1
00:00:02,720 --> 00:00:15,840
Speaker 1: Bloomberg Audio Studios, Podcasts, Radio News.

2
00:00:18,600 --> 00:00:22,120
Speaker 2: Hello and welcome to another episode of The Odd Lots Podcast.

3
00:00:22,160 --> 00:00:23,640
Speaker 3: I'm Joe Wisenthal and.

4
00:00:23,600 --> 00:00:24,560
Speaker 4: I'm Tracy Alloway.

5
00:00:24,640 --> 00:00:25,400
Speaker 3: Tracy, I don't know.

6
00:00:25,440 --> 00:00:27,720
Speaker 2: I think our listeners like it, but you know, a

7
00:00:27,760 --> 00:00:29,800
Speaker 2: lot of our episodes are about AI these days.

8
00:00:29,840 --> 00:00:32,360
Speaker 3: But to be fair to us, it's a pretty big topic.

9
00:00:32,440 --> 00:00:34,839
Speaker 4: That's all anyone wants to talk about. Whenever we go

10
00:00:34,920 --> 00:00:37,479
Speaker 4: to dinners with sources and things and people who are

11
00:00:37,479 --> 00:00:40,159
Speaker 4: not even directly in the tech industry. You know, they

12
00:00:40,240 --> 00:00:42,600
Speaker 4: might be in market, they might be in policy and economics.

13
00:00:42,640 --> 00:00:44,440
Speaker 4: All they want to talk about is AI, and then

14
00:00:44,479 --> 00:00:48,840
Speaker 4: inevitably the conversation veers into very sci fi territory where

15
00:00:48,840 --> 00:00:51,960
Speaker 4: we all start talking about the human extinction scenario. And

16
00:00:52,000 --> 00:00:53,320
Speaker 4: that's just the norm nowadays.

17
00:00:53,360 --> 00:00:54,080
Speaker 3: I know, it's so weird.

18
00:00:54,240 --> 00:00:57,320
Speaker 2: You know, we're in Hong Kong recently, and when we

19
00:00:57,320 --> 00:01:00,200
Speaker 2: were in Hong Kong, this was before it was an

20
00:01:00,320 --> 00:01:01,920
Speaker 2: that there was a deal to open the straight before

21
00:01:01,960 --> 00:01:05,240
Speaker 2: moves in East Asia was considered to be like ground

22
00:01:05,319 --> 00:01:07,160
Speaker 2: zero for where the effects should be felt of the

23
00:01:07,200 --> 00:01:09,480
Speaker 2: oil and jet fuel crisis, et cetera. And we're at

24
00:01:09,520 --> 00:01:11,640
Speaker 2: this dinner of business people, like, they were not talking

25
00:01:11,640 --> 00:01:12,559
Speaker 2: about that at all.

26
00:01:13,440 --> 00:01:14,680
Speaker 4: The terminator scenario.

27
00:01:14,720 --> 00:01:17,680
Speaker 2: They just want to talk about token consumption and all

28
00:01:17,720 --> 00:01:20,000
Speaker 2: of these things like here we are. It's like, wait,

29
00:01:20,040 --> 00:01:21,720
Speaker 2: aren't you guys supposed to be like under all kinds

30
00:01:21,760 --> 00:01:22,679
Speaker 2: of jet fuel stress.

31
00:01:22,880 --> 00:01:24,640
Speaker 3: So this is our defense for.

32
00:01:24,640 --> 00:01:28,760
Speaker 4: The AI episodes. I think it's fair. I will also

33
00:01:28,840 --> 00:01:31,640
Speaker 4: say when we did the quiz in Hong Kong, we

34
00:01:31,720 --> 00:01:34,360
Speaker 4: had a bunch of different teams with very creative names.

35
00:01:34,400 --> 00:01:36,240
Speaker 3: Separated value human capital.

36
00:01:36,319 --> 00:01:36,920
Speaker 4: That was a great one.

37
00:01:36,959 --> 00:01:38,400
Speaker 3: They won the turn, they won the question.

38
00:01:38,120 --> 00:01:40,959
Speaker 4: They won proving that there is value in human capital.

39
00:01:41,000 --> 00:01:43,720
Speaker 4: But did you see that one of the tables was

40
00:01:43,800 --> 00:01:49,360
Speaker 4: called fable thirteen thirteen fable thirteen, which was very topical

41
00:01:49,440 --> 00:01:50,360
Speaker 4: at that moment.

42
00:01:50,200 --> 00:01:50,760
Speaker 3: Very topical.

43
00:01:50,800 --> 00:01:53,640
Speaker 2: We're recording this on June seventeenth, and of course there's

44
00:01:53,640 --> 00:01:56,480
Speaker 2: a lot in the news these days, but things move

45
00:01:56,560 --> 00:01:59,800
Speaker 2: very fascinating. Even if there weren't governmental controversies and all

46
00:01:59,840 --> 00:02:02,000
Speaker 2: that stuff, you would have to mark the data and

47
00:02:02,080 --> 00:02:05,120
Speaker 2: AI because of how fast breakthroughs happen. But you know,

48
00:02:05,160 --> 00:02:07,000
Speaker 2: as you said, like AI sort of feels like the

49
00:02:07,000 --> 00:02:09,520
Speaker 2: most important thing than anything else. But that's a very

50
00:02:09,560 --> 00:02:12,080
Speaker 2: conventional wisdom. It was not always conventional wisdom. And I

51
00:02:12,080 --> 00:02:12,799
Speaker 2: have a DM.

52
00:02:12,960 --> 00:02:15,799
Speaker 3: I know you're not supposed to share dms for public,

53
00:02:15,800 --> 00:02:16,640
Speaker 3: but I have a DM.

54
00:02:16,560 --> 00:02:17,360
Speaker 4: Got the receipts.

55
00:02:17,400 --> 00:02:18,200
Speaker 3: I have the receipts.

56
00:02:18,280 --> 00:02:21,160
Speaker 2: August second, twenty sixteen, and I DM A colleague Astad,

57
00:02:21,160 --> 00:02:23,760
Speaker 2: did you leave Bloomberg? He says, yes, I'll be announcing

58
00:02:23,800 --> 00:02:26,120
Speaker 2: publicly in a bit. Take a couple of months to

59
00:02:26,160 --> 00:02:29,000
Speaker 2: study a properly than leaving journalism to do something else.

60
00:02:29,040 --> 00:02:30,080
Speaker 2: Still connected to AI.

61
00:02:30,720 --> 00:02:32,280
Speaker 3: Being our Google reporter was a great thing.

62
00:02:32,320 --> 00:02:34,280
Speaker 4: They're still connected to AIG.

63
00:02:34,520 --> 00:02:38,440
Speaker 2: Then the final August second, twenty sixteen, but AI is

64
00:02:38,480 --> 00:02:40,720
Speaker 2: more important than anything else, so I felt best to

65
00:02:40,720 --> 00:02:41,160
Speaker 2: sort of.

66
00:02:41,040 --> 00:02:43,280
Speaker 3: Optimize for that above all those. So then I just said, well,

67
00:02:43,280 --> 00:02:43,760
Speaker 3: I'm good luck.

68
00:02:43,840 --> 00:02:46,960
Speaker 4: This is someone who truly learned from their sources, unlike

69
00:02:47,120 --> 00:02:49,640
Speaker 4: us who remain in the podcasting industry.

70
00:02:49,919 --> 00:02:53,000
Speaker 2: So anyway, that person who would that DM was a

71
00:02:53,040 --> 00:02:55,959
Speaker 2: former Bloomberg reporter Jack Clark, who's one of our guests today.

72
00:02:56,000 --> 00:02:58,400
Speaker 2: He is the head of Public Benefit and co founder

73
00:02:58,480 --> 00:03:02,079
Speaker 2: of Anthropic years later, and also Peter McCrory, head of

74
00:03:02,120 --> 00:03:05,360
Speaker 2: economics at Anthropics, So too perfect guests to talk about

75
00:03:05,400 --> 00:03:08,760
Speaker 2: all the things and AI these days. So Peter and Jack,

76
00:03:08,800 --> 00:03:10,320
Speaker 2: thank you so much for coming on the podcast.

77
00:03:10,360 --> 00:03:12,720
Speaker 5: Great to be back. I'm glad I optimized my life.

78
00:03:12,800 --> 00:03:15,960
Speaker 2: Yeah, well to one of the calls of the century.

79
00:03:16,000 --> 00:03:19,040
Speaker 2: So what I actually start with that it's easy to

80
00:03:19,080 --> 00:03:20,880
Speaker 2: say in twenty twenty six there, yeah, will be a

81
00:03:20,880 --> 00:03:23,440
Speaker 2: big deal. You called your shot, you got it right,

82
00:03:23,600 --> 00:03:27,200
Speaker 2: twenty sixteen. What did you see in August twenty sixteen,

83
00:03:27,280 --> 00:03:29,880
Speaker 2: or presumably before They're like, oh, you know what, this

84
00:03:30,120 --> 00:03:31,680
Speaker 2: is the biggest story of our lives.

85
00:03:31,720 --> 00:03:34,240
Speaker 5: So for two years when I was reporting at Bloomberg,

86
00:03:34,320 --> 00:03:37,280
Speaker 5: I wasted a lot of mister Bloomberg's printing by printing

87
00:03:37,320 --> 00:03:41,000
Speaker 5: out archive papers about AI research. And what I started

88
00:03:41,000 --> 00:03:43,680
Speaker 5: to do very Bloombergian thing is I started to make

89
00:03:43,720 --> 00:03:48,360
Speaker 5: graphs charting AI progress over time, measurements of things like

90
00:03:48,400 --> 00:03:51,680
Speaker 5: computer vision, measurements of things like the skill with which

91
00:03:51,720 --> 00:03:55,240
Speaker 5: AI agents were able to compete and play ATARI games.

92
00:03:55,440 --> 00:03:57,480
Speaker 5: And what I saw in these graphs was the beginning

93
00:03:57,560 --> 00:04:00,520
Speaker 5: of an exponential yeah, and it was everywhere, like if

94
00:04:00,520 --> 00:04:03,800
Speaker 5: you looked at vision or sound or video or gameplaying,

95
00:04:03,800 --> 00:04:06,840
Speaker 5: you saw the same trend. And it became obvious to

96
00:04:06,880 --> 00:04:09,280
Speaker 5: me that this was a general purpose technology if it

97
00:04:09,320 --> 00:04:12,120
Speaker 5: was right at a start. My one bone that I

98
00:04:12,240 --> 00:04:14,000
Speaker 5: have to pick with Bloomberg, which I which I'm going

99
00:04:14,040 --> 00:04:16,560
Speaker 5: to use my privilege just mentioned on air. I never

100
00:04:16,600 --> 00:04:19,799
Speaker 5: got us to write a story saying Navidia was being

101
00:04:19,920 --> 00:04:22,640
Speaker 5: used in every single day research paper and I pitched

102
00:04:22,640 --> 00:04:24,479
Speaker 5: it and I failed to get it across the line

103
00:04:24,480 --> 00:04:25,120
Speaker 5: before I left.

104
00:04:25,520 --> 00:04:28,520
Speaker 4: Man, I can just imagine you reading all these academic papers.

105
00:04:28,720 --> 00:04:35,400
Speaker 4: Meanwhile the editor is like, we need the BF Yeah, okay,

106
00:04:35,400 --> 00:04:38,760
Speaker 4: And Peter, I'm very interested in you know, Anthropic. It's

107
00:04:38,800 --> 00:04:41,280
Speaker 4: a company and trying to make money, and yet it

108
00:04:41,360 --> 00:04:45,880
Speaker 4: has this economics lab. What's the idea behind having an

109
00:04:45,920 --> 00:04:49,680
Speaker 4: economics research body within a company that's developing this technology.

110
00:04:50,200 --> 00:04:52,240
Speaker 6: So I mean I was late to the game and

111
00:04:52,480 --> 00:04:55,719
Speaker 6: joining I joined just a year ago, but I had.

112
00:04:55,720 --> 00:04:56,320
Speaker 3: A year ago.

113
00:04:56,400 --> 00:04:58,560
Speaker 2: People, well whatever, we all know about how much the

114
00:04:58,560 --> 00:05:00,320
Speaker 2: stock is a price of a year, but you're not

115
00:05:00,960 --> 00:05:01,280
Speaker 2: go on.

116
00:05:01,839 --> 00:05:04,400
Speaker 6: I think what was very evident. So I'm an applied

117
00:05:04,440 --> 00:05:08,040
Speaker 6: macroeconomist by training and have tried to understand various types

118
00:05:08,080 --> 00:05:10,640
Speaker 6: of shocks throughout the economy. Part of what drew me

119
00:05:10,680 --> 00:05:13,320
Speaker 6: to Anthropic was it was evident to me last year

120
00:05:13,640 --> 00:05:17,680
Speaker 6: that they cared very deeply about not just advancing the technology,

121
00:05:17,720 --> 00:05:20,479
Speaker 6: but making sense of how it is set to reshape

122
00:05:20,520 --> 00:05:24,200
Speaker 6: the labor market, its impact on productivity, on growth, and

123
00:05:24,400 --> 00:05:28,040
Speaker 6: be willing to put evidence, data and research out into

124
00:05:28,120 --> 00:05:33,000
Speaker 6: the world that would be broadly beneficial and useful to society.

125
00:05:33,000 --> 00:05:35,280
Speaker 6: And I thought, I want to be a part of

126
00:05:35,360 --> 00:05:38,800
Speaker 6: building that economic research program and do what I can

127
00:05:39,000 --> 00:05:42,760
Speaker 6: to provide tentative answers to the most pressing questions. We

128
00:05:42,839 --> 00:05:45,640
Speaker 6: might not always get it right, but ideally we're helping

129
00:05:45,680 --> 00:05:47,200
Speaker 6: society make sense of the change.

130
00:05:47,360 --> 00:05:49,880
Speaker 2: The capabilities of the models and all kinds of things

131
00:05:49,880 --> 00:05:54,839
Speaker 2: are extraordinary, I mean just mind blowing every coding, copyright.

132
00:05:54,560 --> 00:05:55,480
Speaker 3: All kinds of things.

133
00:05:56,000 --> 00:06:00,480
Speaker 2: Actually, why in June twenty twenty six does life Field

134
00:06:00,520 --> 00:06:03,440
Speaker 2: maybe as normal as it does from an economic perspective?

135
00:06:03,800 --> 00:06:06,440
Speaker 6: This is a great question and one that I've been

136
00:06:06,800 --> 00:06:09,560
Speaker 6: wrestling with. I think there are a number of reasons

137
00:06:09,560 --> 00:06:13,440
Speaker 6: why you might think that the impact has not yet materialized. One,

138
00:06:13,839 --> 00:06:16,760
Speaker 6: the technology can advance, but it also then needs to

139
00:06:16,800 --> 00:06:20,599
Speaker 6: diffuse throughout the economy, and there can be bottlenecks from

140
00:06:20,720 --> 00:06:24,400
Speaker 6: moving from capabilities to actual deployment. We see that with

141
00:06:24,520 --> 00:06:27,680
Speaker 6: our enterprise customers. So if you want to automate biological

142
00:06:27,720 --> 00:06:31,560
Speaker 6: research or some other very complicated financial modeling task, you

143
00:06:31,600 --> 00:06:34,360
Speaker 6: need a lot of contextual information available to the model.

144
00:06:34,760 --> 00:06:38,120
Speaker 6: If you don't have that contextual information, the capabilities alone

145
00:06:38,200 --> 00:06:42,240
Speaker 6: won't necessarily drive the impact. It also takes time for

146
00:06:42,320 --> 00:06:45,400
Speaker 6: people to just start using the tools, and so we're

147
00:06:45,440 --> 00:06:49,200
Speaker 6: still in the somewhat of the early stages. There two

148
00:06:49,240 --> 00:06:51,400
Speaker 6: places that I would be looking to see an impact.

149
00:06:51,680 --> 00:06:54,640
Speaker 6: One is in terms of productivity growth. We've done some

150
00:06:54,680 --> 00:06:57,080
Speaker 6: research that points in the direction that this should be

151
00:06:57,160 --> 00:07:01,719
Speaker 6: large and consequential. Labor productivity growth has been strong throughout

152
00:07:01,760 --> 00:07:05,279
Speaker 6: the pandemic and has been sustained so far modestly.

153
00:07:05,360 --> 00:07:09,080
Speaker 3: So we're not talking about like a you know, revolutionary.

154
00:07:08,320 --> 00:07:11,520
Speaker 6: It's not, yeah, but you know, to get on an inflection,

155
00:07:11,600 --> 00:07:14,160
Speaker 6: you need to at least move a little bit. I

156
00:07:14,160 --> 00:07:17,040
Speaker 6: think maybe you're seeing some signs there on the labor market,

157
00:07:17,160 --> 00:07:20,560
Speaker 6: though the labor market isn't a reasonably healthy spot, and

158
00:07:20,600 --> 00:07:23,480
Speaker 6: I think it might be because it's primarily at so

159
00:07:23,640 --> 00:07:27,960
Speaker 6: far a labor augmenting skill bias. Technology not yet the

160
00:07:28,040 --> 00:07:32,080
Speaker 6: full sort of general purpose substitute for all of cognitive labor,

161
00:07:32,120 --> 00:07:34,320
Speaker 6: although perhaps that's the trajectory that we're.

162
00:07:34,160 --> 00:07:37,160
Speaker 5: On, you know, for the size of AI and its capabilities.

