17,400 --> 00:22:18,280
Speaker 2: someone who's.

420
00:22:20,200 --> 00:22:22,360
Speaker 3: To talk about semi conduct if he if he's.

421
00:22:22,240 --> 00:22:25,120
Speaker 2: Hardware and chip pilled, it sounds like he would be

422
00:22:25,200 --> 00:22:28,200
Speaker 2: a great guest. So let's talk a little bit more.

423
00:22:28,840 --> 00:22:33,399
Speaker 2: We know that the Biden administration has imposed, along with

424
00:22:33,400 --> 00:22:37,720
Speaker 2: some allies, various limitations on the export of advanced chips

425
00:22:37,720 --> 00:22:40,480
Speaker 2: and so forth. But why don't one of you just

426
00:22:40,520 --> 00:22:43,840
Speaker 2: sort of give us, like the broad outline beyond or

427
00:22:43,880 --> 00:22:50,520
Speaker 2: including the chip limitations of US policy towards you know,

428
00:22:50,720 --> 00:22:54,880
Speaker 2: international AI policy towards China, and how we're actively implementing

429
00:22:54,920 --> 00:22:58,320
Speaker 2: the strategic aspect or the geopolitical aspect of AI.

430
00:22:58,600 --> 00:23:01,200
Speaker 4: Again. You know, I started with this idea that AI

431
00:23:01,400 --> 00:23:05,040
Speaker 4: is a strategic, dual use technology that has big implications

432
00:23:05,040 --> 00:23:07,440
Speaker 4: both for the future of economic growth as well as

433
00:23:07,440 --> 00:23:10,399
Speaker 4: the future of military power. So in September of twenty

434
00:23:10,440 --> 00:23:14,479
Speaker 4: twenty two, a few weeks before the October seventh Export controls,

435
00:23:14,520 --> 00:23:17,120
Speaker 4: Jake Sulivan gave a speech where he said that it's

436
00:23:17,160 --> 00:23:19,719
Speaker 4: not good enough to just be in the lead, but

437
00:23:19,760 --> 00:23:21,960
Speaker 4: that the US needs to have as far a lead

438
00:23:22,000 --> 00:23:26,960
Speaker 4: as possible in critical emerging strategic technologies, AI being first

439
00:23:27,000 --> 00:23:29,600
Speaker 4: and foremost on that list. So we've seen a number

440
00:23:29,640 --> 00:23:32,959
Speaker 4: of different efforts on you know, from both a data

441
00:23:33,000 --> 00:23:35,920
Speaker 4: and algorithmic as well as a compute perspective, to make

442
00:23:35,960 --> 00:23:38,760
Speaker 4: sure that America can continue to run faster than China.

443
00:23:38,880 --> 00:23:43,280
Speaker 4: So from export controls, you have limitations on the types

444
00:23:43,280 --> 00:23:45,840
Speaker 4: of chips which you can be exported into China. We've

445
00:23:45,840 --> 00:23:48,720
Speaker 4: had EUV restrictions so that China can't make the most

446
00:23:49,280 --> 00:23:52,280
Speaker 4: advanced chips domestically, the types of chips that the US

447
00:23:52,359 --> 00:23:55,240
Speaker 4: is no longer allowing Chinese firms to buy in country,

448
00:23:55,400 --> 00:23:58,439
Speaker 4: We're starting to see more and more restrict you know,

449
00:23:59,320 --> 00:24:02,639
Speaker 4: there's movement the water around restrictions of letting Chinese firms

450
00:24:02,680 --> 00:24:07,359
Speaker 4: access those types of chips abroad and with foreign cloud providers.

451
00:24:07,600 --> 00:24:10,680
Speaker 4: And then there's also starting to be some interesting movement

452
00:24:10,800 --> 00:24:14,560
Speaker 4: around restrictions when it comes to both data and algorithms.

453
00:24:14,880 --> 00:24:18,160
Speaker 4: On the data side, you're starting to see new data

454
00:24:18,200 --> 00:24:22,600
Speaker 4: walls be lifted with regards to electric vehicles in particular,

455
00:24:22,680 --> 00:24:25,480
Speaker 4: we'll maybe see more coming forward, and then from an

456
00:24:25,480 --> 00:24:28,880
Speaker 4: algorithmic perspective, you know, two weeks ago open aiy said

457
00:24:28,880 --> 00:24:32,000
Speaker 4: that they were no longer comfortable using a giving API

458
00:24:32,080 --> 00:24:35,520
Speaker 4: access to developers based in the PRC, and that may

459
00:24:35,560 --> 00:24:41,119
Speaker 4: be a harbinger of something larger. Congress recently gave this

460
00:24:41,520 --> 00:24:47,160
Speaker 4: the power to explicitly stop algorithmic exports. So the wall

461
00:24:47,320 --> 00:24:50,800
Speaker 4: slowly but surely is coming up, and the Chinese developers

462
00:24:50,800 --> 00:24:54,399
Speaker 4: and semiconductor ecosystems response to this is really interesting because

463
00:24:54,640 --> 00:24:59,440
Speaker 4: you're seeing very aggressive investment on the hardware side. You

464
00:24:59,520 --> 00:25:03,560
Speaker 4: see stock piling of chips and even billions of dollars

465
00:25:03,560 --> 00:25:08,440
Speaker 4: of the best of what Nvidia can legally export to them. Smuggling, too,

466
00:25:08,520 --> 00:25:10,480
Speaker 4: is starting to be a real thing we recently saw

467
00:25:10,520 --> 00:25:13,800
Speaker 4: some mainstream coverage of, so you know, there will be

468
00:25:14,080 --> 00:25:17,320
Speaker 4: kind of gray areas and people trying to play corner

469
00:25:17,359 --> 00:25:22,600
Speaker 4: cases to get more technology and access. Industrial espionage as

470
00:25:22,640 --> 00:25:24,840
Speaker 4: well as something that starts starting to bubble up a

471
00:25:24,840 --> 00:25:27,439
Speaker 4: bit more. The scandal of sorts with Leopold Ashenbrett are

472
00:25:27,440 --> 00:25:30,119
Speaker 4: getting fired, only for two weeks later Open ai to

473
00:25:30,200 --> 00:25:33,920
Speaker 4: bring on General Nakasoni, the former head of the NSA,

474
00:25:34,000 --> 00:25:37,120
Speaker 4: to shore up their internal controls, I think is an

475
00:25:37,119 --> 00:25:39,640
Speaker 4: interesting leading indicator of what's to come on that front.

476
00:25:55,240 --> 00:25:59,560
Speaker 3: One of the things that you sometimes hear about restrictions on,

477
00:25:59,720 --> 00:26:03,280
Speaker 3: for instance, semiconductor exports to China is the idea that

478
00:26:03,359 --> 00:26:07,600
Speaker 3: by cutting off the country from all this external technology,

479
00:26:07,640 --> 00:26:12,560
Speaker 3: you're only going to accelerate its own technological progress and

480
00:26:12,760 --> 00:26:16,320
Speaker 3: its own development of leading edge chips and things like that.

