1
00:00:02,720 --> 00:00:19,000
Speaker 1: Bloomberg Audio Studios, Podcasts, Radio News. Hello and welcome to

2
00:00:19,040 --> 00:00:21,080
Speaker 1: another episode of The Odd Laws podcast.

3
00:00:21,200 --> 00:00:23,520
Speaker 2: I'm Joe Wisenthal and I'm Tracy Alloway.

4
00:00:23,800 --> 00:00:26,319
Speaker 1: Tracy, have you played around with GPT five much?

5
00:00:26,920 --> 00:00:30,640
Speaker 2: Not really, I've been perplexity pills. Oh that's what your

6
00:00:30,720 --> 00:00:33,360
Speaker 2: main Yeah, that's my main one at the moment. But

7
00:00:33,479 --> 00:00:34,840
Speaker 2: is it good? I hear mixed.

8
00:00:35,080 --> 00:00:37,239
Speaker 1: I use it because I use GPT every day. It

9
00:00:37,280 --> 00:00:41,400
Speaker 1: does not strike me as like obviously better yeah for

10
00:00:41,600 --> 00:00:44,640
Speaker 1: my uses than like the three models, which I've been

11
00:00:44,720 --> 00:00:46,760
Speaker 1: very impressed by because you know, I want to establish them.

12
00:00:46,760 --> 00:00:47,960
Speaker 3: No hater or anything like that.

13
00:00:48,240 --> 00:00:50,040
Speaker 1: But like, it did not strike me as like, oh,

14
00:00:50,080 --> 00:00:50,720
Speaker 1: this is like an.

15
00:00:50,720 --> 00:00:52,839
Speaker 2: Amazing Yeah, this is the thing.

16
00:00:53,040 --> 00:00:54,280
Speaker 3: Step function or whatever.

17
00:00:54,280 --> 00:00:59,080
Speaker 2: It feels like the sort of breakthroughs awe inspiring breakthroughs

18
00:00:59,120 --> 00:01:00,680
Speaker 2: are kind of behind us, and a lot of the

19
00:01:00,720 --> 00:01:04,160
Speaker 2: progress on the models feels very incremental at this point,

20
00:01:04,160 --> 00:01:05,959
Speaker 2: even though people are spending a lot of time and

21
00:01:06,040 --> 00:01:07,240
Speaker 2: resources on doing it.

22
00:01:07,520 --> 00:01:09,880
Speaker 1: The one thing GPG five does prompt me and say, oh,

23
00:01:09,880 --> 00:01:11,560
Speaker 1: that's a great question. Would you like to follow up

24
00:01:11,560 --> 00:01:11,959
Speaker 1: more on that?

25
00:01:12,160 --> 00:01:13,160
Speaker 3: But it's like does it.

26
00:01:13,080 --> 00:01:16,320
Speaker 2: Say, o, Joe, you're so smart? That's such a smart question.

27
00:01:16,440 --> 00:01:17,559
Speaker 3: Say you know what it did? Say?

28
00:01:17,680 --> 00:01:19,560
Speaker 1: I asked to follow up, and it started an answer

29
00:01:19,560 --> 00:01:22,280
Speaker 1: with love it and then love it? Do you want

30
00:01:22,280 --> 00:01:23,200
Speaker 1: me to look into that?

31
00:01:23,720 --> 00:01:24,039
Speaker 4: Yes?

32
00:01:24,200 --> 00:01:26,600
Speaker 2: They are very flattering, aren't they. Actually, that's one thing

33
00:01:26,640 --> 00:01:29,399
Speaker 2: I like about perplexity is it doesn't really flatter you.

34
00:01:29,400 --> 00:01:30,600
Speaker 2: It just spits out an answer.

35
00:01:30,840 --> 00:01:33,800
Speaker 1: So anyway, there's so many questions I have about AI,

36
00:01:33,880 --> 00:01:36,039
Speaker 1: and we talk about the business old fair amount and

37
00:01:36,200 --> 00:01:38,160
Speaker 1: video and all that stuff. We actually don't really talk

38
00:01:38,200 --> 00:01:41,160
Speaker 1: that much about the pure research side as much. But

39
00:01:41,240 --> 00:01:43,160
Speaker 1: it's pretty important, I think, because I think a lot

40
00:01:43,200 --> 00:01:45,360
Speaker 1: of people would agree that if the skills are like

41
00:01:45,400 --> 00:01:47,840
Speaker 1: slowing down, or if there were a wall or something

42
00:01:47,880 --> 00:01:51,000
Speaker 1: like that, that might change some of these business model calculations,

43
00:01:51,040 --> 00:01:53,080
Speaker 1: et cetera. So I think it's good we need to

44
00:01:53,080 --> 00:01:55,280
Speaker 1: get an update on just sort of the state of

45
00:01:55,280 --> 00:01:56,800
Speaker 1: the art the science of AI.

46
00:01:57,080 --> 00:01:57,320
Speaker 4: Yeah.

47
00:01:57,400 --> 00:02:00,600
Speaker 2: Also, it would be nice just to understand what's possible

48
00:02:00,760 --> 00:02:03,080
Speaker 2: in terms of the AI models and what people are

49
00:02:03,080 --> 00:02:06,360
Speaker 2: actually researching, what they're working towards, work like, is it

50
00:02:06,480 --> 00:02:09,480
Speaker 2: mostly about price? Is it mostly about the output? Is

51
00:02:09,520 --> 00:02:12,000
Speaker 2: it mostly about energy use? All those things?

52
00:02:12,080 --> 00:02:13,760
Speaker 1: All those things, Well, I'm really excited to say we

53
00:02:13,800 --> 00:02:16,760
Speaker 1: have the perfect guest, someone who is an AI researcher.

54
00:02:17,000 --> 00:02:19,360
Speaker 1: We're gonna be speaking with Jack Morris. He's currently about

55
00:02:19,400 --> 00:02:21,280
Speaker 1: to finish his PhD.

56
00:02:20,880 --> 00:02:22,400
Speaker 3: At Cornell in AI.

57
00:02:22,440 --> 00:02:27,160
Speaker 1: He's been affiliated with Meta professionally, so presumably he already

58
00:02:27,200 --> 00:02:30,080
Speaker 1: has a one hundred million dollar pay package in the bank.

59
00:02:30,320 --> 00:02:32,840
Speaker 1: Now he's shaking his head, he's not that's a joke.

60
00:02:33,040 --> 00:02:36,079
Speaker 1: But Jack, thank you so much for coming on odd lots.

61
00:02:36,200 --> 00:02:38,040
Speaker 4: Yeah, thanks for having me. This is gonna be fun.

62
00:02:38,120 --> 00:02:39,560
Speaker 1: What do you explain to me, like what you're up to,

63
00:02:39,600 --> 00:02:41,239
Speaker 1: because I don't really understand how.

64
00:02:41,080 --> 00:02:42,040
Speaker 3: It works where people are.

65
00:02:42,040 --> 00:02:44,560
Speaker 1: They're at a university and they're also at a company,

66
00:02:44,639 --> 00:02:47,560
Speaker 1: and this isn't how it works. And much of the world, right,

67
00:02:47,600 --> 00:02:49,880
Speaker 1: people get their degree and then they get a job.

68
00:02:50,080 --> 00:02:52,200
Speaker 1: I get the impression that in the AI world it's

69
00:02:52,240 --> 00:02:56,440
Speaker 1: a little fuzzier in terms of one's affiliations between industry

70
00:02:56,600 --> 00:02:58,040
Speaker 1: and education and stuff like that.

71
00:02:58,280 --> 00:03:01,320
Speaker 4: Yeah, that's definitely true. I think might be on the

72
00:03:01,360 --> 00:03:03,200
Speaker 4: way out, but I can tell you about my situation.

73
00:03:03,400 --> 00:03:06,960
Speaker 4: So there's kind of a public research world and like

74
00:03:06,960 --> 00:03:11,240
Speaker 4: a private research world, and all the academic institutions do

75
00:03:11,280 --> 00:03:17,400
Speaker 4: public research, and the AI labs like Open Ai, Anthropic, Google,

76
00:03:17,440 --> 00:03:19,840
Speaker 4: deep Mind, they essentially do private research where they have

77
00:03:19,960 --> 00:03:23,079
Speaker 4: these people in house that are running experiments and learning

78
00:03:23,080 --> 00:03:25,960
Speaker 4: more about their systems, but they don't publish anything or

79
00:03:25,960 --> 00:03:28,280
Speaker 4: share any of their knowledge. And so a cool thing

80
00:03:28,280 --> 00:03:30,920
Speaker 4: about getting your PhD right now is you can do

81
00:03:31,080 --> 00:03:34,079
Speaker 4: research right about it and then publicize it like put

82
00:03:34,080 --> 00:03:36,400
Speaker 4: it online, I tweet about it. I kind of like

83
00:03:36,520 --> 00:03:38,920
Speaker 4: can talk to you about it. And there's a few

84
00:03:38,960 --> 00:03:42,000
Speaker 4: places left that will still kind of moment, we're never.

85
00:03:41,920 --> 00:03:42,720
Speaker 3: Going to hear from you again.

86
00:03:44,600 --> 00:03:46,640
Speaker 4: Yeah, I'll make sure they have a clause in my

87
00:03:46,720 --> 00:03:50,040
Speaker 4: contract that I can still talk to Joe and Tracy.

88
00:03:49,400 --> 00:03:52,320
Speaker 2: The all thoughts clause. Yes, that would be important. So

89
00:03:52,680 --> 00:03:56,440
Speaker 2: when we say AI research or an AI researcher, what

90
00:03:56,560 --> 00:04:01,480
Speaker 2: exactly does that entail? Can't the AI models just research themselves?

91
00:04:01,600 --> 00:04:02,440
Speaker 2: Just let them do it?

92
00:04:02,720 --> 00:04:06,000
Speaker 4: Yeah, that's actually a very smart idea, and like people

93
00:04:06,120 --> 00:04:08,720
Speaker 4: are really worried about that. Actually, Like if we get

94
00:04:08,720 --> 00:04:13,320
Speaker 4: to the point where the AI can improve itself into researching, yeah,

95
00:04:13,360 --> 00:04:15,120
Speaker 4: then it sort of gets smarter and then it improves

96
00:04:15,120 --> 00:04:17,120
Speaker 4: themself again and it ends up being this kind of

97
00:04:17,400 --> 00:04:21,160
Speaker 4: exponential improvement that ends up with all of our demise.

98
00:04:21,839 --> 00:04:24,840
Speaker 4: But I think right now it's not quite there yet.

99
00:04:25,080 --> 00:04:28,160
Speaker 4: Like maybe you can talk to CHGBT what good Yeah,

100
00:04:28,160 --> 00:04:29,960
Speaker 4: And good news for me too, because it means I

101
00:04:30,000 --> 00:04:33,599
Speaker 4: can still get a degree and be gainfully employed. But

102
00:04:34,200 --> 00:04:36,680
Speaker 4: I think it's it's still helpful, but we still need

103
00:04:36,720 --> 00:04:39,800
Speaker 4: like humans to make these improvements. And in terms of

104
00:04:39,839 --> 00:04:41,479
Speaker 4: what the actual day to day work looks like, I

105
00:04:41,480 --> 00:04:43,800
Speaker 4: think it really varies. Like there's some people working on

106
00:04:44,480 --> 00:04:47,279
Speaker 4: trying to make the models run faster, or trying to

107
00:04:47,279 --> 00:04:50,440
Speaker 4: make the hardware that runs the models run faster more efficiently.

108
00:04:50,880 --> 00:04:52,800
Speaker 4: There's people that try to work on the data, like

109
00:04:52,839 --> 00:04:55,640
Speaker 4: what should we train on more coding problems or more

110
00:04:55,880 --> 00:04:59,600
Speaker 4: textbooks or more Reddit posts, what works best to make

111
00:04:59,640 --> 00:05:01,920
Speaker 4: the model? And then there's a lot more people working

112
00:05:01,960 --> 00:05:04,680
Speaker 4: on different areas of the stack, like training algorithms. I

113
00:05:04,760 --> 00:05:08,240
Speaker 4: kind of have my own little niche and niche. There's

114
00:05:08,279 --> 00:05:11,760
Speaker 4: this old field of information theory from like the twentieth

115
00:05:11,760 --> 00:05:14,360
Speaker 4: century where they talk about bits like a zero or

116
00:05:14,400 --> 00:05:16,600
Speaker 4: a one is a bit and you can add them

117
00:05:16,680 --> 00:05:20,040
Speaker 4: up and have kilobytes and megabytes. And so I've been

118
00:05:20,080 --> 00:05:21,919
Speaker 4: trying to think about what that means in like the

119
00:05:22,000 --> 00:05:24,200
Speaker 4: chat GBT world, if you train a model on a

120
00:05:24,200 --> 00:05:26,880
Speaker 4: certain number of bits, how many bits does it actually learn?