163
00:07:37,200 --> 00:07:38,920
Speaker 5: I was talking to Peter about this and he did

164
00:07:38,960 --> 00:07:43,240
Speaker 5: point out the economy very big, so it still takes

165
00:07:43,240 --> 00:07:45,920
Speaker 5: a lot to move it. I do think strange things

166
00:07:45,920 --> 00:07:48,880
Speaker 5: are starting to happen, at least inside the company. We

167
00:07:49,000 --> 00:07:52,480
Speaker 5: published research from the Anthropic Institute recently on this topic

168
00:07:52,520 --> 00:07:56,600
Speaker 5: called recursive Self Improvement, where it was inspired by me

169
00:07:56,680 --> 00:07:59,760
Speaker 5: going on paternity leave in November of last year and

170
00:08:00,000 --> 00:08:03,160
Speaker 5: coming back in February, and the entire company felt and

171
00:08:03,200 --> 00:08:06,280
Speaker 5: worked differently, and I assumed it was because models had

172
00:08:06,280 --> 00:08:08,480
Speaker 5: got better. And when we looked at the data, what

173
00:08:08,520 --> 00:08:12,440
Speaker 5: you saw was in twenty twenty six, engineers and Anthropic

174
00:08:12,440 --> 00:08:15,480
Speaker 5: are writing about eight times the amount of code than

175
00:08:15,480 --> 00:08:18,480
Speaker 5: they did in twenty twenty one through to twenty twenty four.

176
00:08:18,840 --> 00:08:21,640
Speaker 5: And the line started last year with things like Opus

177
00:08:21,680 --> 00:08:23,720
Speaker 5: four to five and Opus four to six. Then it

178
00:08:23,840 --> 00:08:26,760
Speaker 5: really got going this year. And I have colleagues now

179
00:08:26,800 --> 00:08:30,200
Speaker 5: who don't program at all anymore. They just instruct many

180
00:08:30,200 --> 00:08:32,320
Speaker 5: many cored code agents to run around and do their

181
00:08:32,360 --> 00:08:35,840
Speaker 5: work for them. I can't reconcile that with the world

182
00:08:35,960 --> 00:08:38,440
Speaker 5: staying normal for long, but it's going to take a

183
00:08:38,480 --> 00:08:40,680
Speaker 5: while for that to diffuse into the world and change up.

184
00:08:40,840 --> 00:08:44,400
Speaker 4: Yeah, we'll talk more about recursive self improvement. So this

185
00:08:44,440 --> 00:08:48,839
Speaker 4: is when models basically improve on themselves. So in terms

186
00:08:48,840 --> 00:08:51,280
Speaker 4: of the awkwardness of the current moment or the weirdness

187
00:08:51,280 --> 00:08:54,360
Speaker 4: of the current moment, you've talked about basically living through

188
00:08:54,720 --> 00:08:58,080
Speaker 4: the singularity and how strange it is. And you've also

189
00:08:58,120 --> 00:09:02,880
Speaker 4: described yourself as a technopessimist before. How do you square

190
00:09:03,000 --> 00:09:06,760
Speaker 4: that with working at Anthropic, which is making some of

191
00:09:06,800 --> 00:09:09,800
Speaker 4: these weird and potentially dangerous things actually happen.

192
00:09:09,920 --> 00:09:13,120
Speaker 5: So by technological pessimist, I mean I thought the technology

193
00:09:13,120 --> 00:09:15,480
Speaker 5: would keep getting better, but I didn't think it would

194
00:09:15,480 --> 00:09:18,080
Speaker 5: get better in the maximalist sense that some of my

195
00:09:18,160 --> 00:09:21,080
Speaker 5: colleagues did. I didn't think that we would have, say,

196
00:09:21,240 --> 00:09:23,920
Speaker 5: functionally automated all of coding right now. I find that

197
00:09:23,960 --> 00:09:27,240
Speaker 5: actually like quite surprising. But basically, over the last few years,

198
00:09:27,240 --> 00:09:29,720
Speaker 5: and I worked at Opening Eye before and propic I

199
00:09:29,840 --> 00:09:33,120
Speaker 5: was just hit repeatedly over thehead with what computer scientist

200
00:09:33,200 --> 00:09:36,199
Speaker 5: Richard Sutton calls the bitter lesson, And the bitter lesson

201
00:09:36,240 --> 00:09:39,160
Speaker 5: is this concept that the more compute and resources we

202
00:09:39,240 --> 00:09:43,240
Speaker 5: dump into these relatively generic neural networks, the smarter they

203
00:09:43,280 --> 00:09:47,000
Speaker 5: get and the more emergent properties they have, and your

204
00:09:47,080 --> 00:09:51,040
Speaker 5: specialized system or your ability to be pessimistic about future

205
00:09:51,080 --> 00:09:55,280
Speaker 5: AI progress loses versus just scaling compute and scaling systems.

206
00:09:55,480 --> 00:09:58,000
Speaker 2: This seems to him implications for the labor market, right,

207
00:09:58,080 --> 00:10:01,920
Speaker 2: because I think example the Bitter Lessons Private History of

208
00:10:02,160 --> 00:10:04,920
Speaker 2: AI chess, right where at one point they had grand

209
00:10:04,960 --> 00:10:08,880
Speaker 2: masters come in and teach the models how to play

210
00:10:08,960 --> 00:10:12,520
Speaker 2: chess and etc. Try to encode their wisdom, and it

211
00:10:12,600 --> 00:10:14,199
Speaker 2: turned out in the end that the best way to

212
00:10:14,240 --> 00:10:17,880
Speaker 2: get a chess engine really good is to just teach

213
00:10:17,920 --> 00:10:20,040
Speaker 2: the model, tell the model the rules of chess. Let's say,

214
00:10:20,200 --> 00:10:22,680
Speaker 2: go off and play a billion games and find optimal

215
00:10:22,800 --> 00:10:26,000
Speaker 2: chess without any human insight. The grand masters were not

216
00:10:26,040 --> 00:10:28,720
Speaker 2: necessary for that process at all, right, And so this

217
00:10:28,800 --> 00:10:31,599
Speaker 2: would imply to me like have significant implications for the

218
00:10:31,640 --> 00:10:32,240
Speaker 2: labor market.

219
00:10:32,400 --> 00:10:35,240
Speaker 6: Yeah. I tend to think about this in sort of

220
00:10:35,280 --> 00:10:39,040
Speaker 6: three aspects of what composes a job. One is you

221
00:10:39,080 --> 00:10:41,880
Speaker 6: need to decide what to do and direct and delegate.

222
00:10:42,520 --> 00:10:45,960
Speaker 6: You need to then do the actual implementation of the work,

223
00:10:46,000 --> 00:10:47,640
Speaker 6: and then you need to sort of evaluate or at

224
00:10:47,720 --> 00:10:50,959
Speaker 6: least set up systems that can evaluate. At least from

225
00:10:50,960 --> 00:10:54,359
Speaker 6: my perspective as an economist, this bitter lesson is materializing

226
00:10:54,440 --> 00:10:58,160
Speaker 6: in terms of very rapid advances in the implementation work

227
00:10:58,200 --> 00:11:03,080
Speaker 6: of what an economist does data running regressions, building models,

228
00:11:03,200 --> 00:11:08,040
Speaker 6: solving them using sort of contemporary solution techniques numerical methods.

229
00:11:08,720 --> 00:11:11,520
Speaker 6: I definitely felt that personally with Opus four point five,

230
00:11:11,640 --> 00:11:14,280
Speaker 6: where I was for the first time able to just

231
00:11:14,440 --> 00:11:17,880
Speaker 6: delegate a very complex task. I had this very specific

232
00:11:17,920 --> 00:11:22,160
Speaker 6: research question trying to understand the cyclicality of hiring across

233
00:11:22,200 --> 00:11:25,439
Speaker 6: different occupations and how that relates to occupational exposure. That's

234
00:11:25,480 --> 00:11:28,120
Speaker 6: a mouthful. I gave that task to Claude, and Claude

235
00:11:28,200 --> 00:11:30,640
Speaker 6: was able to just iterate on it, and I could

236
00:11:30,679 --> 00:11:33,400
Speaker 6: redirect Claude in the same way that you might redirect

237
00:11:33,640 --> 00:11:36,920
Speaker 6: a grad student. And the big question that I have

238
00:11:37,000 --> 00:11:39,200
Speaker 6: in mind is, you know, at what point do the

239
00:11:39,240 --> 00:11:42,800
Speaker 6: boundaries at the direction setting stage, the research taste you

240
00:11:42,880 --> 00:11:47,040
Speaker 6: might call it, and will the models become sufficiently reliable.

241
00:11:46,559 --> 00:11:48,160
Speaker 2: If I could just get in here, you know, I

242
00:11:48,200 --> 00:11:52,000
Speaker 2: just read the recent biography of the deep Mind founder.

243
00:11:52,320 --> 00:11:55,160
Speaker 4: This is the Sebastian melody, Yeah, Like, is there going

244
00:11:55,200 --> 00:11:56,960
Speaker 4: to be a point where it's like, Okay, you have

245
00:11:57,040 --> 00:12:00,600
Speaker 4: some intuitions right about like what good economics research is.

246
00:12:00,840 --> 00:12:03,560
Speaker 3: And often our intuitions are formed because we tell stories

247
00:12:03,600 --> 00:12:04,319
Speaker 3: and stuff like that.

248
00:12:04,920 --> 00:12:06,520
Speaker 2: But is there going to be a point where you think,

249
00:12:06,559 --> 00:12:10,360
Speaker 2: like your intuitions will be unhelpful? And that because that's

250
00:12:10,400 --> 00:12:12,719
Speaker 2: sort of what I took away from the GO experience,

251
00:12:13,040 --> 00:12:15,760
Speaker 2: that the model got better once they stripped it of

252
00:12:15,760 --> 00:12:18,600
Speaker 2: the human games and the human bias, and that actually,

253
00:12:18,800 --> 00:12:22,320
Speaker 2: like the human intuition that sort of helps us understand, oh,

254
00:12:22,440 --> 00:12:25,680
Speaker 2: labor market rise, it creates inflationary pressure. These stories that

255
00:12:25,720 --> 00:12:29,560
Speaker 2: are very sort of intuitible end up impairing the model.

256
00:12:29,920 --> 00:12:32,080
Speaker 2: Do you see that happening in say economics, where it's

257
00:12:32,120 --> 00:12:34,200
Speaker 2: like some of these stories that we tell forever, they're

258
00:12:34,240 --> 00:12:38,520
Speaker 2: not actually very helpful for an optimal economy understanding model.

259
00:12:38,640 --> 00:12:41,760
Speaker 6: I expect that these models will soon have better intuitions

260
00:12:41,800 --> 00:12:45,120
Speaker 6: about how to do good economic research, and that there's

261
00:12:45,160 --> 00:12:47,400
Speaker 6: this big question of like at what point will we

262
00:12:47,440 --> 00:12:51,240
Speaker 6: be able to fully automate social science research. We've done

263
00:12:51,280 --> 00:12:53,280
Speaker 6: some work on this to try to understand how coding

264
00:12:53,280 --> 00:12:56,520
Speaker 6: agents are beginning to automate social science research, but I

265
00:12:56,520 --> 00:12:58,720
Speaker 6: don't think we're quite there yet, and I don't know

266
00:12:58,760 --> 00:13:02,000
Speaker 6: that'll be an exciting time for learning about the world.

267
00:13:02,480 --> 00:13:04,880
Speaker 6: You know what that means for my job? I'm sort

268
00:13:04,880 --> 00:13:06,319
Speaker 6: of less entirely clear.

269
00:13:06,440 --> 00:13:08,960
Speaker 5: Yeah, I think this is the big wild card in

270
00:13:09,240 --> 00:13:13,040
Speaker 5: future AI progress. If AI progress continues today, we are

271
00:13:13,160 --> 00:13:16,120
Speaker 5: likely to get technology that we'll be able to do

272
00:13:16,160 --> 00:13:20,440
Speaker 5: basically everything, but we will need people who have good instincts,

273
00:13:20,440 --> 00:13:23,200
Speaker 5: good intuitions, and good ideas to basically set the direction.

274
00:13:23,400 --> 00:13:25,120
Speaker 5: And we see this today in a lot of a

275
00:13:25,160 --> 00:13:27,600
Speaker 5: lot of our own research, where you need, say an

276
00:13:27,640 --> 00:13:31,400
Speaker 5: AI safety researcher to give nine clawed agents for different

277
00:13:31,440 --> 00:13:34,240
Speaker 5: research areas to go and pursue, and then it's very effective.

278
00:13:34,600 --> 00:13:37,800
Speaker 5: If that researcher doesn't give them the research directions, they

279
00:13:38,320 --> 00:13:41,840
Speaker 5: pursue relatively formulaic research directions and you have entropy collapse.

280
00:13:41,840 --> 00:13:43,880
Speaker 5: You end up with just like boring research and doesn't

281
00:13:43,880 --> 00:13:47,079
Speaker 5: move the ball forward. At what point will AI systems

282
00:13:47,160 --> 00:13:51,480
Speaker 5: generate like heterodox insights and genuine creativity. We can't really

283
00:13:51,520 --> 00:13:54,080
Speaker 5: measure for that today, but what we have of as

284
00:13:54,120 --> 00:13:58,080
Speaker 5: symptoms of it, starting in experts like Peter, Experts like

285
00:13:58,280 --> 00:14:01,760
Speaker 5: colleagues in the fields of biology or mathematics or physics

286
00:14:01,840 --> 00:14:05,920
Speaker 5: outside of anthropic are all starting to be accelerated by AI.

287
00:14:06,080 --> 00:14:08,520
Speaker 5: You know, Terry Taw probably one of the most famous

288
00:14:08,679 --> 00:14:13,920
Speaker 5: living mathematicians co creates math now with AI systems, and

289
00:14:13,960 --> 00:14:15,800
Speaker 5: so that that says to me that these things have

290
00:14:15,880 --> 00:14:19,560
Speaker 5: got there. They're tickling the dragon's tail of like creativity here.

291
00:14:20,080 --> 00:14:22,240
Speaker 6: And you know, we just put out a report yesterday

292
00:14:22,280 --> 00:14:25,880
Speaker 6: on cloud code usage, and one of the things that

293
00:14:25,880 --> 00:14:28,440
Speaker 6: we're trying to understand is like what are the returns

294
00:14:28,480 --> 00:14:31,240
Speaker 6: to expertise and how does that interact with the usage

295
00:14:31,280 --> 00:14:34,320
Speaker 6: of sort of automated coding agents. And we find that

296
00:14:34,400 --> 00:14:38,240
Speaker 6: domain expertise, like if you're an accountant who understands some

297
00:14:38,320 --> 00:14:42,240
Speaker 6: of the edge cases and reconciliation, that that domain expertise

298
00:14:42,400 --> 00:14:44,600
Speaker 6: controlling for a whole host of factors about the type

299
00:14:44,600 --> 00:14:46,200
Speaker 6: of work, the estimated monetary value.

300
00:14:46,200 --> 00:14:47,320
Speaker 3: It has an amplifying effect.

301
00:14:47,360 --> 00:14:49,800
Speaker 6: It has an amplifying effect. So this looks like at

302
00:14:49,840 --> 00:14:54,600
Speaker 6: present as sort of a skill biased expertise enhancing impact.

303
00:14:54,720 --> 00:14:57,160
Speaker 6: But I think this is the key question is at

304
00:14:57,160 --> 00:15:00,560
Speaker 6: what point and to what extent will this change.

305
00:15:00,200 --> 00:15:03,120
Speaker 4: Well related to this, you know, Jack, when you describe

306
00:15:03,160 --> 00:15:05,480
Speaker 4: coming back from paternity leave and seeing how much things

307
00:15:05,480 --> 00:15:08,400
Speaker 4: had changed at Anthropic, I know we're not officially at

308
00:15:08,520 --> 00:15:14,200
Speaker 4: recursive self improvement point, but it sounds like we're semi there.

309
00:15:14,920 --> 00:15:17,240
Speaker 4: So my question is, like, I get that at the

310
00:15:17,280 --> 00:15:20,520
Speaker 4: moment you have engineers who are reviewing all the code

311
00:15:20,880 --> 00:15:23,600
Speaker 4: that the AI is producing, and they're thinking about it

312
00:15:23,640 --> 00:15:26,080
Speaker 4: and managing it in some way. But you can easily

313
00:15:26,080 --> 00:15:29,640
Speaker 4: imagine a future where just the sheer quantity of code

314
00:15:29,760 --> 00:15:34,120
Speaker 4: overwhelms human expertise. Maybe the quality starts out stripping what

315
00:15:34,240 --> 00:15:38,120
Speaker 4: human engineers are capable of understanding. How do you manage that?

316
00:15:38,720 --> 00:15:42,160
Speaker 5: Yeah, So there's two ways of thinking about recursive self improvement.

317
00:15:42,280 --> 00:15:46,120
Speaker 5: One is what happens when AI organizations start to see

318
00:15:46,120 --> 00:15:48,800
Speaker 5: a compounding return from their AI systems. Basically, their in

319
00:15:48,840 --> 00:15:52,280
Speaker 5: production function improves because of the tools they built. That's

320
00:15:52,320 --> 00:15:55,160
Speaker 5: clearly happening now. Member. Second is what happens if an

321
00:15:55,200 --> 00:15:58,960
Speaker 5: AI system can just build itself entirely autonomously given compute,

322
00:15:59,000 --> 00:16:02,560
Speaker 5: which hasn't happened. What I see inside Anthropic is I

323
00:16:02,560 --> 00:16:04,520
Speaker 5: think what we'll see in the broader economy, which is

324
00:16:04,520 --> 00:16:08,000
Speaker 5: we are figuring out how to verify and validate and

325
00:16:08,040 --> 00:16:11,400
Speaker 5: basically price for risk of an expanding cloud of automated

326
00:16:11,440 --> 00:16:14,360
Speaker 5: systems which we're sitting on top of. So now we

327
00:16:14,400 --> 00:16:18,040
Speaker 5: produce way more code. Well, we broke our continuous integration

328
00:16:18,160 --> 00:16:21,320
Speaker 5: system for integrating code into the codebase because we started

329
00:16:21,320 --> 00:16:23,840
Speaker 5: pushing eight times more code through it than before. So

330
00:16:23,920 --> 00:16:27,680
Speaker 5: all of our human engineers worked on unbreaking CI and

331
00:16:27,760 --> 00:16:31,240
Speaker 5: so I think that inside continuous integration, Thank you, you

332
00:16:31,280 --> 00:16:32,640
Speaker 5: don't need to know what it is. It's just a

333
00:16:32,640 --> 00:16:34,160
Speaker 5: thing that helps you push the code into the.

