1
00:00:02,720 --> 00:00:14,000
Speaker 1: Bloomberg Audio Studios, Podcasts, Radio News.

2
00:00:17,960 --> 00:00:20,680
Speaker 2: Hello and welcome to another episode of the Odd Thoughts podcast.

3
00:00:20,760 --> 00:00:22,400
Speaker 3: I'm Tracy Alloway and I'm Joe.

4
00:00:22,440 --> 00:00:23,439
Speaker 4: Why isn't thal Joe.

5
00:00:23,440 --> 00:00:26,280
Speaker 2: We like to talk a lot about physical constraints, yeah,

6
00:00:26,360 --> 00:00:29,120
Speaker 2: on this show, right, And this is one reason why

7
00:00:29,160 --> 00:00:31,920
Speaker 2: AI is a really fascinating area for us right now,

8
00:00:31,920 --> 00:00:34,879
Speaker 2: because there are a lot of physical constraints on what

9
00:00:35,040 --> 00:00:39,000
Speaker 2: is ultimately the sort of ephemeral technology. And I think

10
00:00:39,080 --> 00:00:42,519
Speaker 2: that the tension between those two things is really interesting, Right,

11
00:00:42,680 --> 00:00:45,839
Speaker 2: Like you type a prompt into chat, GPT or claude

12
00:00:45,920 --> 00:00:49,560
Speaker 2: or whatever, and it's the sort of like disembodied digital platform.

13
00:00:50,080 --> 00:00:54,560
Speaker 2: You don't necessarily think about the power usage, the real resources,

14
00:00:54,600 --> 00:00:57,480
Speaker 2: the transformers that have to go into data centers to

15
00:00:57,520 --> 00:00:58,200
Speaker 2: get compute.

16
00:00:58,480 --> 00:01:00,680
Speaker 4: The thing that I've been on my mind and lately

17
00:01:00,760 --> 00:01:02,720
Speaker 4: and I've written about it and I plan to write more,

18
00:01:02,960 --> 00:01:07,120
Speaker 4: is this idea that the canonical AI thought experiment is

19
00:01:07,440 --> 00:01:09,600
Speaker 4: what happens if you tell an AI to make a

20
00:01:09,640 --> 00:01:12,360
Speaker 4: lot of paper clips, and then it destroys the world

21
00:01:12,400 --> 00:01:16,479
Speaker 4: because in the pursuit of marshaling all of the world's resources,

22
00:01:16,520 --> 00:01:19,679
Speaker 4: it just turns everything into paper clips. Because it doesn't know, and.

23
00:01:20,040 --> 00:01:22,319
Speaker 2: I have to ask, is this canonical example? Is this

24
00:01:22,360 --> 00:01:24,759
Speaker 2: based on your traumatic fear of Clippy?

25
00:01:26,600 --> 00:01:28,640
Speaker 4: No, but that is you know, it all comes back

26
00:01:28,680 --> 00:01:32,080
Speaker 4: full circle. But what we're seeing in real life is

27
00:01:32,080 --> 00:01:35,920
Speaker 4: that everything from access to electrical grids GPU is being

28
00:01:36,120 --> 00:01:41,440
Speaker 4: the big example, energy, turbines, talent, and now even including

29
00:01:41,440 --> 00:01:44,480
Speaker 4: residential real estate are being reproposed to make more and

30
00:01:44,560 --> 00:01:48,040
Speaker 4: more advanced AI. And in the original paper clip thought experiment,

31
00:01:48,640 --> 00:01:51,920
Speaker 4: the envision, or at least in one version, the philosopher

32
00:01:52,000 --> 00:01:55,160
Speaker 4: Nicolas Bostrom, envisions the AI having exhausted all of the

33
00:01:55,200 --> 00:01:58,840
Speaker 4: world's resources, then sending a probe into outer space to

34
00:01:58,920 --> 00:02:00,880
Speaker 4: consume star inner to build.

35
00:02:00,600 --> 00:02:02,240
Speaker 2: More paper clips eat the universe.

36
00:02:02,400 --> 00:02:05,520
Speaker 4: And to this point we're even talking about going into

37
00:02:05,560 --> 00:02:08,480
Speaker 4: outer space for data centers to build more AI. So

38
00:02:08,600 --> 00:02:12,120
Speaker 4: every version of the thought experiment is being replicated, except

39
00:02:12,160 --> 00:02:14,959
Speaker 4: to just more and more resources to build the AI

40
00:02:15,200 --> 00:02:17,840
Speaker 4: by humans rather than paper clips by AI.

41
00:02:18,160 --> 00:02:22,000
Speaker 2: There's this other connected theme here. So we've talked before

42
00:02:22,040 --> 00:02:24,560
Speaker 2: about how one of the reasons valuations seem to be

43
00:02:24,600 --> 00:02:27,880
Speaker 2: getting insane in the market is because all of this

44
00:02:28,000 --> 00:02:31,960
Speaker 2: activity is being driven by like this existential need to

45
00:02:32,360 --> 00:02:36,519
Speaker 2: become number one in frontier models and this new technology.

46
00:02:36,600 --> 00:02:37,600
Speaker 3: And so if you.

47
00:02:37,600 --> 00:02:40,840
Speaker 2: Say you absolutely have to be the first to invent AGI,

48
00:02:40,960 --> 00:02:44,839
Speaker 2: then you can justify any amount of spending on Earth, right,

49
00:02:44,880 --> 00:02:46,959
Speaker 2: And so what we tend to see is like the

50
00:02:47,000 --> 00:02:49,800
Speaker 2: biggest companies just keep getting bigger. They're the ones that

51
00:02:49,840 --> 00:02:51,480
Speaker 2: can get resources for all this stuff.

52
00:02:51,520 --> 00:02:53,720
Speaker 4: And I think one of the most fascinating things right

53
00:02:53,800 --> 00:02:56,040
Speaker 4: now is that, at least as of right now June fourth,

54
00:02:56,040 --> 00:03:00,120
Speaker 4: twenty twenty six, the frontier models are really close to

55
00:03:00,160 --> 00:03:02,760
Speaker 4: each other, right, Yeah, so and the four point a

56
00:03:03,000 --> 00:03:06,760
Speaker 4: GPT five point five, like they're not that different. And

57
00:03:06,840 --> 00:03:09,600
Speaker 4: one of the things I'm curious about is is there

58
00:03:09,680 --> 00:03:13,160
Speaker 4: something inherent in market dynamics in this space that will

59
00:03:13,200 --> 00:03:16,200
Speaker 4: always keep you know, whether it's not being able to

60
00:03:16,320 --> 00:03:20,400
Speaker 4: distill results from another model and cause I steal them,

61
00:03:20,400 --> 00:03:25,519
Speaker 4: whether it's information sharing among employees. Is there some inherent

62
00:03:25,639 --> 00:03:28,720
Speaker 4: reason why we've seen the stability or could it be

63
00:03:28,800 --> 00:03:31,359
Speaker 4: that at some point one lab just like breaks out

64
00:03:31,480 --> 00:03:34,600
Speaker 4: and establishes permanent still possibility.

65
00:03:34,639 --> 00:03:37,440
Speaker 2: But like I am personally on the side of commodification,

66
00:03:37,600 --> 00:03:41,680
Speaker 2: and everything just becomes kind of basic or whatever basically available.

67
00:03:41,760 --> 00:03:43,640
Speaker 2: I know Okay, well I'm just kidding, all right, thank

68
00:03:43,680 --> 00:03:46,520
Speaker 2: you Joe as a Joe. All right, that's a polite

69
00:03:46,560 --> 00:03:48,720
Speaker 2: prompt to get to the guest. We do, in fact

70
00:03:48,840 --> 00:03:51,480
Speaker 2: have the perfect guest. We're going to be speaking with,

71
00:03:51,800 --> 00:03:55,080
Speaker 2: Anjinay Mitta. He is, of course, a former general partner

72
00:03:55,120 --> 00:03:59,720
Speaker 2: at Anderson Horowitz, a Stanford University visiting scientist who teaches

73
00:03:59,760 --> 00:04:04,240
Speaker 2: the viral AI lecture called Frontier Systems. Also one of

74
00:04:04,240 --> 00:04:07,560
Speaker 2: the first guys to write a check for Anthropic is

75
00:04:07,680 --> 00:04:11,560
Speaker 2: now the founder of a new company called AMPPBC. So

76
00:04:11,880 --> 00:04:14,440
Speaker 2: thank you so much for coming on our thoughts on Thanks.

77
00:04:14,200 --> 00:04:17,039
Speaker 5: For having me. One correction, it's pronounced amp pbc, but

78
00:04:17,120 --> 00:04:18,960
Speaker 5: that's everything else you got corect on the.

79
00:04:18,960 --> 00:04:22,360
Speaker 3: Intro would make sense, wouldn't it as an energy Yeah.

80
00:04:22,600 --> 00:04:25,640
Speaker 2: That, and just remind us the PBC is public benefit

81
00:04:25,680 --> 00:04:28,039
Speaker 2: corporation's right, So you're doing this for the public benefit.

82
00:04:28,120 --> 00:04:30,839
Speaker 5: We're governed by a public benefit charter, which means everything

83
00:04:30,839 --> 00:04:33,719
Speaker 5: we do has to follow our mission. We have a

84
00:04:33,720 --> 00:04:36,560
Speaker 5: public charter mission. We are for profit in the same

85
00:04:36,760 --> 00:04:39,359
Speaker 5: way Ben and Jerry's or OREI and then Tropic our

86
00:04:39,400 --> 00:04:42,599
Speaker 5: public benefits. So we aim to make a healthy, modest

87
00:04:42,720 --> 00:04:45,280
Speaker 5: amount of profits that can sustain our mission. But we

88
00:04:45,640 --> 00:04:48,359
Speaker 5: are we have the flexibility to choose what that margin is.

89
00:04:48,800 --> 00:04:49,479
Speaker 3: Can I just start.

90
00:04:49,520 --> 00:04:51,680
Speaker 2: I want to establish your credentials. Although I feel like

91
00:04:51,720 --> 00:04:53,760
Speaker 2: that very long list did a pretty good job. But

92
00:04:53,839 --> 00:04:56,600
Speaker 2: writing the first check for Anthropic like tell us that

93
00:04:56,680 --> 00:05:00,400
Speaker 2: kind of origin story because the anecdote that you hear

94
00:05:00,600 --> 00:05:04,400
Speaker 2: is like twenty five vcs turned them away initially and

95
00:05:04,440 --> 00:05:05,080
Speaker 2: you said yes.

96
00:05:05,480 --> 00:05:07,400
Speaker 5: It was a little bit of the other way around.

97
00:05:07,440 --> 00:05:09,559
Speaker 5: I said yes. Then we tried to get another twenty

98
00:05:09,560 --> 00:05:12,640
Speaker 5: five dcs to say us and I failed. It was

99
00:05:12,680 --> 00:05:14,320
Speaker 5: a harrowing experience. It was a bit of a wake

100
00:05:14,400 --> 00:05:17,640
Speaker 5: up call. It was late twenty twenty. I had just

101
00:05:17,760 --> 00:05:19,880
Speaker 5: sold my last business. It was called Ubiquity six. It

102
00:05:19,960 --> 00:05:22,359
Speaker 5: was a three D mapping business. An AI is an

103
00:05:22,360 --> 00:05:26,120
Speaker 5: AI business that we had founded in twenty seventeen. And

104
00:05:27,040 --> 00:05:29,880
Speaker 5: I felt like a failure at the time because you know,

105
00:05:29,880 --> 00:05:31,960
Speaker 5: in San Francis. I was in San Francisco. I just

106
00:05:32,000 --> 00:05:35,680
Speaker 5: as a big picture my life stories. I was born India.

107
00:05:36,720 --> 00:05:38,760
Speaker 5: I went to high school in Singapore, and I came

108
00:05:38,800 --> 00:05:40,960
Speaker 5: over to college to the United States at Stanford for

109
00:05:41,080 --> 00:05:43,960
Speaker 5: my undergraduate degree. And then when I arrived at campus

110
00:05:44,080 --> 00:05:47,080
Speaker 5: in twenty eleven. Deep learning had just started taking over

111
00:05:47,120 --> 00:05:50,560
Speaker 5: the world in Silicon Valley. Andre Karpathy was a computer

112
00:05:50,600 --> 00:05:53,479
Speaker 5: science ta to Andrew Ing, who was one of the

113
00:05:54,080 --> 00:05:57,880
Speaker 5: i would say modern sort of founding fathers of deep learning.