481
00:26:16,880 --> 00:26:21,720
Speaker 3: How much credence do you lend to those types of arguments.

482
00:26:22,000 --> 00:26:24,080
Speaker 5: I think it's a moving target right now. I think

483
00:26:24,160 --> 00:26:28,159
Speaker 5: logically you would expect China to put all of this

484
00:26:28,440 --> 00:26:32,879
Speaker 5: resources into building the best GPU possible, right if we

485
00:26:32,920 --> 00:26:34,920
Speaker 5: all think AI is going to be the next a

486
00:26:35,000 --> 00:26:39,159
Speaker 5: big paradigm shift for national competitiveness, not just market competitiveness

487
00:26:39,320 --> 00:26:42,320
Speaker 5: yet right now, because of I will say, the effectiveness

488
00:26:42,359 --> 00:26:45,280
Speaker 5: of US expert control, two rounds of export control, and

489
00:26:45,359 --> 00:26:49,520
Speaker 5: also working with allies around the world to solidify expert control,

490
00:26:49,760 --> 00:26:52,119
Speaker 5: it's actually been a very difficult time for China to

491
00:26:52,280 --> 00:26:55,439
Speaker 5: catch up from the hardware side, both in terms of

492
00:26:55,520 --> 00:27:00,200
Speaker 5: designing the best GPU possible, but also to manufacture the

493
00:27:00,800 --> 00:27:05,560
Speaker 5: chips in bulk using its own chip boundry SMIC because

494
00:27:05,600 --> 00:27:08,240
Speaker 5: you know, China's been cut off from TSMC to be

495
00:27:08,320 --> 00:27:12,560
Speaker 5: able to meet demand that China itself, its own companies

496
00:27:12,640 --> 00:27:15,000
Speaker 5: need to build its own product, let alone expanding to

497
00:27:15,440 --> 00:27:17,880
Speaker 5: you know, other parts around the world. And that's why

498
00:27:17,880 --> 00:27:21,240
Speaker 5: you see all these behaviors still of either Chinese developers

499
00:27:21,320 --> 00:27:24,679
Speaker 5: using VPNs and other ways to access open ais APIs,

500
00:27:25,160 --> 00:27:28,840
Speaker 5: or you have these sort of smuggling of prevented or

501
00:27:28,880 --> 00:27:31,520
Speaker 5: prohibited in media chips into China to be able to

502
00:27:31,640 --> 00:27:35,480
Speaker 5: keep the building of the infrastructure going. And I think

503
00:27:35,480 --> 00:27:38,040
Speaker 5: that goes back to what you were talking about Tracy earlier,

504
00:27:38,280 --> 00:27:41,720
Speaker 5: is that there is a limitation to how much the

505
00:27:41,800 --> 00:27:45,440
Speaker 5: national government wants certain things to happen and whether that

506
00:27:45,480 --> 00:27:48,160
Speaker 5: could actually happen or not when it comes to especially

507
00:27:48,160 --> 00:27:51,040
Speaker 5: building hard tech in China. When it comes to AI

508
00:27:51,119 --> 00:27:54,440
Speaker 5: infrastructure in particular, we see a lot of pretty poor

509
00:27:54,600 --> 00:27:59,600
Speaker 5: examples of government money being wasted in chip manufacturing already

510
00:27:59,640 --> 00:28:02,720
Speaker 5: in a few years before AI was even really a thing,

511
00:28:03,480 --> 00:28:06,760
Speaker 5: and now it's in a very tough spot. I think

512
00:28:06,840 --> 00:28:08,640
Speaker 5: for a lot of people in China to be able

513
00:28:08,640 --> 00:28:12,840
Speaker 5: to overcome the current hurdle, and I think this gap,

514
00:28:13,160 --> 00:28:19,080
Speaker 5: this specifically hardware GPU imposed gap, will only widen when

515
00:28:19,200 --> 00:28:23,200
Speaker 5: Nvidia mass produced. It is a black Well chip, which

516
00:28:23,240 --> 00:28:26,919
Speaker 5: is the next generation of GPU that's going to you know,

517
00:28:27,000 --> 00:28:30,840
Speaker 5: be produced in bulk starting next year, which has a

518
00:28:30,960 --> 00:28:35,200
Speaker 5: huge performance kind of advantage, you know, roughly thirty two

519
00:28:35,240 --> 00:28:38,840
Speaker 5: times more than the A chips, which is what opening

520
00:28:38,920 --> 00:28:41,480
Speaker 5: I used to train GPT for to be able to

521
00:28:42,000 --> 00:28:44,600
Speaker 5: you know, close that gap. So I think unfortunately for

522
00:28:44,640 --> 00:28:47,560
Speaker 5: the Chinese technology builders, the gap will only widen.

523
00:28:48,080 --> 00:28:50,560
Speaker 3: Jordan, I think you've talked about this before as well,

524
00:28:50,600 --> 00:28:53,680
Speaker 3: the idea that if China is restricted in the amount

525
00:28:53,760 --> 00:28:57,200
Speaker 3: of compute that it has access to, maybe there's a

526
00:28:57,240 --> 00:29:01,800
Speaker 3: scenario where its own aid has just become really really

527
00:29:01,840 --> 00:29:03,360
Speaker 3: efficient in one way or another.

528
00:29:03,920 --> 00:29:06,320
Speaker 4: Yeah. I mean this is you know, when you're under

529
00:29:06,360 --> 00:29:08,719
Speaker 4: constraints and there's money to be made, and you know,

530
00:29:09,120 --> 00:29:12,240
Speaker 4: engineers will find a way, right, So I on the

531
00:29:12,280 --> 00:29:15,600
Speaker 4: compute side, right. You know, it's interesting thinking about Blackwell.

532
00:29:15,640 --> 00:29:18,560
Speaker 4: Only about ten percent of the gains that it is

533
00:29:18,640 --> 00:29:22,560
Speaker 4: making from the past generation came from a node size.

534
00:29:22,800 --> 00:29:24,640
Speaker 4: The vast majority of the gains that they've been able

535
00:29:24,680 --> 00:29:27,640
Speaker 4: to squeeze into the Blackwell system are from stuff like

536
00:29:27,640 --> 00:29:32,520
Speaker 4: like optics and interconnects and that stuff. High bandwidth memory.

537
00:29:32,720 --> 00:29:34,880
Speaker 4: That's type of stuff is a lot more difficult to

538
00:29:34,920 --> 00:29:38,719
Speaker 4: export control than just one machine coming out of the Netherlands.

539
00:29:38,720 --> 00:29:42,600
Speaker 4: So you know, it's difficult. There's tens of billions of

540
00:29:42,680 --> 00:29:44,680
Speaker 4: dollars of R and D that's been on this stuff.