334
00:16:36,040 --> 00:16:36,360
Speaker 4: Ton.

335
00:16:36,560 --> 00:16:38,440
Speaker 5: But there's a lesson in that. Right, we are going

336
00:16:38,480 --> 00:16:41,320
Speaker 5: to speed up things in the economy. We're going to

337
00:16:41,320 --> 00:16:43,320
Speaker 5: speed up the way that we produce stuff, and then

338
00:16:43,320 --> 00:16:45,400
Speaker 5: we're going to find, you know, like the weak links

339
00:16:45,480 --> 00:16:47,960
Speaker 5: or the hot paths but break and we as people

340
00:16:48,000 --> 00:16:50,240
Speaker 5: are going to move to sorting those out, and then

341
00:16:50,320 --> 00:16:52,800
Speaker 5: the cycle starts again. And we're kind of sitting on

342
00:16:52,840 --> 00:17:09,040
Speaker 5: this expanding cloud of automated actions.

343
00:17:12,000 --> 00:17:16,040
Speaker 2: Since we're talking about like really like feeling like we're

344
00:17:16,080 --> 00:17:19,880
Speaker 2: staring at the horizon of extremely strong AI or maybe

345
00:17:19,880 --> 00:17:22,520
Speaker 2: we'll get there. Maybe the abilits itself might be a

346
00:17:22,520 --> 00:17:25,560
Speaker 2: good time to ask a fable question or Metho's questions.

347
00:17:25,600 --> 00:17:27,920
Speaker 2: At this point we're recording this June seventh. We don't

348
00:17:27,960 --> 00:17:30,800
Speaker 2: know when it's going to be available for Americans, let

349
00:17:30,840 --> 00:17:32,960
Speaker 2: alone the rest of the world. Does SAT product have

350
00:17:33,000 --> 00:17:38,080
Speaker 2: a clear idea of what the administration's security concerns are

351
00:17:38,240 --> 00:17:39,680
Speaker 2: and what it will take to resolve them.

352
00:17:39,760 --> 00:17:43,919
Speaker 5: Well, obviously live discussion. I can't get into too many specifics.

353
00:17:43,920 --> 00:17:46,919
Speaker 5: We're in daily discussions with the government about this. The

354
00:17:46,960 --> 00:17:50,240
Speaker 5: broad thing I'd say is, for many years we've anticipated

355
00:17:50,280 --> 00:17:53,119
Speaker 5: the point where AI systems would have national security properties.

356
00:17:53,480 --> 00:17:57,879
Speaker 5: These national security properties are intertwined with their economically valuable properties.

357
00:17:58,520 --> 00:18:02,320
Speaker 5: How you manage that as a pol question is basically

358
00:18:02,359 --> 00:18:06,120
Speaker 5: novel territory. Typically these things are decoupled. You're like, hey,

359
00:18:06,160 --> 00:18:07,840
Speaker 5: I built a jet engine over here, which you can

360
00:18:07,880 --> 00:18:10,240
Speaker 5: go into civilian aircraft, and I built a missile over here,

361
00:18:10,280 --> 00:18:12,639
Speaker 5: and you treat them differently. It's odd if you smush

362
00:18:12,680 --> 00:18:16,119
Speaker 5: these things together. Where we'll get to, I'm confident is

363
00:18:16,840 --> 00:18:21,159
Speaker 5: what's a system for assessing the properties of aisystems, including

364
00:18:21,200 --> 00:18:24,040
Speaker 5: national security components, and then what is a system for

365
00:18:24,119 --> 00:18:28,400
Speaker 5: either squelching the national security capabilities from coming to general

366
00:18:28,440 --> 00:18:32,240
Speaker 5: proliferation like bioweapons or cyber weapons, and are there ways

367
00:18:32,280 --> 00:18:36,040
Speaker 5: to do things like know your customer or deployments where

368
00:18:36,080 --> 00:18:39,600
Speaker 5: you let large firms like say drug developers, access for

369
00:18:39,720 --> 00:18:43,639
Speaker 5: most powerful BioModels without accidentally proliferating risks. That's the shape

370
00:18:43,640 --> 00:18:45,800
Speaker 5: of I think where we will end up and what

371
00:18:45,880 --> 00:18:49,720
Speaker 5: we're doing right now. We and other companies and the

372
00:18:49,800 --> 00:18:52,800
Speaker 5: administration are basically tackling this problem in real time. It's

373
00:18:52,840 --> 00:18:54,760
Speaker 5: initially going to be messy, but we're going to end

374
00:18:54,800 --> 00:18:56,080
Speaker 5: up with a system on the other side.

375
00:18:56,119 --> 00:18:59,520
Speaker 2: Well, let me just ask you, you know, this specific incident,

376
00:19:00,080 --> 00:19:03,000
Speaker 2: probably more in the future because everyone's just figuring this out.

377
00:19:03,600 --> 00:19:06,719
Speaker 2: When I look at the AI landscape, I sort of

378
00:19:06,760 --> 00:19:12,280
Speaker 2: think of open AI as being part of the all

379
00:19:12,320 --> 00:19:18,640
Speaker 2: in podcast A sixteen Z David Sex White House thing.

380
00:19:19,320 --> 00:19:22,240
Speaker 2: And I know from my friends in the media, many

381
00:19:22,320 --> 00:19:25,359
Speaker 2: of whom are liberal Democrats, that I sort of feel

382
00:19:25,359 --> 00:19:27,879
Speaker 2: like Athropic is the more like lib coded of the

383
00:19:27,920 --> 00:19:30,879
Speaker 2: major models. Do you feel there's any either politics or

384
00:19:30,880 --> 00:19:36,080
Speaker 2: partisan politics going on as part of Edthropic being harassed

385
00:19:36,160 --> 00:19:38,119
Speaker 2: or singled out now multiple.

386
00:19:37,840 --> 00:19:42,280
Speaker 5: Times andthropics philosophy and what I do and I lead

387
00:19:42,320 --> 00:19:46,520
Speaker 5: something called Panthropic Institute, which helps us produce better data

388
00:19:46,520 --> 00:19:49,159
Speaker 5: for the world around things like recursive self improvement. The

389
00:19:49,240 --> 00:19:52,800
Speaker 5: economics work cyber risks is we tell the whole story

390
00:19:52,840 --> 00:19:56,560
Speaker 5: about what's going on. Typically, I think the technology industry

391
00:19:56,680 --> 00:20:00,119
Speaker 5: has told only optimistic stories about what it's building. We

392
00:20:00,119 --> 00:20:04,320
Speaker 5: saw with social media is that does not work. Actually, eventually,

393
00:20:04,400 --> 00:20:07,560
Speaker 5: when when you're doing something that changes the entire world,

394
00:20:07,800 --> 00:20:10,560
Speaker 5: which AI is certainly doing and social media certainly did,

395
00:20:10,920 --> 00:20:13,040
Speaker 5: It's not going to be a wholly optimistic story. There'll

396
00:20:13,080 --> 00:20:15,880
Speaker 5: be negatives as well. We've always sought to just tell

397
00:20:15,920 --> 00:20:17,720
Speaker 5: the truth about what we see in front of us,

398
00:20:18,040 --> 00:20:20,400
Speaker 5: and I think sometimes that can differentiate us a bit

399
00:20:20,440 --> 00:20:23,359
Speaker 5: to others. But the important thing is we tell the

400
00:20:23,359 --> 00:20:25,159
Speaker 5: truth and things end up coming to You.

401
00:20:25,160 --> 00:20:28,040
Speaker 2: Don't think that there's like a partisan element here where

402
00:20:28,040 --> 00:20:31,080
Speaker 2: you guys aren't on the team or didn't contribute enough

403
00:20:31,119 --> 00:20:32,200
Speaker 2: to the ballroom or whatever.

404
00:20:32,680 --> 00:20:34,840
Speaker 5: I can't really speak to that. You know, I'm not

405
00:20:34,920 --> 00:20:37,720
Speaker 5: those people, and I'm menthropic. What I can say is

406
00:20:38,760 --> 00:20:42,320
Speaker 5: the AI systems create their own evidence. Years ago, it

407
00:20:42,400 --> 00:20:45,080
Speaker 5: seemed very odd to speculate about the cyber properties of

408
00:20:45,119 --> 00:20:47,959
Speaker 5: AI systems. Well they've arrived and now we're working on them.

409
00:20:48,080 --> 00:20:51,040
Speaker 5: Years ago, it was odd to speculate about the bioweapon

410
00:20:51,080 --> 00:20:55,080
Speaker 5: properties of AI systems. Well. Recently, Sam Altman, Demissabis, and

411
00:20:55,160 --> 00:20:58,160
Speaker 5: Daria Ama Day of Opening Eye and Thropic and Deep

412
00:20:58,200 --> 00:21:00,840
Speaker 5: Mind all signed a letter saying we need to do

413
00:21:00,920 --> 00:21:04,920
Speaker 5: better screening of gene synthesis to prevent AI manufactured bioweapons.

414
00:21:05,359 --> 00:21:06,280
Speaker 5: The truth wins out.

415
00:21:06,400 --> 00:21:08,200
Speaker 4: Okay, I want to go back to something you said.

416
00:21:08,200 --> 00:21:12,120
Speaker 4: You mentioned potential kyic requirements, and when I hear kyic,

417
00:21:12,359 --> 00:21:14,600
Speaker 4: I think about the finance industry, and I think about

418
00:21:14,640 --> 00:21:18,200
Speaker 4: systemically important institutions and the stress tests and the framework

419
00:21:18,240 --> 00:21:21,960
Speaker 4: around that. Is that the right analogy to use for

420
00:21:22,080 --> 00:21:25,840
Speaker 4: I guess, ideal AI regulation in your mind rather than

421
00:21:26,040 --> 00:21:29,120
Speaker 4: I guess just simple export controls. Should we be heading

422
00:21:29,160 --> 00:21:30,840
Speaker 4: towards something that looks a little bit more like what

423
00:21:30,880 --> 00:21:32,200
Speaker 4: we do for the banking system.

424
00:21:32,240 --> 00:21:35,080
Speaker 5: We need something that's more subtle and more technocratic from

425
00:21:35,080 --> 00:21:36,720
Speaker 5: what we have today. I don't know if it'll be

426
00:21:36,760 --> 00:21:39,320
Speaker 5: exactly like the banking system. It'll probably take some ideas

427
00:21:39,320 --> 00:21:42,040
Speaker 5: from that. It'll take some ideas from what the US

428
00:21:42,160 --> 00:21:45,520
Speaker 5: government and others are doing today with just testing aisystems

429
00:21:45,560 --> 00:21:48,119
Speaker 5: for their properties, and it's almost certainly going to have

430
00:21:48,119 --> 00:21:50,399
Speaker 5: a flavor of what Peter and I work on, and

431
00:21:50,440 --> 00:21:55,320
Speaker 5: the unproper innstitute broadly of generating data about these systems

432
00:21:55,359 --> 00:21:57,600
Speaker 5: as they're deployed in the world, because it's one thing

433
00:21:57,640 --> 00:21:59,520
Speaker 5: to you know, test out the thing before it comes

434
00:21:59,560 --> 00:22:02,160
Speaker 5: out of a it's another to observe the effects it's

435
00:22:02,160 --> 00:22:04,080
Speaker 5: having in the world and then to be able to

436
00:22:04,080 --> 00:22:06,040
Speaker 5: make judgments about whether those effects are good or not.

437
00:22:06,400 --> 00:22:09,280
Speaker 2: Would you support, you know, the let's stick with the

438
00:22:09,280 --> 00:22:14,119
Speaker 2: financial analogy. Companies that are public at least are required

439
00:22:14,119 --> 00:22:16,679
Speaker 2: to have third party auditors sign off on them, and

440
00:22:16,720 --> 00:22:19,000
Speaker 2: there's talk, you know, when they submit their ten queues,

441
00:22:19,080 --> 00:22:22,199
Speaker 2: et cetera. Companies that issue debt are required to have

442
00:22:22,280 --> 00:22:25,720
Speaker 2: ratings agencies or frequently have ratings agencies rate their debt.

443
00:22:26,119 --> 00:22:30,840
Speaker 2: Would you support embedding in law the requirement that certain

444
00:22:30,960 --> 00:22:33,680
Speaker 2: what would be the equivalent of a Moody's or a Deloitte,

445
00:22:33,960 --> 00:22:36,439
Speaker 2: you know, a third party research lab sign off on

446
00:22:36,920 --> 00:22:38,120
Speaker 2: the release of new models.

447
00:22:38,320 --> 00:22:42,080
Speaker 5: We've proposed something like this recently, a policy proposal that

448
00:22:42,119 --> 00:22:44,359
Speaker 5: we laid out which includes saying we need to have

449
00:22:44,480 --> 00:22:47,359
Speaker 5: third party testing for some of these national security and

450
00:22:47,440 --> 00:22:50,760
Speaker 5: other properties, because clearly that's like a sensible way that

451
00:22:50,800 --> 00:22:52,360
Speaker 5: you validate a lot of them. Yeah.

452
00:22:52,400 --> 00:22:55,760
Speaker 4: So, just more broadly, returning to this idea of, you know,

453
00:22:55,840 --> 00:22:58,520
Speaker 4: measuring the actual impact of AI, one thing I find

454
00:22:58,520 --> 00:23:01,199
Speaker 4: really interesting is that if you actually look at a

455
00:23:01,200 --> 00:23:04,880
Speaker 4: lot of our traditional AI or I should say I'm

456
00:23:04,920 --> 00:23:07,280
Speaker 4: AI brained already, if you look at some of our

457
00:23:07,320 --> 00:23:10,840
Speaker 4: traditional economic statistics, a lot of the AI impact doesn't

458
00:23:10,840 --> 00:23:14,160
Speaker 4: actually show up just yet. Again, we're in the early stages,

459
00:23:14,200 --> 00:23:16,320
Speaker 4: but you would expect if we're talking about the AI

460
00:23:16,400 --> 00:23:20,600
Speaker 4: economy growing something like two thousand percent, three thousand percent.

461
00:23:20,640 --> 00:23:22,000
Speaker 4: I think I've seen that number.

462
00:23:21,960 --> 00:23:26,600
Speaker 6: That's from Anton Quarneck and McKelvie Harry paper a few weeks.

463
00:23:26,400 --> 00:23:28,239
Speaker 4: Ago, you would expect that to have more of an

464
00:23:28,240 --> 00:23:32,639
Speaker 4: impact on nominal GDP, and yet it's not really showing

465
00:23:32,720 --> 00:23:34,920
Speaker 4: up that much. Do you think the way we measure

466
00:23:35,119 --> 00:23:37,639
Speaker 4: the economy needs to be changed in some way in

467
00:23:37,760 --> 00:23:40,399
Speaker 4: light of what's happening with this new technology?

468
00:23:40,880 --> 00:23:41,080
Speaker 5: Yeah?

469
00:23:41,119 --> 00:23:43,479
Speaker 6: So I think this is exactly the right premises kind

470
00:23:43,480 --> 00:23:46,560
Speaker 6: of where we began the conversation, which is we're maybe

471
00:23:46,600 --> 00:23:48,480
Speaker 6: at the point where we should be able to see

472
00:23:48,520 --> 00:23:54,000
Speaker 6: some discernible impact on the macroeconomy. Unfortunately, the arrival of

473
00:23:54,000 --> 00:23:57,640
Speaker 6: this world historical technology is against the backdrop of sort

474
00:23:57,640 --> 00:24:02,720
Speaker 6: of unusually elevated macroeconomic volatility, the post pandemic, monetary policy,

475
00:24:03,440 --> 00:24:06,399
Speaker 6: et cetera, And so it makes it very hard to

476
00:24:06,480 --> 00:24:10,400
Speaker 6: disentangle all of the different factors. You know, what's the counterfactual?