114
00:05:57,880 --> 00:06:01,880
Speaker 5: This idea that you can teach machines to think without

115
00:06:01,920 --> 00:06:04,720
Speaker 5: having to give them prescriptive rules, and so I went

116
00:06:04,760 --> 00:06:06,960
Speaker 5: into sort of machine I got swept up in that

117
00:06:07,400 --> 00:06:09,760
Speaker 5: moment and started studying. A lot of my coursework was

118
00:06:09,760 --> 00:06:14,240
Speaker 5: in machine learning. My primary department at Stanford was in bioinformatics,

119
00:06:14,240 --> 00:06:17,600
Speaker 5: which was machine learning applied to healthcare. I got sidetracked

120
00:06:17,640 --> 00:06:20,480
Speaker 5: to a venture firm called Clina Perkins for about four

121
00:06:20,560 --> 00:06:22,159
Speaker 5: and a half years, where I got the chance to

122
00:06:22,160 --> 00:06:24,680
Speaker 5: work for some of the great investors like John Dore

123
00:06:24,720 --> 00:06:26,599
Speaker 5: and Mary Meeker. And then I left and started my

124
00:06:26,600 --> 00:06:29,360
Speaker 5: own company. And as is the case in Silicon Valley,

125
00:06:29,360 --> 00:06:32,000
Speaker 5: when you start, I mean I was twenty five. I

126
00:06:32,040 --> 00:06:34,840
Speaker 5: went and raised about forty seven or two million dollars

127
00:06:34,920 --> 00:06:37,520
Speaker 5: from some of the usual suspects like Benchmark and Index

128
00:06:37,560 --> 00:06:39,280
Speaker 5: and so on. I thought I was the coolest kid

129
00:06:39,320 --> 00:06:42,440
Speaker 5: in town, and I got the bit out of me

130
00:06:43,680 --> 00:06:48,200
Speaker 5: because we built this incredible technology, which is this AI

131
00:06:48,279 --> 00:06:51,159
Speaker 5: system that could map any location in three D. And

132
00:06:51,200 --> 00:06:54,760
Speaker 5: then the pandemic hit, and so location based mapping, three

133
00:06:54,839 --> 00:06:59,480
Speaker 5: D mapping. The only thing you can control is how

134
00:06:59,520 --> 00:07:02,240
Speaker 5: you react to what happens. And so I did feel

135
00:07:02,240 --> 00:07:03,920
Speaker 5: for a moment like it was bad luck. And then

136
00:07:03,920 --> 00:07:05,160
Speaker 5: you just have to pick up the pieces and make

137
00:07:05,240 --> 00:07:08,080
Speaker 5: the best of it. So I did with my co founder,

138
00:07:08,120 --> 00:07:09,880
Speaker 5: we we figured it out. It was it was a

139
00:07:09,920 --> 00:07:13,560
Speaker 5: tough few years where we had to pivot the business,

140
00:07:14,160 --> 00:07:16,720
Speaker 5: but we landed the plane. We essentially a lot of

141
00:07:16,760 --> 00:07:19,280
Speaker 5: the distributed systems we'd built on the back end side

142
00:07:19,360 --> 00:07:21,040
Speaker 5: ended up being quite valuable. We sold that to a

143
00:07:21,040 --> 00:07:23,520
Speaker 5: company called Discord, which is a chat up for gamers.

144
00:07:24,520 --> 00:07:28,880
Speaker 4: Lost Discord fans in there. Yeah, awesome our listeners.

145
00:07:29,640 --> 00:07:31,480
Speaker 5: So, you know, about a month after I sold the business,

146
00:07:31,520 --> 00:07:33,280
Speaker 5: I got a call from some friends. We're running research

147
00:07:33,280 --> 00:07:34,840
Speaker 5: at Opening Eye, and we'd all been you know, friends

148
00:07:34,880 --> 00:07:37,840
Speaker 5: in the machine learning community in the area, and they said,

149
00:07:37,880 --> 00:07:40,120
Speaker 5: on you know, we've trained a little model called GBT

150
00:07:40,200 --> 00:07:42,680
Speaker 5: three and we think it's the best sense just a

151
00:07:42,720 --> 00:07:45,920
Speaker 5: little model. Yeah, nobody really paid attention. They were like, no,

152
00:07:45,920 --> 00:07:47,600
Speaker 5: nobody cares, but we think it's the best thing since

153
00:07:47,600 --> 00:07:49,760
Speaker 5: sliced bread, and we want to leave and turn this

154
00:07:50,240 --> 00:07:53,360
Speaker 5: into a standalone business. But you know, it'd be helpful

155
00:07:53,440 --> 00:07:55,120
Speaker 5: to get some of your advice on how to do that.

156
00:07:55,840 --> 00:07:57,600
Speaker 5: And I couldn't really come on board full time at

157
00:07:57,600 --> 00:08:01,000
Speaker 5: the time with them because I had to integrate my

158
00:08:01,040 --> 00:08:03,440
Speaker 5: company to the acquire But I came on as their angel,

159
00:08:03,440 --> 00:08:05,440
Speaker 5: and nights and weekends I worked with them on the

160
00:08:05,480 --> 00:08:07,680
Speaker 5: business plan and who we should raise from that. You know,

161
00:08:07,680 --> 00:08:10,160
Speaker 5: that company was anthropic. Dario and Tom and I started

162
00:08:10,160 --> 00:08:12,760
Speaker 5: doing these weekly working sessions in early twenty twenty one,

163
00:08:13,760 --> 00:08:18,840
Speaker 5: and yeah, I assumed that, you know, if we went

164
00:08:18,880 --> 00:08:21,000
Speaker 5: and talked to a bunch of venture capitalists on Central Road,

165
00:08:21,080 --> 00:08:23,520
Speaker 5: especially some of the ones who were involved in the

166
00:08:23,520 --> 00:08:26,720
Speaker 5: biggest hits of the last decade before, that they would

167
00:08:26,760 --> 00:08:28,800
Speaker 5: get it. These are the creators of GPD three. And

168
00:08:28,840 --> 00:08:30,920
Speaker 5: they were like, we just don't get this. We've heard

169
00:08:30,960 --> 00:08:33,719
Speaker 5: the Holy I story before. This whole general Intelligence thing

170
00:08:33,800 --> 00:08:38,200
Speaker 5: is a pipe dream. And it was painful. We tried

171
00:08:38,200 --> 00:08:41,760
Speaker 5: to raise five hundred million dollars. We couldn't. We instead

172
00:08:41,800 --> 00:08:43,880
Speaker 5: scraped toge about one hundred million, which I know sounds

173
00:08:43,880 --> 00:08:45,080
Speaker 5: like a lot bet at the time was a rounding

174
00:08:45,080 --> 00:08:46,760
Speaker 5: era compared to how much Google has spent on the

175
00:08:46,800 --> 00:08:50,520
Speaker 5: same kind of systems. And it was all angels in

176
00:08:50,559 --> 00:08:52,280
Speaker 5: that first round, a bunch of cats and dogs, all

177
00:08:52,320 --> 00:08:54,280
Speaker 5: of us who believed in the mission. And then over

178
00:08:54,320 --> 00:08:58,040
Speaker 5: the next eighteen months, Dario Tom and team put together

179
00:08:58,120 --> 00:09:01,079
Speaker 5: a plan that we kind of worked shopped on, getting

180
00:09:01,080 --> 00:09:03,680
Speaker 5: an Amazon involved as a strategic and that resulted in

181
00:09:03,720 --> 00:09:07,240
Speaker 5: a four billion dollar compute and capital partnership that made

182
00:09:07,280 --> 00:09:12,760
Speaker 5: me realize infrastructure, especially compute infrastructure, was just a key

183
00:09:12,800 --> 00:09:15,840
Speaker 5: requirement to create any kind of modern eye lab. And

184
00:09:15,920 --> 00:09:18,440
Speaker 5: so since then I've spent the past five six years

185
00:09:18,440 --> 00:09:21,240
Speaker 5: figure out how to unblock that compute bottleneck for research teams.

186
00:09:21,480 --> 00:09:24,960
Speaker 4: Amazing well obviously an incredibly well timed.

187
00:09:24,720 --> 00:09:27,679
Speaker 2: It just like emphasizes how much things have changed, right

188
00:09:27,720 --> 00:09:30,480
Speaker 2: where like people are literally throwing money at like almost

189
00:09:30,559 --> 00:09:34,040
Speaker 2: any model now versus like a few years ago, going

190
00:09:34,120 --> 00:09:37,160
Speaker 2: like hi, agi, really know, Well, let.

191
00:09:37,000 --> 00:09:39,720
Speaker 4: Me ask you this question, because this is a very

192
00:09:39,760 --> 00:09:42,160
Speaker 4: top of mind question for me, and we're going to

193
00:09:42,240 --> 00:09:45,319
Speaker 4: we can skip around on the timeline here, but there

194
00:09:45,320 --> 00:09:48,679
Speaker 4: are three labs that are seen as like genuinely at

195
00:09:48,720 --> 00:09:52,640
Speaker 4: the frontier right now, and that is obviously deep mind

196
00:09:52,679 --> 00:09:56,120
Speaker 4: within Google, Open AI and anthropic And then of course

197
00:09:56,280 --> 00:09:58,480
Speaker 4: you know a lot of people say that the Chinese

198
00:09:58,520 --> 00:10:00,880
Speaker 4: labs are very close, if not there, maybe they're a

199
00:10:00,920 --> 00:10:05,000
Speaker 4: few months behind. Is this is there? You know when

200
00:10:05,040 --> 00:10:07,959
Speaker 4: we think about like part of your mission is like

201
00:10:08,200 --> 00:10:10,679
Speaker 4: you say, okay, a new lab should be able to

202
00:10:10,720 --> 00:10:13,080
Speaker 4: get access to compute if you're really bright, like that

203
00:10:13,120 --> 00:10:17,640
Speaker 4: shouldn't be the bottleneck. Does that imply therefore that you

204
00:10:17,760 --> 00:10:21,560
Speaker 4: expect more labs to be able to were they to

205
00:10:21,679 --> 00:10:25,360
Speaker 4: have access to the compute also reached the frontier, and

206
00:10:25,400 --> 00:10:28,040
Speaker 4: that there is something inherent about like this sort of

207
00:10:28,400 --> 00:10:32,440
Speaker 4: seeming stability or parody that we see among frontier models.

208
00:10:32,679 --> 00:10:34,400
Speaker 5: So the answer to your first question is yes, there

209
00:10:34,440 --> 00:10:37,960
Speaker 5: have many frontiers to be conquered, okay, pioneered, and this

210
00:10:38,120 --> 00:10:40,239
Speaker 5: is not just one frontier. I think that's a fundamental

211
00:10:40,520 --> 00:10:43,120
Speaker 5: misunderstanding people have about your frontier. They talk about the

212
00:10:43,240 --> 00:10:47,319
Speaker 5: jagged frontier, jagged intelligence right in a poetic sense, in

213
00:10:47,320 --> 00:10:49,640
Speaker 5: a historical sense, if you think about the wild West

214
00:10:49,720 --> 00:10:53,160
Speaker 5: or the Western frontier, it wasn't just one frontier. There

215
00:10:53,200 --> 00:10:55,160
Speaker 5: was a frontier of gold, and there was a frontier

216
00:10:55,160 --> 00:10:56,880
Speaker 5: of genes. It turns out Levi's you know, it turned

217
00:10:56,880 --> 00:10:59,000
Speaker 5: out to be a new modern behemoth of a company.

218
00:10:59,000 --> 00:11:01,520
Speaker 5: I mean, there were so many new businesses founded in

219
00:11:01,559 --> 00:11:04,959
Speaker 5: the Industrial Revolution, and I think that's that's the reality.

220
00:11:05,040 --> 00:11:08,680
Speaker 5: Is the software engineering frontier, which is where Anthropic is

221
00:11:08,679 --> 00:11:09,920
Speaker 5: clearly leader, is one frontier.

222
00:11:10,160 --> 00:11:10,600
Speaker 2: Yeah.

223
00:11:11,040 --> 00:11:13,760
Speaker 5: I think the you know, chat frontier, the sort of

224
00:11:13,800 --> 00:11:16,800
Speaker 5: consumer chat frontier is another frontier where Opening Eye has

225
00:11:16,800 --> 00:11:17,280
Speaker 5: been a leader.

226
00:11:17,400 --> 00:11:20,800
Speaker 4: Arguably byte Edance is that the video frontier with seed Dance, right.

227
00:11:20,679 --> 00:11:23,440
Speaker 5: Absolutely, yeah, And so I think there's just many many

228
00:11:23,440 --> 00:11:26,640
Speaker 5: frontiers to be conquered are pioneered. Rather, I think Anthropic

229
00:11:26,720 --> 00:11:28,839
Speaker 5: is clearly a role model for the rest of the

230
00:11:28,880 --> 00:11:30,640
Speaker 5: community on how to do it in an efficient way.

231
00:11:31,040 --> 00:11:33,840
Speaker 5: They're you know, I think he more than five thousand people,

232
00:11:34,320 --> 00:11:35,880
Speaker 5: and they've been able to put out state of the

233
00:11:35,960 --> 00:11:38,720
Speaker 5: art models that you know, teams like Google, which have

234
00:11:38,880 --> 00:11:43,360
Speaker 5: sixty thousand people, are close to but not yet quite there.

235
00:11:43,400 --> 00:11:45,880
Speaker 5: So I actually I don't really agree with your with

236
00:11:46,000 --> 00:11:49,199
Speaker 5: your assessment that they're all at parity. If you use

237
00:11:49,200 --> 00:11:51,760
Speaker 5: the models day in and day out, they're quite remarkably

238
00:11:51,800 --> 00:11:54,240
Speaker 5: different in meaningful ways to the person with hands on

239
00:11:54,240 --> 00:11:56,640
Speaker 5: the keyboard, you know, doing the engineering work. And I

240
00:11:56,640 --> 00:12:00,840
Speaker 5: think those those differences reflect the folk of the teams.

241
00:12:01,679 --> 00:12:04,199
Speaker 5: What is the actual mission that the team working on

242
00:12:04,760 --> 00:12:07,920
Speaker 5: that domain cares about day after day after day. So

243
00:12:07,920 --> 00:12:10,920
Speaker 5: in the Stanford Class I teach, the first lecture was

244
00:12:11,000 --> 00:12:14,360
Speaker 5: a breakdown of how frontier models are even created, and

245
00:12:14,400 --> 00:12:17,000
Speaker 5: it's actually quite simple. The recipes is super simple. There's

246
00:12:17,080 --> 00:12:21,280
Speaker 5: basically four steps this pre training, mid training, post training,

247
00:12:21,400 --> 00:12:24,320
Speaker 5: and then what we call the continuous feedback loop. So

248
00:12:24,400 --> 00:12:27,920
Speaker 5: pre training just is just says, hey, you collect a

249
00:12:27,960 --> 00:12:31,360
Speaker 5: bunch of data from the Internet and train a model

250
00:12:31,440 --> 00:12:34,920
Speaker 5: to be a generally good pattern recognition machine. You then

251
00:12:34,960 --> 00:12:37,480
Speaker 5: do mid training, which is to say, in a particular

252
00:12:37,520 --> 00:12:41,319
Speaker 5: domain that you really care about, you inject more capabilities.