121
00:05:27,200 --> 00:05:29,040
Speaker 4: And like can you look at the model and figure

122
00:05:29,040 --> 00:05:30,599
Speaker 4: out like if you have one slice of the model,

123
00:05:30,600 --> 00:05:32,600
Speaker 4: how many bits that is and stuff like that. So

124
00:05:32,640 --> 00:05:34,920
Speaker 4: maybe the easiest way to explain is if you had,

125
00:05:34,960 --> 00:05:37,640
Speaker 4: for some god forsaken reason to use chat GBT as

126
00:05:37,680 --> 00:05:40,560
Speaker 4: like a flash drive, like you had a certain set

127
00:05:40,600 --> 00:05:43,080
Speaker 4: of data and it had to memorize all that data,

128
00:05:43,200 --> 00:05:46,840
Speaker 4: Like how much data could it actually store? That's the

129
00:05:46,960 --> 00:05:48,800
Speaker 4: kind of area I've been working in. And then you know,

130
00:05:48,800 --> 00:05:50,880
Speaker 4: once you're there, you kind of realize we could do this,

131
00:05:51,040 --> 00:05:52,800
Speaker 4: or maybe next semester, if we have time, we could

132
00:05:53,120 --> 00:05:54,839
Speaker 4: try this other thing. And so there's it kind of

133
00:05:54,839 --> 00:05:56,880
Speaker 4: branches out and there's a lot of little problems that

134
00:05:56,920 --> 00:05:57,520
Speaker 4: you can try.

135
00:05:57,920 --> 00:06:01,800
Speaker 1: I mentioned GPT five fine to me, It does not

136
00:06:02,040 --> 00:06:05,479
Speaker 1: strike me as like you know, because actually so the

137
00:06:05,520 --> 00:06:08,360
Speaker 1: first time I use cha GPT is genuinely blown away

138
00:06:08,400 --> 00:06:10,559
Speaker 1: like most people. And then actually I was pretty blown

139
00:06:10,600 --> 00:06:13,600
Speaker 1: away by the three models, in part because of how

140
00:06:13,640 --> 00:06:16,640
Speaker 1: well they could do document search and superior to Google

141
00:06:16,680 --> 00:06:19,680
Speaker 1: Search in many respects and also just the organization of

142
00:06:19,720 --> 00:06:22,279
Speaker 1: a lot of unstructured data, et cetera. Like I didn't

143
00:06:22,279 --> 00:06:25,720
Speaker 1: have like some oh my god wow moment with GPT five.

144
00:06:25,760 --> 00:06:28,960
Speaker 1: It's like, this seems like, how do we measure whether

145
00:06:29,080 --> 00:06:31,359
Speaker 1: AI is getting better all the time.

146
00:06:32,680 --> 00:06:35,360
Speaker 4: Yeah, that's that's a huge question, right.

147
00:06:35,800 --> 00:06:37,200
Speaker 1: Well, let me ask you, Okay, let me ask you

148
00:06:37,320 --> 00:06:41,480
Speaker 1: actually a more specific question. How do the entities that

149
00:06:41,800 --> 00:06:46,320
Speaker 1: test AI models as their job or as their function?

150
00:06:46,800 --> 00:06:50,320
Speaker 1: What does the formal testing process look like to rank

151
00:06:50,400 --> 00:06:52,480
Speaker 1: the quality of AI models?

152
00:06:52,560 --> 00:06:55,240
Speaker 4: Okay, yeah, that's that's more tractable. We can we can

153
00:06:55,320 --> 00:06:57,120
Speaker 4: start there, and then we can talk about three and

154
00:06:57,320 --> 00:07:01,159
Speaker 4: GPT five. So there's essentially two ways people do this

155
00:07:01,240 --> 00:07:05,000
Speaker 4: kind of model evaluation. The main one is just by

156
00:07:05,120 --> 00:07:08,159
Speaker 4: testing them on different data sets. So, for example, there's

157
00:07:08,200 --> 00:07:10,680
Speaker 4: this data set called swee bench that's a bunch of

158
00:07:11,120 --> 00:07:14,680
Speaker 4: software engineering related coding problems and they all have a

159
00:07:14,760 --> 00:07:18,040
Speaker 4: human written solution and tests, and so you can ask

160
00:07:18,120 --> 00:07:20,280
Speaker 4: GPT five, can you write the code for this and

161
00:07:20,320 --> 00:07:22,320
Speaker 4: then run the tests and see if it's right? And

162
00:07:22,400 --> 00:07:24,360
Speaker 4: still the models are pretty bad at that. I think

163
00:07:24,360 --> 00:07:26,800
Speaker 4: they can do about half of them. They're very hard.

164
00:07:26,840 --> 00:07:30,640
Speaker 4: They're like entire days of work for professional software engineers.

165
00:07:30,880 --> 00:07:32,920
Speaker 4: But when a new model comes out, they can say, oh, look,

166
00:07:32,920 --> 00:07:35,360
Speaker 4: we actually got a higher score on sweet bench. And

167
00:07:35,400 --> 00:07:37,520
Speaker 4: there's a ton of different data sets like that. So

168
00:07:37,560 --> 00:07:39,920
Speaker 4: when GBT five comes out, they say, you know, it's

169
00:07:39,920 --> 00:07:42,880
Speaker 4: better at these types of coding tests. And a big

170
00:07:42,880 --> 00:07:46,760
Speaker 4: one that specifically open AI has been advocating for is math,

171
00:07:47,120 --> 00:07:50,200
Speaker 4: like they did the International Math Olympiad, and they said

172
00:07:50,760 --> 00:07:54,520
Speaker 4: essentially GBT five scored at the level of the best

173
00:07:54,640 --> 00:07:59,200
Speaker 4: high school mathematicians, which is pretty cool. But you raise

174
00:07:59,240 --> 00:08:00,960
Speaker 4: a good question of how is that actually map to

175
00:08:01,000 --> 00:08:03,200
Speaker 4: real world usage? And I think this is like a

176
00:08:03,240 --> 00:08:06,680
Speaker 4: really hard problem that people still haven't figured out.

177
00:08:06,960 --> 00:08:11,080
Speaker 2: Does anyone try to capture that sort of like genes sequah?

178
00:08:11,320 --> 00:08:13,640
Speaker 2: I guess when it comes to AI models, is one

179
00:08:13,680 --> 00:08:15,920
Speaker 2: of the tests asking it to I don't know, come

180
00:08:16,000 --> 00:08:17,680
Speaker 2: up with a stupid limerick or something.

181
00:08:18,160 --> 00:08:21,200
Speaker 4: Yeah, there are a lot of tests like that. There's

182
00:08:21,280 --> 00:08:25,200
Speaker 4: some creative writing benchmarks and some poetry related ones. But

183
00:08:25,480 --> 00:08:29,080
Speaker 4: I think you point out something interesting that for example,

184
00:08:29,120 --> 00:08:32,680
Speaker 4: I mostly use Claude from Anthropic and I think Claude

185
00:08:32,760 --> 00:08:36,520
Speaker 4: does have this something to it that's like a little

186
00:08:36,559 --> 00:08:38,960
Speaker 4: bit different, and it's very difficult to characterize. It's just

187
00:08:39,000 --> 00:08:40,400
Speaker 4: sort of the way it speaks to you and the

188
00:08:40,400 --> 00:08:42,760
Speaker 4: way it thinks of itself is I like it a

189
00:08:42,800 --> 00:08:44,960
Speaker 4: lot better, but I don't know how you would design

190
00:08:45,000 --> 00:08:47,320
Speaker 4: like a data set that can really capture that. The

191
00:08:47,360 --> 00:08:50,079
Speaker 4: second way they do the evaluation is by they call

192
00:08:50,080 --> 00:08:53,840
Speaker 4: it it's Elo scores, like in chess. So they, for example,

193
00:08:53,920 --> 00:08:56,440
Speaker 4: ask the two models to write a limerick, and then

194
00:08:56,480 --> 00:08:59,000
Speaker 4: they have humans rank which one is better, and they

195
00:08:59,040 --> 00:09:02,400
Speaker 4: make this kind of lat of Elo rankings for models.

196
00:09:02,640 --> 00:09:05,439
Speaker 4: So I think right now Claude or GPT five or

197
00:09:05,480 --> 00:09:09,440
Speaker 4: maybe the Google model is top on this ladder.

198
00:09:10,000 --> 00:09:12,680
Speaker 1: The algorithm made famous in the social network that Mark

199
00:09:12,760 --> 00:09:17,080
Speaker 1: Zuckerberg used to rate the of his colleagues still the

200
00:09:17,200 --> 00:09:19,400
Speaker 1: workhorse model for comp evaluation.

201
00:09:19,640 --> 00:09:22,800
Speaker 2: That's some good trivia, Joe, very good and no comment. Well,

202
00:09:22,880 --> 00:09:27,440
Speaker 2: I assume just on the hard number evaluation. People are

203
00:09:27,480 --> 00:09:31,839
Speaker 2: also ranking these on data usage, energy, that sort of.

204
00:09:31,760 --> 00:09:32,240
Speaker 4: Thing as well.

205
00:09:32,320 --> 00:09:35,760
Speaker 2: Right speed, speed would be a definitely.

206
00:09:35,960 --> 00:09:38,640
Speaker 4: The AI companies like to use price as a metric,

207
00:09:38,760 --> 00:09:41,120
Speaker 4: which is kind of interesting because there's a lot that

208
00:09:41,160 --> 00:09:43,440
Speaker 4: goes on behind the scenes, including just sort of like

209
00:09:44,280 --> 00:09:47,720
Speaker 4: free money that drives the prices down, but they also

210
00:09:47,760 --> 00:09:50,240
Speaker 4: do benchmark speed, and I think you make a good

211
00:09:50,280 --> 00:09:53,079
Speaker 4: point that the benchmarks can be pretty misleading, Like, for example,

212
00:09:53,080 --> 00:09:56,160
Speaker 4: there's a bunch of recent open source models that came

213
00:09:56,200 --> 00:09:59,000
Speaker 4: from different Chinese AI labs that have really, really high

214
00:09:59,080 --> 00:10:02,400
Speaker 4: scores on certain benchmarks, but people kind of think they're

215
00:10:02,400 --> 00:10:05,680
Speaker 4: not as good for real world usage for whatever reason.

216
00:10:06,360 --> 00:10:08,720
Speaker 1: I've seen people talk about this isn't part of the

217
00:10:08,840 --> 00:10:14,120
Speaker 1: problem with testing AI or evaluating AI. That a lot

218
00:10:14,160 --> 00:10:16,959
Speaker 1: of these problems exist in the real world already, right,

219
00:10:17,200 --> 00:10:19,720
Speaker 1: You see this a lot, and people are always finding this,

220
00:10:19,920 --> 00:10:23,400
Speaker 1: which is that here's an AI model that is amazing

221
00:10:23,559 --> 00:10:27,520
Speaker 1: at math on the math Olympiad, and yet it gets

222
00:10:27,520 --> 00:10:31,280
Speaker 1: tripped up by questions like which is heavier a pound

223
00:10:31,320 --> 00:10:33,880
Speaker 1: of steel or two pounds of feathers, And they'll say

224
00:10:33,920 --> 00:10:35,920
Speaker 1: that that's a trick question. A pound of steel weighs the

225
00:10:35,920 --> 00:10:38,520
Speaker 1: same as two pounds of feathers, which is clearly like

226
00:10:38,840 --> 00:10:41,760
Speaker 1: it was clearly then been trained in some sense to

227
00:10:42,280 --> 00:10:44,960
Speaker 1: recognize these steel versus feathers thing or whatever it is.

228
00:10:45,200 --> 00:10:47,920
Speaker 1: I forget if it's steel, But it also clearly can't

229
00:10:47,960 --> 00:10:49,719
Speaker 1: measure whether one or.

230
00:10:49,679 --> 00:10:50,480
Speaker 3: Two is bigger.