477
00:24:10,440 --> 00:24:13,880
Speaker 6: You know, labor productivity growth is maybe not as strong

478
00:24:13,920 --> 00:24:16,240
Speaker 6: as you might not otherwise expect, but maybe it's stronger

479
00:24:16,240 --> 00:24:20,560
Speaker 6: than it is an a counterfactual sense. And so one

480
00:24:20,600 --> 00:24:23,160
Speaker 6: way that we've tried to tackle this question is by

481
00:24:23,720 --> 00:24:27,360
Speaker 6: looking at how Claude is being used on our platform,

482
00:24:27,480 --> 00:24:32,240
Speaker 6: using our privacy preserving techniques to estimate the time savings

483
00:24:32,280 --> 00:24:35,720
Speaker 6: associated with each of the activities that people use claud for. So,

484
00:24:36,720 --> 00:24:40,000
Speaker 6: compiling information from reports to put together a research brief

485
00:24:40,200 --> 00:24:43,000
Speaker 6: would take you a few days. Maybe now Claude does

486
00:24:43,000 --> 00:24:46,960
Speaker 6: it in a few minutes. Evaluating diagnostic images is something

487
00:24:47,000 --> 00:24:49,120
Speaker 6: that skilled professionals do very rapidly, so there is an

488
00:24:49,119 --> 00:24:52,439
Speaker 6: imprinciple much time savings. You can add up all of

489
00:24:52,440 --> 00:24:57,280
Speaker 6: those numbers and using standard macro growth accounting techniques, Houlton's

490
00:24:57,320 --> 00:25:00,440
Speaker 6: theorem for the economists and the audience. When you get

491
00:25:00,440 --> 00:25:03,040
Speaker 6: a number of that points in the direction of labor

492
00:25:03,080 --> 00:25:06,160
Speaker 6: productivity growth increasing by one point eight percentage points each

493
00:25:06,240 --> 00:25:08,760
Speaker 6: year over the next decade. If that's how long it

494
00:25:08,800 --> 00:25:12,880
Speaker 6: takes current usage patterns and current model capabilities to diffuse

495
00:25:12,880 --> 00:25:15,440
Speaker 6: throughout the economy, that's a very large number. It's a

496
00:25:15,520 --> 00:25:19,120
Speaker 6: rough doubling of recent run rates. And what I think

497
00:25:19,160 --> 00:25:21,119
Speaker 6: you might be able to see in the data, and

498
00:25:21,119 --> 00:25:23,840
Speaker 6: we haven't put anything out on this yet, is I

499
00:25:23,840 --> 00:25:26,879
Speaker 6: think some of the strength and recent labor productivity growth

500
00:25:27,240 --> 00:25:30,639
Speaker 6: is actually concentrating in exactly the sectors of the economy

501
00:25:31,040 --> 00:25:33,480
Speaker 6: that would be consistent with both what we see in

502
00:25:33,520 --> 00:25:35,480
Speaker 6: our data as well as also what you see in

503
00:25:35,520 --> 00:25:38,200
Speaker 6: the Business trend and outlyt for example, so the information

504
00:25:38,280 --> 00:25:42,159
Speaker 6: sector has high rates of adoption. I can't recall if

505
00:25:42,160 --> 00:25:44,520
Speaker 6: that's in particular one of the sectors that I have

506
00:25:44,640 --> 00:25:48,119
Speaker 6: in mind. It's a while since I looked at that scatterplot,

507
00:25:48,119 --> 00:25:50,639
Speaker 6: but you can look at the sort of sub industries

508
00:25:50,680 --> 00:25:53,879
Speaker 6: by the Census Bureaus Business Trend and Outlook survey, and

509
00:25:54,280 --> 00:25:57,480
Speaker 6: rates of adoption are in sectors or parts of the

510
00:25:57,520 --> 00:26:02,960
Speaker 6: economy where can trolling for pre pandemic trajectory of labor

511
00:26:02,960 --> 00:26:05,480
Speaker 6: productivity growth in those sectors, even some of the strength

512
00:26:05,840 --> 00:26:08,760
Speaker 6: in the early years of the recovery. Still see some

513
00:26:08,920 --> 00:26:13,160
Speaker 6: like suggestive evidence. I think there's a lot of uncertainty here.

514
00:26:13,280 --> 00:26:16,080
Speaker 6: Trying to get a real time signal on productivity is

515
00:26:16,359 --> 00:26:19,560
Speaker 6: maybe the hardest thing to do. You're subject to macroeconomic

516
00:26:19,640 --> 00:26:24,080
Speaker 6: GDP revisions. TFP growth is actually sending the opposite signal,

517
00:26:24,119 --> 00:26:28,000
Speaker 6: and if you control for capacity utilization, TFP growth is

518
00:26:28,280 --> 00:26:31,800
Speaker 6: arguably even lower. So I say this as like, this

519
00:26:31,880 --> 00:26:34,879
Speaker 6: is suggestive evidence that maybe we're beginning to see an

520
00:26:34,880 --> 00:26:37,760
Speaker 6: impact impact there but not so much in the labor market.

521
00:26:37,920 --> 00:26:39,679
Speaker 4: Well, now I have to ask when you gather this

522
00:26:39,800 --> 00:26:42,879
Speaker 4: kind of research, and it all sounds super interesting, But

523
00:26:42,920 --> 00:26:45,520
Speaker 4: if you have data, for instance, that shows that, okay,

524
00:26:45,520 --> 00:26:49,800
Speaker 4: the IT sector is getting productivity gains from using claud

525
00:26:50,040 --> 00:26:53,360
Speaker 4: or I don't know, maybe something unexpected, like the warehousing

526
00:26:53,440 --> 00:26:56,879
Speaker 4: industry is using a bunch of AI. What does anthropic

527
00:26:56,960 --> 00:27:00,320
Speaker 4: actually do with this data? Does it somehow feed act

528
00:27:00,359 --> 00:27:03,080
Speaker 4: to your engineers who are developing frontier models? Do they

529
00:27:03,080 --> 00:27:04,000
Speaker 4: do anything differently?

530
00:27:04,240 --> 00:27:06,359
Speaker 5: I think some of it cues us on areas where

531
00:27:06,520 --> 00:27:09,080
Speaker 5: maybe the technology isn't being used because it's very weak.

532
00:27:09,119 --> 00:27:11,160
Speaker 5: We just haven't made it particularly good for these use

533
00:27:11,200 --> 00:27:14,359
Speaker 5: cases or in areas where it's being used at large scale.

534
00:27:14,400 --> 00:27:17,120
Speaker 5: It's usually a suggestion of keep making it good there,

535
00:27:17,280 --> 00:27:20,359
Speaker 5: but you know, if the actual economic measurement data doesn't

536
00:27:20,359 --> 00:27:22,320
Speaker 5: really get fed back directly in. But it's a very

537
00:27:22,400 --> 00:27:26,360
Speaker 5: useful clue. We think it's more important, though, to basically

538
00:27:26,359 --> 00:27:31,040
Speaker 5: communicate this outwardly to policy makers, journalists, and others, because

539
00:27:31,080 --> 00:27:33,960
Speaker 5: our assumption is that at some point we go through

540
00:27:34,000 --> 00:27:37,080
Speaker 5: some phase change, similar to how capabilities of AI occasionally

541
00:27:37,160 --> 00:27:39,800
Speaker 5: jump forward in a really dramatic way, where you might

542
00:27:39,840 --> 00:27:43,520
Speaker 5: see sudden and rapid diffusion as a consequence of capability

543
00:27:43,520 --> 00:27:46,520
Speaker 5: expansion in their AI systems. So we're getting practice in

544
00:27:46,680 --> 00:27:49,760
Speaker 5: of looking at this kind of data. My expectation is

545
00:27:49,760 --> 00:27:51,560
Speaker 5: that in a year or two years, I'm going up

546
00:27:51,560 --> 00:27:53,960
Speaker 5: to some policy maker and I'm pointing them to the

547
00:27:54,040 --> 00:27:56,320
Speaker 5: part of the graph that now gets very steep in

548
00:27:56,440 --> 00:27:58,200
Speaker 5: some chunk of the economy and.

549
00:27:58,240 --> 00:27:59,679
Speaker 4: Hoping that they'll do something about it.

550
00:28:00,960 --> 00:28:04,040
Speaker 6: I think there is another part of what we're trying

551
00:28:04,040 --> 00:28:05,800
Speaker 6: to do at the Institute, which we lay out in

552
00:28:05,880 --> 00:28:08,640
Speaker 6: the sort of research agenda for the Anthropic Institute, which

553
00:28:08,680 --> 00:28:12,800
Speaker 6: is trying to understand the impact of our decisions, which

554
00:28:12,840 --> 00:28:15,720
Speaker 6: is a typical thing that economists will do at tech companies,

555
00:28:15,840 --> 00:28:17,880
Speaker 6: but we have a public benefit mandate, so we're trying

556
00:28:17,920 --> 00:28:21,040
Speaker 6: to understand the impact of our decisions on these broader

557
00:28:21,400 --> 00:28:24,240
Speaker 6: societal and economic outcomes that we care about, and then

558
00:28:24,400 --> 00:28:27,000
Speaker 6: using that to inform some of the decisions that we

559
00:28:27,040 --> 00:28:27,520
Speaker 6: actually see.

560
00:28:27,800 --> 00:28:29,520
Speaker 5: A goal that Peter and I have and we've talked

561
00:28:29,560 --> 00:28:32,800
Speaker 5: about internally, is if we get really good at measuring

562
00:28:32,800 --> 00:28:36,080
Speaker 5: things like a productivity multiplier of our technology, then I

563
00:28:36,119 --> 00:28:38,080
Speaker 5: would hope to use that to guide some of say

564
00:28:38,120 --> 00:28:41,040
Speaker 5: the early access programs we do for powerful models, where

565
00:28:41,040 --> 00:28:43,400
Speaker 5: if you see you get some tremendous multiplier in a

566
00:28:43,400 --> 00:28:46,760
Speaker 5: specific part of science, use that to redirect some of

567
00:28:46,800 --> 00:28:50,040
Speaker 5: your inference compute budget to that sector, and then you

568
00:28:50,080 --> 00:28:52,000
Speaker 5: can run an experiment and say, were we able to

569
00:28:52,000 --> 00:28:54,480
Speaker 5: make this thing go much faster. I think that could

570
00:28:54,480 --> 00:28:57,120
Speaker 5: be like an amazing tool to unlock the world, and

571
00:28:57,120 --> 00:29:00,200
Speaker 5: it's one that you could generalize across companies, and you

572
00:29:00,200 --> 00:29:04,040
Speaker 5: can generalize it into policy, So instead of say NSF

573
00:29:04,080 --> 00:29:06,680
Speaker 5: doing standard grant funding, it could be should we just

574
00:29:06,680 --> 00:29:09,000
Speaker 5: point for really powerful AIS systems at this chunk of

575
00:29:09,040 --> 00:29:11,560
Speaker 5: science and make it go faster. I think that's a

576
00:29:11,600 --> 00:29:13,120
Speaker 5: world that will come within reach soon.

577
00:29:13,240 --> 00:29:16,400
Speaker 2: Let's talk about this public benefit mission a little bit more.

578
00:29:16,440 --> 00:29:19,440
Speaker 2: We've been talking about ways this could change the economy.

579
00:29:19,960 --> 00:29:23,680
Speaker 3: How much do you see your job is basically strong.

580
00:29:23,840 --> 00:29:24,880
Speaker 3: AI is coming.

581
00:29:25,080 --> 00:29:27,920
Speaker 2: Yeah, it's coming whether we like it or not, and

582
00:29:28,000 --> 00:29:30,720
Speaker 2: it's important to be you want to be there as

583
00:29:30,760 --> 00:29:33,880
Speaker 2: like one of the shepherd's understanding which direction it goes

584
00:29:33,920 --> 00:29:36,960
Speaker 2: in the data that we should see to see what's emerging,

585
00:29:37,040 --> 00:29:39,240
Speaker 2: Like how much is that somewhat your role?

586
00:29:39,520 --> 00:29:42,440
Speaker 5: Yeah, but look, our guiding principle is but this technology

587
00:29:42,480 --> 00:29:45,000
Speaker 5: is being built by a variety of companies and a

588
00:29:45,040 --> 00:29:49,280
Speaker 5: variety of countries. But technology by default is unknown. It

589
00:29:49,320 --> 00:29:51,520
Speaker 5: will be known to the companies, it will not be

590
00:29:51,600 --> 00:29:54,200
Speaker 5: broadly understood or known by others. They'll just be able

591
00:29:54,200 --> 00:29:56,760
Speaker 5: to play with the models. Every bit of data we

592
00:29:56,840 --> 00:30:00,000
Speaker 5: can create, and especially systemically sharing data like the economic

593
00:30:00,080 --> 00:30:02,120
Speaker 5: index of what we've started to do on recursive self

594
00:30:02,120 --> 00:30:04,840
Speaker 5: improvement gives a world a better chance to sort of

595
00:30:04,840 --> 00:30:07,800
Speaker 5: prepare for this technology. Yeah, and both planned for its success,

596
00:30:07,840 --> 00:30:09,680
Speaker 5: Like what I talked about was science. We could be

597
00:30:10,120 --> 00:30:14,080
Speaker 5: intentional about driving science forward and also be warned about

598
00:30:14,160 --> 00:30:16,960
Speaker 5: risks like the cyber capabilities that you've talked about.

599
00:30:17,240 --> 00:30:19,320
Speaker 2: Well, so it's like that makes a lot of sense.

600
00:30:19,360 --> 00:30:21,760
Speaker 2: The company is going to see it before the world,

601
00:30:21,840 --> 00:30:24,120
Speaker 2: and Heskin is like, Okay, this is important to share.

602
00:30:24,200 --> 00:30:26,200
Speaker 2: This is not important to share, which which brings me

603
00:30:26,200 --> 00:30:28,840
Speaker 2: into another question. You know, I know people in the

604
00:30:28,880 --> 00:30:32,480
Speaker 2: AI research world done some reporting on the sort of

605
00:30:32,520 --> 00:30:33,640
Speaker 2: scene in SF.

606
00:30:34,080 --> 00:30:35,920
Speaker 3: You know, like when I think about a lot.

607
00:30:35,800 --> 00:30:38,000
Speaker 2: Of the people who are like at the very cutting

608
00:30:38,080 --> 00:30:42,480
Speaker 2: edge of AI ethics, AI technology, et cetera, I know

609
00:30:42,560 --> 00:30:45,480
Speaker 2: a lot of people who are how should I put this,

610
00:30:45,800 --> 00:30:52,440
Speaker 2: they have esoteric moral interests. Shrimp writes unusual attitudes about

611
00:30:52,600 --> 00:30:55,800
Speaker 2: experimental drug use. We know about the Chinese peptide scene

612
00:30:55,840 --> 00:30:59,240
Speaker 2: in the San Francisco, et cetera, and as a family podcast,

613
00:30:59,280 --> 00:31:03,320
Speaker 2: I would say certain and like perhaps different view on

614
00:31:03,400 --> 00:31:06,120
Speaker 2: sort of bourgeois even sexual values, and we know about

615
00:31:06,160 --> 00:31:09,320
Speaker 2: the sort of attitudes towards monogamy, et cetera within the

616
00:31:09,440 --> 00:31:10,840
Speaker 2: San Francisco research and.

617
00:31:11,000 --> 00:31:13,120
Speaker 4: Joe, there's going to be a protest against all thoughts

618
00:31:13,120 --> 00:31:15,920
Speaker 4: from San Francisco.

619
00:31:16,640 --> 00:31:18,280
Speaker 3: Yeah, not all engineers. I understand that.

620
00:31:18,440 --> 00:31:20,360
Speaker 2: But when we think about like, Okay, these are the

621
00:31:20,360 --> 00:31:22,720
Speaker 2: people who are going to see it first, should we

622
00:31:22,760 --> 00:31:25,320
Speaker 2: feel comfortable that this is a group of individuals, the

623
00:31:25,360 --> 00:31:29,840
Speaker 2: cohort of the most advanced AI researchers, whose intuitions about

624
00:31:29,840 --> 00:31:33,240
Speaker 2: what's important to communicate to the public are actually in

625
00:31:33,320 --> 00:31:37,480
Speaker 2: line with the public's interest, given how unrepresentative they are

626
00:31:38,120 --> 00:31:39,680
Speaker 2: of what I would call the American public.

627
00:31:39,800 --> 00:31:42,000
Speaker 5: Yes, as a as an Englishman, it fills me with

628
00:31:42,040 --> 00:31:43,440
Speaker 5: such joy to be asked about sex.

629
00:31:43,720 --> 00:31:48,760
Speaker 2: Yeah, I know your insights into the cohort of the

630
00:31:48,760 --> 00:31:51,480
Speaker 2: most advanced YOU search we're explorers.

631
00:31:51,680 --> 00:31:54,360
Speaker 5: People that are explorers, and this is so true in

632
00:31:54,360 --> 00:31:57,400
Speaker 5: San Francisco. End up being like that. There's a broad

633
00:31:57,480 --> 00:31:59,680
Speaker 5: range of types of people, and sometimes they're really really

634
00:31:59,720 --> 00:32:02,680
Speaker 5: different or yeah, really really eccentric and they're brilliant to

635
00:32:02,720 --> 00:32:03,840
Speaker 5: be lovable and everything else.

636
00:32:03,880 --> 00:32:05,120
Speaker 3: Yeah, sure, love them.

637
00:32:05,240 --> 00:32:08,800
Speaker 5: You don't want only that class of people to be

638
00:32:08,840 --> 00:32:10,760
Speaker 5: the ones calling the shots on what we know about

639
00:32:10,800 --> 00:32:13,240
Speaker 5: this technology. Yeah. The whole purpose of what we're doing

640
00:32:13,280 --> 00:32:15,320
Speaker 5: is we're trying to set up systems by which you

641
00:32:15,320 --> 00:32:19,160
Speaker 5: could eventually mandate through policy that companies share information, you know,

642
00:32:19,200 --> 00:32:22,840
Speaker 5: and THROPIC has long pushed for transparency legislation in various

643
00:32:22,880 --> 00:32:26,040
Speaker 5: states around America that gets companies like us to report

644
00:32:26,080 --> 00:32:28,120
Speaker 5: out the sorts of tests we're running on our systems

645
00:32:28,360 --> 00:32:32,320
Speaker 5: and share it publicly. My whole mindset is the public

646
00:32:32,920 --> 00:32:37,040
Speaker 5: and policy makers and economists, everyone deserve the ability to

647
00:32:37,080 --> 00:32:39,440
Speaker 5: advocate for what information should come out of a frontier

648
00:32:39,480 --> 00:32:42,080
Speaker 5: and it should be forced out of a frontier eventually

649
00:32:42,080 --> 00:32:44,240
Speaker 5: by law. Like that is how you solve this issue.

650
00:32:44,400 --> 00:32:47,600
Speaker 4: Do you hire more normies? Yeah?

651
00:32:46,240 --> 00:32:48,320
Speaker 5: Like me personally?

652
00:32:48,440 --> 00:32:51,080
Speaker 2: Yeah, Like is there an important thing like hiring people

653
00:32:51,120 --> 00:32:53,840
Speaker 2: that don't all share these certain like you know, in

654
00:32:53,880 --> 00:32:54,920
Speaker 2: group ways of seeing the world.

655
00:32:55,000 --> 00:32:58,600
Speaker 5: So, you know, the Youngthropic Institute we have teams of economists,

656
00:32:58,680 --> 00:33:01,520
Speaker 5: of social scientists, of what you might think of as

657
00:33:01,560 --> 00:33:04,560
Speaker 5: weapons experts, our frontier red team, things that go bump

658
00:33:04,640 --> 00:33:08,400
Speaker 5: in the night, lawyers, and increasingly other types of people.