541
00:29:44,880 --> 00:29:47,680
Speaker 4: But it's sort of an open question to me how

542
00:29:47,720 --> 00:29:51,360
Speaker 4: difficult re engineering all the innovations that went into Blackwell

543
00:29:51,640 --> 00:29:54,400
Speaker 4: is going to be versus re engineering an e UV machine,

544
00:29:54,400 --> 00:29:57,200
Speaker 4: which like seems to be next to impossible. This is

545
00:29:57,240 --> 00:30:00,240
Speaker 4: an important point just to jump in here. Like when

546
00:30:00,240 --> 00:30:03,960
Speaker 4: people when in pop culture, when people talk about chips,

547
00:30:04,080 --> 00:30:06,880
Speaker 4: it's almost always like node size and smaller and smaller.

548
00:30:07,120 --> 00:30:10,600
Speaker 3: But this idea of like the vibrant pop culture of

549
00:30:10,640 --> 00:30:12,760
Speaker 3: semi people. But this is I'm not sure we're quite there.

550
00:30:13,000 --> 00:30:15,320
Speaker 2: This is how like ignorant people like myself if I

551
00:30:15,360 --> 00:30:17,920
Speaker 2: think like, oh, well, you know, the cutting edge chip,

552
00:30:18,400 --> 00:30:20,520
Speaker 2: it's just in my mind smaller, nodest. But this is

553
00:30:20,520 --> 00:30:22,080
Speaker 2: like a really important point that you hit on that

554
00:30:22,200 --> 00:30:25,760
Speaker 2: when actually when we're talking about say in video GPUs,

555
00:30:26,200 --> 00:30:29,280
Speaker 2: there are other dimensions that are actually much more important

556
00:30:29,280 --> 00:30:30,040
Speaker 2: than node size.

557
00:30:30,080 --> 00:30:31,920
Speaker 5: That's right. And you know, if you look at all

558
00:30:32,000 --> 00:30:34,480
Speaker 5: the new products in the video is releasing to the

559
00:30:34,560 --> 00:30:37,240
Speaker 5: market is not a chip. Of course, you can buy

560
00:30:37,280 --> 00:30:40,120
Speaker 5: a chip, but you know it's really eight chips together

561
00:30:40,280 --> 00:30:43,760
Speaker 5: connected with high bandwidth memory and the interconnects and all

562
00:30:43,760 --> 00:30:46,560
Speaker 5: the other elements that goes into a system. And then

563
00:30:46,600 --> 00:30:49,520
Speaker 5: next generation the packaging will be thirty six chip and

564
00:30:49,560 --> 00:30:51,880
Speaker 5: then seventy two chip. And of course there's a good

565
00:30:51,920 --> 00:30:55,360
Speaker 5: reason for that because all these models, if you believe

566
00:30:55,440 --> 00:30:59,360
Speaker 5: that larger models become smarter models, then they need massive

567
00:30:59,360 --> 00:31:02,320
Speaker 5: amounts of all processing to be able to train all

568
00:31:02,320 --> 00:31:05,080
Speaker 5: these model officially and to deploy them officially. So that's

569
00:31:05,120 --> 00:31:08,600
Speaker 5: why you tie up all these high performing GPUs together,

570
00:31:09,000 --> 00:31:12,480
Speaker 5: and I think that's what makes this hardware gap even

571
00:31:12,520 --> 00:31:15,880
Speaker 5: more difficult from a system level, for even a cloud

572
00:31:16,000 --> 00:31:19,640
Speaker 5: level to really overcome. I think, you know, before, we've

573
00:31:19,640 --> 00:31:22,160
Speaker 5: been talking a lot about the chip war between different countries.

574
00:31:22,240 --> 00:31:25,320
Speaker 5: I think we're vastly quickly evolving into a cloud war

575
00:31:25,560 --> 00:31:29,600
Speaker 5: between different countries competing on the AI and technology in general,

576
00:31:29,680 --> 00:31:32,280
Speaker 5: because it's all about the cloud and the system and

577
00:31:32,400 --> 00:31:35,520
Speaker 5: how officially can you acquire and then utilize them. That

578
00:31:35,560 --> 00:31:36,520
Speaker 5: makes a big difference.

579
00:31:37,320 --> 00:31:40,000
Speaker 3: Tell me if I'm wrong here, But my sense also

580
00:31:40,160 --> 00:31:44,720
Speaker 3: with like semiconductor rivalry, maybe to a lesser extent, but

581
00:31:44,800 --> 00:31:48,600
Speaker 3: certainly with the AI rivalry, is that it's mostly a

582
00:31:48,640 --> 00:31:52,000
Speaker 3: competition between the US and China. Like, as far as

583
00:31:52,000 --> 00:31:57,840
Speaker 3: I can tell, Europe isn't necessarily obsessing about developing national

584
00:31:57,960 --> 00:32:01,240
Speaker 3: chat GPT versions as much as you know the US

585
00:32:01,360 --> 00:32:04,000
Speaker 3: and China are. Why Why is that so?

586
00:32:04,160 --> 00:32:06,880
Speaker 4: You know, I think from a from the semiconductor perspective, right,

587
00:32:07,040 --> 00:32:09,959
Speaker 4: like it's really China and the rest of the world.

588
00:32:10,360 --> 00:32:13,320
Speaker 4: There is a very integrated global ecosystem in which you

589
00:32:13,400 --> 00:32:16,920
Speaker 4: have European players, East Asian players, South Asian players, the

590
00:32:17,040 --> 00:32:21,480
Speaker 4: US that are all kind of working together to push,

591
00:32:21,560 --> 00:32:24,880
Speaker 4: you know, to push node sides and improve the sort

592
00:32:24,960 --> 00:32:27,560
Speaker 4: and develop the sorts of technologies which are going into

593
00:32:27,560 --> 00:32:30,680
Speaker 4: something like Blackwell and China, you know, has kind of

594
00:32:30,720 --> 00:32:33,920
Speaker 4: been cut off or is at some levels being cut

595
00:32:33,920 --> 00:32:37,000
Speaker 4: off from that broader ecosystem and having to start to

596
00:32:37,080 --> 00:32:39,920
Speaker 4: domesticate a lot of the technologies that they were able

597
00:32:39,960 --> 00:32:43,200
Speaker 4: to formally rely on from other countries. Now, Kevin, I

598
00:32:43,200 --> 00:32:44,680
Speaker 4: don't know if you want to talk to the like

599
00:32:44,920 --> 00:32:47,200
Speaker 4: every country with its own model and own cloud thing,

600
00:32:47,200 --> 00:32:49,560
Speaker 4: because I think that's a really interesting I'm curious about that.