231
00:10:50,840 --> 00:10:54,960
Speaker 4: Yeah, that's a really good example. I think they kind

232
00:10:54,960 --> 00:10:58,720
Speaker 4: of successively include these kinds of things in more rounds

233
00:10:58,720 --> 00:11:00,760
Speaker 4: of training data, and so every time a new model

234
00:11:00,800 --> 00:11:03,640
Speaker 4: comes out, they kind of patch little holes that appeared

235
00:11:03,640 --> 00:11:06,040
Speaker 4: in the previous models. So you're pointing to this, like

236
00:11:06,080 --> 00:11:08,280
Speaker 4: they probably started with the classic riddle that's like a

237
00:11:08,320 --> 00:11:10,200
Speaker 4: pound of bricks or a pound of feathers bricks and

238
00:11:10,240 --> 00:11:13,120
Speaker 4: they're equal, but then like the models got that wrong,

239
00:11:13,160 --> 00:11:14,040
Speaker 4: and so they added to.

240
00:11:13,960 --> 00:11:19,080
Speaker 1: Something a very efficient way to achieve intelligence, like, oh yeah,

241
00:11:19,080 --> 00:11:19,960
Speaker 1: we should have included that.

242
00:11:20,000 --> 00:11:21,640
Speaker 3: Oh yeah, we got to include that trick. Oh yeah,

243
00:11:21,640 --> 00:11:22,320
Speaker 3: we gotta have right.

244
00:11:22,360 --> 00:11:26,480
Speaker 1: Like ever, like going that does not speak to me

245
00:11:26,880 --> 00:11:30,200
Speaker 1: of a line towards something that we would call anything

246
00:11:30,280 --> 00:11:32,280
Speaker 1: resembling human intelligence.

247
00:11:32,400 --> 00:11:35,760
Speaker 4: I definitely agree. I think one counter example is people

248
00:11:35,760 --> 00:11:37,880
Speaker 4: said this for a long time about self driving cars,

249
00:11:38,240 --> 00:11:40,480
Speaker 4: Like everyone was really excited about them for a long time,

250
00:11:40,520 --> 00:11:42,760
Speaker 4: and then they kind of didn't really work, like eight

251
00:11:42,880 --> 00:11:45,360
Speaker 4: or so years ago, and there was this period where

252
00:11:45,360 --> 00:11:47,959
Speaker 4: they were saying, oh, the models can't do green cones.

253
00:11:48,040 --> 00:11:50,400
Speaker 4: We're going out there trying to take videos of green cones,

254
00:11:50,440 --> 00:11:55,640
Speaker 4: and yeah, they can't do snow. I'm saying that it

255
00:11:55,720 --> 00:11:59,240
Speaker 4: worked for them, and so it might be possible. But

256
00:11:59,720 --> 00:12:01,960
Speaker 4: in the case of language models, there's something a little

257
00:12:02,000 --> 00:12:05,880
Speaker 4: more interesting happening, because we now have two ways to learn.

258
00:12:06,280 --> 00:12:07,760
Speaker 4: If you guys are ready, we could we could get

259
00:12:07,760 --> 00:12:10,040
Speaker 4: into something a little technical, which I think gives you

260
00:12:10,080 --> 00:12:13,280
Speaker 4: some insights. So there's essentially two ways you can teach

261
00:12:13,360 --> 00:12:16,680
Speaker 4: machines to learn from data. One is called supervised learning,

262
00:12:16,920 --> 00:12:19,640
Speaker 4: where the computer will copy what you did, which is

263
00:12:19,640 --> 00:12:22,040
Speaker 4: like basically what we're talking about now, and the other

264
00:12:22,160 --> 00:12:25,199
Speaker 4: is called reinforcement learning, where the computer just does something

265
00:12:25,280 --> 00:12:27,120
Speaker 4: and then you give it a reward if it does

266
00:12:27,160 --> 00:12:30,360
Speaker 4: something well. And so for a long time, like the

267
00:12:30,400 --> 00:12:34,640
Speaker 4: original chat GBT was mostly just trained with supervised learning,

268
00:12:34,720 --> 00:12:37,120
Speaker 4: like it would just copy the text from all of

269
00:12:37,160 --> 00:12:39,680
Speaker 4: the Internet, and so the best it could ever do

270
00:12:39,880 --> 00:12:44,280
Speaker 4: is emulate Reddit posts very well. And there was a

271
00:12:44,320 --> 00:12:47,439
Speaker 4: tiny bit of reinforcement learning, but people didn't know how

272
00:12:47,480 --> 00:12:50,040
Speaker 4: to do it right. And then you mentioned this three model,

273
00:12:50,040 --> 00:12:52,839
Speaker 4: which is kind of in some ways like a big jump,

274
00:12:52,960 --> 00:12:55,040
Speaker 4: like it made the models much better at math, much

275
00:12:55,040 --> 00:12:57,760
Speaker 4: better at certain things. And the way they did that

276
00:12:57,840 --> 00:13:00,760
Speaker 4: is actually through reinforcement learning. Found out a way to

277
00:13:00,840 --> 00:13:02,760
Speaker 4: kind of like let the model think for a while

278
00:13:03,240 --> 00:13:05,280
Speaker 4: and then give it a reward when it gets the

279
00:13:05,360 --> 00:13:07,600
Speaker 4: answer at the end. It's kind of scary.

280
00:13:07,840 --> 00:13:10,199
Speaker 2: Yeah, when you say give it a reward, is.

281
00:13:10,120 --> 00:13:13,680
Speaker 3: It like take a cookie paying robots?

282
00:13:13,920 --> 00:13:14,120
Speaker 1: Yeah?

283
00:13:14,240 --> 00:13:16,920
Speaker 2: Well no, genuinely, like what is the reward? You just

284
00:13:16,920 --> 00:13:18,080
Speaker 2: tell it it did a good job.

285
00:13:18,440 --> 00:13:20,199
Speaker 4: You just give it like a higher number. Okay, and

286
00:13:20,240 --> 00:13:21,559
Speaker 4: that makes you happy, all right.

287
00:13:22,120 --> 00:13:24,520
Speaker 2: I'd get a little bit worried when we're like giving

288
00:13:24,520 --> 00:13:27,520
Speaker 2: it cupcakes or something like here you go, good job.

289
00:13:28,440 --> 00:13:30,240
Speaker 2: Just going back to the intro, you know, we were

290
00:13:30,240 --> 00:13:32,880
Speaker 2: talking about how it feels like a lot of the

291
00:13:32,920 --> 00:13:36,520
Speaker 2: progress on AI models is a little bit more incremental,

292
00:13:36,960 --> 00:13:39,000
Speaker 2: and I guess it's hard to tell whether that's just

293
00:13:39,200 --> 00:13:41,840
Speaker 2: personal bias because now we're used to them and the

294
00:13:41,880 --> 00:13:44,720
Speaker 2: sort of wow moment has passed. But what does it

295
00:13:44,760 --> 00:13:47,440
Speaker 2: feel like to you in terms of improvements? Are we

296
00:13:47,600 --> 00:13:52,040
Speaker 2: seeing the improvement cycle accelerate or decelerate at this point?

297
00:13:52,240 --> 00:13:55,000
Speaker 4: I think it's kind of like the market, where it's

298
00:13:55,040 --> 00:13:57,679
Speaker 4: like always it gets faster for a little while, and

299
00:13:57,679 --> 00:14:00,560
Speaker 4: then it feels like things have slowed down and the

300
00:14:00,600 --> 00:14:02,920
Speaker 4: progress is never quite in the areas that you expect

301
00:14:03,000 --> 00:14:06,720
Speaker 4: as one example, people really thought this year was the

302
00:14:06,840 --> 00:14:10,680
Speaker 4: year when the assistance would start being able to act

303
00:14:10,720 --> 00:14:13,560
Speaker 4: like actual assistants, like the Year of agents. People actually

304
00:14:13,640 --> 00:14:15,800
Speaker 4: coined that term, I think, like the year of agents,

305
00:14:16,000 --> 00:14:19,080
Speaker 4: and it really it didn't happen for whatever reason. Maybe

306
00:14:19,080 --> 00:14:21,160
Speaker 4: it will in the next three months. But the agents

307
00:14:21,160 --> 00:14:23,200
Speaker 4: are still pretty bad the ones that you can use.

308
00:14:23,440 --> 00:14:25,920
Speaker 4: But they did get way better at competitive math, Like

309
00:14:25,960 --> 00:14:29,600
Speaker 4: now they can do these like world class proofs that

310
00:14:29,640 --> 00:14:33,080
Speaker 4: they couldn't do before. So it's almost unpredictable, like which

311
00:14:33,160 --> 00:14:36,120
Speaker 4: areas the AI will kind of conquer next, But it

312
00:14:36,160 --> 00:14:38,920
Speaker 4: does feel like progress is continuing.

313
00:14:39,320 --> 00:14:42,920
Speaker 1: Actually, what happened with agents? I've never had a successful

314
00:14:43,280 --> 00:14:45,840
Speaker 1: agent experience, even basic things like come up with a

315
00:14:45,880 --> 00:14:49,120
Speaker 1: list of every past odd Lots guests, yeah and put

316
00:14:49,160 --> 00:14:52,120
Speaker 1: it in a file or something like that, which just

317
00:14:52,760 --> 00:14:55,200
Speaker 1: there's an RSS feed that exists for odd Lots. This

318
00:14:55,200 --> 00:14:57,480
Speaker 1: should be ray stick for it all around, and then

319
00:14:58,040 --> 00:15:00,880
Speaker 1: something will happen or it'll get lazy. Here's like here's

320
00:15:00,920 --> 00:15:04,560
Speaker 1: fifteen and what is actually this is thought leaders love

321
00:15:04,600 --> 00:15:06,560
Speaker 1: this stuff. They love to talking about the agents. So

322
00:15:06,600 --> 00:15:09,720
Speaker 1: what actually happened with agents? Maybe they'll get there, but

323
00:15:09,800 --> 00:15:11,400
Speaker 1: what do you use to what is the roadblock there.

324
00:15:11,880 --> 00:15:14,680
Speaker 4: I don't think there's any conceptual roadblock, Like there's no

325
00:15:14,800 --> 00:15:17,400
Speaker 4: reason why you couldn't collect data for that and train

326
00:15:17,480 --> 00:15:20,600
Speaker 4: them either in a supervised way or using reinforcement learning.

327
00:15:20,920 --> 00:15:23,520
Speaker 4: It just hasn't happened yet. So I think maybe behind

328
00:15:23,520 --> 00:15:25,400
Speaker 4: the scenes it turned out that the problem was harder

329
00:15:25,400 --> 00:15:28,560
Speaker 4: than people thought, Like getting data from all those scenarios

330
00:15:28,640 --> 00:15:31,640
Speaker 4: is really hard. And there have been some stories from

331
00:15:31,800 --> 00:15:34,040
Speaker 4: like people that I've heard of that found these little

332
00:15:34,080 --> 00:15:37,840
Speaker 4: companies in San Francisco and they build these tiny environments

333
00:15:37,880 --> 00:15:41,240
Speaker 4: for the AI labs to do reinforcement learning on for agents,

334
00:15:41,320 --> 00:15:44,000
Speaker 4: like for example, doing a calendar. They'll build like a

335
00:15:44,000 --> 00:15:47,120
Speaker 4: little calendar app, but make it have rewards so you

336
00:15:47,120 --> 00:15:49,440
Speaker 4: can do reinforcement learning, and they can just sell that

337
00:15:49,520 --> 00:15:51,760
Speaker 4: for like hundreds of thousands of dollars. So I think

338
00:15:51,920 --> 00:15:54,600
Speaker 4: the progress is ongoing behind the scenes, Like there's a

339
00:15:54,600 --> 00:15:58,080
Speaker 4: whole ecosystem built around it. It just hasn't really manifested

340
00:15:58,080 --> 00:15:59,400
Speaker 4: in the products that we use.

341
00:16:00,000 --> 00:16:02,880
Speaker 2: I was going to ask, how much of the difficulty is,

342
00:16:03,360 --> 00:16:06,360
Speaker 2: you know, the actual development of the models, the thinking part,

343
00:16:06,440 --> 00:16:10,640
Speaker 2: versus just getting them to plug in seamlessly with other applications.