253
00:12:41,559 --> 00:12:43,840
Speaker 5: So if you want this model to reason about science

254
00:12:43,920 --> 00:12:45,920
Speaker 5: or math or physics, then you give it science or

255
00:12:45,920 --> 00:12:47,839
Speaker 5: matter physics data and then you get a pretty good

256
00:12:47,880 --> 00:12:50,480
Speaker 5: model that specialized in that domain. And then you deployed

257
00:12:50,480 --> 00:12:52,520
Speaker 5: to the real world where you have people using it.

258
00:12:52,840 --> 00:12:56,839
Speaker 5: And the context feedback, which is when the model is

259
00:12:56,920 --> 00:12:59,400
Speaker 5: able to do a task well or not and you

260
00:12:59,520 --> 00:13:03,080
Speaker 5: can verify whether that task was done correctly, gives the

261
00:13:03,400 --> 00:13:06,120
Speaker 5: model the data it needs to keep improving on that task,

262
00:13:06,400 --> 00:13:07,240
Speaker 5: on that distribution.

263
00:13:22,960 --> 00:13:25,640
Speaker 2: This is slightly tangential, but like, I give a lot

264
00:13:25,640 --> 00:13:29,080
Speaker 2: of feedback to the models because Joe made me paranoid

265
00:13:29,120 --> 00:13:31,600
Speaker 2: about the basilisk theory. So I want the models to

266
00:13:31,600 --> 00:13:34,160
Speaker 2: appreciate me once they take over the world. But when

267
00:13:34,160 --> 00:13:36,199
Speaker 2: you give them feedback, like if they spit out a

268
00:13:36,200 --> 00:13:40,319
Speaker 2: wrong answer and you say that's wrong, they immediately apologize

269
00:13:40,400 --> 00:13:42,800
Speaker 2: and fall over themselves to say that they're sorry. But

270
00:13:42,840 --> 00:13:46,199
Speaker 2: then you ask them, like give me another output, or

271
00:13:46,280 --> 00:13:48,439
Speaker 2: like would you do it again the same way, and

272
00:13:48,520 --> 00:13:50,840
Speaker 2: they like they often say yes or they give like

273
00:13:50,880 --> 00:13:53,439
Speaker 2: a very similar answer. They don't seem to be responding

274
00:13:53,520 --> 00:13:54,280
Speaker 2: in real.

275
00:13:54,040 --> 00:13:54,760
Speaker 4: Time, correct.

276
00:13:54,840 --> 00:13:56,720
Speaker 5: So when I say feedback, I'm in a very specific

277
00:13:56,760 --> 00:13:59,760
Speaker 5: kind of feedback, which I call verifiable feedback. So when

278
00:13:59,760 --> 00:14:03,000
Speaker 5: you say that wasn't right or that was wrong, that's

279
00:14:03,040 --> 00:14:05,960
Speaker 5: an opinion, okay. Verifiable feedback is when you can have

280
00:14:06,080 --> 00:14:08,280
Speaker 5: as close to factual verification as possible.

281
00:14:08,280 --> 00:14:10,200
Speaker 2: The reason So what does that actually look like?

282
00:14:10,320 --> 00:14:13,240
Speaker 5: That's a great question. So let's take reason by example

283
00:14:13,320 --> 00:14:15,840
Speaker 5: in two or three cases. In the case of software engineering,

284
00:14:16,000 --> 00:14:18,280
Speaker 5: in the way software engineer is actually code is you

285
00:14:18,280 --> 00:14:20,760
Speaker 5: write a piece of code and then you submit it

286
00:14:21,040 --> 00:14:22,880
Speaker 5: to the main code base, and then you usually have

287
00:14:22,960 --> 00:14:25,600
Speaker 5: a peer on your team review the code and approve

288
00:14:25,640 --> 00:14:28,600
Speaker 5: it or reject it, and if it gets approved that

289
00:14:29,040 --> 00:14:31,520
Speaker 5: that's the first step that's called a PR a pull request.

290
00:14:31,800 --> 00:14:34,240
Speaker 5: And if another human on your team that you trust

291
00:14:34,520 --> 00:14:37,320
Speaker 5: approved it, that's one kind of verification of quality. And

292
00:14:37,360 --> 00:14:40,840
Speaker 5: then two before that piece of code usually gets deployed

293
00:14:40,880 --> 00:14:43,560
Speaker 5: to a production system, you have unit tests, and those

294
00:14:43,600 --> 00:14:48,360
Speaker 5: are quite objective tests of is this code performing the

295
00:14:48,440 --> 00:14:51,080
Speaker 5: function we need it to And if it passes both

296
00:14:51,120 --> 00:14:54,160
Speaker 5: those tests, it's a verifiable piece of code that accomplished

297
00:14:54,160 --> 00:14:57,200
Speaker 5: the goal. So in software engineering, the reason we've seen

298
00:14:57,280 --> 00:15:00,560
Speaker 5: such a dramatic improvement and capabilities is that a lot

299
00:15:00,640 --> 00:15:04,080
Speaker 5: of these labs are using feedback from that verification loop.

300
00:15:04,840 --> 00:15:07,880
Speaker 5: In the case of another lab I incubated called Periodic Labs,

301
00:15:07,880 --> 00:15:10,160
Speaker 5: which we started a year ago, and you should come

302
00:15:10,160 --> 00:15:12,640
Speaker 5: by sometime, we've got forty thousand square feet in mental

303
00:15:12,640 --> 00:15:15,760
Speaker 5: Park where we've got AI models that are predicting new

304
00:15:16,160 --> 00:15:18,280
Speaker 5: The goal is to try to find a room temperature superconductor,

305
00:15:18,320 --> 00:15:20,359
Speaker 5: and so these models predicted.

306
00:15:20,280 --> 00:15:24,080
Speaker 4: I forgot about superconductor I forgot about that.

307
00:15:23,880 --> 00:15:25,000
Speaker 3: That was a fun summer.

308
00:15:25,400 --> 00:15:28,880
Speaker 5: Yes, this time we will verify the if we ever

309
00:15:28,920 --> 00:15:30,240
Speaker 5: put something on you, and you will know it's not

310
00:15:31,720 --> 00:15:34,840
Speaker 5: that's not going to be us. But the AI system

311
00:15:34,840 --> 00:15:39,400
Speaker 5: predicts new materials candidates. Then we have robots that synthesize

312
00:15:39,400 --> 00:15:41,600
Speaker 5: the new material in the lab and then use X

313
00:15:41,680 --> 00:15:43,920
Speaker 5: ray diffraction machines to test whether the material has the

314
00:15:43,960 --> 00:15:48,000
Speaker 5: properties they said it would. And that's verifiable feedback from reality,

315
00:15:48,280 --> 00:15:50,800
Speaker 5: from physics, and then we pipe that data back into

316
00:15:50,800 --> 00:15:53,640
Speaker 5: the training loop over and over again. That context feedback

317
00:15:54,080 --> 00:15:58,000
Speaker 5: is very factually verifiable, and that's where progress is the

318
00:15:58,080 --> 00:16:02,040
Speaker 5: fastest today because that that feed back doesn't result in

319
00:16:02,080 --> 00:16:04,800
Speaker 5: the kind of hallucinations that you often experience with these

320
00:16:04,840 --> 00:16:07,120
Speaker 5: models on more subjective desks. It's also, by the way,

321
00:16:07,120 --> 00:16:08,960
Speaker 5: why the models is a terriblest objective desk, a creative

322
00:16:08,960 --> 00:16:12,600
Speaker 5: writing and sometimes they can get quite toxic, to be honest,

323
00:16:12,600 --> 00:16:13,920
Speaker 5: if you get them don the wrong loop. I don't

324
00:16:13,920 --> 00:16:15,600
Speaker 5: know if you've been using it as a therapy, bought

325
00:16:15,640 --> 00:16:16,000
Speaker 5: and so on.

326
00:16:16,760 --> 00:16:17,280
Speaker 2: I have not.

327
00:16:17,480 --> 00:16:18,920
Speaker 5: Just for the right, that's great.

328
00:16:19,080 --> 00:16:21,520
Speaker 2: It did ask me to defy the laws of gravity

329
00:16:21,640 --> 00:16:23,560
Speaker 2: at one point because I was trying to create something

330
00:16:23,600 --> 00:16:25,120
Speaker 2: in my backyard and I was asking you how to

331
00:16:25,160 --> 00:16:27,120
Speaker 2: do it, and it was like, then just set this

332
00:16:27,240 --> 00:16:29,240
Speaker 2: up like a following away and I was like, that's

333
00:16:29,280 --> 00:16:32,080
Speaker 2: not within the laws of physics. Whatever.

334
00:16:33,480 --> 00:16:36,000
Speaker 4: No good, Well, what's interesting And this is actually a

335
00:16:36,120 --> 00:16:39,280
Speaker 4: trillion dollar question from just a very broad standpoint is,

336
00:16:39,320 --> 00:16:42,720
Speaker 4: as you point out, even prior to AI, the field

337
00:16:42,760 --> 00:16:47,120
Speaker 4: of coding had a very systematized approach to the feedback

338
00:16:47,160 --> 00:16:49,680
Speaker 4: loops already yes, and so then it's like a I

339
00:16:49,720 --> 00:16:52,000
Speaker 4: could sort of replicate that anyone who's done you vibe

340
00:16:52,000 --> 00:16:55,320
Speaker 4: coding can see in the chain of thoughts sometimes that

341
00:16:55,360 --> 00:16:57,280
Speaker 4: didn't work. Let me try this, that didn't work, let

342
00:16:57,320 --> 00:17:01,200
Speaker 4: me try this. Most fields don't really have that. By

343
00:17:01,240 --> 00:17:04,119
Speaker 4: and large, journalism doesn't have that. I mean, there are things,

344
00:17:04,200 --> 00:17:06,600
Speaker 4: there are outputs that are better and worse, we don't

345
00:17:06,640 --> 00:17:10,560
Speaker 4: really have that that sort of like formalized approach to

346
00:17:10,720 --> 00:17:15,920
Speaker 4: the yes no. Does that just zooming out? To my mind,

347
00:17:15,920 --> 00:17:18,479
Speaker 4: that would apply that maybe at least to some extent,

348
00:17:18,800 --> 00:17:21,720
Speaker 4: coding is a little bit special from a sort of

349
00:17:21,720 --> 00:17:25,520
Speaker 4: white collar knowledge work that in terms of like is

350
00:17:25,560 --> 00:17:29,160
Speaker 4: it going to be as good as say, I don't

351
00:17:29,160 --> 00:17:32,560
Speaker 4: know sales or something like that, because it has a

352
00:17:32,880 --> 00:17:34,720
Speaker 4: coding as a long history.

353
00:17:34,680 --> 00:17:37,400
Speaker 2: It's formulated and structure that structured pipeline.

354
00:17:37,440 --> 00:17:39,560
Speaker 5: Yeah, this is a great point. So where progress will

355
00:17:39,600 --> 00:17:43,560
Speaker 5: be made most predictably is in parts of knowledge work,

356
00:17:43,880 --> 00:17:48,080
Speaker 5: where the task is essentially a workflow that's fairly structured. Yeah,

357
00:17:48,119 --> 00:17:50,040
Speaker 5: and so somebody who spends most of their day in

358
00:17:50,119 --> 00:17:53,560
Speaker 5: putting cells into an Excel spreadsheet, well, that part of

359
00:17:53,600 --> 00:17:56,360
Speaker 5: the job will get automated pretty fast because that's actually verifiable,

360
00:17:56,400 --> 00:17:58,439
Speaker 5: and you know what, that's frankly often the most tedious

361
00:17:58,440 --> 00:18:01,800
Speaker 5: part of the job anyway. So I'm quite excited to

362
00:18:01,880 --> 00:18:05,520
Speaker 5: see that progress because I'm terrible a spreadsheets, and I

363
00:18:05,560 --> 00:18:07,560
Speaker 5: think if we could free up more of my time

364
00:18:07,560 --> 00:18:09,800
Speaker 5: and hopefully other people's time, to focus on the art

365
00:18:09,960 --> 00:18:13,400
Speaker 5: of the spreadsheet, not the deediest.

366
00:18:13,000 --> 00:18:15,960
Speaker 4: Part of the entry, the entry and retrieval.

367
00:18:16,040 --> 00:18:19,600
Speaker 5: Yeah, yeah, exactly, you know. And in journalism, I think

368
00:18:19,600 --> 00:18:22,359
Speaker 5: it's the same thing. There's so much craft that gets

369
00:18:22,440 --> 00:18:25,159
Speaker 5: that goes into the verification of a story before it

370
00:18:25,200 --> 00:18:27,920
Speaker 5: goes out that's not legible to the world. You know.