601
00:32:49,840 --> 00:32:51,600
Speaker 5: Yeah, I think you know not to beat on Europe

602
00:32:51,640 --> 00:32:55,360
Speaker 5: too much, but I do think given just the regulatory

603
00:32:55,440 --> 00:32:59,520
Speaker 5: I was ay bias right to wars releasing new technology

604
00:32:59,680 --> 00:33:03,880
Speaker 5: really quickly. In general, it just isn't a very hospitable

605
00:33:03,960 --> 00:33:07,600
Speaker 5: environment to be able to build a product legit CHATGPT,

606
00:33:07,800 --> 00:33:11,080
Speaker 5: which was really released I believe on a whim or

607
00:33:11,080 --> 00:33:14,000
Speaker 5: on a quick competitive note. It wasn't something that's super

608
00:33:14,040 --> 00:33:16,160
Speaker 5: thought through. But because of the environment that we have

609
00:33:16,200 --> 00:33:18,720
Speaker 5: in America, we allow that sort of things to happen

610
00:33:18,800 --> 00:33:21,640
Speaker 5: and then we think about regulation later. Coming back to

611
00:33:21,680 --> 00:33:25,600
Speaker 5: the national angle, I think there is a strong sort

612
00:33:25,640 --> 00:33:29,280
Speaker 5: of motivation, coming back to the cultural point that every

613
00:33:29,320 --> 00:33:32,720
Speaker 5: country does want its own national champion to some extent.

614
00:33:32,760 --> 00:33:35,800
Speaker 5: I think French France has Misstraw, which it kind of

615
00:33:35,800 --> 00:33:38,040
Speaker 5: puts up as its own national champion. It's a very

616
00:33:38,080 --> 00:33:41,680
Speaker 5: strong open source AI model maker that has a lot

617
00:33:41,680 --> 00:33:44,720
Speaker 5: of strong relationship with Microsoft as well. I think there's

618
00:33:44,760 --> 00:33:47,760
Speaker 5: another company out of Germany that also is a model maker.

619
00:33:48,040 --> 00:33:50,760
Speaker 5: So there are a few examples out there. Even Japan

620
00:33:50,880 --> 00:33:54,480
Speaker 5: has a I believe sakana Ai, which is its own

621
00:33:54,640 --> 00:33:59,360
Speaker 5: possible startup national champion to model, Japan's you know, culture

622
00:33:59,480 --> 00:34:02,480
Speaker 5: or truth or knowledge or whatnot into a national model,

623
00:34:02,640 --> 00:34:04,880
Speaker 5: and India has a few others as well. And the

624
00:34:04,880 --> 00:34:07,480
Speaker 5: Middle East is another very important region that is playing

625
00:34:07,480 --> 00:34:10,800
Speaker 5: an increasingly large role, I think, as the government actually

626
00:34:10,880 --> 00:34:13,480
Speaker 5: pushed out some really good open source model about a

627
00:34:13,560 --> 00:34:16,919
Speaker 5: year ago called Falcon that series which has been able

628
00:34:16,960 --> 00:34:18,960
Speaker 5: to put the Middle East on the map when it

629
00:34:19,000 --> 00:34:21,680
Speaker 5: comes to national development of AM models. So there is

630
00:34:21,719 --> 00:34:24,200
Speaker 5: a lot going on, but certainly not nearly as much

631
00:34:24,560 --> 00:34:28,959
Speaker 5: as the US China competition. Besides that, actually.

632
00:34:28,600 --> 00:34:31,480
Speaker 3: You just reminded me this is something I've always wondered

633
00:34:31,520 --> 00:34:34,640
Speaker 3: as well, but what is the state of open source

634
00:34:34,920 --> 00:34:37,359
Speaker 3: in China? Like, is it as much of a thing

635
00:34:37,480 --> 00:34:40,200
Speaker 3: as it is in the States or elsewhere? What does

636
00:34:40,200 --> 00:34:42,279
Speaker 3: that actually look like in terms of sharing code and

637
00:34:42,320 --> 00:34:42,919
Speaker 3: things like that.

638
00:34:43,600 --> 00:34:46,279
Speaker 5: So it's very much a thing in fact. So I

639
00:34:46,400 --> 00:34:49,520
Speaker 5: used to work at a GitHub, which you know, is

640
00:34:49,560 --> 00:34:52,320
Speaker 5: one of the largest a code repository of open source

641
00:34:52,360 --> 00:34:54,880
Speaker 5: code and to drive in a natural expansion, so I

642
00:34:54,920 --> 00:34:57,440
Speaker 5: got to work with a lot of Chinese open source

643
00:34:57,760 --> 00:35:01,240
Speaker 5: developers and it's actually the very few, i would say

644
00:35:01,880 --> 00:35:05,480
Speaker 5: vibrant communities that really embrace a lot of openness transparency.

645
00:35:05,840 --> 00:35:09,720
Speaker 5: They have this really nerdy holiday on October twenty fourth,

646
00:35:10,120 --> 00:35:12,759
Speaker 5: which if you you know binary is you know to

647
00:35:12,760 --> 00:35:16,720
Speaker 5: to Day, It's ten exactly, But basically it's a called

648
00:35:16,960 --> 00:35:21,880
Speaker 5: Programmer Day and it's a completely grassroots holiday in China

649
00:35:22,200 --> 00:35:24,719
Speaker 5: that celebrates open source development. They have, you know, a

650
00:35:24,800 --> 00:35:27,960
Speaker 5: huge party, a huge convention. They talk about open source sharing.

651
00:35:28,280 --> 00:35:31,239
Speaker 5: They all use you know, GitHub or hugging Face or

652
00:35:31,360 --> 00:35:36,160
Speaker 5: all your your usual open source collaboration platform. So the

653
00:35:36,200 --> 00:35:39,160
Speaker 5: open source kind of movement, if you will, has really

654
00:35:39,239 --> 00:35:42,600
Speaker 5: deep roots in China and that has been around for

655
00:35:42,680 --> 00:35:45,719
Speaker 5: a couple of decades already before we have this kind

656
00:35:45,719 --> 00:35:50,840
Speaker 5: of open source closed source model dichotomy that's happening in AI.

657
00:35:50,960 --> 00:35:53,480
Speaker 5: In fact, a few of the nonprofits that are kind

658
00:35:53,480 --> 00:35:56,920
Speaker 5: of government affiliated in China has released open source models

659
00:35:57,160 --> 00:35:59,880
Speaker 5: as well as a way to showcase what they have

660
00:36:00,200 --> 00:36:02,160
Speaker 5: done as a way to kind of both compete but

661
00:36:02,200 --> 00:36:05,200
Speaker 5: also to collaborate with academics and researchers elsewhere.

662
00:36:05,719 --> 00:36:08,520
Speaker 4: Yeah, you know, just coming a little more on the

663
00:36:08,560 --> 00:36:12,400
Speaker 4: government open source connection. You know, it's it's easier to

664
00:36:12,480 --> 00:36:15,319
Speaker 4: be like open source is a good strategy, particularly if

665
00:36:15,360 --> 00:36:18,040
Speaker 4: you're behind. We've seen a lot of investment on the

666
00:36:18,080 --> 00:36:21,279
Speaker 4: hardware side with Risk five as well as on as

667
00:36:21,280 --> 00:36:24,000
Speaker 4: well as as Kevin mentioned, on the software side. And

668
00:36:24,280 --> 00:36:26,400
Speaker 4: you know, I think it's it's still an open question,

669
00:36:26,520 --> 00:36:28,759
Speaker 4: but a lot of there's a lot of chatter on

670
00:36:28,800 --> 00:36:33,240
Speaker 4: the Chinese Internet about just to what extent the Chinese

671
00:36:33,239 --> 00:36:37,840
Speaker 4: model makers are dependent on Meta continuing to drop open

672
00:36:37,880 --> 00:36:41,279
Speaker 4: source models in the Lama series. And you know, we

673
00:36:41,320 --> 00:36:45,160
Speaker 4: may if and when Zuckerberg decides that he's done playing

674
00:36:45,200 --> 00:36:47,600
Speaker 4: the open source game, we may see a very different

675
00:36:47,600 --> 00:36:51,000
Speaker 4: trajectory for the Chinese model makers if they're not able

676
00:36:51,080 --> 00:36:54,400
Speaker 4: to kind of you know, get their algorithmic and hardware

677
00:36:54,400 --> 00:36:55,480
Speaker 4: shops in order before then.