344
00:16:11,160 --> 00:16:15,440
Speaker 4: Yeah, I think the second thing is probably the biggest

345
00:16:15,440 --> 00:16:17,920
Speaker 4: barrier in terms of time, Like it just takes a

346
00:16:17,920 --> 00:16:20,520
Speaker 4: really long time to figure out what data you need

347
00:16:20,640 --> 00:16:23,160
Speaker 4: and collect it properly and actually train the models on

348
00:16:23,200 --> 00:16:25,680
Speaker 4: that data. But at the same time, there are people

349
00:16:26,120 --> 00:16:27,920
Speaker 4: like me who are trying to work on better like

350
00:16:28,000 --> 00:16:31,400
Speaker 4: conceptual frameworks for training the models. So to go back

351
00:16:31,400 --> 00:16:37,280
Speaker 4: to the three example, doing reinforcement learning on CHATGBT, like

352
00:16:37,320 --> 00:16:40,240
Speaker 4: that seems to me like a huge breakthrough, Like we

353
00:16:40,280 --> 00:16:42,760
Speaker 4: didn't know how to do that before. It unlocks all

354
00:16:42,800 --> 00:16:45,800
Speaker 4: sorts of doors and ways to train the models. So

355
00:16:45,960 --> 00:16:48,400
Speaker 4: even if maybe you don't think that model was that

356
00:16:48,480 --> 00:16:50,880
Speaker 4: much better than the previous one, it seems like it

357
00:16:50,960 --> 00:16:54,160
Speaker 4: will give us huge improvements in the future.

358
00:17:10,040 --> 00:17:14,879
Speaker 1: So you mentioned at the intro that it's possible, hopefully

359
00:17:14,920 --> 00:17:16,919
Speaker 1: you'll get a close but you might end up in

360
00:17:16,960 --> 00:17:19,879
Speaker 1: a situation which you go to work for some frontier

361
00:17:20,040 --> 00:17:22,800
Speaker 1: AI lab and we never hear from you again, or

362
00:17:22,840 --> 00:17:25,480
Speaker 1: you just post cryptic tweets like oh no idea, what's coming,

363
00:17:25,880 --> 00:17:26,680
Speaker 1: Oh it's gonna.

364
00:17:26,440 --> 00:17:29,760
Speaker 3: Be so over or whatever. Yeah, an the death Star, Yeah,

365
00:17:29,760 --> 00:17:30,560
Speaker 3: it's very annoying.

366
00:17:30,640 --> 00:17:33,320
Speaker 1: The way they all tweet, it's possible talk to us

367
00:17:33,359 --> 00:17:36,400
Speaker 1: about like why not work on an open source project?

368
00:17:36,880 --> 00:17:38,880
Speaker 1: And this is of course when people talk about deep

369
00:17:38,920 --> 00:17:40,680
Speaker 1: seek and a lot of the Chinese models that the

370
00:17:40,760 --> 00:17:43,520
Speaker 1: US competes with, a lot of those are open source.

371
00:17:43,840 --> 00:17:46,960
Speaker 1: Presumably you could keep coming on odd lads over and

372
00:17:47,000 --> 00:17:50,639
Speaker 1: over again, why like what is even the case for

373
00:17:50,800 --> 00:17:52,960
Speaker 1: the best and the brightest to work on a closed

374
00:17:52,960 --> 00:17:54,600
Speaker 1: source frontier models.

375
00:17:54,840 --> 00:17:58,080
Speaker 4: Yeah, it's a really hard question, Like I've I've struggled

376
00:17:58,119 --> 00:18:00,399
Speaker 4: with this in my own personal decision making. I was

377
00:18:00,560 --> 00:18:03,080
Speaker 4: originally thinking, Oh, I'd love to become a professor and

378
00:18:03,119 --> 00:18:07,479
Speaker 4: mentor younger students and get a whole like group of

379
00:18:07,520 --> 00:18:11,160
Speaker 4: these ideas going and start working on similar related problems

380
00:18:11,200 --> 00:18:12,920
Speaker 4: to the stuff I was talking about. And I still

381
00:18:12,920 --> 00:18:15,639
Speaker 4: think that would be fun. But there's a big gap

382
00:18:15,680 --> 00:18:18,520
Speaker 4: in terms of the things we can do at Cornell

383
00:18:18,640 --> 00:18:20,600
Speaker 4: and the things that you can do at open AI.

384
00:18:20,800 --> 00:18:24,919
Speaker 4: Like they just have like crazy infrastructure for training models

385
00:18:24,960 --> 00:18:29,480
Speaker 4: really easily and data and a ton of really good data.

386
00:18:29,960 --> 00:18:32,720
Speaker 4: And so I think as that gap has widened, I've

387
00:18:32,720 --> 00:18:34,760
Speaker 4: felt like a lot of what we're doing is like

388
00:18:35,080 --> 00:18:38,080
Speaker 4: kind of devising these toy scenarios where we can study

389
00:18:38,080 --> 00:18:41,240
Speaker 4: interesting things, but I feel a bit disconnected from the

390
00:18:41,280 --> 00:18:45,399
Speaker 4: real like progress of humanity. You know, like if you

391
00:18:45,480 --> 00:18:47,879
Speaker 4: really agree that this is like the biggest problem of

392
00:18:47,920 --> 00:18:49,760
Speaker 4: our time. I don't want to say it's like the

393
00:18:49,800 --> 00:18:52,439
Speaker 4: Manhattan Project, but like, what's more like trying to go

394
00:18:52,480 --> 00:18:54,680
Speaker 4: to the Moon in the sixties? The space race. It's

395
00:18:54,760 --> 00:18:56,720
Speaker 4: kind of like a space race going on in these

396
00:18:56,760 --> 00:18:58,840
Speaker 4: different private labs. You want to be a part of it.

397
00:18:58,880 --> 00:19:02,320
Speaker 4: Like there's crazy energy that it has huge implications for

398
00:19:02,359 --> 00:19:05,880
Speaker 4: the future of society. So I think I am interested

399
00:19:05,880 --> 00:19:09,760
Speaker 4: in participating in that. My big question is like, if

400
00:19:09,760 --> 00:19:13,560
Speaker 4: you think that the reinforcement learning thing was the most

401
00:19:13,600 --> 00:19:16,800
Speaker 4: recent big scientific breakthrough, like oh one, and then three,

402
00:19:17,240 --> 00:19:20,440
Speaker 4: what's next? And then like where will that actually be happening.

403
00:19:20,480 --> 00:19:22,800
Speaker 4: That's kind of what I'm thinking about right now.

404
00:19:22,880 --> 00:19:26,280
Speaker 2: Just on the data point. I was reading your excellent

405
00:19:26,480 --> 00:19:29,920
Speaker 2: substack and you argue that there's probably an upper bound

406
00:19:30,040 --> 00:19:33,320
Speaker 2: to what you can get out of a given data set,

407
00:19:34,160 --> 00:19:38,440
Speaker 2: and at some point, like the training starts to look

408
00:19:38,520 --> 00:19:42,520
Speaker 2: pretty similar, right, and the data becomes the differentiating factor.

409
00:19:43,520 --> 00:19:48,000
Speaker 2: How important are data sets to AI research? And I guess, like,

410
00:19:48,119 --> 00:19:50,359
Speaker 2: how do you go about finding really cool ones and

411
00:19:50,359 --> 00:19:53,520
Speaker 2: what's left. Because I feel like, you know, using the

412
00:19:53,560 --> 00:19:57,320
Speaker 2: space race analogy, everyone has been running so fast on this.

413
00:19:57,600 --> 00:19:59,800
Speaker 2: It feels like all the data sets must have been

414
00:19:59,840 --> 00:20:02,159
Speaker 2: a explored by now, but I guess they haven't.

415
00:20:02,520 --> 00:20:05,520
Speaker 4: Yeah, yeah, I think this is really getting to the

416
00:20:05,560 --> 00:20:08,280
Speaker 4: heart of what people are trying to figure out right

417
00:20:08,280 --> 00:20:11,880
Speaker 4: now in all these different labs. So I think you're

418
00:20:12,000 --> 00:20:15,800
Speaker 4: pretty much right that all of the public data sets

419
00:20:15,840 --> 00:20:21,159
Speaker 4: we have are pretty much used to TRAIN three or

420
00:20:21,520 --> 00:20:24,040
Speaker 4: GPT five or whatever. If there is a really good

421
00:20:24,400 --> 00:20:27,280
Speaker 4: website that should have been scraped and downloaded into the model,

422
00:20:27,320 --> 00:20:30,320
Speaker 4: it should probably be used. But there apparently is a

423
00:20:30,400 --> 00:20:33,679
Speaker 4: much larger amount of private data than public data. I mean,

424
00:20:33,760 --> 00:20:36,959
Speaker 4: you all work for Bloomberg, so you're probably intimately familiar

425
00:20:37,040 --> 00:20:39,119
Speaker 4: with this. But if you think about the different AI

426
00:20:39,200 --> 00:20:41,639
Speaker 4: labs that exist, they actually now do kind of have

427
00:20:41,760 --> 00:20:45,600
Speaker 4: different data related modes. Like XAI, they have all of

428
00:20:45,640 --> 00:20:49,280
Speaker 4: the Twitter data that's basically impossible to get elsewhere. CHADGBT

429
00:20:49,520 --> 00:20:52,480
Speaker 4: now has all of the user conversations with CHATGBT, which

430
00:20:52,520 --> 00:20:55,040
Speaker 4: are really useful. Claude has a ton of coding data

431
00:20:55,040 --> 00:20:57,720
Speaker 4: that other people don't have. Google has YouTube, which some

432
00:20:57,760 --> 00:21:00,760
Speaker 4: people think might be like the next source of making

433
00:21:00,800 --> 00:21:03,520
Speaker 4: really good models, and none of those things are really included,

434
00:21:03,880 --> 00:21:06,120
Speaker 4: at least not much in today's models.

435
00:21:06,560 --> 00:21:07,680
Speaker 3: This is really important.

436
00:21:07,800 --> 00:21:11,920
Speaker 1: Like once a lab builds some sort of base, whether

437
00:21:12,040 --> 00:21:16,160
Speaker 1: it's anthropic encoding or maybe cursor encoding, even though they're

438
00:21:16,200 --> 00:21:19,960
Speaker 1: not like a core lab, et cetera, like they become

439
00:21:20,119 --> 00:21:22,920
Speaker 1: a source of their own data that literally nobody else has.

440
00:21:23,240 --> 00:21:26,320
Speaker 4: Yeah, actually Cursor is a great example. So they are

441
00:21:26,440 --> 00:21:29,320
Speaker 4: very technical, they have really smart people. They're very small,

442
00:21:29,440 --> 00:21:32,760
Speaker 4: so they haven't quite scaled to at least in terms

443
00:21:32,760 --> 00:21:33,400
Speaker 4: of the number of people.

444
00:21:33,400 --> 00:21:34,960
Speaker 1: But I think about this like every time I was like,

445
00:21:35,440 --> 00:21:37,000
Speaker 1: when I've played with this is like this is good,

446
00:21:37,040 --> 00:21:40,040
Speaker 1: this is bad. I'm constantly teaching their model to get better,

447
00:21:40,119 --> 00:21:41,480
Speaker 1: right right, right right.

448
00:21:41,880 --> 00:21:43,479
Speaker 4: They're in a problem where they have the data. They

449
00:21:43,520 --> 00:21:46,000
Speaker 4: just have to take the right algorithms and scale it

450
00:21:46,080 --> 00:21:47,879
Speaker 4: up to train the model to be as good as

451
00:21:48,040 --> 00:21:50,720
Speaker 4: Claude is. But that actually seems a lot more feasible

452
00:21:50,800 --> 00:21:53,080
Speaker 4: than other companies that have no data and want to

453
00:21:53,080 --> 00:21:55,080
Speaker 4: train good models, even if they know how, it seems

454
00:21:55,200 --> 00:21:55,800
Speaker 4: very difficult.

455
00:21:56,640 --> 00:22:01,439
Speaker 2: How closely are AI researchers working or talking to I

456
00:22:01,440 --> 00:22:04,840
Speaker 2: guess other parts of the AI ecosystem, so you know,

457
00:22:05,080 --> 00:22:09,159
Speaker 2: chip makers, maybe cloud providers, that sort of thing. Is

458
00:22:09,160 --> 00:22:10,760
Speaker 2: there a lot of dialogue or not really.

459
00:22:11,000 --> 00:22:14,240
Speaker 4: I think certain people talk all the time to the

460
00:22:14,280 --> 00:22:17,720
Speaker 4: chip makers, Like there's a big community of people. You know,

461
00:22:17,800 --> 00:22:21,080
Speaker 4: the AI models all run on GPUs, and there are

462
00:22:21,160 --> 00:22:23,040
Speaker 4: a lot of people that are getting really good at

463
00:22:23,080 --> 00:22:26,760
Speaker 4: writing fast GPU code. It's called kernels, and all those

464
00:22:26,760 --> 00:22:29,280
Speaker 4: people who work on kernels talk to the chip makers

465
00:22:29,320 --> 00:22:32,040
Speaker 4: all the time. Like Amazon's making their own chips, Google

466
00:22:32,080 --> 00:22:35,000
Speaker 4: has their own chip. Now all the hyperscalers are making chips,

467
00:22:35,000 --> 00:22:36,919
Speaker 4: and I think they're all trying to talk to the

468
00:22:36,920 --> 00:22:39,240
Speaker 4: people that actually write the fast code that runs on

469
00:22:39,320 --> 00:22:41,480
Speaker 4: chips to figure out I think they call it hardware

470
00:22:41,520 --> 00:22:43,919
Speaker 4: software code design, Like everyone's kind of getting together and

471
00:22:43,960 --> 00:22:45,560
Speaker 4: trying to figure out what the best way is to

472
00:22:45,640 --> 00:22:47,639
Speaker 4: design the next round of GPUs.