371
00:18:27,920 --> 00:18:30,080
Speaker 5: I've had a chance to spend some time with some

372
00:18:30,119 --> 00:18:33,520
Speaker 5: of the journalistic institutions of the barrier like Cade, Metz

373
00:18:33,600 --> 00:18:35,840
Speaker 5: or b Als and at the Journal, and as you

374
00:18:35,840 --> 00:18:38,440
Speaker 5: spend time with them you realize, I mean they're verifying

375
00:18:38,720 --> 00:18:42,280
Speaker 5: every sentence that goes into each fact checking, right, so

376
00:18:42,400 --> 00:18:45,200
Speaker 5: fact check. That's an example where I think we should

377
00:18:45,240 --> 00:18:47,479
Speaker 5: be leaning on these tools and you should expect more

378
00:18:47,520 --> 00:18:52,199
Speaker 5: progress and the parts then that will be more to

379
00:18:52,240 --> 00:18:55,960
Speaker 5: borrow jagget frontier framing. There, that's we will be in

380
00:18:55,960 --> 00:18:59,879
Speaker 5: a regime of jagged frontier progress where wherever parts of

381
00:19:00,080 --> 00:19:05,600
Speaker 5: workflows that are verifiable factually will essentially you'll see progress

382
00:19:05,640 --> 00:19:08,600
Speaker 5: they're very predictably over the next few years. And consequently,

383
00:19:08,600 --> 00:19:12,560
Speaker 5: wherever that progress that the workflows are not verifiable is

384
00:19:12,560 --> 00:19:14,480
Speaker 5: actually where humans are going to shine. And I think

385
00:19:14,480 --> 00:19:16,280
Speaker 5: that's where parts of the economy are. You're going to

386
00:19:16,280 --> 00:19:21,160
Speaker 5: see extraordinary gains in the wages of humans who have

387
00:19:21,240 --> 00:19:26,359
Speaker 5: creativity and craft that are not typically verifiable you know,

388
00:19:26,520 --> 00:19:28,760
Speaker 5: through traditional objective means. Does that make sense?

389
00:19:28,960 --> 00:19:31,679
Speaker 2: Yeah, it does, and it dovetails with a lot of

390
00:19:31,680 --> 00:19:33,800
Speaker 2: what we've been talking about on the show recently. Just

391
00:19:33,840 --> 00:19:37,359
Speaker 2: going back to verifiable feedback. So, Okay, the model spits

392
00:19:37,400 --> 00:19:39,879
Speaker 2: out something and you can check whether it's right or wrong.

393
00:19:40,760 --> 00:19:43,760
Speaker 2: Is it important to understand how the model actually got

394
00:19:43,960 --> 00:19:47,920
Speaker 2: to that answer, because we have discussions with like big

395
00:19:47,960 --> 00:19:51,760
Speaker 2: bank CEOs who are using more AI, and their response

396
00:19:51,760 --> 00:19:53,840
Speaker 2: to this question is always like, well, if we can

397
00:19:54,200 --> 00:19:57,080
Speaker 2: put restrictions around the AI, if we make sure that

398
00:19:57,119 --> 00:19:59,880
Speaker 2: it's like released into a sandbox before it's released into

399
00:20:00,119 --> 00:20:03,840
Speaker 2: wider world, we're all set from a regulatory perspective, and

400
00:20:04,000 --> 00:20:07,280
Speaker 2: regulators don't actually need to know what's in the black

401
00:20:07,320 --> 00:20:09,720
Speaker 2: box model and how it's working. But like this seems

402
00:20:09,760 --> 00:20:10,960
Speaker 2: a bit concerning to me.

403
00:20:11,240 --> 00:20:14,600
Speaker 5: Yeah, No, I'm quite strongly opinionated about this one, which

404
00:20:14,680 --> 00:20:17,520
Speaker 5: is that technical literacy should be non negotiable. It's the

405
00:20:17,560 --> 00:20:19,880
Speaker 5: reason I spend so much time teaching this class at Stanford,

406
00:20:19,920 --> 00:20:21,960
Speaker 5: putting it up online, and the idea of the Frontier

407
00:20:21,960 --> 00:20:25,200
Speaker 5: system's classes that end to end. It's a full simple

408
00:20:25,320 --> 00:20:29,040
Speaker 5: but first principles of breakdown of how these systems are

409
00:20:29,080 --> 00:20:32,400
Speaker 5: built from scratch, from land PowerShell, like the energy, where

410
00:20:32,400 --> 00:20:33,960
Speaker 5: do we get them the data centers? Then how do

411
00:20:33,960 --> 00:20:35,679
Speaker 5: we train the models? And the final project of the

412
00:20:35,680 --> 00:20:38,360
Speaker 5: class with the kids was actually the one person Frontier Lab,

413
00:20:38,400 --> 00:20:41,080
Speaker 5: which is at the end they're creating their own models

414
00:20:41,080 --> 00:20:43,639
Speaker 5: and so on. Because the idea is that a person

415
00:20:43,680 --> 00:20:45,919
Speaker 5: with the right tools today can scale themselves infinitely, but

416
00:20:45,960 --> 00:20:47,840
Speaker 5: they need to know how to use the tools, what

417
00:20:47,880 --> 00:20:49,919
Speaker 5: the limitations are, when to lean on them versus not.

418
00:20:50,359 --> 00:20:52,840
Speaker 5: And I think this is a generalizable piece of technical

419
00:20:52,840 --> 00:20:56,239
Speaker 5: literacy that all leaders should have. It's like saying, you know,

420
00:20:57,359 --> 00:21:01,040
Speaker 5: I in the nineties, I imagine if if you knew

421
00:21:01,240 --> 00:21:03,760
Speaker 5: you could use the Internet without really knowing how it worked,

422
00:21:04,000 --> 00:21:06,439
Speaker 5: but you know, on the margins when like the page

423
00:21:06,480 --> 00:21:10,640
Speaker 5: doesn't like refresh or you're like this this cookie thing

424
00:21:10,720 --> 00:21:13,480
Speaker 5: is annoying me. Like over time, people who are more

425
00:21:13,480 --> 00:21:18,080
Speaker 5: technically literate just realized sometimes you've got a debug, you know,

426
00:21:18,160 --> 00:21:21,400
Speaker 5: the browser, and those of us who've learned over time

427
00:21:21,400 --> 00:21:25,159
Speaker 5: to do knowledge work are more adapt at leaning on

428
00:21:25,160 --> 00:21:26,960
Speaker 5: them versus not. Like just now, when I was trying

429
00:21:27,000 --> 00:21:28,760
Speaker 5: to get onto the Internet, I realized, okay, there's this

430
00:21:29,080 --> 00:21:31,840
Speaker 5: you know, Wi Fi password, whatever, And then you don't

431
00:21:32,040 --> 00:21:34,840
Speaker 5: end up relying on them in ways that they can't

432
00:21:34,840 --> 00:21:37,560
Speaker 5: fulfill your need anyway. And what's a little bit more

433
00:21:37,640 --> 00:21:41,600
Speaker 5: dangerous with these systems is because we tend to anthropomorphize

434
00:21:41,640 --> 00:21:45,560
Speaker 5: them without the technical literacy that that I wish all

435
00:21:45,680 --> 00:21:48,560
Speaker 5: leaders had about reasoning about how these systems were built.

436
00:21:49,720 --> 00:21:52,360
Speaker 5: What you end up doing is projecting out in your

437
00:21:52,440 --> 00:21:55,480
Speaker 5: mind what the capabilities are in ways that are inaccurate.

438
00:21:55,720 --> 00:21:58,520
Speaker 5: You project out their impact and society that are not accurate.

439
00:21:58,720 --> 00:22:01,000
Speaker 5: You project out their business in a way that are

440
00:22:01,040 --> 00:22:02,439
Speaker 5: not Actually, I mean in the very fact that when

441
00:22:02,480 --> 00:22:04,400
Speaker 5: you started this conversation, I don't blame you for it.

442
00:22:04,560 --> 00:22:08,919
Speaker 5: You're like, there's three models at the frontier. Yeah, I'm like, well,

443
00:22:09,240 --> 00:22:12,320
Speaker 5: which frontier and which three models? Because from where I'm sitting,

444
00:22:12,359 --> 00:22:15,320
Speaker 5: there's like seventeen different frontiers right now. There's four different

445
00:22:15,320 --> 00:22:17,679
Speaker 5: players in each one, and the businesses of all of

446
00:22:17,680 --> 00:22:22,160
Speaker 5: them are kind of breathtaking. So I think that technical

447
00:22:22,200 --> 00:22:26,040
Speaker 5: literacy should always for leaders be a basic requirement. And

448
00:22:26,080 --> 00:22:29,399
Speaker 5: then if you're deploying your these systems at Goldman Sacks,

449
00:22:29,720 --> 00:22:32,639
Speaker 5: you won't oversimplify and get tripped up later when you know,

450
00:22:32,680 --> 00:22:35,119
Speaker 5: two years later, you realize half your employee base has

451
00:22:35,119 --> 00:22:38,480
Speaker 5: been leaning on this like sandbox framing, when in reality,

452
00:22:38,840 --> 00:22:41,600
Speaker 5: inside the sandbox they were doing all kinds of They

453
00:22:41,600 --> 00:22:44,280
Speaker 5: were using the tools in ways that were prone to hallucination,

454
00:22:44,440 --> 00:22:46,959
Speaker 5: prone to risks, prompt injection, they were leaning on it

455
00:22:47,000 --> 00:22:50,440
Speaker 5: in ways that were not informed in the appropriate ways.

456
00:22:50,880 --> 00:22:51,520
Speaker 5: Is this making sense?

457
00:22:51,600 --> 00:22:53,520
Speaker 2: Yeah, Like at a minimum they would not be using

458
00:22:53,520 --> 00:22:55,080
Speaker 2: it in the optimal way.

459
00:22:55,000 --> 00:23:00,359
Speaker 5: Right or relying too much on it. It's the can't

460
00:23:00,400 --> 00:23:04,480
Speaker 5: outsource your understanding to a model. Can you can outsource

461
00:23:04,520 --> 00:23:08,159
Speaker 5: your thinking? You can outsource part of the tedious workflows,

462
00:23:08,400 --> 00:23:11,080
Speaker 5: but you can't outsource your understanding. Yeah. And if you

463
00:23:11,160 --> 00:23:13,800
Speaker 5: keep thinking, if you say, if you create these simplistic

464
00:23:13,840 --> 00:23:17,120
Speaker 5: frameworks of oh, here's a sandbox and this is safe,

465
00:23:18,160 --> 00:23:21,440
Speaker 5: you have to you have to use that sandbox in

466
00:23:21,480 --> 00:23:23,920
Speaker 5: the right way. Because if you say, well, now, everything

467
00:23:23,960 --> 00:23:25,480
Speaker 5: that happens at the sandbox is totally fine. That the

468
00:23:25,520 --> 00:23:27,760
Speaker 5: model says, use the spreadsheet. The spreadsheet is good, it's

469
00:23:27,760 --> 00:23:30,320
Speaker 5: deployed in our servers, but you didn't actually check the

470
00:23:30,840 --> 00:23:33,119
Speaker 5: spreadsheet and what went into the spreadsheet. And then the

471
00:23:33,160 --> 00:23:36,080
Speaker 5: model actually understand the particular structure of the business, the

472
00:23:36,119 --> 00:23:38,200
Speaker 5: physics of the business that you're trying to model out.

473
00:23:38,520 --> 00:23:41,600
Speaker 5: Then you've outsourced your understanding to it. Does that make sense?

474
00:23:41,840 --> 00:23:47,480
Speaker 4: Absolutely? Let's talk about AMP Yes. And because one, you know,

475
00:23:47,520 --> 00:23:49,840
Speaker 4: you're never going to get the frontier in anything unless

476
00:23:49,840 --> 00:23:52,680
Speaker 4: you have access to compute. It seems pretty obvious, and

477
00:23:53,000 --> 00:23:56,640
Speaker 4: there are various arrangements for acquiring compute. You have companies

478
00:23:56,640 --> 00:24:00,159
Speaker 4: building their own data centers, you have smaller labs, and

479
00:24:00,240 --> 00:24:03,080
Speaker 4: maybe they use someone else's data centers or a neo cloud, right,

480
00:24:03,119 --> 00:24:06,560
Speaker 4: et cetera. What are you building at AMP such that

481
00:24:06,680 --> 00:24:09,359
Speaker 4: at least as part of this story, is trying to

482
00:24:09,440 --> 00:24:10,920
Speaker 4: solve the compute bottleneck.

483
00:24:10,960 --> 00:24:14,520
Speaker 5: Specifically, yeah, we are. It's very simple what we're what

484
00:24:14,560 --> 00:24:17,000
Speaker 5: we're doing at AMP. We're doing two things. We are

485
00:24:17,080 --> 00:24:20,760
Speaker 5: trying to standardize the format for compute, which today is

486
00:24:20,800 --> 00:24:24,000
Speaker 5: super fragmented. So in the history of infrastructure, if you

487
00:24:24,040 --> 00:24:29,119
Speaker 5: look at whether it was the Industrial Revolution, the Internet streaming,

488
00:24:29,600 --> 00:24:33,680
Speaker 5: there were usually formats of inputs that were quite heterogeneous.

489
00:24:33,800 --> 00:24:40,040
Speaker 5: They were fragmented, and then to unlock productivity you had

490
00:24:40,080 --> 00:24:43,240
Speaker 5: to standardize a format. So in the case of electricity,

491
00:24:43,760 --> 00:24:48,360
Speaker 5: until it was until ACDC was standardized, right, megawatts would

492
00:24:48,359 --> 00:24:52,120
Speaker 5: just sit in stranded pockets around the United States being unused.

493
00:24:52,880 --> 00:24:57,000
Speaker 5: And then once we standardized the format to ACDC, then

494
00:24:57,040 --> 00:24:59,359
Speaker 5: the question was okay, great, now we've turned all these

495
00:24:59,480 --> 00:25:04,120
Speaker 5: stranded dockets of electricity into one sort of interoperable universal format.

496
00:25:04,800 --> 00:25:06,679
Speaker 5: Now how do we distribute it to everybody who needs it?