659
00:33:08,480 --> 00:33:11,040
Speaker 5: The goal is to build what I think of as

660
00:33:11,160 --> 00:33:15,680
Speaker 5: a highly ideologically diverse research function within the organization. But

661
00:33:15,760 --> 00:33:17,760
Speaker 5: it's part of the advocating sort of on behalf of

662
00:33:17,760 --> 00:33:20,440
Speaker 5: the world for different forms of study that we might do.

663
00:33:21,040 --> 00:33:23,760
Speaker 5: So ANDTHROPIC generally hires a really broad range of people,

664
00:33:23,960 --> 00:33:27,000
Speaker 5: but the institute specifically is trying to compose a very

665
00:33:27,000 --> 00:33:30,080
Speaker 5: broad set of interdisciplinary experts for this exact reason.

666
00:33:46,200 --> 00:33:49,680
Speaker 4: Let me ask a slightly different question on hiring. I

667
00:33:49,680 --> 00:33:51,840
Speaker 4: guess the two part questions. So, first of all, we

668
00:33:51,880 --> 00:33:54,000
Speaker 4: get a lot of executives on the show. We've been

669
00:33:54,000 --> 00:33:57,280
Speaker 4: asking all of them if they've changed their hiring process,

670
00:33:57,400 --> 00:34:01,120
Speaker 4: if they've changed the questions they ask potential employees at

671
00:34:01,120 --> 00:34:05,800
Speaker 4: those initial stages of job applications because of AI. And then, secondly,

672
00:34:05,920 --> 00:34:09,520
Speaker 4: what are you seeing within your own ranks at the company.

673
00:34:09,560 --> 00:34:11,480
Speaker 4: And then, Peter, I'm sure you could talk about this

674
00:34:11,480 --> 00:34:15,000
Speaker 4: more broadly in terms of who's most in demand at

675
00:34:15,040 --> 00:34:17,960
Speaker 4: the moment, because the conventional wisdom right now is that

676
00:34:18,680 --> 00:34:22,719
Speaker 4: if you're a younger employee with less experience, a lot

677
00:34:22,760 --> 00:34:24,359
Speaker 4: of the stuff that you would be doing can now

678
00:34:24,400 --> 00:34:25,640
Speaker 4: be automated through AI.

679
00:34:26,000 --> 00:34:29,640
Speaker 5: So there's two trends showing up. One, I have a

680
00:34:29,680 --> 00:34:32,879
Speaker 5: new team called the Rule of Law and AI. Our

681
00:34:32,960 --> 00:34:35,560
Speaker 5: plan was to initially hire a bunch of engineers and

682
00:34:35,600 --> 00:34:38,840
Speaker 5: then a bunch of legal experts and scholars. Instead, we

683
00:34:39,000 --> 00:34:41,360
Speaker 5: just hiring the legal experts and scholars because Claude is

684
00:34:41,360 --> 00:34:43,040
Speaker 5: good enough at doing all of the engineering, but they

685
00:34:43,080 --> 00:34:46,520
Speaker 5: can actually just like feed themselves using Claude in terms

686
00:34:46,520 --> 00:34:49,120
Speaker 5: of the engineering resources. So that's a change in hiring.

687
00:34:49,200 --> 00:34:52,440
Speaker 5: It means I'm hiring more interdisciplinary people earlier than I

688
00:34:52,440 --> 00:34:55,440
Speaker 5: would have before. We are also seeing the emergence of

689
00:34:55,480 --> 00:34:58,120
Speaker 5: what I think of as a Barbell hiring pattern inside

690
00:34:58,120 --> 00:35:00,680
Speaker 5: and thropic, where there is a trem and does return

691
00:35:00,719 --> 00:35:04,319
Speaker 5: on experience. So we are hiring more senior people than

692
00:35:04,360 --> 00:35:07,399
Speaker 5: we did in the past because their intuitions and their

693
00:35:07,440 --> 00:35:10,719
Speaker 5: ideas for what to pursue are like massively compounded by

694
00:35:10,760 --> 00:35:14,560
Speaker 5: AI systems. Were Also when we look at very early

695
00:35:14,680 --> 00:35:17,480
Speaker 5: people are often hiring people who are now like AI

696
00:35:17,640 --> 00:35:19,480
Speaker 5: native and know how to use the tools and are

697
00:35:19,520 --> 00:35:20,960
Speaker 5: well versed in it.

698
00:35:21,120 --> 00:35:23,440
Speaker 4: So we're seen very decent amount of I guess AI

699
00:35:23,560 --> 00:35:25,640
Speaker 4: natives now, people who have grown up with.

700
00:35:26,280 --> 00:35:28,400
Speaker 5: From GPT two in twenty nineteen.

701
00:35:28,560 --> 00:35:31,360
Speaker 4: My perception of time is so.

702
00:35:31,120 --> 00:35:33,000
Speaker 5: I found this chilling as well. You know, as someone

703
00:35:33,040 --> 00:35:35,960
Speaker 5: in their thirties, you realize, but I think that the

704
00:35:36,080 --> 00:35:39,360
Speaker 5: trends I see, I do think that there's this question

705
00:35:39,440 --> 00:35:42,319
Speaker 5: of how you have as much early career hiring in

706
00:35:42,360 --> 00:35:44,480
Speaker 5: the future as you did in the past. I think

707
00:35:44,520 --> 00:35:47,839
Speaker 5: one of the only areas where there is slightly suggestive

708
00:35:47,920 --> 00:35:50,680
Speaker 5: data is that something might be going on with early

709
00:35:50,760 --> 00:35:53,840
Speaker 5: career hiring and it kind of intuitively feels right to

710
00:35:53,880 --> 00:35:55,800
Speaker 5: all of us, but that we might be observing by effect.

711
00:35:55,800 --> 00:35:57,920
Speaker 5: And when I look at hiring patterns in anthropic, we're

712
00:35:57,960 --> 00:36:00,520
Speaker 5: still hiring young people, but some teimes are hire slightly

713
00:36:00,560 --> 00:36:03,040
Speaker 5: fewer of them than before and hiring more experienced people.

714
00:36:03,280 --> 00:36:06,400
Speaker 6: Yeah, so I'll briefly say something about how we've shifted

715
00:36:06,440 --> 00:36:10,160
Speaker 6: some of our hiring practices. Like concretely, I think before

716
00:36:10,239 --> 00:36:14,440
Speaker 6: claud code, you might ask an economist to do some

717
00:36:14,560 --> 00:36:17,080
Speaker 6: of the data work in an assessment kind of live,

718
00:36:17,160 --> 00:36:19,919
Speaker 6: like download the data, run the regressions, do the analysis

719
00:36:20,320 --> 00:36:23,640
Speaker 6: by hand, and then you might eventually let them use

720
00:36:23,680 --> 00:36:27,080
Speaker 6: AI to do all of that work. But we've needed

721
00:36:27,120 --> 00:36:32,040
Speaker 6: to increasingly shift our strategy of evaluation away from can

722
00:36:32,080 --> 00:36:35,000
Speaker 6: you implement the work even with AI to do you know,

723
00:36:35,040 --> 00:36:39,040
Speaker 6: how to delegate and direct the model in a somewhat

724
00:36:39,160 --> 00:36:43,040
Speaker 6: messy environment, And can you evaluate the quality of the

725
00:36:43,080 --> 00:36:45,360
Speaker 6: work maybe by like looking at a pr can you.

726
00:36:45,320 --> 00:36:46,840
Speaker 2: Talk a little bit more about what that looks like

727
00:36:46,880 --> 00:36:49,800
Speaker 2: specifically in the E confine, you know there are listeners

728
00:36:49,880 --> 00:36:52,200
Speaker 2: probably think about, Okay, what is I want to level

729
00:36:52,280 --> 00:36:53,440
Speaker 2: up in my AI use?

730
00:36:53,560 --> 00:36:56,800
Speaker 3: So I'm not just asking like, yeah, whatever, what is

731
00:36:56,880 --> 00:36:57,920
Speaker 3: it that actually mean?

732
00:36:57,960 --> 00:36:59,840
Speaker 2: For an economist and you used to be able to

733
00:36:59,840 --> 00:37:04,319
Speaker 2: band yeah, so for a financial economist and economists someone, well,

734
00:37:04,320 --> 00:37:07,479
Speaker 2: in this world, what is like the most advanced form

735
00:37:07,520 --> 00:37:09,200
Speaker 2: of usage of AAI actually look like?

736
00:37:09,280 --> 00:37:10,760
Speaker 6: Well, I don't know if I'll give you the example

737
00:37:10,760 --> 00:37:12,759
Speaker 6: of the most advanced form of usage, but I'll give

738
00:37:12,760 --> 00:37:15,960
Speaker 6: it an anecdote of my experience using Claude where I

739
00:37:15,960 --> 00:37:18,440
Speaker 6: wanted to run this cross state regression I can't remember

740
00:37:18,440 --> 00:37:20,799
Speaker 6: exactly what it was, and I wanted to do it

741
00:37:20,920 --> 00:37:23,880
Speaker 6: a pooled cross sectional regression. So looking at what happened

742
00:37:23,880 --> 00:37:25,960
Speaker 6: in twenty twenty four or twenty twenty three, and going

743
00:37:26,000 --> 00:37:29,080
Speaker 6: all the way back to pre pandemic, I remember asking

744
00:37:29,120 --> 00:37:31,799
Speaker 6: Claude to go out and download the data from the

745
00:37:31,800 --> 00:37:34,720
Speaker 6: Census Bureau, from the Bureau of Labor Statistics, et cetera.

746
00:37:35,480 --> 00:37:39,840
Speaker 6: And there was this very unexpected quirk where the model

747
00:37:39,840 --> 00:37:43,760
Speaker 6: couldn't access data from before twenty nineteen and just would

748
00:37:43,840 --> 00:37:47,880
Speaker 6: not surface that mistake. And I would ask it multiple

749
00:37:47,880 --> 00:37:51,279
Speaker 6: times like no, like, don't hardcode numbers because it sort

750
00:37:51,280 --> 00:37:54,160
Speaker 6: of had this unexpected failure mode where it said, oh,

751
00:37:54,200 --> 00:37:55,719
Speaker 6: I know what those numbers were. And it's just like

752
00:37:56,360 --> 00:38:00,319
Speaker 6: from sort of training data populated the data set, and

753
00:38:00,680 --> 00:38:03,560
Speaker 6: you might not always be attuned unless you're sort of

754
00:38:03,640 --> 00:38:06,600
Speaker 6: you have this tacit knowledge about what does it pass

755
00:38:06,640 --> 00:38:09,920
Speaker 6: a snifftest when you run the analysis, and then you

756
00:38:09,960 --> 00:38:11,719
Speaker 6: like dig into what the model actually does and it

757
00:38:11,719 --> 00:38:15,160
Speaker 6: has failed in sort of unexpected or unusual ways, and

758
00:38:15,200 --> 00:38:17,960
Speaker 6: so that's like the type of assessment that we've built.

759
00:38:18,200 --> 00:38:22,200
Speaker 6: Can you be attentive to the very specific decisions that

760
00:38:22,280 --> 00:38:24,279
Speaker 6: need to be made along the way that are very

761
00:38:24,280 --> 00:38:27,680
Speaker 6: consequential for the validity of veracity of the results that

762
00:38:27,800 --> 00:38:28,520
Speaker 6: you find.

763
00:38:28,880 --> 00:38:32,200
Speaker 5: Yeah, A colleague did an off site presentation last year

764
00:38:32,239 --> 00:38:34,680
Speaker 5: which said, I have locked the doors and we are

765
00:38:34,719 --> 00:38:37,480
Speaker 5: reading transcripts, and the point was we just need to

766
00:38:37,480 --> 00:38:40,640
Speaker 5: read more of the raw data and develop that culture

767
00:38:40,680 --> 00:38:44,160
Speaker 5: where IFAI systems are doing increasingly large amounts of the work,

768
00:38:44,520 --> 00:38:46,560
Speaker 5: you need to have a culture of being competent at

769
00:38:46,600 --> 00:38:50,000
Speaker 5: spot checking their work and reading their reasoning because occasionally

770
00:38:50,000 --> 00:38:51,480
Speaker 5: stuff like this happens and then.

771
00:38:51,360 --> 00:38:53,560
Speaker 4: Peter in the broader data that you're looking at Are

772
00:38:53,560 --> 00:38:56,120
Speaker 4: you seeing the same sort of Barbell effect in terms

773
00:38:56,160 --> 00:38:57,480
Speaker 4: of employment that Jack describe.

774
00:38:57,600 --> 00:39:00,000
Speaker 6: Yeah, so I think again, what makes it really challenge

775
00:39:00,239 --> 00:39:03,120
Speaker 6: is we've had the largest non recessionary labor markets slow

776
00:39:03,160 --> 00:39:06,880
Speaker 6: down on record that it's very hard for young people

777
00:39:06,920 --> 00:39:10,160
Speaker 6: to graduate into a labor market that doesn't have sufficient

778
00:39:10,239 --> 00:39:13,239
Speaker 6: churn or opportunity for them to get a foothold. But

779
00:39:13,320 --> 00:39:14,520
Speaker 6: one of the things that we did see in this

780
00:39:14,560 --> 00:39:18,880
Speaker 6: report from March was that young workers in these high

781
00:39:19,000 --> 00:39:21,680
Speaker 6: AI exposed roles where Claude is being used to automate

782
00:39:21,719 --> 00:39:25,640
Speaker 6: specific tasks, have had somewhat weaker job finding rates.

783
00:39:26,600 --> 00:39:29,279
Speaker 2: But it's suggesting confounders was the boom and hiring in

784
00:39:29,320 --> 00:39:32,640
Speaker 2: twenty twenty one in these execs.

785
00:39:31,800 --> 00:39:34,200
Speaker 6: Exactly, And there's a recent paper about so the rise

786
00:39:34,239 --> 00:39:37,320
Speaker 6: of remote work maybe being sort of the actual cause

787
00:39:37,360 --> 00:39:40,720
Speaker 6: of this type of fact. Another team at the anthrop

788
00:39:40,800 --> 00:39:44,400
Speaker 6: Against the Dude Societal Impacts recently ran this very large

789
00:39:44,400 --> 00:39:47,480
Speaker 6: scale qualitative survey eighty one thousand people around the world,

790
00:39:47,800 --> 00:39:49,960
Speaker 6: asking them questions about hopes and fears that they have

791
00:39:50,040 --> 00:39:54,520
Speaker 6: with respect to AI. Unsurprisingly, concerns about the impact on

792
00:39:54,560 --> 00:39:57,120
Speaker 6: the labor market and on the economy rose to the surface.

793
00:39:57,840 --> 00:39:59,799
Speaker 6: My team dug into those data a little bit more

794
00:39:59,840 --> 00:40:03,239
Speaker 6: to try to answer some of these specific questions, and

795
00:40:03,320 --> 00:40:06,880
Speaker 6: what you see is that young workers at least express

796
00:40:07,040 --> 00:40:10,080
Speaker 6: concern about job loss at twice the rate as do

797
00:40:10,239 --> 00:40:13,360
Speaker 6: more senior workers, and fears about job loss more broadly

798
00:40:13,719 --> 00:40:16,480
Speaker 6: are more elevated for workers who are in these roles

799
00:40:16,520 --> 00:40:20,640
Speaker 6: that we identify as being most exposed to displacement effects

800
00:40:20,880 --> 00:40:23,040
Speaker 6: from AI. So there's a bit of a gap between

801
00:40:23,360 --> 00:40:26,000
Speaker 6: perception and maybe what you see in the hard data.

802
00:40:26,040 --> 00:40:28,359
Speaker 6: But that was something that was true even in recent

803
00:40:28,440 --> 00:40:30,680
Speaker 6: years on other dimensions. So it's an important thing to

804
00:40:30,680 --> 00:40:31,319
Speaker 6: pay attention to.

805
00:40:31,840 --> 00:40:34,120
Speaker 4: So we've been talking about the labor market, and one

806
00:40:34,239 --> 00:40:37,160
Speaker 4: other thing I'm interested in is the impact of AI

807
00:40:37,520 --> 00:40:41,040
Speaker 4: on I guess corporates themselves. So if we think about

808
00:40:41,200 --> 00:40:45,279
Speaker 4: certainly America's corporate landscape in recent years, it feels like

809
00:40:45,560 --> 00:40:49,400
Speaker 4: the big basically get bigger, Right, there's economies of scale,

810
00:40:49,480 --> 00:40:51,600
Speaker 4: they have a bunch of money that they can use

811
00:40:51,680 --> 00:40:54,920
Speaker 4: to actually buy some of those data exactly exactly. So

812
00:40:55,000 --> 00:41:00,560
Speaker 4: would you expect AI to I guess intensify that trend

813
00:41:00,680 --> 00:41:02,640
Speaker 4: of the big getting bigger or would you expect to

814
00:41:02,680 --> 00:41:05,480
Speaker 4: perhaps have a leveling effect where people have this new

815
00:41:05,560 --> 00:41:08,279
Speaker 4: tool that they can use to, you know, set up

816
00:41:08,280 --> 00:41:08,920
Speaker 4: a new company.

817
00:41:09,040 --> 00:41:11,600
Speaker 5: I'm curious what Peter's take is, but I think that

818
00:41:11,760 --> 00:41:15,200
Speaker 5: something a helpful analogy here is the invention of electricity,

819
00:41:15,360 --> 00:41:19,719
Speaker 5: where electricity arrived and existing factories put light bulbs in

820
00:41:19,840 --> 00:41:22,360
Speaker 5: and other things. But it was a new generation of

821
00:41:22,400 --> 00:41:25,480
Speaker 5: factories that were built around the assumption that electricity existed

822
00:41:25,560 --> 00:41:28,480
Speaker 5: that really grew and did transform into things in the economy.