473
00:22:48,359 --> 00:22:52,760
Speaker 1: So you mentioned, okay, Google might have an advantage because

474
00:22:52,960 --> 00:22:56,199
Speaker 1: it owns YouTube and there's just tons of obviously just

475
00:22:56,280 --> 00:22:56,840
Speaker 1: tons of.

476
00:22:57,119 --> 00:22:57,800
Speaker 3: Data in there.

477
00:22:57,840 --> 00:23:00,560
Speaker 1: So one way you could get access to the YouTube

478
00:23:00,640 --> 00:23:03,439
Speaker 1: data is to literally be Google and own it. But

479
00:23:03,560 --> 00:23:06,640
Speaker 1: another way that maybe you could get access to YouTube

480
00:23:06,720 --> 00:23:10,560
Speaker 1: data is operate in China where there are no laws

481
00:23:10,680 --> 00:23:13,320
Speaker 1: about this type of thing, or no, they're not beholding

482
00:23:13,320 --> 00:23:16,920
Speaker 1: the US copyright and just sort of scrape at all. Again,

483
00:23:17,080 --> 00:23:20,360
Speaker 1: since most of the Chinese AI labs are open to source,

484
00:23:21,040 --> 00:23:24,560
Speaker 1: why isn't this just a huge advantage for the Chinese

485
00:23:24,600 --> 00:23:27,359
Speaker 1: labs that they're really not going to be Hey, open

486
00:23:27,400 --> 00:23:29,080
Speaker 1: AI they get super at the New York Times all

487
00:23:29,119 --> 00:23:32,440
Speaker 1: these deepseek isn't having to deal with all these headaches?

488
00:23:33,000 --> 00:23:38,919
Speaker 4: Yeah, I think the American AI labs will probably do

489
00:23:39,080 --> 00:23:41,720
Speaker 4: things behind the scenes that they wouldn't tell you about

490
00:23:41,800 --> 00:23:45,439
Speaker 4: to get good data solution. Just don't so Yeah, Like

491
00:23:45,520 --> 00:23:48,680
Speaker 4: I think they wouldn't release the models that are potentially

492
00:23:48,720 --> 00:23:51,639
Speaker 4: trained on scraped or copyrighted data. But if that's the

493
00:23:51,680 --> 00:23:55,000
Speaker 4: way to get better math Olympiad scores, then people will

494
00:23:55,160 --> 00:23:57,560
Speaker 4: I think I would guess do that. But you're right

495
00:23:57,600 --> 00:24:00,159
Speaker 4: that like the Chinese, the Chinese model makers can to

496
00:24:00,160 --> 00:24:02,480
Speaker 4: sort of take all the books that they can pirate

497
00:24:02,560 --> 00:24:04,600
Speaker 4: from the Internet and train on them and they're not

498
00:24:04,680 --> 00:24:06,920
Speaker 4: violating any laws and they can release the model to

499
00:24:06,960 --> 00:24:09,520
Speaker 4: the public and it's all fine, which is honestly great

500
00:24:09,520 --> 00:24:12,800
Speaker 4: for us because then people like me could probably download

501
00:24:12,800 --> 00:24:15,240
Speaker 4: a model that's better than we would get otherwise.

502
00:24:15,640 --> 00:24:18,400
Speaker 2: What was your impression of deep Seek when it came out?

503
00:24:18,680 --> 00:24:19,080
Speaker 2: And now?

504
00:24:20,119 --> 00:24:24,000
Speaker 4: I was pretty surprised at how much of a splash

505
00:24:24,080 --> 00:24:26,920
Speaker 4: they made. The model is really good, and I think

506
00:24:26,960 --> 00:24:30,639
Speaker 4: a lot of people are building on it, including me,

507
00:24:30,840 --> 00:24:33,399
Speaker 4: and like most people that are at AI companies that

508
00:24:33,440 --> 00:24:36,639
Speaker 4: aren't super super big are building on deep Seek. But

509
00:24:37,840 --> 00:24:40,960
Speaker 4: it was surprising, like what a huge deal it was

510
00:24:41,000 --> 00:24:43,280
Speaker 4: to people, like my mom's asking me about deep Seek.

511
00:24:43,280 --> 00:24:45,439
Speaker 4: I think my grandma knew about deep Seek, and she

512
00:24:45,520 --> 00:24:47,040
Speaker 4: barely knew about chat GBT.

513
00:24:47,200 --> 00:24:50,800
Speaker 2: So that's when you know it's gone mainstream when starts.

514
00:24:50,480 --> 00:24:53,399
Speaker 4: Asking you and there was nothing else so far. I

515
00:24:53,480 --> 00:24:55,879
Speaker 4: think in the AI space that's made quite that much news.

516
00:24:55,960 --> 00:24:58,200
Speaker 1: But it sounds like what you're saying is that it's

517
00:24:58,240 --> 00:25:01,520
Speaker 1: a very good model, but that on the actual specs

518
00:25:02,359 --> 00:25:05,800
Speaker 1: from your perspective, it didn't quite deserve is much attention,

519
00:25:05,920 --> 00:25:08,680
Speaker 1: Like it was good, but like in your view, it's

520
00:25:08,720 --> 00:25:11,720
Speaker 1: not so good that everyone needed to be talking about it.

521
00:25:12,000 --> 00:25:16,360
Speaker 4: Yeah, I think it's really useful because they released all

522
00:25:16,400 --> 00:25:19,080
Speaker 4: their model weights and they said exactly what they did

523
00:25:19,080 --> 00:25:21,119
Speaker 4: to train it. Although they didn't say what the data was,

524
00:25:21,640 --> 00:25:24,240
Speaker 4: but it gave me the impression of there maybe six

525
00:25:24,280 --> 00:25:26,840
Speaker 4: to twelve months behind the American AI labs in terms

526
00:25:26,880 --> 00:25:29,040
Speaker 4: of how well they can do the training and stuff.

527
00:25:29,200 --> 00:25:31,560
Speaker 4: But it still was a pretty big update for me

528
00:25:31,680 --> 00:25:34,080
Speaker 4: to know that, Wow, there are one hundred people that

529
00:25:34,160 --> 00:25:36,520
Speaker 4: don't have PhDs working at a Chinese hedge fund that

530
00:25:36,560 --> 00:25:39,320
Speaker 4: are training these like cutting edge models. Like it is

531
00:25:39,359 --> 00:25:41,480
Speaker 4: incredible and they work very hard, they're very good.

532
00:25:57,600 --> 00:26:00,119
Speaker 2: Do you have pressure or do you feel pressure or

533
00:26:00,600 --> 00:26:05,280
Speaker 2: do AI researchers in general fuel pressure to consider monetization

534
00:26:05,520 --> 00:26:08,680
Speaker 2: when they're researching things or is it you know, mostly

535
00:26:09,080 --> 00:26:12,960
Speaker 2: still curiosity driven, that sort of old school Silicon Valley

536
00:26:13,000 --> 00:26:15,399
Speaker 2: we're improving the world kind of thing. Or is it

537
00:26:15,560 --> 00:26:19,280
Speaker 2: much more mercenary given that all of these big companies

538
00:26:19,320 --> 00:26:21,680
Speaker 2: seem to be competing in the same space.

539
00:26:22,359 --> 00:26:27,240
Speaker 4: Yeah. I think that over time it's gotten harder and

540
00:26:27,359 --> 00:26:30,400
Speaker 4: harder to do things that are just like cool ideas

541
00:26:30,560 --> 00:26:35,320
Speaker 4: or seem cute but don't have any necessary application, and

542
00:26:35,520 --> 00:26:37,639
Speaker 4: things are getting closer and closer to products, you know,

543
00:26:37,760 --> 00:26:40,720
Speaker 4: even like the language models that power CHAGBT. I was

544
00:26:40,760 --> 00:26:43,720
Speaker 4: working in those before CHAGBT, and they had some uses,

545
00:26:43,760 --> 00:26:47,439
Speaker 4: but also they're intellectually interesting and like fun to build.

546
00:26:47,960 --> 00:26:50,560
Speaker 4: But now if I came up with a better way

547
00:26:50,600 --> 00:26:54,360
Speaker 4: to train CHAGBT, that's like a multi billion dollar innovation.

548
00:26:54,560 --> 00:26:55,440
Speaker 2: The stakes are higher.

549
00:26:55,640 --> 00:26:57,760
Speaker 4: Yeah, I'd be like an asset to like the United

550
00:26:57,800 --> 00:27:00,120
Speaker 4: States government or something if I knew how to do that.

551
00:27:00,240 --> 00:27:02,800
Speaker 4: So I guess it depends on what kind of problems

552
00:27:02,840 --> 00:27:05,560
Speaker 4: you work on. Like, I'm more interested in understanding how

553
00:27:05,600 --> 00:27:10,080
Speaker 4: things work, so it becomes a bit less financially dire.

554
00:27:10,200 --> 00:27:14,440
Speaker 1: I think that six to twelve month gap between what

555
00:27:14,480 --> 00:27:17,640
Speaker 1: was that that was a January deep segment. Yeah, everyone

556
00:27:18,200 --> 00:27:20,320
Speaker 1: was in December that they first got at attention, then

557
00:27:20,359 --> 00:27:22,399
Speaker 1: for some reason really hit in January. Is that a

558
00:27:22,440 --> 00:27:26,160
Speaker 1: sustainable gap? Is there something either in access to data,

559
00:27:26,400 --> 00:27:30,240
Speaker 1: access to talent, access to compute, access to chips, whatever,

560
00:27:30,320 --> 00:27:34,800
Speaker 1: access to energy that in your view will allow us

561
00:27:34,840 --> 00:27:37,680
Speaker 1: frontier lebs to maintain some sort of six to twelve

562
00:27:37,720 --> 00:27:38,720
Speaker 1: month gap for a while.

563
00:27:39,119 --> 00:27:41,560
Speaker 4: It's pretty unclear to me. I think there are different

564
00:27:41,600 --> 00:27:43,920
Speaker 4: beliefs you can have. You can believe that the ideas

565
00:27:44,000 --> 00:27:46,959
Speaker 4: and the people are really the thing that differentiates the models,

566
00:27:47,040 --> 00:27:49,160
Speaker 4: and in that case, I think we haven't so far

567
00:27:49,240 --> 00:27:52,920
Speaker 4: seen a lot of like the top USAI researchers going

568
00:27:52,920 --> 00:27:56,800
Speaker 4: to work at Chinese labs, so that seems stable. You

569
00:27:56,800 --> 00:27:59,080
Speaker 4: could think that chips really matter, and in that case

570
00:27:59,600 --> 00:28:02,360
Speaker 4: the chip race is really happening between big American companies.

571
00:28:02,400 --> 00:28:04,960
Speaker 4: Like I think, actually China has a pretty big deficit

572
00:28:05,000 --> 00:28:08,440
Speaker 4: coming up in terms of like the GPUs we're exporting,

573
00:28:09,080 --> 00:28:10,919
Speaker 4: or you can think that the data matters, and I

574
00:28:10,920 --> 00:28:15,439
Speaker 4: guess actually any of these point in the favor of

575
00:28:15,480 --> 00:28:18,040
Speaker 4: the US. I think if you think the data really matters,

576
00:28:18,600 --> 00:28:21,520
Speaker 4: maybe the data they gather through like deepseek dot com

577
00:28:21,600 --> 00:28:23,359
Speaker 4: usage is really good and they can use it to

578
00:28:23,400 --> 00:28:26,119
Speaker 4: like bootstrap a better model. But I think the American

579
00:28:26,160 --> 00:28:28,639
Speaker 4: companies really do have an advantage. Like you all might

580
00:28:28,680 --> 00:28:31,600
Speaker 4: have heard this story just as an anecdote. Apparently at

581
00:28:31,640 --> 00:28:36,240
Speaker 4: Anthropic they've been buying and scanning thousands of old books

582
00:28:36,480 --> 00:28:38,520
Speaker 4: for several years, so they have this division. I think

583
00:28:38,520 --> 00:28:41,640
Speaker 4: they're based in New York that buys like shipping containers

584
00:28:41,680 --> 00:28:44,760
Speaker 4: full of old manuscripts, cuts off the spines and puts

585
00:28:44,800 --> 00:28:47,000
Speaker 4: them in these scanning machines and then they turn them

586
00:28:47,040 --> 00:28:49,840
Speaker 4: into like really high quality text. And so I'm noting

587
00:28:49,880 --> 00:28:53,040
Speaker 4: Claude has this like weird aspect to it. Maybe part

588
00:28:53,080 --> 00:28:57,160
Speaker 4: of the reason is they've gathered like trillions of words

589
00:28:57,280 --> 00:29:00,360
Speaker 4: worth of like old book data over many years, and

590
00:29:00,400 --> 00:29:03,080
Speaker 4: that's pretty hard to replicate elsewhere. So I think that

591
00:29:03,160 --> 00:29:05,040
Speaker 4: head start really does mean a lot.