497
00:25:06,720 --> 00:25:08,159
Speaker 5: And we came up with this distribution layer in the

498
00:25:08,240 --> 00:25:11,359
Speaker 5: United States called the grid. That's that's all we're doing.

499
00:25:11,480 --> 00:25:13,160
Speaker 3: So building a grid for compute.

500
00:25:13,200 --> 00:25:16,440
Speaker 5: Correct, we're standardizing. We're trying to standardize the compute layer today,

501
00:25:16,840 --> 00:25:20,520
Speaker 5: different chip types, different manufacturers, different clouds. I mean, it's

502
00:25:20,520 --> 00:25:23,320
Speaker 5: a complete mess. And if you're go ahead.

503
00:25:23,280 --> 00:25:25,919
Speaker 2: Say more about how we plan to do this, because

504
00:25:25,920 --> 00:25:29,240
Speaker 2: we've talked before about you know, there are various people

505
00:25:29,240 --> 00:25:33,280
Speaker 2: out there that want to create indices of compute futures

506
00:25:33,359 --> 00:25:36,240
Speaker 2: potentially on compute and the issue that always comes.

507
00:25:36,080 --> 00:25:39,040
Speaker 5: Up is fungibility, right exactly. So we've got a couple

508
00:25:39,080 --> 00:25:40,680
Speaker 5: of ways we solve the fung ability problem. This is

509
00:25:40,720 --> 00:25:45,240
Speaker 5: a pretty thorny challenge. We solve it in two or

510
00:25:45,240 --> 00:25:48,160
Speaker 5: three ways. The first is we have a system called

511
00:25:48,200 --> 00:25:52,240
Speaker 5: the grid, which actually makes the compute fungible at a

512
00:25:52,280 --> 00:25:54,600
Speaker 5: consumption layer. So under the hood, we have a bunch

513
00:25:54,640 --> 00:25:58,120
Speaker 5: of different chip types, we support various different manufacturers, and

514
00:25:58,280 --> 00:26:01,320
Speaker 5: there's a system that was built to do this already

515
00:26:01,560 --> 00:26:04,119
Speaker 5: inside a little company called Google, and one of the

516
00:26:04,119 --> 00:26:06,760
Speaker 5: technical leads on that on that project was called BORG

517
00:26:06,920 --> 00:26:09,480
Speaker 5: internally at Google is my co founder, Sebastian Lobo. He

518
00:26:09,520 --> 00:26:12,200
Speaker 5: was my roommate at Stanford fourteen years ago. He's engineering

519
00:26:12,240 --> 00:26:14,240
Speaker 5: co founder, and we're building BORG for everybody else, which

520
00:26:14,280 --> 00:26:17,119
Speaker 5: is essentially a translation layer that says, no matter what

521
00:26:17,160 --> 00:26:20,240
Speaker 5: the underlying chip type is, the machine learning researcher who's

522
00:26:20,320 --> 00:26:22,760
Speaker 5: using the chip just has to worry about the workload

523
00:26:22,800 --> 00:26:24,560
Speaker 5: and we handle everything else underneath the hood.

524
00:26:24,680 --> 00:26:29,320
Speaker 2: When you say system, is this hardware or software that's doing.

525
00:26:29,119 --> 00:26:31,439
Speaker 5: That, it's all software. Okay, Yeah, So we handle that

526
00:26:31,440 --> 00:26:35,520
Speaker 5: translation layer in software, and it's a pretty gnarly challenge.

527
00:26:35,560 --> 00:26:39,119
Speaker 5: But today we're able to do that in ways that

528
00:26:39,240 --> 00:26:42,919
Speaker 5: improved utilization sometimes from fifty sixty percent at labs that

529
00:26:42,960 --> 00:26:45,280
Speaker 5: we have incubated or on the grid to close to

530
00:26:45,359 --> 00:26:47,840
Speaker 5: ninety five ninety six percent. At Google, the utilization is

531
00:26:47,920 --> 00:26:51,119
Speaker 5: roughly ninety nine percent. When Sebastian arrived at Google, it

532
00:26:51,160 --> 00:26:53,840
Speaker 5: was about sixty two percent. By the time he left,

533
00:26:53,840 --> 00:26:55,760
Speaker 5: it was roughly at ninety nine percent at Google. If

534
00:26:55,960 --> 00:26:59,520
Speaker 5: utilization is at ninety six percent, that's considered a major outage. Today,

535
00:26:59,560 --> 00:27:02,679
Speaker 5: the average data center in the industry, in the ecosystem,

536
00:27:02,680 --> 00:27:04,640
Speaker 5: in the independent ecosystem is running at less than seventy

537
00:27:04,640 --> 00:27:08,760
Speaker 5: percent utilization. The Colossus two, which is running in Memphis,

538
00:27:08,800 --> 00:27:12,000
Speaker 5: Elon's five hundred thousand, five hundred thousand GB three hundreds

539
00:27:12,320 --> 00:27:14,720
Speaker 5: was running at less than sixty percent note utilization and

540
00:27:14,800 --> 00:27:18,640
Speaker 5: less than eleven percent MFU model flop utilization is how

541
00:27:18,720 --> 00:27:20,840
Speaker 5: much of the chip is actually being used. So there's

542
00:27:20,840 --> 00:27:23,119
Speaker 5: two kinds of utilization people care about in the data center.

543
00:27:23,280 --> 00:27:26,600
Speaker 5: First is how many chips are being used. That's the highest.

544
00:27:26,880 --> 00:27:29,720
Speaker 5: That's just the most naive measure. If that number is

545
00:27:29,720 --> 00:27:33,080
Speaker 5: not a ninety plus percent, no excuses. So you have

546
00:27:33,119 --> 00:27:35,240
Speaker 5: the chips, they should at least at least be doing something.

547
00:27:35,680 --> 00:27:38,480
Speaker 5: And then within the chip during how much of the

548
00:27:38,560 --> 00:27:41,920
Speaker 5: chip is being used within a workload. That number is

549
00:27:41,960 --> 00:27:42,680
Speaker 5: usually much lower.

550
00:27:42,720 --> 00:27:46,600
Speaker 4: I'm very intrigued by this latter point about that even

551
00:27:46,600 --> 00:27:49,800
Speaker 4: if even like the chip itself may not be even

552
00:27:50,119 --> 00:27:55,160
Speaker 4: used at full pacity. Because I see these numbers and

553
00:27:55,200 --> 00:27:58,439
Speaker 4: you say, like a lab has like a we have

554
00:27:58,480 --> 00:28:01,600
Speaker 4: two hundred chips, we've acquired eight hundred GPUs, et cetera.

555
00:28:02,040 --> 00:28:06,119
Speaker 4: And when I see these headlines, I assumed that ali

556
00:28:06,560 --> 00:28:10,520
Speaker 4: optimal utilization techniques must be so good. That you can

557
00:28:10,560 --> 00:28:14,520
Speaker 4: infer someone's capabilities simply by how many and video GPUs

558
00:28:14,560 --> 00:28:18,000
Speaker 4: they've acquired. But you're saying is that there is actually

559
00:28:18,119 --> 00:28:21,840
Speaker 4: quite a bit of heterogeneity about the techniques and approaches

560
00:28:22,119 --> 00:28:23,679
Speaker 4: to getting the most juice out.

561
00:28:23,560 --> 00:28:25,600
Speaker 5: Of ad chip. Yes, you have to measure what matters,

562
00:28:25,600 --> 00:28:28,840
Speaker 5: and what matters is output. Okay, you're when anytime I

563
00:28:28,840 --> 00:28:30,920
Speaker 5: start a new lab with the in the case of

564
00:28:30,960 --> 00:28:35,000
Speaker 5: periodic labs, it was we started with Liam Fettis was

565
00:28:35,000 --> 00:28:37,320
Speaker 5: the co creative chat Epte and Dostubbuk who led the

566
00:28:37,320 --> 00:28:39,480
Speaker 5: physics teams a deep mind. And when we sat down

567
00:28:39,480 --> 00:28:42,960
Speaker 5: and we planned out the company's roadmap, the most important

568
00:28:43,000 --> 00:28:44,520
Speaker 5: thing to us to measure it was not the number

569
00:28:44,520 --> 00:28:45,200
Speaker 5: of chips we had.

570
00:28:45,440 --> 00:28:48,080
Speaker 4: Yeah, it's the the eval what we call all this

571
00:28:48,280 --> 00:28:50,560
Speaker 4: chip bragging. They're like, oh, we acquired it is just

572
00:28:50,600 --> 00:28:54,160
Speaker 4: a sort of bravado. Yeah, all right, this is helpful.

573
00:28:54,440 --> 00:28:56,840
Speaker 5: You don't measure the inputs you should be No, I agree,

574
00:28:57,000 --> 00:28:57,440
Speaker 5: But I'm.

575
00:28:57,360 --> 00:29:00,760
Speaker 2: Actually fascinated that, like there is a saw toware solution

576
00:29:01,000 --> 00:29:03,120
Speaker 2: to what I perceived in my head as like a

577
00:29:03,240 --> 00:29:05,240
Speaker 2: very physical constraint.

578
00:29:05,840 --> 00:29:07,120
Speaker 3: How does this actually work?

579
00:29:07,360 --> 00:29:09,200
Speaker 2: Like feel free to get technical here.

580
00:29:09,320 --> 00:29:10,719
Speaker 3: Like, I want to understand the system.

581
00:29:10,800 --> 00:29:12,680
Speaker 5: Yes, so let me give you the technological answer and

582
00:29:12,680 --> 00:29:15,320
Speaker 5: the economic answer. The economic answer actually is a simple

583
00:29:15,360 --> 00:29:20,520
Speaker 5: one reason about the way the compute business works today

584
00:29:20,640 --> 00:29:24,520
Speaker 5: is primarily on the construct of the atomic unit of

585
00:29:24,560 --> 00:29:28,680
Speaker 5: long term leases. So I'm a researcher, I need some compute.

586
00:29:28,880 --> 00:29:32,160
Speaker 5: I show up to a compute provider and say hello,

587
00:29:32,320 --> 00:29:34,880
Speaker 5: I would like some compute please, And the computer provider says,

588
00:29:34,880 --> 00:29:38,360
Speaker 5: no problem. Here's you know, five hundred AMD chips or

589
00:29:38,360 --> 00:29:40,800
Speaker 5: in video chips that you can lease from me on

590
00:29:40,880 --> 00:29:43,360
Speaker 5: various time scales, and you've got to pay for twenty

591
00:29:43,400 --> 00:29:46,200
Speaker 5: four seven. It's like leasing an apartment, and whether you

592
00:29:46,320 --> 00:29:48,400
Speaker 5: use it or not, that's your problem. But it's two

593
00:29:48,440 --> 00:29:50,880
Speaker 5: dollars fifty per hour, three dollars an hour. So instead

594
00:29:50,920 --> 00:29:53,800
Speaker 5: you take a long term lease. And now the cloud provider,

595
00:29:53,840 --> 00:29:56,880
Speaker 5: the comput provider said great, I just booked revenue for

596
00:29:57,080 --> 00:29:59,800
Speaker 5: the next two years that this guy rented. Yeah, Now

597
00:30:00,240 --> 00:30:03,080
Speaker 5: what happens with that compute, whether it's used or not,

598
00:30:03,360 --> 00:30:07,080
Speaker 5: is the researchers problem. They've outsourced that problem as a

599
00:30:07,120 --> 00:30:09,960
Speaker 5: result of this wastage that we're talking about. And I'm

600
00:30:09,960 --> 00:30:12,960
Speaker 5: happy to go into why it's hard for individual teams

601
00:30:12,960 --> 00:30:16,240
Speaker 5: to utilize most of the capacity. The primary reason it's

602
00:30:16,520 --> 00:30:20,400
Speaker 5: because research is spiky. It's hard to forecast, so you

603
00:30:20,480 --> 00:30:23,640
Speaker 5: put over provision for your peak, not your baseload. Because

604
00:30:23,680 --> 00:30:26,520
Speaker 5: what happens you're researching on these algorithms, and the minute

605
00:30:26,880 --> 00:30:29,600
Speaker 5: like one is working, you go, guys, let's scale. We

606
00:30:29,640 --> 00:30:31,800
Speaker 5: want to ship this thing, so let's improve, throw as

607
00:30:31,800 --> 00:30:34,480
Speaker 5: many chips at it, and then once we ship it,

608
00:30:35,000 --> 00:30:37,840
Speaker 5: the needs go down. So between these spikes there's just

609
00:30:37,920 --> 00:30:42,520
Speaker 5: huge pockets of unused compute. As a result, the effective

610
00:30:42,680 --> 00:30:45,959
Speaker 5: price per hour that you're paying is closer to twenty

611
00:30:46,000 --> 00:30:48,760
Speaker 5: five to twenty eight dollars, whereas the marketed rate that

612
00:30:48,840 --> 00:30:51,800
Speaker 5: you think you're paying is two dollars fifty. Yeah, so

613
00:30:51,960 --> 00:30:57,440
Speaker 5: that's spread due to waste. It is just insane. So

614
00:30:57,440 --> 00:31:00,400
Speaker 5: from an economic perspective, that's the wastage, that's the dead loss.

615
00:31:00,480 --> 00:31:03,200
Speaker 5: Yeah right, okay, So now how do we from a

616
00:31:03,240 --> 00:31:06,800
Speaker 5: technological perspective, how do we utilize that opportunity? Literally all

617
00:31:06,800 --> 00:31:09,920
Speaker 5: we do is from a software perspective, we take all

618
00:31:09,960 --> 00:31:13,040
Speaker 5: of that unutilized compute, no matter what format it is.