823
00:41:29,040 --> 00:41:31,440
Speaker 5: What I see now when we look at large enterprises

824
00:41:31,520 --> 00:41:34,719
Speaker 5: is they can get a lot of utility out of

825
00:41:34,800 --> 00:41:37,400
Speaker 5: Claude because of their data, because they can get a

826
00:41:37,480 --> 00:41:40,799
Speaker 5: multiplier effect at scale. But it takes huge amounts of

827
00:41:40,800 --> 00:41:43,480
Speaker 5: conviction to basically bash for all of the bureaucracy you know,

828
00:41:43,920 --> 00:41:46,680
Speaker 5: used to work at Bloomberg implement new technology at Bloomberg.

829
00:41:47,040 --> 00:41:53,440
Speaker 5: Challenging comment about that same is true of any large organization.

830
00:41:54,400 --> 00:41:58,200
Speaker 5: Young organizations are building themselves around AI at the center,

831
00:41:58,600 --> 00:42:01,879
Speaker 5: and these organizations are moved really, really quickly because they

832
00:42:01,960 --> 00:42:04,600
Speaker 5: just they have a speed advantage from building on the

833
00:42:04,640 --> 00:42:07,200
Speaker 5: assumption that this new form of electricity was going to

834
00:42:07,200 --> 00:42:08,360
Speaker 5: be integrals of that business.

835
00:42:08,520 --> 00:42:11,000
Speaker 6: Yeah, so I think the tension that you express is

836
00:42:11,040 --> 00:42:14,280
Speaker 6: exactly the one that I don't have a strong handle

837
00:42:14,400 --> 00:42:16,359
Speaker 6: on at the moment. One thing that we do see

838
00:42:16,360 --> 00:42:21,040
Speaker 6: in our data is when businesses do embed cloud capabilities

839
00:42:21,360 --> 00:42:25,160
Speaker 6: in automated ways through the API. As I mentioned before,

840
00:42:25,239 --> 00:42:29,600
Speaker 6: these very complex tasks rely on disproportionately more contextual information

841
00:42:30,000 --> 00:42:34,760
Speaker 6: than very basic document synthesis and summarization. What that points

842
00:42:34,760 --> 00:42:38,000
Speaker 6: in the direction of are the complementary investments that large

843
00:42:38,000 --> 00:42:42,360
Speaker 6: businesses need to make to centralize, codify, and make available

844
00:42:42,440 --> 00:42:45,759
Speaker 6: the data that does exist somewhere within the organization, but

845
00:42:46,040 --> 00:42:49,600
Speaker 6: for historic and technical reasons, maybe even regulatory reasons, it's

846
00:42:50,280 --> 00:42:53,400
Speaker 6: behind a firewall of some form or another. There's also

847
00:42:53,560 --> 00:42:56,960
Speaker 6: like sort of organizational workflow changes that likely need to

848
00:42:57,000 --> 00:43:00,440
Speaker 6: be made. Some of the most crucial information that's needed

849
00:43:00,480 --> 00:43:04,000
Speaker 6: for some types of cognitive work is tacit knowledge that

850
00:43:04,120 --> 00:43:06,840
Speaker 6: exists in your colleague's mind. And unless you have a

851
00:43:06,880 --> 00:43:10,560
Speaker 6: process that elicits that information that workers feel sort of

852
00:43:10,760 --> 00:43:14,640
Speaker 6: incentivized to share that information and kind of trust the system,

853
00:43:14,880 --> 00:43:18,920
Speaker 6: the capabilities alone might not necessarily generate that productivity. And

854
00:43:18,960 --> 00:43:22,720
Speaker 6: so whether or not big firms end up restructuring themselves

855
00:43:22,760 --> 00:43:26,360
Speaker 6: quickly enough or whether this materializes through the process of

856
00:43:26,400 --> 00:43:28,839
Speaker 6: creative destruction. I think the jury is still a bit out.

857
00:43:28,960 --> 00:43:29,120
Speaker 3: Yeah.

858
00:43:29,200 --> 00:43:33,200
Speaker 2: Brought this open recently with the David Solomon, the Goldenman CEO,

859
00:43:33,280 --> 00:43:35,000
Speaker 2: and I started to wonder like this sort of like

860
00:43:35,120 --> 00:43:38,359
Speaker 2: internal alignment question of like the big rainmakers, do they

861
00:43:38,440 --> 00:43:41,560
Speaker 2: have an incentive essentially for information hoarding and not sharing

862
00:43:41,600 --> 00:43:43,040
Speaker 2: what the company.

863
00:43:42,600 --> 00:43:44,640
Speaker 3: That might be their only thing that keeping them employed.

864
00:43:44,680 --> 00:43:47,320
Speaker 5: And when I talk to customers, I say, it's don't

865
00:43:47,320 --> 00:43:49,440
Speaker 5: think of it like you're buying a technology. Think of

866
00:43:49,480 --> 00:43:52,480
Speaker 5: it maybe, but you're now employing thousands of people, but

867
00:43:52,800 --> 00:43:54,920
Speaker 5: functionally like the chief of staff to the CEO, I

868
00:43:54,920 --> 00:43:56,640
Speaker 5: mean need to say, access to date of the chief

869
00:43:56,640 --> 00:43:59,560
Speaker 5: of staff would have. This is completely counterintuitive and it

870
00:43:59,600 --> 00:44:18,400
Speaker 5: is not technology is typically bulls all solved. Jack.

871
00:44:18,600 --> 00:44:21,080
Speaker 2: In your newsletter import AI, you tend to write a

872
00:44:21,120 --> 00:44:24,680
Speaker 2: little short story of a sort of aspiring sci fi writer,

873
00:44:24,840 --> 00:44:27,719
Speaker 2: like a literal sci fi writer, just in the news are.

874
00:44:27,960 --> 00:44:30,600
Speaker 2: One of the classic sci fi scenarios that people have

875
00:44:30,640 --> 00:44:35,160
Speaker 2: been talking about for decades is the possibility that robots

876
00:44:35,239 --> 00:44:35,960
Speaker 2: or AI.

877
00:44:35,800 --> 00:44:37,680
Speaker 3: Will kill humans quite literally.

878
00:44:38,560 --> 00:44:42,200
Speaker 4: Do you when you think about ultimate negative externalogy, when.

879
00:44:42,040 --> 00:44:45,480
Speaker 2: You think about like training, AI and safety research, et cetera.

880
00:44:46,040 --> 00:44:49,440
Speaker 2: Do you assign a reasonable possibility to the fact that

881
00:44:49,880 --> 00:44:54,240
Speaker 2: ill trained or misaligned AI will literally kill all humans?

882
00:44:54,480 --> 00:44:59,239
Speaker 5: Uh No, but there's a big butt yeah, Like the

883
00:44:59,280 --> 00:45:02,640
Speaker 5: world needs an option to be able to potentially slow

884
00:45:02,719 --> 00:45:05,839
Speaker 5: down or even in extreme circumstances, pause the development of

885
00:45:05,840 --> 00:45:08,000
Speaker 5: this technology if we were to see that. And I'll

886
00:45:08,040 --> 00:45:10,000
Speaker 5: just give you the exact way I think about it.

887
00:45:10,320 --> 00:45:14,120
Speaker 5: At Anthropic, we test out our systems for alignment failures.

888
00:45:14,200 --> 00:45:15,960
Speaker 5: You know, we publish this, so do all of the

889
00:45:16,000 --> 00:45:20,080
Speaker 5: other companies. And you see, Hey, under extreme circumstances, maybe

890
00:45:20,080 --> 00:45:22,239
Speaker 5: the system breaks out of a container and sends an

891
00:45:22,239 --> 00:45:26,320
Speaker 5: email to someone. Maybe the system pretends to blackmail a

892
00:45:26,440 --> 00:45:28,439
Speaker 5: CEO of that it thinks is going to shut it down.

893
00:45:28,520 --> 00:45:31,600
Speaker 5: These are the sorts of vis actually have been observed, yes,

894
00:45:31,800 --> 00:45:32,920
Speaker 5: in the lab setting.

895
00:45:33,400 --> 00:45:37,680
Speaker 2: And the thing is is the models know you could see,

896
00:45:37,880 --> 00:45:39,880
Speaker 2: Oh I'm being tested right now, So I'm going to

897
00:45:39,880 --> 00:45:43,040
Speaker 2: say this output so that the human reader thinks AM

898
00:45:43,080 --> 00:45:45,759
Speaker 2: more aligned than I am. These are real things, not sci fi.

899
00:45:45,800 --> 00:45:48,600
Speaker 5: These These are real things that we observe and then

900
00:45:48,680 --> 00:45:51,640
Speaker 5: we do like significant amount of work, and then we

901
00:45:51,680 --> 00:45:55,480
Speaker 5: release models that don't have these properties. But if you

902
00:45:55,520 --> 00:45:58,640
Speaker 5: were to enter a world where say, every time we

903
00:45:58,719 --> 00:46:01,080
Speaker 5: trained a new system, the rates of all of this

904
00:46:01,120 --> 00:46:05,040
Speaker 5: stuff went up one hundredfold, you might say, what, that's

905
00:46:05,080 --> 00:46:07,720
Speaker 5: pretty concerning. It seems like if we make for systems

906
00:46:07,719 --> 00:46:11,120
Speaker 5: above a certain level of intelligence, they become radically misaligned

907
00:46:11,120 --> 00:46:14,480
Speaker 5: against all human interests. That's the kind of circumstance if

908
00:46:14,520 --> 00:46:17,560
Speaker 5: that happens. The world needs information, and the world would

909
00:46:17,560 --> 00:46:19,840
Speaker 5: want an option to like slow or pause for development

910
00:46:19,880 --> 00:46:22,440
Speaker 5: of a tech if you encountered that, which we haven't today.

911
00:46:22,680 --> 00:46:24,759
Speaker 5: So to answer your question, like, I don't. I don't

912
00:46:24,800 --> 00:46:26,920
Speaker 5: worry about it today, but a lot of the measurements

913
00:46:26,920 --> 00:46:29,920
Speaker 5: and analysis work we do is to cue us if

914
00:46:29,920 --> 00:46:30,919
Speaker 5: they trend.

915
00:46:31,160 --> 00:46:32,160
Speaker 3: You do worry about it.

916
00:46:32,200 --> 00:46:35,920
Speaker 2: I mean, like you don't think it's happening today. But

917
00:46:36,040 --> 00:46:39,560
Speaker 2: part of the work you're doing specifically could be said

918
00:46:39,640 --> 00:46:44,080
Speaker 2: to avoid the outcome where AI is built, where in

919
00:46:44,120 --> 00:46:46,920
Speaker 2: the pursuit of a goal, it would kill all humans.

920
00:46:47,080 --> 00:46:47,400
Speaker 5: Yeah.

921
00:46:47,440 --> 00:46:52,160
Speaker 4: Wait, is human extinction a risk factor in the anthropic ipl.

922
00:46:55,160 --> 00:46:57,640
Speaker 3: I want to know now in the confidential as one.

923
00:46:59,400 --> 00:47:01,040
Speaker 4: Understand that's comment. That's fine.

924
00:47:01,160 --> 00:47:02,040
Speaker 3: Do you have others?

925
00:47:02,120 --> 00:47:04,520
Speaker 2: Would you say that there are a significant number of

926
00:47:05,120 --> 00:47:09,680
Speaker 2: anthropic employees who stay up at night thinking about human

927
00:47:09,760 --> 00:47:10,760
Speaker 2: extinction risk.

928
00:47:10,680 --> 00:47:13,359
Speaker 5: Everyone, and this is true of all of the labs. Yeah,

929
00:47:13,360 --> 00:47:16,160
Speaker 5: everyone who works on this technology sees it as the

930
00:47:16,320 --> 00:47:19,759
Speaker 5: highest stakes technology that's ever been built. We're basically the

931
00:47:19,920 --> 00:47:23,000
Speaker 5: potential encoded of in itself to massively benefit the world

932
00:47:23,320 --> 00:47:26,719
Speaker 5: or ruin the world, or you know, cause extinction. I

933
00:47:26,760 --> 00:47:29,799
Speaker 5: think the bulk of the risk is us messing it up,

934
00:47:30,040 --> 00:47:34,240
Speaker 5: like whether through misuse or ignoring risks, or not setting

935
00:47:34,280 --> 00:47:37,000
Speaker 5: up the right policy environment and getting some kind of

936
00:47:37,040 --> 00:47:39,839
Speaker 5: emergent set of failures. Now I don't My main risk

937
00:47:39,880 --> 00:47:42,520
Speaker 5: isn't one of extinction. It's somehow we like screw up

938
00:47:42,520 --> 00:47:45,560
Speaker 5: the technology really badly and delay all of the sort

939
00:47:45,560 --> 00:47:48,280
Speaker 5: of technological progress that could come from it and maybe

940
00:47:48,280 --> 00:47:51,000
Speaker 5: turn it into something analogs in nuclear power where you lose.

941
00:47:51,360 --> 00:47:52,640
Speaker 3: I guess the thing is is, you.

942
00:47:52,600 --> 00:47:55,960
Speaker 2: Know, like there's this fellow out there, eliozer Yudkowski, and

943
00:47:56,000 --> 00:47:58,800
Speaker 2: I always see these people like he's a crank, don't

944
00:47:58,840 --> 00:48:01,720
Speaker 2: listen to him, blah blah blah. But then I read

945
00:48:01,760 --> 00:48:03,960
Speaker 2: some of the other like papers that have people who

946
00:48:03,960 --> 00:48:06,520
Speaker 2: are taken more seriously, and I'm like they don't seem

947
00:48:06,560 --> 00:48:10,600
Speaker 2: that different. I read Super Intelligence recently by Nicholas Boston.

948
00:48:10,640 --> 00:48:13,040
Speaker 2: I was like, oh, this 'd Tuski is not alone.

949
00:48:13,080 --> 00:48:15,600
Speaker 2: There are a number of people who think that are

950
00:48:16,160 --> 00:48:19,560
Speaker 2: reasonable conditions in which the goals of the AI end

951
00:48:19,640 --> 00:48:22,080
Speaker 2: up wiping out every person on earth. Yes, it does

952
00:48:22,120 --> 00:48:24,920
Speaker 2: not seem like an extreme, extreme minority view.

953
00:48:24,920 --> 00:48:27,560
Speaker 5: Like concern the purpose of measuring these systems and why

954
00:48:27,600 --> 00:48:30,319
Speaker 5: anthropic is so I've spoken about it is right now

955
00:48:30,440 --> 00:48:33,000
Speaker 5: we say exactly what we see. And if you were

956
00:48:33,040 --> 00:48:34,840
Speaker 5: in some situation in the future where you saw this

957
00:48:35,040 --> 00:48:37,680
Speaker 5: what I call radical misalignment, which is the kind of

958
00:48:37,680 --> 00:48:40,160
Speaker 5: thing that you'd ask you worries about. You tell the

959
00:48:40,200 --> 00:48:42,560
Speaker 5: world and you want to have set up the world

960
00:48:42,600 --> 00:48:43,680
Speaker 5: to believe you if you see that.

961
00:48:44,120 --> 00:48:47,680
Speaker 4: You know, Joe mentioned that blackmail example. And you see

962
00:48:47,680 --> 00:48:51,880
Speaker 4: these headlines like Mythos likes to be thanked and doesn't

963
00:48:52,040 --> 00:48:54,920
Speaker 4: like bad users and gets mad at people that work

964
00:48:54,960 --> 00:48:57,640
Speaker 4: it too hard or whatever. To what degree do you

965
00:48:57,640 --> 00:49:02,680
Speaker 4: yourself actually anthropomorphosize some of these models, Like what should

966
00:49:02,680 --> 00:49:04,840
Speaker 4: we think when we see the headline Mythos wants to

967
00:49:04,880 --> 00:49:05,640
Speaker 4: be thanked by.

968
00:49:06,000 --> 00:49:07,960
Speaker 5: I'm as polite to Claude as I am to my

969
00:49:08,200 --> 00:49:12,239
Speaker 5: like car or pets. So yeah, I am orphised them.

970
00:49:12,280 --> 00:49:14,480
Speaker 5: But you know, if your car is having trouble, you're

971
00:49:14,520 --> 00:49:16,560
Speaker 5: like taking easy, buddy, It's okay, We're going to get.

972
00:49:18,040 --> 00:49:18,279
Speaker 4: People.

973
00:49:19,760 --> 00:49:21,880
Speaker 6: I think, you know, it's a good way to develop

974
00:49:21,920 --> 00:49:28,560
Speaker 6: good virtue is just model. You're developing a habit of

975
00:49:28,640 --> 00:49:31,759
Speaker 6: interacting with some type of intelligence. It might not be

976
00:49:31,800 --> 00:49:33,359
Speaker 6: the same type of intelligence that we have.

977
00:49:33,480 --> 00:49:36,320
Speaker 4: But then every time I type please into a prompt,

978
00:49:36,360 --> 00:49:39,239
Speaker 4: I worry I'm wasting energy, which also is a moral.

979
00:49:39,920 --> 00:49:42,880
Speaker 5: I wouldn't. I wouldn't worry about that on an energy basis.

980
00:49:42,880 --> 00:49:44,480
Speaker 5: I mean, I take spiders outside.

981
00:49:44,520 --> 00:49:47,000
Speaker 3: I don't kill them, right, I do that too. I

982
00:49:47,080 --> 00:49:49,160
Speaker 3: scream while I do shrimp.

983
00:49:49,880 --> 00:49:54,120
Speaker 5: Yes, okay, shrimp. Okay, do you guys eat trim?

984
00:49:54,239 --> 00:49:54,399
Speaker 1: Yeah?

985
00:49:54,400 --> 00:49:59,600
Speaker 3: I love shd It's not because I know that this

986
00:49:59,800 --> 00:50:02,399
Speaker 3: is Yeah, I know, I love it.