592
00:29:06,160 --> 00:29:08,600
Speaker 2: What are you most excited about at the moment? The

593
00:29:08,640 --> 00:29:12,520
Speaker 2: book thing sounds very cool, but what is getting all

594
00:29:12,520 --> 00:29:13,960
Speaker 2: your attention right now?

595
00:29:14,160 --> 00:29:18,800
Speaker 4: Thanks for asking. I think I mentioned before I'm really

596
00:29:18,840 --> 00:29:21,560
Speaker 4: trying to figure out what's coming next. There are some

597
00:29:21,640 --> 00:29:24,800
Speaker 4: obvious things like we can get computer usage data and

598
00:29:25,160 --> 00:29:27,720
Speaker 4: train better agents, or we can get more coding data

599
00:29:27,800 --> 00:29:30,280
Speaker 4: and make them better coding or writing gp code or whatever,

600
00:29:30,560 --> 00:29:35,280
Speaker 4: But like, what are the non obvious advancements? And my

601
00:29:35,480 --> 00:29:39,640
Speaker 4: personal opinion is that the next round of improvements and

602
00:29:39,680 --> 00:29:44,240
Speaker 4: AI models will come from some type of personalization and

603
00:29:44,360 --> 00:29:48,560
Speaker 4: online learning, which means like models that one are trained

604
00:29:48,640 --> 00:29:50,880
Speaker 4: like per person or per company. So like you could

605
00:29:50,880 --> 00:29:54,280
Speaker 4: think of like CHADGBT is the same model that gets

606
00:29:54,280 --> 00:29:57,680
Speaker 4: served to everyone, so it has to store information about

607
00:29:57,920 --> 00:30:02,080
Speaker 4: random restaurants and like countries you never go to. But

608
00:30:02,200 --> 00:30:04,880
Speaker 4: instead if you had a CHAGBT that's specific to Bloomberg

609
00:30:04,960 --> 00:30:07,719
Speaker 4: or specific to your work, it might be able to

610
00:30:07,760 --> 00:30:10,280
Speaker 4: like use more of its brain to do work for you.

611
00:30:10,760 --> 00:30:13,040
Speaker 4: And then the second thing is if it was updated

612
00:30:13,120 --> 00:30:14,960
Speaker 4: every day, so like if you ask it to make

613
00:30:15,000 --> 00:30:19,160
Speaker 4: your odd lots calendar, yeah, or RSS feed and you're like, no,

614
00:30:19,360 --> 00:30:21,160
Speaker 4: that was wrong, Like you did it wrong for this

615
00:30:21,240 --> 00:30:23,960
Speaker 4: reason this reason, and you try again tomorrow, it'll still

616
00:30:24,000 --> 00:30:27,560
Speaker 4: break tomorrow because it doesn't like continuously improve its capabilities.

617
00:30:28,080 --> 00:30:31,520
Speaker 4: So oh yeah, I think that's the direction things are going.

618
00:30:31,640 --> 00:30:33,520
Speaker 3: I've heard people talk about this now.

619
00:30:33,560 --> 00:30:36,600
Speaker 1: Granted, models are getting better over time, but you know,

620
00:30:36,640 --> 00:30:40,880
Speaker 1: people might compare a coding model to a beginning software

621
00:30:41,000 --> 00:30:43,080
Speaker 1: engineer and say, the coding model is better, but that

622
00:30:43,200 --> 00:30:45,360
Speaker 1: software engineer is going to start getting better the next

623
00:30:45,400 --> 00:30:47,040
Speaker 1: day they're on the job, and every day for the

624
00:30:47,040 --> 00:30:49,240
Speaker 1: rest of their career, they're probably going to be a

625
00:30:49,280 --> 00:30:52,600
Speaker 1: better software engineer than they were the day before, whereas

626
00:30:52,640 --> 00:30:56,800
Speaker 1: at least that version of the model will not be better.

627
00:30:56,840 --> 00:30:58,200
Speaker 3: That is that right? Yeah?

628
00:30:58,240 --> 00:31:00,000
Speaker 1: Yeah, that seems like an issue that people talk about

629
00:31:00,280 --> 00:31:00,800
Speaker 1: in your world.

630
00:31:01,040 --> 00:31:03,160
Speaker 4: Yeah, yeah, I think this is a big problem. It's

631
00:31:03,200 --> 00:31:06,000
Speaker 4: like we have to wait six months for the chat

632
00:31:06,040 --> 00:31:09,000
Speaker 4: GPT five point one to come out, and then maybe

633
00:31:09,040 --> 00:31:11,640
Speaker 4: they'll include your problems as the training data, and so

634
00:31:11,680 --> 00:31:14,520
Speaker 4: maybe it'll get better, but it might not. And instead,

635
00:31:14,560 --> 00:31:17,280
Speaker 4: I think people need to think about ways to do

636
00:31:17,320 --> 00:31:20,360
Speaker 4: that update more dynamically, like every time you talk to it,

637
00:31:20,600 --> 00:31:22,360
Speaker 4: or maybe every night when you go to sleep, the

638
00:31:22,400 --> 00:31:24,880
Speaker 4: model kind of like gets to work and studies what

639
00:31:24,920 --> 00:31:27,040
Speaker 4: it was talking to you about and crafts better tests

640
00:31:27,040 --> 00:31:28,920
Speaker 4: for itself and then learns and then when you wake up,

641
00:31:29,000 --> 00:31:30,120
Speaker 4: the model's actually better.

642
00:31:30,600 --> 00:31:33,120
Speaker 1: The other big question that I have and is kind

643
00:31:33,160 --> 00:31:36,280
Speaker 1: of related to this, especially when we're talking about AI

644
00:31:36,440 --> 00:31:40,520
Speaker 1: replacing the humans in certain forms of labor, is that

645
00:31:40,720 --> 00:31:44,360
Speaker 1: like do we need really really advanced aid like in

646
00:31:44,400 --> 00:31:47,320
Speaker 1: other words, like there is a lot of again, the

647
00:31:47,440 --> 00:31:51,560
Speaker 1: existing models are extremely impressive, Like in your view, do

648
00:31:51,640 --> 00:31:54,880
Speaker 1: we need to get a lot better technically for them

649
00:31:54,920 --> 00:31:58,280
Speaker 1: to have economic impact? And since these are in many

650
00:31:58,280 --> 00:32:01,600
Speaker 1: cases businesses at the end of the day, is it

651
00:32:01,760 --> 00:32:05,400
Speaker 1: necessary that there's so much work being done towards advancing

652
00:32:05,880 --> 00:32:06,600
Speaker 1: the cutting edge?

653
00:32:06,920 --> 00:32:09,400
Speaker 4: Yeah, yeah, that's a great question, Like we could have

654
00:32:10,280 --> 00:32:13,720
Speaker 4: really good interns without ever getting better scores on the

655
00:32:13,720 --> 00:32:16,960
Speaker 4: Math Olympiad, Like that's not necessarily something that we ever

656
00:32:17,040 --> 00:32:19,680
Speaker 4: had to go after. I think part of the reason

657
00:32:19,720 --> 00:32:21,680
Speaker 4: for that is that AI labs are engaged in this

658
00:32:21,800 --> 00:32:24,360
Speaker 4: kind of neck and neck race to have the smartest model.

659
00:32:24,720 --> 00:32:28,640
Speaker 4: But I totally agree that AI could be economically transformative

660
00:32:29,040 --> 00:32:32,080
Speaker 4: without having a higher ceiling in terms of what it

661
00:32:32,080 --> 00:32:33,520
Speaker 4: can do. It's more like it needs to be more

662
00:32:33,560 --> 00:32:36,240
Speaker 4: consistent or like dependable than actually smarter.

663
00:32:37,320 --> 00:32:39,440
Speaker 2: This might be a weird question, but once you've made

664
00:32:39,640 --> 00:32:43,760
Speaker 2: a sort of foundational improvement to a particular model, how

665
00:32:43,800 --> 00:32:47,400
Speaker 2: easy or difficult is it to rewind if you need to.

666
00:32:47,760 --> 00:32:50,400
Speaker 2: And one of the reasons I ask is because you know,

667
00:32:50,480 --> 00:32:53,880
Speaker 2: some people have been complaining that they've been training chat

668
00:32:53,920 --> 00:32:56,760
Speaker 2: GPT to I don't know, be their boyfriend or whatever,

669
00:32:56,960 --> 00:33:01,040
Speaker 2: be their therapist topic. Yeah, and then it gets upgraded

670
00:33:01,360 --> 00:33:04,640
Speaker 2: and all of that training suddenly disappears and the personality

671
00:33:04,760 --> 00:33:06,280
Speaker 2: of the model changes.

672
00:33:07,160 --> 00:33:09,840
Speaker 4: Yeah, that was a really interesting story. So I think

673
00:33:09,880 --> 00:33:13,560
Speaker 4: the model before GPT five was four to zero. And

674
00:33:13,600 --> 00:33:17,280
Speaker 4: they said that they thought internally, like all the scientists

675
00:33:17,400 --> 00:33:20,600
Speaker 4: encoder people, that the new model was superior in every way.

676
00:33:20,640 --> 00:33:23,280
Speaker 4: It gives you shorter responses, it's a bit nicer, it's

677
00:33:23,360 --> 00:33:26,640
Speaker 4: much smarter. And then people got really upset because they

678
00:33:26,680 --> 00:33:28,720
Speaker 4: had spent so much time talking to the old model

679
00:33:28,760 --> 00:33:32,160
Speaker 4: that they felt like they'd experience like a serious loss

680
00:33:32,280 --> 00:33:33,080
Speaker 4: in their life.

681
00:33:33,240 --> 00:33:37,080
Speaker 2: Joe would miss the love it love it No.

682
00:33:37,160 --> 00:33:40,800
Speaker 1: But for real, this is un Ironically this strikes me

683
00:33:40,840 --> 00:33:44,440
Speaker 1: as another example for open source, which is that if

684
00:33:44,520 --> 00:33:47,160
Speaker 1: I'm going to form a I don't see it. I'm

685
00:33:47,160 --> 00:33:49,320
Speaker 1: forty five, I'm too old for that. But if someone

686
00:33:49,440 --> 00:33:51,800
Speaker 1: is going to form like some sort of friendship with

687
00:33:51,840 --> 00:33:54,480
Speaker 1: an AI model, I don't want it to be at

688
00:33:54,480 --> 00:33:57,400
Speaker 1: the whim of Sam Altman deciding it was like, oh

689
00:33:57,440 --> 00:34:00,160
Speaker 1: there's an upgrade. I would like to be friends, so

690
00:34:00,200 --> 00:34:02,120
Speaker 1: weird to be friends with the model that I know

691
00:34:02,200 --> 00:34:06,480
Speaker 1: that I can run in perpetuity and it will never change.

692
00:34:06,720 --> 00:34:09,359
Speaker 4: Yeah. I think that's definitely a good argument for why

693
00:34:09,400 --> 00:34:12,440
Speaker 4: open source is important, And if you ever fall in

694
00:34:12,480 --> 00:34:14,200
Speaker 4: love with a model, you should fall in love with

695
00:34:14,239 --> 00:34:14,880
Speaker 4: an openness.

696
00:34:16,280 --> 00:34:18,320
Speaker 2: That's good life advice, practical life.

697
00:34:18,120 --> 00:34:19,480
Speaker 3: Advice, really good life advice.

698
00:34:19,680 --> 00:34:22,399
Speaker 2: Well, speaking of open source, you know, I know programmers

699
00:34:22,719 --> 00:34:26,080
Speaker 2: tend to like open source for obvious reasons, but are

700
00:34:26,160 --> 00:34:30,680
Speaker 2: there any downsides to open source for AI specifically?