619
00:31:13,240 --> 00:31:14,840
Speaker 5: It might be in video, it might be MD we

620
00:31:14,880 --> 00:31:17,200
Speaker 5: love AMD, it might be some other chip, and we

621
00:31:17,240 --> 00:31:21,800
Speaker 5: turn it into one one fungible resource and that we

622
00:31:22,040 --> 00:31:24,360
Speaker 5: standard as a format on something we call grid credits,

623
00:31:24,560 --> 00:31:27,920
Speaker 5: So researchers don't even need to think about what chip

624
00:31:27,960 --> 00:31:29,600
Speaker 5: type is under the hood. You know, they're just paying

625
00:31:29,600 --> 00:31:32,160
Speaker 5: what they need or what they use. And so from

626
00:31:32,200 --> 00:31:34,840
Speaker 5: a fiducial perspective, I'm on seven boards as an investor,

627
00:31:34,840 --> 00:31:37,640
Speaker 5: I get very excited when when teams switch from this

628
00:31:37,720 --> 00:31:39,600
Speaker 5: sort of long term lease model where they're paying twenty

629
00:31:39,640 --> 00:31:42,200
Speaker 5: five twenty six dollars per GP hour now they're actually

630
00:31:42,200 --> 00:31:44,520
Speaker 5: only paying the two dollars fifty that was marketed because

631
00:31:44,520 --> 00:31:47,240
Speaker 5: everything they're not using gets reallocated to the grid and

632
00:31:47,320 --> 00:31:49,560
Speaker 5: other research labs can use that resource.

633
00:32:05,440 --> 00:32:09,040
Speaker 4: This is the problem you're mostly solving for. Is the

634
00:32:09,160 --> 00:32:12,280
Speaker 4: training part because they're training right training? Or is it

635
00:32:12,360 --> 00:32:12,920
Speaker 4: both both?

636
00:32:13,040 --> 00:32:16,400
Speaker 5: Yeah, that's the So the beauty about having diverse types

637
00:32:16,440 --> 00:32:19,800
Speaker 5: of compute on our grid is that once you make

638
00:32:19,880 --> 00:32:23,280
Speaker 5: the resource fungible, you can do any workload. You just

639
00:32:23,320 --> 00:32:26,479
Speaker 5: fill all the unutilized pockets with inference and then all

640
00:32:26,520 --> 00:32:28,040
Speaker 5: the reservations with training.

641
00:32:28,080 --> 00:32:33,160
Speaker 4: Yeah, so can you explain, like why is it that

642
00:32:33,400 --> 00:32:38,160
Speaker 4: every lab also seems interested right now in customized silicon,

643
00:32:38,320 --> 00:32:41,560
Speaker 4: including Microsoft announcing a chip that says like, oh, our

644
00:32:41,640 --> 00:32:44,280
Speaker 4: new ma I don't know Maya, I don't know how

645
00:32:44,320 --> 00:32:44,960
Speaker 4: it's pronounced.

646
00:32:45,840 --> 00:32:48,400
Speaker 5: I believe we had class yesterday at Stanford and he

647
00:32:48,440 --> 00:32:49,280
Speaker 5: pronounced it as Mai.

648
00:32:49,480 --> 00:32:52,080
Speaker 4: Okay, their new MAI model. And he's like, oh, we

649
00:32:52,120 --> 00:32:54,440
Speaker 4: also have a new Maya two hundred chip or something

650
00:32:54,480 --> 00:32:58,080
Speaker 4: that's optimized with it. Why is it that so many

651
00:32:58,640 --> 00:33:03,640
Speaker 4: labs or lab I guess, feel impelled to like also

652
00:33:03,760 --> 00:33:06,840
Speaker 4: design a chip that go along, goes along with the model,

653
00:33:06,920 --> 00:33:10,120
Speaker 4: and long term is what you're doing saying like this

654
00:33:10,280 --> 00:33:12,760
Speaker 4: really is not necessary to have that sort of model

655
00:33:12,800 --> 00:33:13,480
Speaker 4: chip alignment.

656
00:33:13,600 --> 00:33:17,960
Speaker 5: Yeah, there's two technological reasons and two economic reasons.

657
00:33:18,000 --> 00:33:18,240
Speaker 1: Okay.

658
00:33:18,440 --> 00:33:22,280
Speaker 5: The first is from an economic perspective, about eighty cents

659
00:33:22,320 --> 00:33:24,720
Speaker 5: of every dollar a lab spends today on their R

660
00:33:24,760 --> 00:33:28,640
Speaker 5: and D flows to a chip provider like Nvidia. Okay, right,

661
00:33:28,680 --> 00:33:30,840
Speaker 5: And so as a result, your margins are just super

662
00:33:30,880 --> 00:33:33,400
Speaker 5: super rough. So from a unit economic perspective, you want

663
00:33:33,400 --> 00:33:35,880
Speaker 5: more control over your margins. And therefore, when you look

664
00:33:35,920 --> 00:33:38,800
Speaker 5: at your unit economics, you're going, wait, wait a minute,

665
00:33:38,840 --> 00:33:41,680
Speaker 5: For every dollar we make, there's this massive chunk that's

666
00:33:41,720 --> 00:33:42,560
Speaker 5: going to somebody else.

667
00:33:42,960 --> 00:33:46,560
Speaker 4: So instead of spending eighty cents to Nvidia, you spend

668
00:33:46,560 --> 00:33:49,120
Speaker 4: seventy eight cents to t subs and keep that two

669
00:33:49,240 --> 00:33:50,000
Speaker 4: cents for yourself.

670
00:33:50,840 --> 00:33:54,080
Speaker 5: Well, I think that the better our software gets, the

671
00:33:54,120 --> 00:33:57,280
Speaker 5: more that margins should flow actually to the researcher. Okay,

672
00:33:57,320 --> 00:33:59,560
Speaker 5: because that's where the value will be captured.

673
00:33:59,600 --> 00:34:01,480
Speaker 4: But like, but wait, sorry, you were going to say,

674
00:34:01,840 --> 00:34:05,800
Speaker 4: what's the technical reason why they're trying to do optimal

675
00:34:05,880 --> 00:34:06,960
Speaker 4: model chip alignment.

676
00:34:07,040 --> 00:34:09,439
Speaker 5: On the technical side, the primary reason is you want

677
00:34:09,440 --> 00:34:12,480
Speaker 5: control over your supply chain because today in a compute

678
00:34:12,840 --> 00:34:15,279
Speaker 5: well we've been in a compute capacity constrained world now

679
00:34:15,280 --> 00:34:17,640
Speaker 5: for at least four or five years. But if you

680
00:34:17,680 --> 00:34:19,920
Speaker 5: can't get the chips you need, you're not in control

681
00:34:19,960 --> 00:34:22,040
Speaker 5: of your own supply chain. So you're dependent on computer

682
00:34:22,080 --> 00:34:26,080
Speaker 5: allocations that the compute manufacturer thinks is optimal. Right. By

683
00:34:26,120 --> 00:34:27,959
Speaker 5: the way, that's how it works at the foundry level today,

684
00:34:28,000 --> 00:34:33,040
Speaker 5: TSMC gets to decide which compute provider. Compute provider's business

685
00:34:33,040 --> 00:34:35,239
Speaker 5: grows or not, yeah, because they only have so much

686
00:34:35,239 --> 00:34:38,600
Speaker 5: production capacity. And so the technological reason is you want

687
00:34:38,680 --> 00:34:42,040
Speaker 5: supply chain independence. And so when you want economic independence,

688
00:34:42,120 --> 00:34:44,880
Speaker 5: unit economic independence, and you want supply chain independence. You

689
00:34:44,880 --> 00:34:46,719
Speaker 5: want as much control over your own chip.

690
00:34:46,840 --> 00:34:48,240
Speaker 4: But Microsoft doesn't have a FAB.

691
00:34:48,480 --> 00:34:50,480
Speaker 5: That's not what I'm saying. What I'm saying is an

692
00:34:50,520 --> 00:34:53,680
Speaker 5: inference for example, Yeah, Satcha would like more control over

693
00:34:53,719 --> 00:34:56,719
Speaker 5: his unit economics, so he's making an inference chip, right,

694
00:34:56,760 --> 00:34:58,759
Speaker 5: Because if you're dependent on a third party to give

695
00:34:58,800 --> 00:35:00,920
Speaker 5: you the inference chip, okay, and you need and if

696
00:35:00,960 --> 00:35:03,200
Speaker 5: you don't have an inferenceship, you can't sell more more

697
00:35:03,239 --> 00:35:04,960
Speaker 5: product you want more?

698
00:35:05,600 --> 00:35:08,080
Speaker 4: Is it about having a predictable supply of jest for

699
00:35:08,239 --> 00:35:11,600
Speaker 4: you rather than a predictable supply of okay?

700
00:35:11,719 --> 00:35:12,040
Speaker 5: Yes.

701
00:35:12,520 --> 00:35:15,000
Speaker 2: So there's a lot of discussion right now about more

702
00:35:15,040 --> 00:35:18,120
Speaker 2: efficient model allocation. So this idea that like, you do

703
00:35:18,200 --> 00:35:21,000
Speaker 2: not have to be using the latest model to ask

704
00:35:21,160 --> 00:35:23,360
Speaker 2: like what the weather is going to be tomorrow or

705
00:35:23,440 --> 00:35:25,960
Speaker 2: something like that. And you also don't want to blow

706
00:35:26,000 --> 00:35:28,920
Speaker 2: through your entire one year token budget in the space

707
00:35:28,960 --> 00:35:32,560
Speaker 2: of four months, as Uber apparently did. So the spikes

708
00:35:32,680 --> 00:35:36,120
Speaker 2: in usage that you're seeing that allow you to do

709
00:35:36,280 --> 00:35:39,680
Speaker 2: the system and you have grid credits, does some of

710
00:35:39,719 --> 00:35:43,440
Speaker 2: that go away if people become smarter about which models

711
00:35:43,480 --> 00:35:44,360
Speaker 2: they're actually using.

712
00:35:44,520 --> 00:35:47,200
Speaker 5: Okay, so there's an embedded assumption. I think I should

713
00:35:47,160 --> 00:35:50,000
Speaker 5: tease a part in your question. Usage is different from

714
00:35:50,040 --> 00:35:53,080
Speaker 5: the production of the model. So what's happening right in

715
00:35:53,160 --> 00:35:55,640
Speaker 5: terms of the pipeline is you use the grid to

716
00:35:55,760 --> 00:36:00,000
Speaker 5: produce the system, the model, and then the model produces tokens.

717
00:36:01,600 --> 00:36:05,680
Speaker 5: If the end user is only using tokens, then as

718
00:36:05,680 --> 00:36:08,319
Speaker 5: long as everybody we have enough diversity in the end

719
00:36:08,440 --> 00:36:11,600
Speaker 5: user base using models hosted on the grid, things actually

720
00:36:11,640 --> 00:36:14,680
Speaker 5: even out. Okay, that's cyclicality in the same way electricity

721
00:36:14,719 --> 00:36:17,719
Speaker 5: in America evens out if you have enough scale at

722
00:36:17,800 --> 00:36:20,839
Speaker 5: scale basically as part except when like.

723
00:36:20,880 --> 00:36:21,600
Speaker 4: There's a heat wave.

724
00:36:21,840 --> 00:36:24,279
Speaker 5: There's a heat wave exactly, so some of that infrastructure

725
00:36:24,840 --> 00:36:26,800
Speaker 5: we are having to reboot. But you can think about

726
00:36:26,800 --> 00:36:30,600
Speaker 5: AMP in the broadest sense as a utility company where

727
00:36:30,600 --> 00:36:33,080
Speaker 5: what's called an independent system operator of the grid. So

728
00:36:33,120 --> 00:36:35,719
Speaker 5: we don't own our own data centers, we don't own

729
00:36:35,760 --> 00:36:38,480
Speaker 5: our own labs, but we coordinate the capacity needs across

730
00:36:38,520 --> 00:36:42,800
Speaker 5: different parties and a sufficient scale that those usage patterns

731
00:36:42,840 --> 00:36:45,560
Speaker 5: actually just gets evened out. Does that make sense?

732
00:36:45,719 --> 00:36:48,520
Speaker 4: Yeah, Well you are an investor, Well, you are an

733
00:36:48,560 --> 00:36:50,680
Speaker 4: investor in an open Router I believe, which I think

734
00:36:50,719 --> 00:36:54,080
Speaker 4: is an interesting company. Do you see seting aside AMP

735
00:36:54,120 --> 00:36:57,040
Speaker 4: for a second. Do you think that there is a

736
00:36:57,040 --> 00:37:00,680
Speaker 4: at this point still within say corporate a Mariera, a

737
00:37:00,680 --> 00:37:05,240
Speaker 4: certain lack of savviness about knowing which model to route

738
00:37:05,280 --> 00:37:07,719
Speaker 4: to for the query, and that there will be an

739
00:37:07,800 --> 00:37:13,160
Speaker 4: improvement in learning within companies, within users so that you

740
00:37:13,200 --> 00:37:16,640
Speaker 4: don't have these incidents for like massive token consumption, because

741
00:37:16,680 --> 00:37:19,880
Speaker 4: perhaps everyone was using the wrong the Cadillac model and

742
00:37:19,960 --> 00:37:21,719
Speaker 4: the Ford model would have been just as fine for

743
00:37:21,760 --> 00:37:22,280
Speaker 4: that purpose.

744
00:37:22,320 --> 00:37:26,560
Speaker 5: Oh yeah, we're absolutely in the medieval ages of this technology.