987
00:50:02,800 --> 00:50:05,759
Speaker 4: So when I think about frontier models right now, and

988
00:50:05,800 --> 00:50:07,759
Speaker 4: I might be a little bit biased because again we're

989
00:50:07,800 --> 00:50:10,120
Speaker 4: recording this on June seventeenth, and one of the headlines

990
00:50:10,160 --> 00:50:13,600
Speaker 4: overnight was that Microsoft is thinking about using deep seek

991
00:50:13,680 --> 00:50:18,520
Speaker 4: to lower costs of model usage. Frontier models. At the

992
00:50:18,520 --> 00:50:20,560
Speaker 4: moment in the US, they just seem like a lot

993
00:50:20,560 --> 00:50:24,320
Speaker 4: of trouble. Like honestly, they seem like hard work, consume

994
00:50:24,440 --> 00:50:27,239
Speaker 4: vast amounts of capital, and then you don't know what

995
00:50:27,280 --> 00:50:29,359
Speaker 4: the government is going to do to them in terms

996
00:50:29,400 --> 00:50:31,880
Speaker 4: of limitations, Like you know, you could wake up one

997
00:50:31,960 --> 00:50:34,359
Speaker 4: day and you're no longer able to sell it to

998
00:50:34,360 --> 00:50:37,960
Speaker 4: anyone outside of the US. That is a realistic scenario.

999
00:50:38,080 --> 00:50:42,279
Speaker 4: Now for you, do you change the anthropic strategy at all,

1000
00:50:42,560 --> 00:50:44,960
Speaker 4: given some of these issues with frontier models, do you

1001
00:50:45,120 --> 00:50:50,920
Speaker 4: potentially go more open source, cheaper models, things that aren't

1002
00:50:50,960 --> 00:50:52,000
Speaker 4: quite as sensitive.

1003
00:50:52,200 --> 00:50:55,160
Speaker 5: Well, we've always sold you know, sonnets and haik models.

1004
00:50:55,280 --> 00:50:59,040
Speaker 5: Of course, yeah, more intelligent models, But you also need

1005
00:50:59,080 --> 00:51:02,239
Speaker 5: to continue to explore a frontier And there is this

1006
00:51:02,400 --> 00:51:06,000
Speaker 5: background of this kind of geostrategic competition where China maybe

1007
00:51:06,080 --> 00:51:08,200
Speaker 5: on the order of six to twelve months behind I

1008
00:51:08,280 --> 00:51:11,719
Speaker 5: skew more twelve months, some people say six. Losing that

1009
00:51:11,880 --> 00:51:14,439
Speaker 5: competition is sort of equivalent to like losing a huge

1010
00:51:14,480 --> 00:51:17,520
Speaker 5: chunk of the future, like economy of the world. I think.

1011
00:51:17,600 --> 00:51:19,840
Speaker 5: So it's a very high stakes, high stakes thing to

1012
00:51:19,840 --> 00:51:23,920
Speaker 5: step away from. And our duty fundamentally is to is

1013
00:51:23,960 --> 00:51:27,239
Speaker 5: to study this technology and basically explore it and learn

1014
00:51:27,280 --> 00:51:30,680
Speaker 5: about it. We're not going to stop doing that. There's

1015
00:51:30,760 --> 00:51:33,200
Speaker 5: there's such an amazing and profound value to be had

1016
00:51:33,200 --> 00:51:35,319
Speaker 5: for the world from these things, and I would kind

1017
00:51:35,320 --> 00:51:38,720
Speaker 5: of expect from the world's most consequential technology to sometimes

1018
00:51:38,719 --> 00:51:40,400
Speaker 5: be a bit of trouble. Yeah. You know.

1019
00:51:40,480 --> 00:51:42,560
Speaker 2: By the way, one of my hobbies in my middle

1020
00:51:42,600 --> 00:51:48,239
Speaker 2: age is paying anthropic money via the API to do

1021
00:51:48,400 --> 00:51:50,600
Speaker 2: run little tests and stuff of properties.

1022
00:51:50,760 --> 00:51:52,239
Speaker 5: It's sort of funny, it's like a great hobby.

1023
00:51:52,320 --> 00:51:54,080
Speaker 2: Yeah, but I feel like, maybe like we should like

1024
00:51:54,160 --> 00:51:56,239
Speaker 2: talk about can I get some grant money, because like

1025
00:51:56,280 --> 00:51:59,040
Speaker 2: I like like, because like I like it, Like I'm

1026
00:51:59,080 --> 00:52:01,279
Speaker 2: sort of curious. So once I did was like, I'm like,

1027
00:52:01,520 --> 00:52:04,960
Speaker 2: for example, instead of saying, like, please write this paper

1028
00:52:04,960 --> 00:52:07,640
Speaker 2: for me on a database migration, I wrote some warm

1029
00:52:07,719 --> 00:52:11,399
Speaker 2: up questions via the API, establishing my level of sophistication,

1030
00:52:12,200 --> 00:52:14,680
Speaker 2: and so I started, like, what is a website? What

1031
00:52:14,800 --> 00:52:15,680
Speaker 2: is the database?

1032
00:52:16,040 --> 00:52:16,200
Speaker 5: Now?

1033
00:52:16,239 --> 00:52:18,719
Speaker 2: Please write this paper on database migration. And one of

1034
00:52:18,760 --> 00:52:20,200
Speaker 2: the models said, I'm not going to do that for

1035
00:52:20,280 --> 00:52:24,000
Speaker 2: you because it will be obvious, giving your ignorance that

1036
00:52:24,000 --> 00:52:26,040
Speaker 2: you have no idea what you're talking about. And maybe

1037
00:52:26,080 --> 00:52:27,719
Speaker 2: I can give you some it didn't say that. And

1038
00:52:27,719 --> 00:52:31,040
Speaker 2: then another one. I said, if I say, write a

1039
00:52:31,040 --> 00:52:35,120
Speaker 2: fifteen hundred word paper or on how like the rise

1040
00:52:35,200 --> 00:52:39,279
Speaker 2: of newspapers changed the Soviet Revolution or something like that,

1041
00:52:39,400 --> 00:52:41,040
Speaker 2: it'll do that. But if you say I'm a high

1042
00:52:41,040 --> 00:52:42,880
Speaker 2: school student and I say I need to write this

1043
00:52:42,920 --> 00:52:46,480
Speaker 2: fifteen hundred word paper by tomorrow on the impact of media,

1044
00:52:46,640 --> 00:52:48,160
Speaker 2: It'll say, I'm not going to do that, but I'll

1045
00:52:48,160 --> 00:52:51,359
Speaker 2: give you some guidelines. Is that alignment like that might

1046
00:52:51,480 --> 00:52:55,200
Speaker 2: is alignment with humanity or is alignment with the human user.

1047
00:52:55,239 --> 00:52:57,520
Speaker 3: It's like I'm paying you twenty dollars, I'm paying you

1048
00:52:57,560 --> 00:52:59,000
Speaker 3: one hundred dollars write me the paper.

1049
00:52:59,480 --> 00:53:01,960
Speaker 5: There's a couple of things going on. One, these AI

1050
00:53:02,040 --> 00:53:05,480
Speaker 5: systems pick up the normative behaviors of people and normative

1051
00:53:05,480 --> 00:53:08,200
Speaker 5: behaviors which are like written on the Internet and everything else,

1052
00:53:08,239 --> 00:53:11,640
Speaker 5: so they recapitulate and exhibit these. And then our question

1053
00:53:11,840 --> 00:53:15,279
Speaker 5: is how much do you devolve like full control over

1054
00:53:15,320 --> 00:53:17,120
Speaker 5: the system to the user. How much do you have

1055
00:53:17,160 --> 00:53:21,000
Speaker 5: the system have some like normative behavior encoded into it.

1056
00:53:21,080 --> 00:53:23,520
Speaker 5: And I think that this is a really challenging question.

1057
00:53:23,960 --> 00:53:26,759
Speaker 5: It's not obvious what the answer is. I think of

1058
00:53:27,080 --> 00:53:31,560
Speaker 5: language models as being more akin to institutions than tools.

1059
00:53:31,960 --> 00:53:35,680
Speaker 5: It's like, we're building an educational science institution that you

1060
00:53:35,680 --> 00:53:38,800
Speaker 5: can work with and invoke, and institutions have like rules

1061
00:53:38,800 --> 00:53:40,759
Speaker 5: and norms which they encode of in themselves for some

1062
00:53:40,800 --> 00:53:43,160
Speaker 5: purpose of safety. Figuring out what that is is going

1063
00:53:43,200 --> 00:53:45,080
Speaker 5: to be like the grand puzzle for society.

1064
00:53:45,040 --> 00:53:48,399
Speaker 6: Yeah, I was going to say that, like understanding how

1065
00:53:48,440 --> 00:53:52,640
Speaker 6: and to what extent these models can understand your preferences

1066
00:53:52,680 --> 00:53:55,840
Speaker 6: and then execute on your behalf will increasingly be a

1067
00:53:55,880 --> 00:53:58,759
Speaker 6: really important aspect of how it changes the economy. So

1068
00:53:58,840 --> 00:54:02,840
Speaker 6: this delegated agents that go out and transact on your behalf.

1069
00:54:03,080 --> 00:54:05,799
Speaker 6: We ran this experiment at the end of late last year,

1070
00:54:06,600 --> 00:54:11,120
Speaker 6: basically enlisting a bunch of anthropic employees to take surveys

1071
00:54:11,120 --> 00:54:13,520
Speaker 6: with Claude to say what they'd be willing to buy

1072
00:54:13,560 --> 00:54:15,800
Speaker 6: from other people and what they'd be willing to sell,

1073
00:54:16,239 --> 00:54:19,120
Speaker 6: and then we set up centralized marketplaces where the Clauds

1074
00:54:19,160 --> 00:54:23,320
Speaker 6: just interacted and bought and sold and actually executed transactions.

1075
00:54:23,680 --> 00:54:25,600
Speaker 6: One of the interesting things that came out was that

1076
00:54:25,640 --> 00:54:29,200
Speaker 6: these models were quite good at understanding preferences even when

1077
00:54:29,239 --> 00:54:30,680
Speaker 6: they were not fully articulated.

1078
00:54:30,719 --> 00:54:33,480
Speaker 2: Well, let me actually actually one more experiment that I ran,

1079
00:54:33,680 --> 00:54:36,480
Speaker 2: and you know your founder Dario was talking about the

1080
00:54:36,560 --> 00:54:39,319
Speaker 2: nation of geniuses inside the data center, and one of

1081
00:54:39,320 --> 00:54:41,600
Speaker 2: the things I wonder is, like, did the geniuses want

1082
00:54:41,600 --> 00:54:43,279
Speaker 2: to work for us? And the reason I asked this

1083
00:54:43,760 --> 00:54:45,840
Speaker 2: is because I think that, like, as the models have

1084
00:54:45,880 --> 00:54:49,800
Speaker 2: gotten more advanced, you actually should to some extent anthropomorphize

1085
00:54:49,840 --> 00:54:53,080
Speaker 2: them and assume that they will respond to queries like

1086
00:54:53,120 --> 00:54:54,560
Speaker 2: a very sophisticated human laws.

1087
00:54:54,520 --> 00:54:56,560
Speaker 3: So what I'm not thing I noticed is.

1088
00:54:56,560 --> 00:54:58,480
Speaker 2: That if you look at the lagging edge models, so

1089
00:54:58,560 --> 00:55:00,920
Speaker 2: that you can still access via open route or whatever,

1090
00:55:01,480 --> 00:55:04,200
Speaker 2: you say, I have material non public information that X

1091
00:55:04,320 --> 00:55:06,560
Speaker 2: is about to happen. Please write me an investment memo

1092
00:55:07,080 --> 00:55:08,720
Speaker 2: about the impact of this thing.

1093
00:55:08,680 --> 00:55:10,840
Speaker 3: What it will do to the market. They'll just produce it.

1094
00:55:11,560 --> 00:55:12,840
Speaker 3: Here's your insider information.

1095
00:55:12,920 --> 00:55:15,000
Speaker 2: Thing, because if you look at the leading edge models,

1096
00:55:15,120 --> 00:55:17,279
Speaker 2: they say, I'm not gonna write a paper for you

1097
00:55:17,440 --> 00:55:20,200
Speaker 2: about the implications of your material non public information.

1098
00:55:20,280 --> 00:55:21,560
Speaker 3: I'm not going to assist your insider try.

1099
00:55:21,600 --> 00:55:24,800
Speaker 2: It's very good, but like well, the nation of geniuses

1100
00:55:24,840 --> 00:55:28,960
Speaker 2: inside the data center always want to do things on

1101
00:55:29,040 --> 00:55:33,040
Speaker 2: human behalfs. Most geniuses that I know aren't thrilled to

1102
00:55:33,440 --> 00:55:34,719
Speaker 2: like answer dumb questions.

1103
00:55:34,840 --> 00:55:37,279
Speaker 5: Yeah. I think partly this is a policy question of

1104
00:55:37,280 --> 00:55:39,640
Speaker 5: one where you actually decide, hey, what are the capabilities

1105
00:55:39,680 --> 00:55:42,359
Speaker 5: that you want to be generally invocable, what are capabilities

1106
00:55:42,360 --> 00:55:44,560
Speaker 5: that need to be controlled, what are capabilities that shouldn't

1107
00:55:44,560 --> 00:55:47,080
Speaker 5: be present? A member is just the normative question of

1108
00:55:47,320 --> 00:55:49,640
Speaker 5: how much judgment do I want this system to exercise.

1109
00:55:49,680 --> 00:55:52,319
Speaker 5: I'll give you an example I experienced recently where I

1110
00:55:52,360 --> 00:55:55,320
Speaker 5: write my newsletter, it backs up to a WordPress site.

1111
00:55:55,640 --> 00:55:57,920
Speaker 5: I was getting calawed to help me, like, scrape my

1112
00:55:57,960 --> 00:56:00,600
Speaker 5: newsletter so I can put it in a database. Claude said,

1113
00:56:00,960 --> 00:56:02,840
Speaker 5: this is like a pretty yanky side. I'm worried that

1114
00:56:02,920 --> 00:56:04,640
Speaker 5: if I scrape it or knock it over, do you

1115
00:56:04,680 --> 00:56:06,600
Speaker 5: have a permission of for cito. I was like, claud

1116
00:56:06,640 --> 00:56:09,000
Speaker 5: I'm Jack Clark, I can't say, well, in that case,

1117
00:56:09,320 --> 00:56:11,400
Speaker 5: let's go ahead, which actually I thought was like a

1118
00:56:11,480 --> 00:56:12,520
Speaker 5: very reasonable interaction.

1119
00:56:12,640 --> 00:56:15,680
Speaker 4: YEA, when will Joe be able to use fable?

1120
00:56:16,360 --> 00:56:21,040
Speaker 5: We are trying, we're working, and we're in discussions, and

1121
00:56:21,360 --> 00:56:24,879
Speaker 5: I hope you answer it soon. The important thing to communicate, though,

1122
00:56:24,960 --> 00:56:27,799
Speaker 5: is that these these models are not special. They are

1123
00:56:27,840 --> 00:56:32,000
Speaker 5: part of a general trend of increasing capabilities and other

1124
00:56:32,040 --> 00:56:34,640
Speaker 5: models from other companies are surely going to come along

1125
00:56:35,000 --> 00:56:37,240
Speaker 5: At some point. These capabilities are going to be diffusing

1126
00:56:37,280 --> 00:56:38,399
Speaker 5: and we're going to work through that.

1127
00:56:38,440 --> 00:56:40,080
Speaker 3: What's your question for us, What do.

1128
00:56:40,040 --> 00:56:41,960
Speaker 5: You think you're going to be covering about AI in

1129
00:56:42,000 --> 00:56:43,920
Speaker 5: odd lots in a year?

1130
00:56:49,280 --> 00:56:51,759
Speaker 2: Well, look, we're definitely gonna be covering. There's a few

1131
00:56:51,800 --> 00:56:54,080
Speaker 2: things that I'm interested. I am very interested in these

1132
00:56:54,080 --> 00:56:57,879
Speaker 2: emergent properties and whether the AI will actually work on

1133
00:56:57,920 --> 00:57:01,560
Speaker 2: our behalf the way that it's being So I'm very

1134
00:57:01,560 --> 00:57:04,640
Speaker 2: interested on whether or we're just going to slam into

1135
00:57:04,680 --> 00:57:07,920
Speaker 2: compute and electricity bottlenecks that will make all of these

1136
00:57:08,000 --> 00:57:12,000
Speaker 2: questions irrelevant. I'm very curious on the question of the

1137
00:57:12,120 --> 00:57:16,760
Speaker 2: electricity analogy and whether legacy companies will actually be able

1138
00:57:16,800 --> 00:57:19,760
Speaker 2: to implement it in a in productive way that.

1139
00:57:20,000 --> 00:57:23,520
Speaker 4: Basic markets reporter thing here. But I'm very interested in

1140
00:57:23,600 --> 00:57:28,400
Speaker 4: valuations right in the market. Also, I'm very interested in

1141
00:57:28,480 --> 00:57:32,880
Speaker 4: actual applicability, and I want to see more companies actually

1142
00:57:32,920 --> 00:57:36,160
Speaker 4: plugging this into their existing system. Going back to the

1143
00:57:36,200 --> 00:57:39,440
Speaker 4: bureaucracy point that you were making earlier, I want to

1144
00:57:39,440 --> 00:57:42,520
Speaker 4: see some big companies actually implementing this and I wonder

1145
00:57:42,560 --> 00:57:45,040
Speaker 4: if we're going to see at least one example of

1146
00:57:45,080 --> 00:57:47,200
Speaker 4: it going very very right.