701
00:34:31,080 --> 00:34:33,080
Speaker 4: I think if you're running a company, there are a

702
00:34:33,120 --> 00:34:36,000
Speaker 4: lot of downsides potentially to open source. If you have

703
00:34:36,200 --> 00:34:41,120
Speaker 4: some brand new, fancy way of doing computation inside the

704
00:34:41,160 --> 00:34:43,319
Speaker 4: model that's actually better, you might want to keep that

705
00:34:43,360 --> 00:34:45,680
Speaker 4: information to yourself. And when you release the model, to

706
00:34:45,719 --> 00:34:47,920
Speaker 4: make it runnable, you have to release all the code

707
00:34:47,960 --> 00:34:50,600
Speaker 4: to run the model, which might contain like your secrets,

708
00:34:50,640 --> 00:34:52,520
Speaker 4: and so I think that's why people are hesitant to

709
00:34:52,520 --> 00:34:55,520
Speaker 4: do it. The other reason is because when you release

710
00:34:55,600 --> 00:34:59,480
Speaker 4: the model, it actually contains quite a lot of residual

711
00:34:59,480 --> 00:35:02,319
Speaker 4: information about how you actually trained it, Like you might

712
00:35:02,360 --> 00:35:04,400
Speaker 4: be able to infer what the data set was and

713
00:35:04,440 --> 00:35:08,080
Speaker 4: what the training process was, or even reconstruct the entire

714
00:35:08,200 --> 00:35:10,760
Speaker 4: training data set given just the weights of the model.

715
00:35:11,080 --> 00:35:15,160
Speaker 4: And so if you're worried about people finding out that

716
00:35:15,200 --> 00:35:17,400
Speaker 4: a certain thing was in your training data, you probably

717
00:35:17,440 --> 00:35:19,040
Speaker 4: can't release that model open source.

718
00:35:19,760 --> 00:35:23,520
Speaker 2: That reminds me how much of an AI researcher's day

719
00:35:23,520 --> 00:35:27,040
Speaker 2: to day life is just like looking at other model,

720
00:35:27,120 --> 00:35:30,359
Speaker 2: other people's models, and trying to, like I guess, pull

721
00:35:30,400 --> 00:35:32,799
Speaker 2: them apart and figure out how they were made and

722
00:35:32,800 --> 00:35:33,920
Speaker 2: sort of work backwards.

723
00:35:34,960 --> 00:35:38,080
Speaker 4: That definitely happens from time to time. I think usually

724
00:35:38,120 --> 00:35:41,000
Speaker 4: the scientific process is something like you start with other

725
00:35:41,040 --> 00:35:44,360
Speaker 4: people's models, and you run them and you see what happens,

726
00:35:44,400 --> 00:35:46,960
Speaker 4: and then you decide on some part of that process

727
00:35:47,000 --> 00:35:49,680
Speaker 4: that you think could be improved or could be explored further,

728
00:35:50,040 --> 00:35:51,960
Speaker 4: and you make some tiny changes to it, and then

729
00:35:52,000 --> 00:35:54,640
Speaker 4: you run it again and you compare like numbers, or

730
00:35:54,680 --> 00:35:57,320
Speaker 4: you make graphs of what happened before and what happens after.

731
00:35:57,680 --> 00:35:59,960
Speaker 4: So actually quite a bit of it, like, for example,

732
00:36:00,040 --> 00:36:02,520
Speaker 4: pull the GPT two model from open Ai, which was

733
00:36:03,280 --> 00:36:06,840
Speaker 4: twenty nineteen or something, their first kind of really larger

734
00:36:06,920 --> 00:36:10,279
Speaker 4: scale chatbot. Like I've spent hundreds of hours kind of

735
00:36:10,320 --> 00:36:12,520
Speaker 4: like playing with that code and talking to the model

736
00:36:12,560 --> 00:36:15,400
Speaker 4: and stuff like that. So thank goodness for open source.

737
00:36:15,440 --> 00:36:18,600
Speaker 1: For that reason, I joked in the beginning about you

738
00:36:18,680 --> 00:36:21,879
Speaker 1: having one hundred million dollar salary, but for real, as

739
00:36:21,920 --> 00:36:24,960
Speaker 1: you think about your career, and I hope you do

740
00:36:24,960 --> 00:36:28,040
Speaker 1: get a hundred million dollar salary, but as you think

741
00:36:28,080 --> 00:36:30,799
Speaker 1: about your career, what excites you?

742
00:36:30,880 --> 00:36:32,240
Speaker 3: And how much is it money?

743
00:36:32,440 --> 00:36:34,520
Speaker 1: But the reason I think about this is like they're

744
00:36:34,600 --> 00:36:38,000
Speaker 1: huge checks out there, but maybe some things are more.

745
00:36:38,040 --> 00:36:42,160
Speaker 1: Maybe achieving AGI is more excited than making an ad

746
00:36:42,200 --> 00:36:46,040
Speaker 1: network more efficient. Maybe something there's something more exciting than

747
00:36:46,640 --> 00:36:50,360
Speaker 1: shaving off a billionth of a second in terms of

748
00:36:50,360 --> 00:36:53,320
Speaker 1: a trade execution, all these things like how much is

749
00:36:53,360 --> 00:36:57,399
Speaker 1: it about exploring the frontiers of science, the new space race,

750
00:36:57,480 --> 00:36:59,840
Speaker 1: landing on the Moon versus the paycheck?

751
00:37:00,160 --> 00:37:03,000
Speaker 4: It's all about the paycheck. I'm just kidding. No, no,

752
00:37:03,160 --> 00:37:05,680
Speaker 4: not at all. Yeah, it's funny you ask. So this

753
00:37:05,719 --> 00:37:08,160
Speaker 4: hasn't happened to me, But just in the past two

754
00:37:08,200 --> 00:37:10,759
Speaker 4: weeks or so, a good friend of mine has been

755
00:37:10,800 --> 00:37:13,520
Speaker 4: dealing with this problem because she got an offer on

756
00:37:13,600 --> 00:37:15,839
Speaker 4: the order of like tens of millions of dollars per

757
00:37:15,920 --> 00:37:20,680
Speaker 4: year from a big AI company and she wasn't sure

758
00:37:20,680 --> 00:37:23,279
Speaker 4: if she wanted to work there, and I think originally

759
00:37:23,320 --> 00:37:26,799
Speaker 4: she said no, and then they doubled her offer, and

760
00:37:26,840 --> 00:37:28,920
Speaker 4: then like it's the exact same amount of cash, but

761
00:37:28,920 --> 00:37:31,120
Speaker 4: twice as much per year for certain number of years.

762
00:37:31,640 --> 00:37:34,600
Speaker 4: And you know, we were talking amongst ourselves like what

763
00:37:34,600 --> 00:37:37,120
Speaker 4: does this even mean at this point, Like you're, you know,

764
00:37:37,160 --> 00:37:40,279
Speaker 4: a twenty eight year old computer scientist that's been coming

765
00:37:40,320 --> 00:37:42,239
Speaker 4: from a PhD. So you make more on the order

766
00:37:42,280 --> 00:37:45,440
Speaker 4: of tens of thousands of dollars per year. I honestly

767
00:37:45,520 --> 00:37:49,120
Speaker 4: think personally, the marginal difference between having like ten and

768
00:37:49,160 --> 00:37:51,640
Speaker 4: twenty million dollars is like very low, Like I don't

769
00:37:51,640 --> 00:37:53,440
Speaker 4: even know what I would do with this.

770
00:37:53,520 --> 00:37:57,479
Speaker 1: Is this is my experience for me making ten million

771
00:37:57,520 --> 00:37:58,920
Speaker 1: twenty mine has basically.

772
00:37:58,520 --> 00:37:59,160
Speaker 3: Been the same to me.

773
00:37:59,360 --> 00:38:05,120
Speaker 4: Yeah, congratulations, but so yeah, I think there's more of

774
00:38:05,160 --> 00:38:08,040
Speaker 4: a desire to like be there the next time something

775
00:38:08,080 --> 00:38:12,040
Speaker 4: really interesting happens, and that kind of supersedes the money.

776
00:38:12,120 --> 00:38:14,120
Speaker 4: Like any of these places will pay you what's like

777
00:38:14,160 --> 00:38:16,319
Speaker 4: a really good salary to live on, and so it's

778
00:38:16,360 --> 00:38:19,399
Speaker 4: actually not a big consideration. It only becomes complicated when

779
00:38:19,400 --> 00:38:21,920
Speaker 4: you have like one option that's going to pay you

780
00:38:21,960 --> 00:38:24,319
Speaker 4: like forty times more than the other option, and then

781
00:38:24,760 --> 00:38:25,799
Speaker 4: things get confusing.

782
00:38:26,320 --> 00:38:28,879
Speaker 2: No, this isn't this should actually I was just thinking

783
00:38:28,920 --> 00:38:29,919
Speaker 2: about making twenty million.

784
00:38:30,000 --> 00:38:30,439
Speaker 3: No, I think.

785
00:38:32,080 --> 00:38:33,920
Speaker 1: Because I think about, Okay, what if you have this

786
00:38:33,960 --> 00:38:37,120
Speaker 1: great salary and you're like can live very easily in

787
00:38:37,120 --> 00:38:40,160
Speaker 1: New York City and have a really great life, or

788
00:38:40,600 --> 00:38:43,120
Speaker 1: you could make ten times that, which is a stupid

789
00:38:43,400 --> 00:38:45,800
Speaker 1: insane salary, right, but you don't write like your job.

790
00:38:46,040 --> 00:38:46,680
Speaker 3: But it's so.

791
00:38:46,800 --> 00:38:51,600
Speaker 1: Much money that strikes me is like not a trivial life.

792
00:38:52,400 --> 00:38:54,239
Speaker 1: You only live one time. There's like a different so

793
00:38:54,239 --> 00:38:55,480
Speaker 1: it could be a difficult question.

794
00:38:55,800 --> 00:38:58,400
Speaker 4: Yeah, yeah, but you can remind yourself that, like the

795
00:38:58,520 --> 00:39:02,120
Speaker 4: job you take once isn't the job that defines you forever.

796
00:39:02,280 --> 00:39:04,040
Speaker 4: Maybe maybe the right thing to do is to take

797
00:39:04,080 --> 00:39:05,520
Speaker 4: it for a few years but not the whole time,

798
00:39:05,520 --> 00:39:06,279
Speaker 4: and then go do something.

799
00:39:06,320 --> 00:39:09,680
Speaker 1: Everyone says they're going to do that and then.

800
00:39:09,560 --> 00:39:14,080
Speaker 2: They get locked in. Speaking of insanely large salaries, we

801
00:39:14,160 --> 00:39:16,520
Speaker 2: know that people are earning these salaries because they're like

802
00:39:16,719 --> 00:39:23,320
Speaker 2: star AI researchers. How much does personality play into where

803
00:39:23,360 --> 00:39:25,279
Speaker 2: you want to go work? Would you want to go

804
00:39:25,320 --> 00:39:30,399
Speaker 2: work somewhere specifically because there's an absolutely amazing researcher, or

805
00:39:30,520 --> 00:39:32,920
Speaker 2: does it tend to be again more about the paycheck,

806
00:39:32,960 --> 00:39:35,160
Speaker 2: maybe more about the data that's available to you, or

807
00:39:35,200 --> 00:39:37,440
Speaker 2: maybe more about the specific project that you're going to

808
00:39:37,480 --> 00:39:38,000
Speaker 2: be working on.

809
00:39:38,560 --> 00:39:42,520
Speaker 4: Yeah, I think different people assign different amounts of weight

810
00:39:42,600 --> 00:39:45,759
Speaker 4: to each of those things. In my experience, like most

811
00:39:45,760 --> 00:39:47,759
Speaker 4: of the people I know come from academia, which means

812
00:39:47,760 --> 00:39:49,920
Speaker 4: they already kind of gave up more of a salary

813
00:39:49,960 --> 00:39:52,800
Speaker 4: to do study things more deeply for several years. So

814
00:39:52,840 --> 00:39:55,239
Speaker 4: I think people that I know are more biased against money.

815
00:39:55,280 --> 00:39:58,040
Speaker 4: But like people do care about that. But I think

816
00:39:58,080 --> 00:40:00,640
Speaker 4: that the ego thing really matters. Some people want to

817
00:40:00,640 --> 00:40:02,640
Speaker 4: feel like they're very important and they're working on a

818
00:40:02,640 --> 00:40:05,840
Speaker 4: problem that matters. One way some companies are able to

819
00:40:05,840 --> 00:40:08,560
Speaker 4: pull researchers away from other companies is by saying, we'll

820
00:40:08,760 --> 00:40:11,200
Speaker 4: sign you more importance in your role and we'll give.