745
00:37:27,280 --> 00:37:30,040
Speaker 5: I think what will happen is increasingly, based on my

746
00:37:30,080 --> 00:37:35,240
Speaker 5: conversations with corporate American leaders and corporate leaders across the world,

747
00:37:35,520 --> 00:37:38,319
Speaker 5: they don't really care about the models. They don't care

748
00:37:38,360 --> 00:37:40,560
Speaker 5: about the underlying model the technology. They just don't care.

749
00:37:40,560 --> 00:37:43,640
Speaker 5: It's too much complexity. We just want the work done.

750
00:37:43,840 --> 00:37:45,440
Speaker 5: Can you guys please figure out how to get the

751
00:37:45,440 --> 00:37:48,040
Speaker 5: work done in the cheapest way, in the most efficient way,

752
00:37:48,040 --> 00:37:50,880
Speaker 5: in the most secure and trusted way. And increasingly what

753
00:37:50,960 --> 00:37:54,920
Speaker 5: you'll find is that which particular model is helping you

754
00:37:54,960 --> 00:37:57,600
Speaker 5: out in a particular task will just be abstracted you

755
00:37:57,600 --> 00:38:00,000
Speaker 5: won't even think about that. It'll just be a companion

756
00:38:00,080 --> 00:38:01,759
Speaker 5: and you're just going to talk to It will be

757
00:38:01,760 --> 00:38:05,000
Speaker 5: a companion provided by a brand new trust, and under

758
00:38:05,040 --> 00:38:07,200
Speaker 5: the hood, it might be using the two hundred different

759
00:38:07,200 --> 00:38:10,560
Speaker 5: models to orchestrate your ask and over time that efficiency

760
00:38:10,600 --> 00:38:13,080
Speaker 5: will get better and better and better and better. And

761
00:38:13,160 --> 00:38:15,280
Speaker 5: that's why I just don't think there's only three frontier

762
00:38:15,280 --> 00:38:17,080
Speaker 5: models that are going to win. Well, it's going to

763
00:38:17,080 --> 00:38:18,000
Speaker 5: be an ecosystem.

764
00:38:18,040 --> 00:38:20,239
Speaker 2: This is I know you don't want my take, Joe, but.

765
00:38:22,360 --> 00:38:25,960
Speaker 4: Coffee, I love you take.

766
00:38:26,239 --> 00:38:28,880
Speaker 2: I'll save it for the outro. Actually, on this note,

767
00:38:29,160 --> 00:38:32,240
Speaker 2: we have seen some headlines recently. Obviously there's the Uber

768
00:38:32,280 --> 00:38:34,680
Speaker 2: one about token spending and I think it was the

769
00:38:34,719 --> 00:38:37,959
Speaker 2: CEO said he wasn't sure if the ROI was there

770
00:38:38,400 --> 00:38:41,960
Speaker 2: on Uber's AI usage. And we've seen there was a

771
00:38:41,960 --> 00:38:46,240
Speaker 2: good Vox article recently about a corporate reckoning with AI spend.

772
00:38:46,400 --> 00:38:50,000
Speaker 2: Since you're going out and talking to CEOs, do you

773
00:38:50,040 --> 00:38:53,520
Speaker 2: see any like Has anything shifted in the past couple

774
00:38:53,640 --> 00:38:56,280
Speaker 2: months or so in the way people are thinking about

775
00:38:56,320 --> 00:38:58,960
Speaker 2: the return on this initial investment or the return on

776
00:38:59,080 --> 00:39:00,680
Speaker 2: spending on tokens.

777
00:39:01,480 --> 00:39:05,040
Speaker 5: Yes, I think it's a Barbelle distribution. So there's two

778
00:39:05,080 --> 00:39:08,440
Speaker 5: types of CEOs broadly speaking. The first is the CEOs

779
00:39:08,440 --> 00:39:12,840
Speaker 5: who are using the tools themselves, and those folks are going, aha,

780
00:39:13,760 --> 00:39:17,640
Speaker 5: I understand the jagged frontier. When they understand the jagged frontier,

781
00:39:17,680 --> 00:39:21,520
Speaker 5: we talked about their strategies, their questions they ask me

782
00:39:22,200 --> 00:39:25,719
Speaker 5: are completely different from the CEOs who are outsourcing their understanding.

783
00:39:26,200 --> 00:39:30,040
Speaker 5: They're not trying the tools. They're mostly asking their kids like, hey, kiddo,

784
00:39:30,480 --> 00:39:34,080
Speaker 5: this chat GPT thing, Like it's good, right, and your

785
00:39:34,160 --> 00:39:37,120
Speaker 5: kid is like, yeah, it's pretty good, dad, And then.

786
00:39:37,400 --> 00:39:38,919
Speaker 4: Kids think it's really dumb by the way.

787
00:39:39,000 --> 00:39:41,479
Speaker 5: Yeah, So that's the other thing, right, So the kids

788
00:39:41,480 --> 00:39:44,000
Speaker 5: are super smart and they're using the tools. They're like,

789
00:39:44,080 --> 00:39:46,200
Speaker 5: it's good at this thing, but not at that. So

790
00:39:46,239 --> 00:39:47,759
Speaker 5: they understand the jagged frontier part.

791
00:39:48,040 --> 00:39:50,440
Speaker 4: You know what, they think I'm dumb for using it.

792
00:39:50,520 --> 00:39:53,240
Speaker 4: They're like, Dad, like you're not doing anything smart.

793
00:39:53,239 --> 00:39:53,480
Speaker 5: You know.

794
00:39:54,040 --> 00:39:55,439
Speaker 4: They don't think the models are dumb.

795
00:39:55,440 --> 00:39:58,000
Speaker 5: They they might be exactly, they might be going the

796
00:39:58,040 --> 00:40:00,720
Speaker 5: way you're using it exactly not optimal.

797
00:40:01,600 --> 00:40:03,960
Speaker 4: So so the you know what I'm saying is my

798
00:40:04,040 --> 00:40:06,080
Speaker 4: kids are sick and they have no intent and they

799
00:40:06,120 --> 00:40:08,240
Speaker 4: have no idea about anything, and they just think I'm dumb.

800
00:40:08,440 --> 00:40:11,680
Speaker 4: That's the that just might be a generalizable that's really

801
00:40:11,719 --> 00:40:13,120
Speaker 4: the I see.

802
00:40:13,200 --> 00:40:15,160
Speaker 5: Okay, well you can send them over to me anytime.

803
00:40:15,200 --> 00:40:16,320
Speaker 5: I'm happy to be the fun uncle.

804
00:40:16,480 --> 00:40:18,520
Speaker 4: Yeah, that would be great. You can show them that

805
00:40:18,600 --> 00:40:19,680
Speaker 4: actually this is fun to.

806
00:40:20,120 --> 00:40:22,839
Speaker 5: My wife and I am happy to host hert That's

807
00:40:22,840 --> 00:40:25,879
Speaker 5: really what I'm trying to get it. It's it's the summer, right.

808
00:40:25,880 --> 00:40:27,480
Speaker 5: We have two nieces in London and we call it

809
00:40:27,520 --> 00:40:30,240
Speaker 5: Camp Mida Shan. My last name is Mida, my wife's

810
00:40:30,280 --> 00:40:32,960
Speaker 5: name is Shan, and so you're you're welcome to send

811
00:40:32,960 --> 00:40:34,120
Speaker 5: them to camp anytime.

812
00:40:34,640 --> 00:40:36,360
Speaker 3: That's amazing.

813
00:40:36,480 --> 00:40:39,000
Speaker 5: But that's that's the bifurcation. As leaders who are actually

814
00:40:39,000 --> 00:40:41,680
Speaker 5: trying the tools out, they realize they're extraordinary at some

815
00:40:41,760 --> 00:40:44,920
Speaker 5: things and not at others. And so the depending on

816
00:40:44,960 --> 00:40:47,880
Speaker 5: whether you get it or not, or you're actually getting

817
00:40:47,880 --> 00:40:50,239
Speaker 5: your hands dirty or not, I find the questions are

818
00:40:50,239 --> 00:40:50,880
Speaker 5: completely different.

819
00:40:50,960 --> 00:40:54,319
Speaker 4: So this has been incredibly helpful conversation in terms of

820
00:40:54,360 --> 00:40:58,759
Speaker 4: like understanding basically the problem of it's actually tons of

821
00:40:58,760 --> 00:41:01,680
Speaker 4: money is being spent, and your thesis is that it's

822
00:41:01,719 --> 00:41:06,759
Speaker 4: massively suboptimally used up and down the stack. You mentioned this, Okay,

823
00:41:06,840 --> 00:41:10,000
Speaker 4: you get a credit et cetera. Like, do you actually

824
00:41:10,000 --> 00:41:13,680
Speaker 4: see that being financialized in a way? I mean, okay,

825
00:41:13,680 --> 00:41:15,799
Speaker 4: I'm I bought this capacity. I have a lot of

826
00:41:15,840 --> 00:41:18,600
Speaker 4: un new time. I mean, don't always have a research

827
00:41:18,640 --> 00:41:21,680
Speaker 4: idea that it's going to require a big model, run

828
00:41:22,080 --> 00:41:24,640
Speaker 4: test I can resell that. Is that something that you

829
00:41:24,680 --> 00:41:27,720
Speaker 4: see like something that genuinely resembles a financial market?

830
00:41:28,040 --> 00:41:31,919
Speaker 5: I hope not, because when you had when you add

831
00:41:31,920 --> 00:41:37,920
Speaker 5: speculation to you know, production goods, it creates scarcity of

832
00:41:37,920 --> 00:41:40,760
Speaker 5: a different kind, right, because then you have financial traders

833
00:41:40,840 --> 00:41:44,600
Speaker 5: and markets trying to trade the speculative value of the asset,

834
00:41:45,080 --> 00:41:47,000
Speaker 5: and that's going to hurt a lot of our research

835
00:41:47,000 --> 00:41:49,920
Speaker 5: teams in technology. On the other hand, I think that

836
00:41:50,000 --> 00:41:53,400
Speaker 5: creates a need for innovation inside of the research teams.

837
00:41:53,480 --> 00:41:56,080
Speaker 5: And so one of the one of the cooperating functions

838
00:41:56,080 --> 00:41:59,960
Speaker 5: we have inside of our business is a forecasting capability

839
00:42:00,080 --> 00:42:02,320
Speaker 5: where we have a team that's very similar to actually

840
00:42:02,480 --> 00:42:04,680
Speaker 5: the kind of forecasting team you'd have inside of a

841
00:42:04,680 --> 00:42:07,640
Speaker 5: hedge fund. We're constantly predicting demand and supply, and then

842
00:42:07,680 --> 00:42:10,799
Speaker 5: we're actually procuring capacity in advance through call options. On

843
00:42:10,840 --> 00:42:13,919
Speaker 5: compute clusters. But our needs are similar to the kind

844
00:42:13,960 --> 00:42:17,520
Speaker 5: of internal trading desk you'd have inside of a large

845
00:42:17,880 --> 00:42:20,480
Speaker 5: steel company right where they need to lock up iron

846
00:42:20,520 --> 00:42:23,800
Speaker 5: ore and so on for their production needs. So I'm

847
00:42:23,840 --> 00:42:27,000
Speaker 5: a big fan of efficient markets, and I'm trying to

848
00:42:27,000 --> 00:42:29,839
Speaker 5: actively invest in and help entrepreneurs out and teams out

849
00:42:29,800 --> 00:42:31,880
Speaker 5: who are trying to drive more efficiency in the service

850
00:42:31,920 --> 00:42:35,279
Speaker 5: of more productivity in science and engineering. I'm not that

851
00:42:36,000 --> 00:42:40,120
Speaker 5: thrilled about the financialization of these products if it ultimately

852
00:42:40,200 --> 00:42:42,520
Speaker 5: results in more speculation. Does that make sense?

853
00:42:42,800 --> 00:42:46,160
Speaker 2: Yeah, I'm just curious since you're tracking demand in that way,

854
00:42:46,280 --> 00:42:48,880
Speaker 2: Like if you were going to describe the slope of

855
00:42:48,960 --> 00:42:51,279
Speaker 2: demand right now versus say.

856
00:42:51,080 --> 00:42:52,880
Speaker 3: Like a year ago, is it steeper?

857
00:42:53,040 --> 00:42:56,040
Speaker 2: Is it starting to plateau perpendicular?

858
00:42:56,520 --> 00:42:56,880
Speaker 5: Wow?

859
00:42:56,960 --> 00:42:57,280
Speaker 2: Okay.

860
00:42:57,560 --> 00:43:00,399
Speaker 5: If you look at the compute prices of long term

861
00:43:00,440 --> 00:43:04,840
Speaker 5: rentals over the last six months from between January and now,

862
00:43:05,280 --> 00:43:07,960
Speaker 5: they're trading up du x. So we start, for example,

863
00:43:07,960 --> 00:43:12,080
Speaker 5: for twenty twenty six, we started securing our capacity in

864
00:43:12,160 --> 00:43:15,200
Speaker 5: January at these long term rates, we could resell that

865
00:43:15,600 --> 00:43:18,080
Speaker 5: at a duo X markup if we wanted to.

866
00:43:18,600 --> 00:43:21,240
Speaker 4: Part of the reason that twenty twenty six has become

867
00:43:21,680 --> 00:43:25,480
Speaker 4: just totally AI has consumed everyone's mind. I think is

868
00:43:25,520 --> 00:43:29,920
Speaker 4: because people got very excited about Claude code specifically. But

869
00:43:29,960 --> 00:43:33,000
Speaker 4: that was a breakthrough at the harness level, not the

870
00:43:33,080 --> 00:43:36,840
Speaker 4: model level, right suddenly, like the really exciting like wow,

871
00:43:36,880 --> 00:43:39,080
Speaker 4: this is just so fun, it's just so easy of

872
00:43:39,120 --> 00:43:42,719
Speaker 4: a computer inside your computer. That was a harness breakthrough.