678
00:36:56,280 --> 00:36:59,960
Speaker 2: So we just have a few minutes left. But you know,

679
00:37:00,080 --> 00:37:04,200
Speaker 2: we talked about the current state of US restrictions towards

680
00:37:04,280 --> 00:37:06,759
Speaker 2: China or how the US thinks about AI and the

681
00:37:06,760 --> 00:37:11,800
Speaker 2: geopolitical context. There are chip export controls, there's more being done,

682
00:37:11,960 --> 00:37:15,239
Speaker 2: you know, when it comes to cracking down an industrial espionage.

683
00:37:15,560 --> 00:37:18,200
Speaker 2: There have been a lot of stories, including several broken

684
00:37:18,200 --> 00:37:22,880
Speaker 2: by Bloomberg about you know, academic collaboration and you know,

685
00:37:23,080 --> 00:37:25,520
Speaker 2: trying limiting who can talk to whom. And we know

686
00:37:25,600 --> 00:37:28,080
Speaker 2: that there's a lot of issues in academia related to

687
00:37:28,080 --> 00:37:31,040
Speaker 2: computer science and tech and Chinese researchers in the US,

688
00:37:31,520 --> 00:37:34,680
Speaker 2: et cetera. We have this framework, but it might all

689
00:37:34,760 --> 00:37:37,919
Speaker 2: change because the president, the party of the presidency might

690
00:37:38,000 --> 00:37:41,440
Speaker 2: change in a few months. What do we know, if anything,

691
00:37:42,120 --> 00:37:46,400
Speaker 2: about the Trump view towards AI and AI in the

692
00:37:46,440 --> 00:37:47,520
Speaker 2: geopolitical context.

693
00:37:48,520 --> 00:37:52,320
Speaker 5: So I think the interesting about Trump is that during

694
00:37:52,360 --> 00:37:56,640
Speaker 5: his presidency, his first term, he actually did have worked

695
00:37:56,680 --> 00:37:59,520
Speaker 5: on national AI initiative of something which I believe is

696
00:37:59,520 --> 00:38:02,920
Speaker 5: also by Eric Schmid, the former CEO of Google, and

697
00:38:03,080 --> 00:38:05,680
Speaker 5: he signed on to some kind of a national initiative

698
00:38:05,680 --> 00:38:08,000
Speaker 5: plan in twenty twenty, but it was mid of COVID,

699
00:38:08,120 --> 00:38:09,960
Speaker 5: none of us had any time to pay attention that.

700
00:38:10,000 --> 00:38:12,520
Speaker 5: It's like didn't go anywhere. But he did have some

701
00:38:12,960 --> 00:38:16,479
Speaker 5: kind of history with AI. But based on the most

702
00:38:16,480 --> 00:38:21,640
Speaker 5: recent kind of the Republican National Conventions Platform just being released,

703
00:38:22,080 --> 00:38:24,640
Speaker 5: one of the lining items is that they will overturn

704
00:38:25,239 --> 00:38:29,360
Speaker 5: Biden's AI executive order, so Trump can put his own spin,

705
00:38:29,680 --> 00:38:32,440
Speaker 5: or his own branding, or his own substance on AI

706
00:38:32,560 --> 00:38:36,319
Speaker 5: when he if he does become a president again. So

707
00:38:36,360 --> 00:38:39,880
Speaker 5: we can expect a lot of kind of volatility on

708
00:38:39,920 --> 00:38:42,480
Speaker 5: that front. And another thing that I would say that

709
00:38:42,640 --> 00:38:44,799
Speaker 5: I think is the most consistent thing about Trump is

710
00:38:44,880 --> 00:38:48,480
Speaker 5: his unilateralism. Right He's always been about America first, American only,

711
00:38:48,840 --> 00:38:51,640
Speaker 5: and right now there is a pretty good rhythm of

712
00:38:51,719 --> 00:38:55,399
Speaker 5: global AI governance that's happening, starting with in the UK

713
00:38:55,480 --> 00:38:59,879
Speaker 5: a the Bletchley Park Declaration, to a summit in South

714
00:39:00,040 --> 00:39:03,440
Speaker 5: Korea earlier this year, to then another summit in Paris

715
00:39:03,640 --> 00:39:07,200
Speaker 5: in February of next year where you have world leaders

716
00:39:07,239 --> 00:39:10,920
Speaker 5: including China and all the other relevant departments and ministries

717
00:39:10,960 --> 00:39:14,640
Speaker 5: around the world coming together to talk about AI governance

718
00:39:14,800 --> 00:39:18,000
Speaker 5: in a global sense. If I were to put my

719
00:39:18,360 --> 00:39:20,680
Speaker 5: money in the betting market about Trump, I think he

720
00:39:20,719 --> 00:39:22,239
Speaker 5: will probably blow a lot of that out of the

721
00:39:22,280 --> 00:39:25,799
Speaker 5: water to go to his unilateral instinct and really put

722
00:39:25,840 --> 00:39:28,600
Speaker 5: America stamp on this, and it's hard to predict what

723
00:39:28,600 --> 00:39:29,279
Speaker 5: that will look like.

724
00:39:29,520 --> 00:39:32,600
Speaker 4: Trump went on the All In podcast, maybe the podcast

725
00:39:32,600 --> 00:39:33,680
Speaker 4: that must not be named one.

726
00:39:33,800 --> 00:39:34,759
Speaker 2: No, it's fun, but.

727
00:39:36,200 --> 00:39:36,480
Speaker 5: It was.