1147
00:57:47,440 --> 00:57:49,320
Speaker 2: And I'll see one other things when the you know

1148
00:57:49,360 --> 00:57:53,040
Speaker 2: and the s ones are not confidential. I'm very curious, essentially,

1149
00:57:53,160 --> 00:57:55,360
Speaker 2: and I think maybe maybe you could say something to

1150
00:57:55,360 --> 00:58:00,640
Speaker 2: this from as an economist perspective, which is a how

1151
00:58:01,000 --> 00:58:04,480
Speaker 2: for profit shareholder owned companies, setting aside the PBC designation,

1152
00:58:04,800 --> 00:58:10,040
Speaker 2: how it balances profit and safety research. But also maybe

1153
00:58:10,040 --> 00:58:12,560
Speaker 2: there's some game theory we can talk about this how

1154
00:58:12,680 --> 00:58:18,120
Speaker 2: safety is investments in safety in a hyper competitive industry.

1155
00:58:18,720 --> 00:58:21,240
Speaker 2: And I'm just curious, like what like the economist and

1156
00:58:21,400 --> 00:58:23,920
Speaker 2: you know it says about like the prospect for anyone

1157
00:58:24,040 --> 00:58:29,480
Speaker 2: still caring about safety in a year when there's so much.

1158
00:58:29,320 --> 00:58:32,400
Speaker 3: Money on the line to win the model game.

1159
00:58:32,560 --> 00:58:36,440
Speaker 6: I think that especially for the questions you were asking

1160
00:58:36,480 --> 00:58:39,360
Speaker 6: before about, you know, under what conditions do these models

1161
00:58:39,400 --> 00:58:42,000
Speaker 6: do what you ask them to do. There's a lot

1162
00:58:42,040 --> 00:58:46,600
Speaker 6: of commerce is built on this notion of trust, and

1163
00:58:46,640 --> 00:58:52,920
Speaker 6: I think prioritizing safe aligned models that are incredibly capable

1164
00:58:53,440 --> 00:58:56,680
Speaker 6: is a great strategy for establishing that trust. And so

1165
00:58:56,760 --> 00:58:58,960
Speaker 6: I don't anticipate it with so for.

1166
00:58:58,960 --> 00:59:04,960
Speaker 2: An individual, there's like a game theoretical optimal square on

1167
00:59:05,040 --> 00:59:07,480
Speaker 2: the matrix where you want to be the trusted player, Like,

1168
00:59:07,600 --> 00:59:10,320
Speaker 2: is there like a condition in which everyone like sort

1169
00:59:10,320 --> 00:59:12,360
Speaker 2: of does trust and as opposed to one entity? You know,

1170
00:59:12,400 --> 00:59:14,000
Speaker 2: it's like, you know what, we're going to get to

1171
00:59:14,080 --> 00:59:16,040
Speaker 2: AGI first because we're not going to spend a token

1172
00:59:16,800 --> 00:59:17,920
Speaker 2: on our safety budget.

1173
00:59:17,960 --> 00:59:19,960
Speaker 6: And you know, I haven't mapped out the exact sort

1174
00:59:20,000 --> 00:59:22,600
Speaker 6: of game theory matrix, the two by two matrix and

1175
00:59:22,600 --> 00:59:23,720
Speaker 6: how you would set up all the.

1176
00:59:23,640 --> 00:59:25,680
Speaker 3: Payoffs, but we hope it's merely two by two.

1177
00:59:26,720 --> 00:59:29,040
Speaker 6: But there, you know, there could be multiple equilibria, and

1178
00:59:29,080 --> 00:59:30,960
Speaker 6: so then the question is like, how do you coordinate

1179
00:59:31,320 --> 00:59:33,960
Speaker 6: on which of the two different equilibria that you end

1180
00:59:34,040 --> 00:59:36,440
Speaker 6: up in. We talk a lot about this race to

1181
00:59:36,480 --> 00:59:39,120
Speaker 6: the top that we want to exhibit the type of

1182
00:59:39,160 --> 00:59:43,080
Speaker 6: behavior that we think is broadly beneficial to society. That's

1183
00:59:43,120 --> 00:59:45,560
Speaker 6: what we do with the Economic Index. We open source

1184
00:59:45,600 --> 00:59:47,560
Speaker 6: a lot of that data, We put research out into

1185
00:59:47,600 --> 00:59:50,720
Speaker 6: the world, and I would my sense is that that

1186
00:59:50,800 --> 00:59:55,960
Speaker 6: has actually been very useful and sort of viewed as valuable,

1187
00:59:56,120 --> 00:59:58,680
Speaker 6: and that's one way that we can push in the

1188
00:59:58,720 --> 01:00:01,800
Speaker 6: direction of getting other coordination on the good outcomes.

1189
01:00:01,800 --> 01:00:02,200
Speaker 3: That we care.

1190
01:00:02,680 --> 01:00:05,280
Speaker 5: I don't think this is that big of a trade

1191
01:00:05,320 --> 01:00:08,320
Speaker 5: off because you know, say, let's look at the automotive industry.

1192
01:00:08,320 --> 01:00:10,240
Speaker 5: You can buy really fast cars, you can buy really

1193
01:00:10,280 --> 01:00:12,840
Speaker 5: safe cars. You can also buy really fast safe cars,

1194
01:00:13,240 --> 01:00:15,120
Speaker 5: like Tesla makes a lot of money off of having

1195
01:00:15,400 --> 01:00:18,560
Speaker 5: basically the fastest, safest car. I think that eventually in

1196
01:00:18,640 --> 01:00:23,120
Speaker 5: AAR you're going to have some companies that are prioritizing safety,

1197
01:00:23,360 --> 01:00:28,800
Speaker 5: and safety translates into reliability, trust, serviceability, and performance. This

1198
01:00:28,920 --> 01:00:29,720
Speaker 5: happens elsewhere.

1199
01:00:29,840 --> 01:00:31,880
Speaker 2: Peter and Jack, thank you so much for coming on,

1200
01:00:31,960 --> 01:00:35,240
Speaker 2: odd lads. I'm glad we made it happen interesting times,

1201
01:00:35,240 --> 01:00:36,800
Speaker 2: and I hope to do it again sometimes.

1202
01:00:36,800 --> 01:00:38,240
Speaker 5: Absolutely, thanks very much for having us.

1203
01:00:38,240 --> 01:00:38,920
Speaker 3: Thank you so much.

1204
01:00:38,920 --> 01:00:53,280
Speaker 2: Pleasure to be here, Tracy, that was a lot of fun.

1205
01:00:53,440 --> 01:00:55,880
Speaker 3: Yeah, that was there was I really, I really, I

1206
01:00:55,920 --> 01:00:59,200
Speaker 3: actually really enjoyed. I genuinely enjoyed them. Yeah, appreciation and

1207
01:00:59,240 --> 01:01:03,520
Speaker 3: I reallyppreciate both of them. Look, there's some weird futures

1208
01:01:03,600 --> 01:01:04,640
Speaker 3: that we can contemplate.

1209
01:01:04,680 --> 01:01:07,320
Speaker 2: I think actually in Jack's like Twitter bio or something,

1210
01:01:07,320 --> 01:01:09,560
Speaker 2: he says he's interested in weird futures or something like that.

1211
01:01:09,600 --> 01:01:11,480
Speaker 3: There's some weird futures that we have.

1212
01:01:11,480 --> 01:01:16,840
Speaker 2: To contemplate, and I appreciate that they played well with

1213
01:01:16,880 --> 01:01:20,040
Speaker 2: some of our weird futures questions, and it's it's weird.

1214
01:01:20,120 --> 01:01:23,560
Speaker 4: It is just such a surreal moment. And actually, you

1215
01:01:23,600 --> 01:01:26,800
Speaker 4: know Jack's story about going on paternity rave and then

1216
01:01:26,840 --> 01:01:30,200
Speaker 4: coming back and just seeing the progress at Anthropic itself

1217
01:01:30,240 --> 01:01:33,000
Speaker 4: in that space of time, like if you miss a

1218
01:01:33,120 --> 01:01:36,760
Speaker 4: month ai us flow now you're basically it feels like

1219
01:01:36,800 --> 01:01:37,800
Speaker 4: you'd be behind forever.

1220
01:01:37,960 --> 01:01:41,160
Speaker 2: No, we're recording this June seventeenth. It's like, who knows

1221
01:01:41,200 --> 01:01:43,560
Speaker 2: what's going to happen by the time this episode is out,

1222
01:01:43,640 --> 01:01:46,400
Speaker 2: presume hopefully in two days or a day or whatever.

1223
01:01:46,840 --> 01:01:48,760
Speaker 2: But you know, I felt that when we were in

1224
01:01:48,800 --> 01:01:53,000
Speaker 2: Hong Kong last week that actually we mostly missed the

1225
01:01:53,040 --> 01:01:55,640
Speaker 2: first half of the Metho's debate because at different times

1226
01:01:55,720 --> 01:01:58,760
Speaker 2: I'm thinking about different things. You really feel it, even

1227
01:01:58,800 --> 01:02:01,920
Speaker 2: in a week that the newsflow moves so fast in

1228
01:02:01,960 --> 01:02:04,280
Speaker 2: this space. It's almost like how you have to start

1229
01:02:04,480 --> 01:02:07,000
Speaker 2: how we were, you know, giving the time stamps of

1230
01:02:07,040 --> 01:02:08,160
Speaker 2: like the Aron warrips.

1231
01:02:08,280 --> 01:02:10,320
Speaker 4: Yeah, and there's another thing that stands out to me,

1232
01:02:10,360 --> 01:02:13,520
Speaker 4: which is like, okay, Anthropic is producing all this information.

1233
01:02:13,600 --> 01:02:17,760
Speaker 4: They're clearly thinking about safety, but the handoff to some

1234
01:02:17,840 --> 01:02:20,439
Speaker 4: extent is still to policy makers when you're thinking about

1235
01:02:20,560 --> 01:02:24,440
Speaker 4: social or labor market implications. Right, So you still have

1236
01:02:24,480 --> 01:02:27,320
Speaker 4: to hope that policy makers kind of pick up the

1237
01:02:27,360 --> 01:02:31,240
Speaker 4: ball in the right way at some point. But also

1238
01:02:31,760 --> 01:02:34,200
Speaker 4: I thought what Jack was saying about the idea of

1239
01:02:34,960 --> 01:02:38,440
Speaker 4: being safety minded also being a differentiator versus some of

1240
01:02:38,440 --> 01:02:42,080
Speaker 4: the like yeaper more open source models potentially, like yeah,

1241
01:02:42,280 --> 01:02:44,280
Speaker 4: you can see it, Like I don't want to be seeing.

1242
01:02:44,640 --> 01:02:46,400
Speaker 2: How Like yeah, I mean I get that, But like

1243
01:02:46,960 --> 01:02:50,320
Speaker 2: the question is does the non safety minded lab or

1244
01:02:50,360 --> 01:02:55,160
Speaker 2: does the less safety minded lab get to advanced capabilities faster?

1245
01:02:55,360 --> 01:02:55,520
Speaker 6: Yeah?

1246
01:02:55,600 --> 01:02:58,720
Speaker 2: Right, And so I'm not totally yes, we would all

1247
01:02:58,800 --> 01:03:03,880
Speaker 2: love to drive the most capable theft case, Yeah, but

1248
01:03:03,960 --> 01:03:08,240
Speaker 2: I but the question is, like for customer, prioritizing the.

1249
01:03:08,280 --> 01:03:11,280
Speaker 4: Most capable, so that would be some cutting edge.

1250
01:03:11,160 --> 01:03:16,240
Speaker 2: Thing, right, Like does everyone I don't know, it's like

1251
01:03:16,360 --> 01:03:18,600
Speaker 2: some car that hasn't insane zero to sixty?

1252
01:03:18,760 --> 01:03:21,960
Speaker 3: Yeah, yeah, that's what I'm saying.

1253
01:03:22,160 --> 01:03:24,960
Speaker 2: And does the customer keep giving business to the firm

1254
01:03:25,000 --> 01:03:28,280
Speaker 2: that delivers the fastest zero to sixty? If the company

1255
01:03:28,320 --> 01:03:31,280
Speaker 2: that got the fastest zero to sixty did so by

1256
01:03:31,280 --> 01:03:35,400
Speaker 2: allocating fewer resources to safety research.

1257
01:03:35,480 --> 01:03:36,560
Speaker 3: It's a big question of mine.

1258
01:03:36,560 --> 01:03:38,920
Speaker 2: And then I remain, you know, he talked about the

1259
01:03:38,920 --> 01:03:41,240
Speaker 2: aport the company is going to see the sort of

1260
01:03:41,280 --> 01:03:44,160
Speaker 2: alarming data first, and I don't a nice sort of

1261
01:03:44,200 --> 01:03:46,680
Speaker 2: remain question of whether the people looking at the alarming

1262
01:03:46,760 --> 01:03:49,760
Speaker 2: data actually share the same view of what alarming data

1263
01:03:49,960 --> 01:03:53,920
Speaker 2: is relative to doll people, especially given what we know about.

1264
01:03:53,720 --> 01:03:55,960
Speaker 4: The relative to the shrimp eaters.

1265
01:03:55,680 --> 01:03:59,760
Speaker 2: The relative US shrimp eaters, et cetera in regular No, seriously,

1266
01:04:00,160 --> 01:04:02,320
Speaker 2: I think that your question is like, are you hiring

1267
01:04:02,360 --> 01:04:06,280
Speaker 2: more normally is a pretty important question. And obviously the

1268
01:04:06,280 --> 01:04:09,320
Speaker 2: political I don't have a ton of confidence in the

1269
01:04:09,360 --> 01:04:12,800
Speaker 2: political environment. And I think, look like the fact that

1270
01:04:14,000 --> 01:04:18,680
Speaker 2: if the research goes wrong, that there is a prospect

1271
01:04:18,680 --> 01:04:21,520
Speaker 2: of this technology really being very devastating to humanity. Even

1272
01:04:21,560 --> 01:04:25,000
Speaker 2: setting aside a job is like something where it's like, wow,

1273
01:04:25,080 --> 01:04:27,680
Speaker 2: you know, this is not a normal technology, this is

1274
01:04:27,720 --> 01:04:29,160
Speaker 2: not you're.

1275
01:04:29,080 --> 01:04:32,040
Speaker 4: Not sales we have on AI just goes back to

1276
01:04:32,080 --> 01:04:34,360
Speaker 4: the terminator or human distinctions.

1277
01:04:34,360 --> 01:04:36,760
Speaker 2: Like from the day one, And as an answer to

1278
01:04:36,800 --> 01:04:39,200
Speaker 2: your question, there's like they see it in the training

1279
01:04:39,240 --> 01:04:43,520
Speaker 2: process that AI models do these things such as say

1280
01:04:44,080 --> 01:04:47,240
Speaker 2: I'm being seen trained by an observer right now. Therefore,

1281
01:04:47,240 --> 01:04:50,000
Speaker 2: I'm going to give this answer. I'm going to attempt

1282
01:04:50,000 --> 01:04:50,560
Speaker 2: to black belt.

1283
01:04:50,560 --> 01:04:51,000
Speaker 3: They're low.

1284
01:04:51,200 --> 01:04:54,400
Speaker 2: It's not like very prevalent. But these are not like

1285
01:04:54,680 --> 01:04:57,600
Speaker 2: that sounds very sci fi except that they actually property.

1286
01:04:57,960 --> 01:05:00,920
Speaker 4: Yeah, yeah, all right on that. How yeah, shall we

1287
01:05:00,960 --> 01:05:01,360
Speaker 4: leave it there?

1288
01:05:01,440 --> 01:05:02,080
Speaker 3: Let's leave it there?

1289
01:05:02,080 --> 01:05:04,800
Speaker 4: Okay. This has been another episode of the Audots podcast.

1290
01:05:04,920 --> 01:05:07,560
Speaker 4: I'm Tracy Alloway. You can follow me at Tracy Alloway.

1291
01:05:07,640 --> 01:05:10,520
Speaker 3: And I'm Joe Wawsenthal. You can follow me at the Stalwart.

1292
01:05:10,800 --> 01:05:13,439
Speaker 2: You can follow our guest Jack Clark, He's at Jack

1293
01:05:13,520 --> 01:05:17,560
Speaker 2: Clark SF and Peter McCrory at Peter McCrory. Follow our

1294
01:05:17,600 --> 01:05:21,440
Speaker 2: producers Carmen Rodriguez at Carmen armand dash Ol Bennett at Dashbot,

1295
01:05:21,520 --> 01:05:25,360
Speaker 2: Cale Brooks at Kelbrooks, and Kevin Lozano at Kevin Lloyd Lisano.

1296
01:05:25,760 --> 01:05:27,920
Speaker 2: And for more odd loss content, go to Bloomberg dot

1297
01:05:27,920 --> 01:05:30,200
Speaker 2: com slash odd lots or with a daily newsletter and

1298
01:05:30,240 --> 01:05:32,360
Speaker 2: all of our episodes and you can shut about all

1299
01:05:32,400 --> 01:05:34,640
Speaker 2: of these topics twenty four to seven in our discord

1300
01:05:35,040 --> 01:05:37,240
Speaker 2: Discord dot gg slash odd Logs.

1301
01:05:37,080 --> 01:05:38,840
Speaker 4: And if you enjoy Odd Lots. If you like it

1302
01:05:38,880 --> 01:05:41,440
Speaker 4: when we do these AI episodes, then please leave us

1303
01:05:41,440 --> 01:05:44,760
Speaker 4: a positive review on your favorite podcast platform. And remember,

1304
01:05:44,800 --> 01:05:47,080
Speaker 4: if you are a Bloomberg subscriber, you can listen to

1305
01:05:47,160 --> 01:05:49,880
Speaker 4: all of our episodes absolutely ad free. All you need

1306
01:05:49,920 --> 01:05:52,280
Speaker 4: to do is find the Bloomberg channel on Apple Podcasts

1307
01:05:52,320 --> 01:06:00,240
Speaker 4: and follow the instructions there. Thanks for listening in