873
00:43:43,520 --> 00:43:47,400
Speaker 4: Do you see, like when you think about investment among

874
00:43:47,520 --> 00:43:52,000
Speaker 4: AI labs, do you see any shift in allocation away

875
00:43:52,200 --> 00:43:56,880
Speaker 4: from pure scaling and improving the model towards sort of

876
00:43:56,920 --> 00:43:59,680
Speaker 4: like tooling and harnesses as a way to get more

877
00:43:59,800 --> 00:44:01,120
Speaker 4: gu out of the models.

878
00:44:01,280 --> 00:44:03,200
Speaker 5: No, I'm sorry, I have to correct you there. It

879
00:44:03,280 --> 00:44:06,200
Speaker 5: was not just a harness innovation. Those two things go

880
00:44:06,280 --> 00:44:09,960
Speaker 5: hand in hand. It's a symphony of improvement between It's

881
00:44:09,960 --> 00:44:12,640
Speaker 5: a dialectic between the model capability and the harness. That

882
00:44:12,640 --> 00:44:16,239
Speaker 5: harness was designed specifically for the capabilities that the new

883
00:44:16,239 --> 00:44:18,560
Speaker 5: model was going to have, and so when you design

884
00:44:18,680 --> 00:44:22,320
Speaker 5: these things in the industry, we call this code design. Okay,

885
00:44:22,440 --> 00:44:26,040
Speaker 5: so you have the harness designed side by side with

886
00:44:26,200 --> 00:44:29,200
Speaker 5: the researcher who's designed the next generation capabilities in the model,

887
00:44:29,480 --> 00:44:31,120
Speaker 5: and you get a little bit of visibility and where

888
00:44:31,160 --> 00:44:32,800
Speaker 5: the model is going to be good because as I

889
00:44:32,840 --> 00:44:35,920
Speaker 5: described earlier, the pipeline is actually quite predictable pre training,

890
00:44:36,040 --> 00:44:39,760
Speaker 5: mid training, continuous feedback loop. Once you have that visibility,

891
00:44:39,760 --> 00:44:44,200
Speaker 5: you go, aha, we specifically want to improve the capabilities

892
00:44:44,200 --> 00:44:46,600
Speaker 5: on this type of task. It's going to take us

893
00:44:46,600 --> 00:44:49,560
Speaker 5: about three months to get there, start designing the harness

894
00:44:49,719 --> 00:44:52,759
Speaker 5: for that improvement. By the time they show up, then

895
00:44:52,800 --> 00:44:56,160
Speaker 5: you can have the harness assume that the model will

896
00:44:56,160 --> 00:44:58,000
Speaker 5: be able to do X, y Z on its own,

897
00:44:58,320 --> 00:45:00,840
Speaker 5: whereas ABC is going to need the third party tools.

898
00:45:01,040 --> 00:45:03,799
Speaker 5: So then the harness says, remember that three months ago,

899
00:45:04,120 --> 00:45:07,719
Speaker 5: you were terrible at understanding a spreadsheet. Yeah, so then

900
00:45:07,760 --> 00:45:10,239
Speaker 5: we had to go right, like, go use a third

901
00:45:10,239 --> 00:45:14,280
Speaker 5: party tool to use a spreadsheet. In the last three months,

902
00:45:14,320 --> 00:45:17,160
Speaker 5: what we've done is added the ability to actually reason

903
00:45:17,160 --> 00:45:19,080
Speaker 5: about a spreadsheet in the model and the model not.

904
00:45:19,200 --> 00:45:21,439
Speaker 5: So now you don't need to use a third party spreadsheet, okay.

905
00:45:21,440 --> 00:45:24,120
Speaker 5: And so then the harness gets updated to say, don't

906
00:45:24,200 --> 00:45:26,200
Speaker 5: go out and use a third party spreadsheet, which, by

907
00:45:26,200 --> 00:45:28,319
Speaker 5: the way, collapses the time required to do that task

908
00:45:28,400 --> 00:45:31,240
Speaker 5: by like sometimes a minute to two minutes. Now suddenly

909
00:45:31,280 --> 00:45:34,359
Speaker 5: I've improved the user experience, and that's when things really sing.

910
00:45:34,800 --> 00:45:37,040
Speaker 5: It's when both of those those parts, the model and

911
00:45:37,080 --> 00:45:40,440
Speaker 5: the harness are co designed to create a symphony. Does

912
00:45:40,480 --> 00:45:41,160
Speaker 5: that make sense? Yeah?

913
00:45:41,160 --> 00:45:46,120
Speaker 2: Absolutely, all right, Anjinay Mitta of AMP PBC, thank you

914
00:45:46,160 --> 00:45:48,319
Speaker 2: so much for coming on od lots really appreciate it.

915
00:45:48,320 --> 00:45:49,080
Speaker 5: Thanks for having me.

916
00:45:49,000 --> 00:45:51,520
Speaker 2: And everyone go out and check out the Stanford Lecture series.

917
00:45:51,560 --> 00:45:52,600
Speaker 2: It's on YouTube right.

918
00:45:52,480 --> 00:45:55,480
Speaker 5: It is CS one fifty three dot Stanford dot edu. Perfect.

919
00:45:55,600 --> 00:45:57,560
Speaker 4: I'll I have a big flight coming up, so I'll

920
00:45:57,600 --> 00:45:57,880
Speaker 4: watch it.

921
00:45:57,920 --> 00:45:59,879
Speaker 5: Then you should download all the lectures. It's quite a few.

922
00:46:00,120 --> 00:46:02,200
Speaker 4: Thank you so much. ONNJ. That was fantastic.

923
00:46:02,239 --> 00:46:17,680
Speaker 3: That was great, all right, Joe, that was a great discussion.

924
00:46:17,880 --> 00:46:20,719
Speaker 2: Yeah, I should emphasize just how big a deal that

925
00:46:20,800 --> 00:46:24,120
Speaker 2: lecture series actually is at Stanford, Like students are beating

926
00:46:24,120 --> 00:46:26,719
Speaker 2: down the door basically to get into that. And if

927
00:46:26,760 --> 00:46:29,000
Speaker 2: it's free on YouTube, you should definitely check it out.

928
00:46:29,400 --> 00:46:32,319
Speaker 4: I just want to establish that if I had given

929
00:46:32,360 --> 00:46:34,680
Speaker 4: you the Avian simulation that I didn't want to hear

930
00:46:34,719 --> 00:46:37,160
Speaker 4: your take, or be the idea that I would have

931
00:46:37,200 --> 00:46:40,240
Speaker 4: wanted to hear Anna's take instead of yours, I want.

932
00:46:40,040 --> 00:46:42,800
Speaker 2: To hear your take no, it's fine, Joe. I realized

933
00:46:42,840 --> 00:46:45,319
Speaker 2: that most listeners are here for the guest takes. I

934
00:46:45,360 --> 00:46:48,200
Speaker 2: get it, but I thought his point about the jagged

935
00:46:48,200 --> 00:46:51,279
Speaker 2: front frontier was an important one. And this idea that

936
00:46:51,360 --> 00:46:54,040
Speaker 2: like maybe the future it's not going to be a

937
00:46:54,080 --> 00:46:56,840
Speaker 2: winner takes all things. In terms of models, you're going

938
00:46:56,880 --> 00:46:59,480
Speaker 2: to have a bunch of different models doing different things

939
00:46:59,560 --> 00:47:02,680
Speaker 2: that might suit different companies. And also the idea that

940
00:47:02,719 --> 00:47:05,200
Speaker 2: like a lot of companies aren't going to care about

941
00:47:05,239 --> 00:47:08,320
Speaker 2: which specific model they're using, they just want the cheapest

942
00:47:08,320 --> 00:47:12,680
Speaker 2: one that basically gets the job done. In my mind,

943
00:47:13,040 --> 00:47:17,759
Speaker 2: that sounds like more of a is commodified the yeah,

944
00:47:17,800 --> 00:47:21,560
Speaker 2: a commodified market, right, rather than like, oh, people are

945
00:47:21,640 --> 00:47:24,759
Speaker 2: going to pay up for as you said, the Cadillac.

946
00:47:24,880 --> 00:47:28,960
Speaker 4: Well, so what I would say is, by in listening

947
00:47:29,200 --> 00:47:32,399
Speaker 4: to Angina and amp is that people will want a

948
00:47:32,520 --> 00:47:36,680
Speaker 4: commodified service, but that under the hood, I mean, this

949
00:47:36,840 --> 00:47:39,000
Speaker 4: just sounds like what she's really trying to solve. And

950
00:47:39,040 --> 00:47:42,600
Speaker 4: it's very interesting I as a user or a company

951
00:47:42,640 --> 00:47:47,440
Speaker 4: by a commodified service. But under the hood, the commodity

952
00:47:47,640 --> 00:47:50,879
Speaker 4: has an incredible amount of variety of models through which

953
00:47:50,920 --> 00:47:53,440
Speaker 4: it can rouse sure some of which will be the Cadillac,

954
00:47:53,800 --> 00:47:55,400
Speaker 4: some of it will some of it will be the

955
00:47:55,480 --> 00:47:56,399
Speaker 4: current coffee cup.

956
00:47:56,719 --> 00:48:00,720
Speaker 2: Yeah, absolutely, But like my point is maybe in terms

957
00:48:00,719 --> 00:48:04,640
Speaker 2: of valuations, sure, right, Like if everyone is assuming that

958
00:48:04,719 --> 00:48:07,680
Speaker 2: the Cadillac is going to be like the one that

959
00:48:07,840 --> 00:48:10,480
Speaker 2: everyone is going to get, and the total available market

960
00:48:10,520 --> 00:48:13,880
Speaker 2: the TAM infamously is like not just the world, but

961
00:48:13,920 --> 00:48:17,160
Speaker 2: potentially the universe. Like that seems a stretch to.

962
00:48:17,120 --> 00:48:20,280
Speaker 4: Me totally, and just generally I thought it was super

963
00:48:20,320 --> 00:48:22,680
Speaker 4: interesting and the idea this is we've done a couple

964
00:48:22,760 --> 00:48:27,359
Speaker 4: episodes recently, specifically learning more about both chip level and

965
00:48:27,560 --> 00:48:31,560
Speaker 4: box level optimizations, both how many chips you're using and

966
00:48:31,640 --> 00:48:34,719
Speaker 4: how well you're using ad chip Definitely when more to.

967
00:48:34,680 --> 00:48:35,160
Speaker 5: Do on that.

968
00:48:35,360 --> 00:48:37,320
Speaker 2: It still blows my mind that this is a problem

969
00:48:37,320 --> 00:48:41,800
Speaker 2: that can be solved with software rather than like something physical.

970
00:48:42,440 --> 00:48:45,480
Speaker 2: You just come up with a way to efficiently allocate

971
00:48:45,680 --> 00:48:48,640
Speaker 2: the compute. Yeah, because in my mind, like it's it's

972
00:48:48,680 --> 00:48:51,080
Speaker 2: such a physical problem. And we've talked to you know,

973
00:48:51,440 --> 00:48:55,759
Speaker 2: previous DEI market participants like Brandon McBee at Core. We've

974
00:48:56,040 --> 00:48:58,319
Speaker 2: and they talk about like, oh, it's difficult to standardize

975
00:48:58,360 --> 00:49:01,200
Speaker 2: because of the configurations of chip and things like that.

976
00:49:01,320 --> 00:49:03,759
Speaker 2: But if you could solve it just through a software system.

977
00:49:04,200 --> 00:49:06,919
Speaker 3: That's pretty crazy. I guess Google's already done it. Yeah,

978
00:49:07,000 --> 00:49:07,960
Speaker 3: all right, shall we leave it there?

979
00:49:08,040 --> 00:49:08,759
Speaker 4: Let's save it there.

980
00:49:09,160 --> 00:49:11,880
Speaker 2: This has been another episode of the Odd Lots podcast.

981
00:49:11,960 --> 00:49:15,040
Speaker 2: I'm Tracy Alloway. You can follow me at Tracy Alloway.

982
00:49:14,920 --> 00:49:18,040
Speaker 4: And I'm Jill Wisenthal. You can follow me at the Stalwart.

983
00:49:18,239 --> 00:49:21,080
Speaker 4: Follow our guest A Mida at a Jena Midda, Follow

984
00:49:21,080 --> 00:49:24,320
Speaker 4: our producers Carmen Rodriguez at Carmen armat Dash, Ol Bennett

985
00:49:24,360 --> 00:49:28,319
Speaker 4: at Dashbod, Calebrooks at Calebrooks, and Kevin Lozano at Kevin

986
00:49:28,400 --> 00:49:29,120
Speaker 4: Lloyd Lozano.

987
00:49:29,440 --> 00:49:31,440
Speaker 2: And for more Odd Lots content, you should check out

988
00:49:31,480 --> 00:49:33,879
Speaker 2: our daily newsletter. You can find that at Bloomberg dot

989
00:49:33,920 --> 00:49:35,640
Speaker 2: com forward slash odd Lots.

990
00:49:35,400 --> 00:49:37,360
Speaker 4: And you can shout about all of these topics twenty

991
00:49:37,360 --> 00:49:40,880
Speaker 4: four to seven in our discord Discord dot gg slash

992
00:49:40,880 --> 00:49:41,359
Speaker 4: od loots.

993
00:49:41,600 --> 00:49:44,400
Speaker 2: And if you enjoyed this conversation, then please leave a

994
00:49:44,440 --> 00:49:47,160
Speaker 2: comment or like the video, or better yet, subscribe.

995
00:49:47,200 --> 00:50:04,880
Speaker 4: Thanks for listening in