821
00:40:11,040 --> 00:40:12,440
Speaker 2: You we'll give you a really big title.

822
00:40:12,680 --> 00:40:16,400
Speaker 4: Yeah, exactly. Seriously, the title is like, Okay, maybe before

823
00:40:16,400 --> 00:40:17,960
Speaker 4: you were like a researcher or not. You get to

824
00:40:18,000 --> 00:40:19,840
Speaker 4: be like a head researcher. You get to have people

825
00:40:19,920 --> 00:40:22,000
Speaker 4: under you, or you're a chief scientist, and all these

826
00:40:22,000 --> 00:40:23,160
Speaker 4: things do matter to people.

827
00:40:23,719 --> 00:40:25,680
Speaker 3: It's a very good book about it.

828
00:40:26,000 --> 00:40:29,720
Speaker 1: Pursuing a mission in the realm of like a driven

829
00:40:29,840 --> 00:40:32,880
Speaker 1: visionary even when it's commercially.

830
00:40:32,600 --> 00:40:35,600
Speaker 2: Just say it, just say yeah, that's right. No.

831
00:40:35,719 --> 00:40:37,920
Speaker 1: I think about this all the time. Do you want

832
00:40:37,920 --> 00:40:39,120
Speaker 1: to work for Ilia or do you want to work

833
00:40:39,120 --> 00:40:40,919
Speaker 1: for Sam? And which one is the ahab and which

834
00:40:40,920 --> 00:40:44,160
Speaker 1: one is just trying to make an honest living selling ads.

835
00:40:44,239 --> 00:40:47,760
Speaker 1: I find this to be like a genuinely interesting, interesting

836
00:40:47,840 --> 00:40:50,680
Speaker 1: question for any individual to have to reckon with in

837
00:40:50,719 --> 00:40:51,200
Speaker 1: this career.

838
00:40:51,320 --> 00:40:52,960
Speaker 4: Oh. Absolutely, And sometimes it can be.

839
00:40:53,000 --> 00:40:55,719
Speaker 1: Very difficult to tell Jack Morris, thank you so much

840
00:40:55,719 --> 00:40:58,399
Speaker 1: for coming on. Please pursue a career that will allow

841
00:40:58,440 --> 00:40:59,840
Speaker 1: you to come back on a log.

842
00:41:00,200 --> 00:41:03,960
Speaker 2: Or insert the odd lots close when you're negotiating your

843
00:41:03,960 --> 00:41:06,239
Speaker 2: one hundred million dollar salary, or.

844
00:41:06,280 --> 00:41:08,560
Speaker 1: Take the fifty so you know what, fifty million, but

845
00:41:08,640 --> 00:41:10,880
Speaker 1: let me I don't need one hundred million, fifty million.

846
00:41:10,680 --> 00:41:11,359
Speaker 3: But keep the album.

847
00:41:11,880 --> 00:41:13,200
Speaker 4: Yeah, that would be fine with me.

848
00:41:13,360 --> 00:41:14,799
Speaker 3: All right, great, Well, thank you so much.

849
00:41:15,200 --> 00:41:16,680
Speaker 2: Yeah, thanks, thank you so much.

850
00:41:16,719 --> 00:41:17,239
Speaker 4: That was great.

851
00:41:29,680 --> 00:41:31,960
Speaker 1: Appreciate I think about that sometimes, like what if you

852
00:41:32,040 --> 00:41:35,479
Speaker 1: got like an insane salary like that, you just could

853
00:41:35,560 --> 00:41:37,480
Speaker 1: you would be insane to say no to But like

854
00:41:37,640 --> 00:41:39,000
Speaker 1: I don't know, that's I mean.

855
00:41:39,080 --> 00:41:41,960
Speaker 3: It's not our problem, but like, wouldn't it be fun?

856
00:41:42,200 --> 00:41:44,560
Speaker 1: You know? It's like, oh, but you're gonna be working

857
00:41:44,560 --> 00:41:47,600
Speaker 1: on ad optimization or whatever and you're not going to

858
00:41:47,680 --> 00:41:49,280
Speaker 1: be there when they land.

859
00:41:49,080 --> 00:41:51,640
Speaker 3: On the moon. But you got paid ten times.

860
00:41:51,360 --> 00:41:53,520
Speaker 1: More than the people at the Bay station working on

861
00:41:53,640 --> 00:41:55,560
Speaker 1: landing on the moon. That strins me as a kind

862
00:41:55,560 --> 00:41:56,520
Speaker 1: of a tough life choice.

863
00:41:56,520 --> 00:41:58,400
Speaker 2: I think you're using up a lot of brain power

864
00:41:58,400 --> 00:42:00,840
Speaker 2: and energy on a problem which will Jem said is

865
00:42:00,880 --> 00:42:01,239
Speaker 2: not you.

866
00:42:02,120 --> 00:42:03,000
Speaker 3: That's exactly right.

867
00:42:03,120 --> 00:42:06,600
Speaker 2: No, that conversation was really fun. Nice to talk to

868
00:42:06,640 --> 00:42:10,239
Speaker 2: an actual researcher just doing stuff in the space. One

869
00:42:10,239 --> 00:42:12,400
Speaker 2: thing I thought was very interesting was this idea that

870
00:42:12,480 --> 00:42:16,319
Speaker 2: everyone gets excited about a specific improvement in AI, and

871
00:42:16,360 --> 00:42:20,640
Speaker 2: then it seems like that particular one doesn't materialize and

872
00:42:20,760 --> 00:42:24,000
Speaker 2: instead something else emerges, as like the big breakthrough. So

873
00:42:24,360 --> 00:42:27,239
Speaker 2: instead of agents, we have math.

874
00:42:27,320 --> 00:42:29,640
Speaker 1: And math which none of us will ever. I would

875
00:42:29,840 --> 00:42:32,839
Speaker 1: really like for an agent to do something simple. I'm

876
00:42:32,880 --> 00:42:35,120
Speaker 1: going to a city book on the trip or whatever.

877
00:42:35,239 --> 00:42:37,000
Speaker 1: Or change my flight. Oh my god, I tried to.

878
00:42:37,160 --> 00:42:38,600
Speaker 2: That would be amazing.

879
00:42:38,080 --> 00:42:42,080
Speaker 1: Recently change my flight. Here's my information. I don't I

880
00:42:42,120 --> 00:42:45,239
Speaker 1: would like that. I do not need the math olympiad.

881
00:42:45,560 --> 00:42:46,440
Speaker 1: I am very impressed.

882
00:42:46,440 --> 00:42:47,160
Speaker 3: I don't need it.

883
00:42:47,640 --> 00:42:51,480
Speaker 2: Also, I am now very very intrigued by reinforced learning

884
00:42:51,880 --> 00:42:55,880
Speaker 2: and how you actually reward the computers for doing good stuff.

885
00:42:55,920 --> 00:42:58,440
Speaker 2: I feel like, actually that would be a really interesting

886
00:42:59,040 --> 00:43:03,200
Speaker 2: area to mine. Which is motivating motivating the models to

887
00:43:03,320 --> 00:43:03,880
Speaker 2: do better?

888
00:43:04,360 --> 00:43:06,680
Speaker 1: Yeah, I've thought about that, like in chess, like how

889
00:43:06,760 --> 00:43:08,359
Speaker 1: do how do the computers know.

890
00:43:08,360 --> 00:43:08,960
Speaker 3: They want to win?

891
00:43:09,160 --> 00:43:09,359
Speaker 4: Yeah?

892
00:43:09,400 --> 00:43:10,560
Speaker 3: You know, like why do they care?

893
00:43:10,760 --> 00:43:12,759
Speaker 2: You know, all they're saying anyway, why are they here?

894
00:43:12,920 --> 00:43:13,759
Speaker 2: Why are we here?

895
00:43:14,480 --> 00:43:16,240
Speaker 3: That's the thing with AI conversations.

896
00:43:16,400 --> 00:43:18,120
Speaker 2: That's existential fact, something.

897
00:43:17,880 --> 00:43:19,960
Speaker 1: We didn't talk about, which I am interested. No one

898
00:43:20,000 --> 00:43:22,600
Speaker 1: really talks about AI safety anymore. If you notice, like

899
00:43:22,640 --> 00:43:25,319
Speaker 1: they like very little, like for better or worse. You

900
00:43:25,360 --> 00:43:28,080
Speaker 1: don't hear people just all money and they don't really

901
00:43:28,120 --> 00:43:30,719
Speaker 1: talk about what the AI kill us all one day.

902
00:43:30,840 --> 00:43:32,879
Speaker 3: But one thing I did wonder about.

903
00:43:32,960 --> 00:43:35,440
Speaker 1: So when Deep Seat came out, one of its breakthroughs

904
00:43:35,520 --> 00:43:37,640
Speaker 1: was it showed the whole chain of thought, right, you

905
00:43:37,680 --> 00:43:40,200
Speaker 1: could see that, which prior to that open AI or

906
00:43:40,239 --> 00:43:42,319
Speaker 1: chatchybt's chain of thought model didn't show you.

907
00:43:42,280 --> 00:43:42,680
Speaker 4: That, right.

908
00:43:42,920 --> 00:43:44,920
Speaker 1: And it does strike me that if there are certain

909
00:43:45,000 --> 00:43:48,520
Speaker 1: things that are for safety reasons or whatever held back

910
00:43:48,600 --> 00:43:50,480
Speaker 1: or they don't want to do this, the nature of

911
00:43:50,480 --> 00:43:53,840
Speaker 1: competition means all the guardrails are coming off of Actually,

912
00:43:53,920 --> 00:43:56,279
Speaker 1: like that's if there's some guardrail you you have on

913
00:43:56,680 --> 00:43:59,480
Speaker 1: someone's going to open source whatever it is and they're

914
00:43:59,520 --> 00:44:00,520
Speaker 1: going to all give it up.

915
00:44:00,640 --> 00:44:04,080
Speaker 2: Yeah, both on the guardrails and on the data use ys.

916
00:44:04,600 --> 00:44:06,080
Speaker 2: All right, well shall we leave it there.

917
00:44:06,160 --> 00:44:06,879
Speaker 3: Let's leave it there.

918
00:44:07,040 --> 00:44:09,560
Speaker 2: This has been another episode of the aud Loots podcast.

919
00:44:09,600 --> 00:44:12,480
Speaker 2: I'm Tracy Alloway. You can follow me at Tracy Alloway.

920
00:44:12,640 --> 00:44:15,240
Speaker 1: And I'm Jill Wisenthal. You can follow me at the Stalwart.

921
00:44:15,360 --> 00:44:19,680
Speaker 1: Follow our guest Jack Morris, He's at j xmnop. Follow

922
00:44:19,760 --> 00:44:23,000
Speaker 1: our producers Kerman Rodriguez at Kerman armand dash O Bennett

923
00:44:23,000 --> 00:44:26,480
Speaker 1: at Dashbod and kil Brooks at Kilbrooks. More odd Loss content,

924
00:44:26,480 --> 00:44:28,520
Speaker 1: go to Bloomberg dot com slash od Lots with the

925
00:44:28,600 --> 00:44:31,319
Speaker 1: daily newsletter and all of our episodes, and you can

926
00:44:31,400 --> 00:44:33,359
Speaker 1: chat about all of these topics twenty four to seven

927
00:44:33,480 --> 00:44:36,520
Speaker 1: in our discord Discord dot gg slash.

928
00:44:36,200 --> 00:44:38,799
Speaker 2: Odd Lots And if you enjoy odd Lots, if you

929
00:44:38,960 --> 00:44:41,400
Speaker 2: like it when we talk about twenty million dollars salaries

930
00:44:41,440 --> 00:44:43,840
Speaker 2: that will never be ours, then please leave us a

931
00:44:43,920 --> 00:44:47,600
Speaker 2: positive review on your favorite podcast platform. And remember, if

932
00:44:47,640 --> 00:44:50,040
Speaker 2: you are a Bloomberg subscriber, you can listen to all

933
00:44:50,080 --> 00:44:52,960
Speaker 2: of our episodes absolutely ad free. All you need to

934
00:44:53,000 --> 00:44:55,520
Speaker 2: do is find the Bloomberg channel on Apple Podcasts and

935
00:44:55,600 --> 00:45:14,840
Speaker 2: follow the instructions there. Thanks for listening it