728
00:39:36,560 --> 00:39:38,919
Speaker 4: It was really remarkable because they asked him about China

729
00:39:38,960 --> 00:39:42,200
Speaker 4: and AI and he had, like, you know, a moderately

730
00:39:42,239 --> 00:39:46,799
Speaker 4: sophisticated answer, sort of shocking. He you know, identified that

731
00:39:46,840 --> 00:39:49,080
Speaker 4: it's a race between the US and China. And he

732
00:39:49,200 --> 00:39:53,000
Speaker 4: also said that energy is really important to data centers,

733
00:39:53,000 --> 00:39:54,960
Speaker 4: and we got to make that happen. You know, if

734
00:39:54,960 --> 00:39:57,920
Speaker 4: you edited down his comments from five minutes to like

735
00:39:58,000 --> 00:40:01,040
Speaker 4: thirty seconds, it was pretty coherent. But it was a

736
00:40:01,080 --> 00:40:04,080
Speaker 4: remarkable thing that, like he was talking to someone about

737
00:40:04,120 --> 00:40:07,120
Speaker 4: this and they got like two halfway decent talking points

738
00:40:07,160 --> 00:40:09,279
Speaker 4: in his head. So, you know, how he's going to

739
00:40:09,360 --> 00:40:12,160
Speaker 4: approach technology in China, as I think, is a really

740
00:40:12,239 --> 00:40:15,200
Speaker 4: open question. On the one hand, we had this very

741
00:40:15,719 --> 00:40:19,239
Speaker 4: like well thought out, sophisticated playbook that he ran to

742
00:40:19,320 --> 00:40:21,960
Speaker 4: basically get the world to turn on Huawei. On the

743
00:40:22,000 --> 00:40:23,640
Speaker 4: other hand, you know, we had a lot of weird,

744
00:40:23,680 --> 00:40:27,600
Speaker 4: trumpy stuff, like with Zte, the Commerce Department's basically going

745
00:40:27,680 --> 00:40:30,040
Speaker 4: to kill it. She called him, said that it was

746
00:40:30,080 --> 00:40:32,480
Speaker 4: going to cost too many Chinese jobs, and then he

747
00:40:32,520 --> 00:40:35,120
Speaker 4: went on and tweeted the next day saying too many

748
00:40:35,200 --> 00:40:37,560
Speaker 4: jobs lost in China, like we're going to give Zte

749
00:40:37,719 --> 00:40:40,200
Speaker 4: a reprieve. So he's really of two minds on all

750
00:40:40,239 --> 00:40:42,880
Speaker 4: this stuff. Maybe someone talks him into being scared about inflation,

751
00:40:43,400 --> 00:40:45,919
Speaker 4: who knows, But I think there's a very wide range

752
00:40:45,960 --> 00:40:48,960
Speaker 4: of outcomes both on the sort of his domestic approach

753
00:40:48,960 --> 00:40:51,480
Speaker 4: to artificial intelligence as well as he's as well as

754
00:40:51,480 --> 00:40:54,080
Speaker 4: how he's going to think about technology competition in China.

755
00:40:54,239 --> 00:40:56,920
Speaker 2: You said something there that actually I wanted to ask,

756
00:40:57,040 --> 00:40:59,759
Speaker 2: and this is just about the sort of technological arms

757
00:40:59,800 --> 00:41:03,799
Speaker 2: rate again, but obviously within the within the US and

758
00:41:03,840 --> 00:41:06,920
Speaker 2: the data centers, there's a lot of anxiety about access

759
00:41:06,960 --> 00:41:09,279
Speaker 2: to electricity and how are we going to supply that?

760
00:41:09,440 --> 00:41:12,120
Speaker 2: And we all know the US is not in a

761
00:41:12,120 --> 00:41:14,680
Speaker 2: good state these days, are like building things and adding

762
00:41:14,719 --> 00:41:17,759
Speaker 2: a lot of energy capacity. Is that a when you

763
00:41:17,800 --> 00:41:20,880
Speaker 2: think about US versus China in that respect? Could that

764
00:41:21,000 --> 00:41:24,560
Speaker 2: be a comparative advantage for the Chinese companies? I mean,

765
00:41:24,600 --> 00:41:28,279
Speaker 2: I know China continues to add more nuclear generation and

766
00:41:28,280 --> 00:41:30,920
Speaker 2: they're probably more liberal in terms of adding more coal

767
00:41:30,960 --> 00:41:34,239
Speaker 2: power and other things like that. Could they end up

768
00:41:34,280 --> 00:41:37,760
Speaker 2: within an advantage based on their electricity grid on solving

769
00:41:37,800 --> 00:41:38,879
Speaker 2: the power side of AI.

770
00:41:39,120 --> 00:41:41,440
Speaker 4: You know, this might be a bit of a red herring.

771
00:41:41,560 --> 00:41:45,319
Speaker 4: I am like cautiously optimistic that Republicans and Democrats will

772
00:41:45,400 --> 00:41:48,640
Speaker 4: end up kind of giving some you know, NIPA exceptions

773
00:41:48,680 --> 00:41:51,720
Speaker 4: for these data centers, which, like everyone is now agreeing

774
00:41:51,880 --> 00:41:55,239
Speaker 4: is this like incredibly important national strategic resource. I mean,

775
00:41:55,560 --> 00:41:57,959
Speaker 4: I think the bookcase with China, right is if they're

776
00:41:58,000 --> 00:42:00,399
Speaker 4: stuck with less efficient ships and just how to build

777
00:42:00,440 --> 00:42:03,440
Speaker 4: data centers twice the size that are twice as expensive,

778
00:42:03,440 --> 00:42:05,239
Speaker 4: then like, yeah, maybe they can like throw on a

779
00:42:05,280 --> 00:42:08,440
Speaker 4: few more coal plants and nuclear nuclear reactors to get

780
00:42:08,480 --> 00:42:13,399
Speaker 4: them to get them going. But I am cautiously optimistic

781
00:42:13,480 --> 00:42:17,640
Speaker 4: that in video will continue to make increasingly power efficient

782
00:42:17,719 --> 00:42:20,960
Speaker 4: chips and you know, state, local and national regulators will

783
00:42:21,040 --> 00:42:24,640
Speaker 4: understand that this is something worth bending the rules on,

784
00:42:24,800 --> 00:42:27,040
Speaker 4: which is sort of what you've seen on the sort

785
00:42:27,040 --> 00:42:29,799
Speaker 4: of semiconductor fabrication build out over the past few years

786
00:42:29,840 --> 00:42:30,440
Speaker 4: with the Chips Act.

787
00:42:30,480 --> 00:42:32,200
Speaker 5: And if I can add one more points to what

788
00:42:32,280 --> 00:42:34,480
Speaker 5: you just said, Jordan, I think a lot of the

789
00:42:34,520 --> 00:42:37,400
Speaker 5: focus that we've talked about has been the algorithm, the

790
00:42:37,520 --> 00:42:40,200
Speaker 5: model how power for the GPU is. But a lot

791
00:42:40,239 --> 00:42:43,799
Speaker 5: of engineering sources has been devoted and will continue to

792
00:42:43,840 --> 00:42:48,520
Speaker 5: be to reduce the power hungriness of AI model because

793
00:42:48,560 --> 00:42:51,120
Speaker 5: one of the big things that's a hugely hotly debated

794
00:42:51,160 --> 00:42:54,120
Speaker 5: topic is that all these tokens that we're generating for

795
00:42:54,160 --> 00:42:56,440
Speaker 5: all these apps are just way too expensive, right, the

796
00:42:56,480 --> 00:42:59,040
Speaker 5: expense comes from the compute and then the power that's

797
00:42:59,040 --> 00:43:01,799
Speaker 5: being consumed to it. So there will be a lot

798
00:43:01,880 --> 00:43:05,279
Speaker 5: of engineering that is going to accomplish when it comes

799
00:43:05,280 --> 00:43:08,160
Speaker 5: to reducing the costs and the need for energy over time.

800
00:43:08,520 --> 00:43:11,359
Speaker 5: So I don't think the energy trade is a long

801
00:43:11,440 --> 00:43:14,919
Speaker 5: term investment thesis, but maybe a short term arbitrage as

802
00:43:14,960 --> 00:43:17,839
Speaker 5: far as how much extra power we needed just to

803
00:43:17,880 --> 00:43:20,720
Speaker 5: feed the hungry AI Denta center when all the engineering

804
00:43:20,760 --> 00:43:23,839
Speaker 5: is go into reducing that dependence, because everyone sees how

805
00:43:23,880 --> 00:43:25,560
Speaker 5: that could be detrimental in a lot of ways.

806
00:43:25,760 --> 00:43:28,279
Speaker 2: Jordan and Kevin, thank you so much for coming on

807
00:43:28,280 --> 00:43:31,080
Speaker 2: a latch. That was a fantastic our goal for the

808
00:43:31,080 --> 00:43:33,920
Speaker 2: episode of Understanding this more I feel I do. I

809
00:43:33,960 --> 00:43:36,480
Speaker 2: do so thank you both so much for coming on that.

810
00:43:36,719 --> 00:43:48,399
Speaker 2: Happy Tracy, I thought that was great. I thought, I mean,

811
00:43:48,480 --> 00:43:50,799
Speaker 2: there were just a bunch of things that like, I,

812
00:43:50,840 --> 00:43:53,279
Speaker 2: you know, even setting aside the geopolitical aspects that I

813
00:43:53,360 --> 00:43:57,480
Speaker 2: just didn't know about what the AI you know, space

814
00:43:57,520 --> 00:44:01,160
Speaker 2: looks like in China, and if that was very helpful, right, I'm.

815
00:44:01,000 --> 00:44:03,919
Speaker 3: Sort of obsessed with the programmer's day now. I want

816
00:44:03,920 --> 00:44:06,160
Speaker 3: to learn more about the holiday. Do you think all

817
00:44:06,200 --> 00:44:08,880
Speaker 3: the open source programmers they like all get together to

818
00:44:08,920 --> 00:44:11,760
Speaker 3: celebrate and do a dance, maybe a byte dance.

819
00:44:11,960 --> 00:44:16,359
Speaker 2: Yeah, oh boom, come on, okay, Tracy, you get We'll

820
00:44:16,440 --> 00:44:17,200
Speaker 2: leave that one in.

821
00:44:17,239 --> 00:44:20,920
Speaker 3: Thank you, so the producers will add that cricket noise

822
00:44:21,040 --> 00:44:22,040
Speaker 3: that they have on hand.

823
00:44:23,560 --> 00:44:26,560
Speaker 2: I am also interested in the holiday. I do think

824
00:44:26,800 --> 00:44:30,120
Speaker 2: so Kevin's first answer, I just thought it was like

825
00:44:30,280 --> 00:44:35,400
Speaker 2: really clarifying about why a country might think that having

826
00:44:35,440 --> 00:44:39,120
Speaker 2: a national version is important right now, because if and

827
00:44:39,160 --> 00:44:41,920
Speaker 2: I hadn't really thought about it in those terms, because

828
00:44:41,960 --> 00:44:44,359
Speaker 2: on some level, right, like it's easy enough, like if

829
00:44:44,400 --> 00:44:47,680
Speaker 2: it's you know, Lama, you can use that anywhere on

830
00:44:47,920 --> 00:44:51,000
Speaker 2: you know, various chips and stuff. But the idea that

831
00:44:51,080 --> 00:44:54,959
Speaker 2: like the AI is going to like codify truth as

832
00:44:55,000 --> 00:44:58,560
Speaker 2: it means in the context of that country, and you know,

833
00:44:58,640 --> 00:45:02,080
Speaker 2: of China their version of what's true. Kevin gave the

834
00:45:02,120 --> 00:45:03,960
Speaker 2: Jackmaw example, which is that one day he may be

835
00:45:04,000 --> 00:45:05,600
Speaker 2: a hero and one day he may have a villain

836
00:45:06,040 --> 00:45:08,560
Speaker 2: or maybe a villain, like that's always changing, and so

837
00:45:08,600 --> 00:45:11,239
Speaker 2: there may be some impulse to have the model sort

838
00:45:11,280 --> 00:45:14,920
Speaker 2: of correctly reflect what is state policy. But this idea

839
00:45:14,960 --> 00:45:18,680
Speaker 2: that like it captures you know, creates this version of

840
00:45:18,719 --> 00:45:21,480
Speaker 2: the national identity, culture and language is very interesting.

841
00:45:21,680 --> 00:45:24,360
Speaker 3: No, absolutely, and I think China is sort of the

842
00:45:24,680 --> 00:45:27,000
Speaker 3: extreme example of that. And that's sort of what I

843
00:45:27,040 --> 00:45:30,440
Speaker 3: was alluding to, not very well in the intro when

844
00:45:30,480 --> 00:45:32,600
Speaker 3: I was talking about data control, but like, there is

845
00:45:33,000 --> 00:45:37,240
Speaker 3: this cultural element to AI where you are using something

846
00:45:37,280 --> 00:45:40,360
Speaker 3: like chat GPT, or maybe not at the moment, but

847
00:45:40,400 --> 00:45:43,200
Speaker 3: you could imagine a future where you're using something like

848
00:45:43,280 --> 00:45:45,600
Speaker 3: chat GPT in the same way that you're using something

849
00:45:45,640 --> 00:45:49,920
Speaker 3: like Wikipedia, where it's almost like the codified public version

850
00:45:50,120 --> 00:45:52,640
Speaker 3: of events, right, And so if you are a government

851
00:45:52,760 --> 00:45:57,440
Speaker 3: like China, why wouldn't you want to have some degree

852
00:45:57,480 --> 00:46:00,800
Speaker 3: of control over what that's spitting out and the narrative

853
00:46:00,960 --> 00:46:01,800
Speaker 3: that it's building.

854
00:46:02,239 --> 00:46:04,239
Speaker 2: I was also really intrigued that, you know, we talk

855
00:46:04,280 --> 00:46:06,960
Speaker 2: about chip wars all the time, but the idea that

856
00:46:06,960 --> 00:46:09,640
Speaker 2: there is a distinction that is useful to make between

857
00:46:09,680 --> 00:46:12,480
Speaker 2: a chip war and a cloud war, and that okay,

858
00:46:12,520 --> 00:46:14,320
Speaker 2: there is you know, we all know that, like it

859
00:46:14,360 --> 00:46:18,160
Speaker 2: was supposedly you know, smaller nodes or whatever are better,

860
00:46:18,520 --> 00:46:21,280
Speaker 2: although I think in the context of GPUs that's less important,

861
00:46:21,360 --> 00:46:24,600
Speaker 2: but that so much of the technology is about the interconnect,

862
00:46:24,640 --> 00:46:26,680
Speaker 2: and we talked about this with Corewave, that there is

863
00:46:26,719 --> 00:46:31,640
Speaker 2: a distinct element to building up a very powerful GPU

864
00:46:31,719 --> 00:46:35,680
Speaker 2: cloud versus a traditional cloud because because of the cables

865
00:46:35,719 --> 00:46:38,360
Speaker 2: and the packaging, et cetera, and so the idea that

866
00:46:38,400 --> 00:46:42,440
Speaker 2: it's more about more than just like the chip itself

867
00:46:42,920 --> 00:46:45,840
Speaker 2: and you know, the size of that chip so to speak.

868
00:46:45,960 --> 00:46:48,719
Speaker 3: No, absolutely, And one thing that really stood out from

869
00:46:48,719 --> 00:46:51,680
Speaker 3: that conversation, and I'm thinking back, this actually happened when

870
00:46:51,719 --> 00:46:53,680
Speaker 3: I was in Hong Kong and I was, you know,

871
00:46:53,719 --> 00:46:56,680
Speaker 3: coordinating some of the news coverage there. But I remember

872
00:46:57,040 --> 00:47:00,760
Speaker 3: the big crackdown on the sort of like retail internet

873
00:47:00,880 --> 00:47:03,760
Speaker 3: companies like Ali Baba, and I remember at the time

874
00:47:04,320 --> 00:47:09,120
Speaker 3: policymakers were very very specific and very deliberate and very

875
00:47:09,160 --> 00:47:12,839
Speaker 3: open about trying to direct more capital, you know, less

876
00:47:12,880 --> 00:47:17,799
Speaker 3: disorderly capital flowing into SaaS, more capital going into hardware

877
00:47:17,920 --> 00:47:21,040
Speaker 3: and making things like chips. It was almost, you know,

878
00:47:21,200 --> 00:47:23,360
Speaker 3: for lack of a better tech analogy, it was like

879
00:47:23,400 --> 00:47:25,879
Speaker 3: flipping a switch, right. They basically said, like, no more

880
00:47:25,920 --> 00:47:30,480
Speaker 3: money going into retail tech, everything going into semiconductor manufacturing

881
00:47:30,480 --> 00:47:32,920
Speaker 3: and all that stuff. So it was it was interesting

882
00:47:32,960 --> 00:47:37,080
Speaker 3: to hear like both the challenges and the opportunities from

883
00:47:37,080 --> 00:47:38,400
Speaker 3: that perspective in China.

884
00:47:38,480 --> 00:47:40,200
Speaker 2: By the way, I don't know if it was The

885
00:47:40,239 --> 00:47:43,680
Speaker 2: Financial Times, but that was the first whoever came up

886
00:47:43,719 --> 00:47:46,920
Speaker 2: with chech shept. That was very clever. That was very

887
00:47:46,920 --> 00:47:50,880
Speaker 2: well done. I first saw that in that that word

888
00:47:51,080 --> 00:47:53,759
Speaker 2: that in the ft. But that's a pretty clever term.

889
00:47:53,800 --> 00:47:54,480
Speaker 2: That was a good one.

890
00:47:54,560 --> 00:47:57,839
Speaker 3: I'm disappointed to find out it's actually vaporware, although no,

891
00:47:57,880 --> 00:48:01,120
Speaker 3: I know, perhaps not surprised. Okay, shall we leave it there.

892
00:48:01,239 --> 00:48:01,919
Speaker 2: Let's leave it there.

893
00:48:01,960 --> 00:48:04,640
Speaker 3: This has been another episode of the auth Lots podcast.

894
00:48:04,719 --> 00:48:07,560
Speaker 3: I'm Tracy Alloway. You can follow me at Tracy Alloway.

895
00:48:07,800 --> 00:48:11,000
Speaker 2: And I'm Joe Wisenthal. You can follow me at The Stalwart.

896
00:48:11,320 --> 00:48:15,080
Speaker 2: Follow our guests Jordan Schneider He's at Jordan sc h

897
00:48:15,320 --> 00:48:18,520
Speaker 2: NYC and check out his China Talk newsletter and podcast.

898
00:48:18,920 --> 00:48:22,879
Speaker 2: And Kevin Schoue He's at Kevin Sshoue and check out

899
00:48:22,880 --> 00:48:27,200
Speaker 2: his writing at Interconnected and follow our producers Carmen Rodriguez

900
00:48:27,239 --> 00:48:30,040
Speaker 2: at Carmen armand dash Ol Bennett at Dashbot and kel

901
00:48:30,080 --> 00:48:32,840
Speaker 2: Brooks at kel Brooks. Thank you to our producer Moses

902
00:48:32,880 --> 00:48:35,719
Speaker 2: Ondam and from our Oddlots content go to Bloomberg dot

903
00:48:35,719 --> 00:48:38,080
Speaker 2: com slash od lots where you have transcripts, a blog,

904
00:48:38,120 --> 00:48:40,239
Speaker 2: and a newsletter and you can chat about all of

905
00:48:40,280 --> 00:48:44,280
Speaker 2: these topics twenty four to seven in our discord discord

906
00:48:44,320 --> 00:48:46,719
Speaker 2: dot gg slash odd lots, we have an AI room

907
00:48:46,760 --> 00:48:48,879
Speaker 2: in there. Maybe when this comes out, we'll see if

908
00:48:49,520 --> 00:48:51,880
Speaker 2: either Kevin or Jordan or both would be up for

909
00:48:51,960 --> 00:48:54,080
Speaker 2: doing an AMA in there. The day it comes out.

910
00:48:54,120 --> 00:48:56,160
Speaker 2: I'm sure people have a lot of questions, so we'll

911
00:48:56,160 --> 00:48:57,880
Speaker 2: try to make that. Try to make that happen.

912
00:48:57,960 --> 00:49:00,960
Speaker 3: And if you enjoy authlots, if if you want us to,

913
00:49:01,239 --> 00:49:05,000
Speaker 3: I don't know, start a pop culture semiconductor spin off show,

914
00:49:05,239 --> 00:49:08,200
Speaker 3: then please leave us a positive review on your favorite

915
00:49:08,200 --> 00:49:11,879
Speaker 3: podcast platform. And remember, if you are a Bloomberg subscriber,

916
00:49:11,920 --> 00:49:14,839
Speaker 3: you can listen to all of our episodes absolutely ad free.

917
00:49:14,960 --> 00:49:17,160
Speaker 3: All you need to do is connect your Bloomberg account

918
00:49:17,280 --> 00:49:19,960
Speaker 3: with Apple Podcasts. In order to do that, just find

919
00:49:20,000 --> 00:49:23,799
Speaker 3: the Bloomberg channel on Apple Podcasts and follow the instructions there.

920
00:49:24,080 --> 00:49:24,880
Speaker 3: Thanks for listening.