1
00:00:07,600 --> 00:00:11,039
Speaker 1: As artificial intelligence surge is ahead, Can we trust open

2
00:00:11,080 --> 00:00:15,120
Speaker 1: ai and its boss Sam Altman. I'm a Zimazar and

3
00:00:15,160 --> 00:00:18,560
Speaker 1: in this episode I'm talking to Sam. Welcome to the

4
00:00:18,640 --> 00:00:27,880
Speaker 1: Exponentially podcast Now. Sam is a rock and roll star

5
00:00:27,960 --> 00:00:31,280
Speaker 1: of AI. He runs open ai and they built chat GPT.

6
00:00:31,880 --> 00:00:35,320
Speaker 1: He's raised billions of dollars from Microsoft. His early backers

7
00:00:35,320 --> 00:00:39,000
Speaker 1: include Elon Musk and Reid Hoffman. It's been quite the journey,

8
00:00:39,159 --> 00:00:41,639
Speaker 1: but the more we know about AI, the more questions

9
00:00:41,640 --> 00:00:44,720
Speaker 1: of technology raises. I caught up with Sam at the

10
00:00:44,720 --> 00:00:47,400
Speaker 1: beginning of a world tour that would cover twenty countries

11
00:00:47,440 --> 00:00:49,839
Speaker 1: in just thirty days. We spoke to each other at

12
00:00:49,960 --> 00:00:52,760
Speaker 1: University College London in front of a live audience of

13
00:00:52,840 --> 00:00:58,520
Speaker 1: nearly one thousand people. You must be rushed off your feet.

14
00:00:58,560 --> 00:01:02,040
Speaker 1: You're in the middle of a enormous world tool. How

15
00:01:02,040 --> 00:01:02,480
Speaker 1: are you doing?

16
00:01:02,800 --> 00:01:04,959
Speaker 2: It's been super great and I didn't I wasn't sure

17
00:01:05,000 --> 00:01:06,319
Speaker 2: how much fun I was going to have. I really

18
00:01:06,319 --> 00:01:07,800
Speaker 2: wanted to do it because I think this sort of

19
00:01:07,800 --> 00:01:12,560
Speaker 2: San Francisco echo chamber is not a great thing and

20
00:01:13,040 --> 00:01:16,040
Speaker 2: I have never found a replacement for getting my airplanes

21
00:01:16,040 --> 00:01:18,880
Speaker 2: and meeting people and the feedback We've gotten about what

22
00:01:19,120 --> 00:01:20,959
Speaker 2: people want us to do, how they're thinking about AI,

23
00:01:21,000 --> 00:01:23,000
Speaker 2: what they're excited about, what they're nervous about. It's been

24
00:01:23,040 --> 00:01:24,959
Speaker 2: even more useful than I expected, and I've had a

25
00:01:24,959 --> 00:01:25,360
Speaker 2: great time.

26
00:01:25,480 --> 00:01:28,200
Speaker 1: I've seen you taking notes by hand in a notebook

27
00:01:28,200 --> 00:01:29,840
Speaker 1: because you hear it from people as well.

28
00:01:30,360 --> 00:01:32,080
Speaker 2: BE still taking my notes by hand. I do my

29
00:01:32,160 --> 00:01:32,840
Speaker 2: to do this by hand.

30
00:01:33,200 --> 00:01:35,319
Speaker 1: There's going to be a lesson in there for many

31
00:01:35,319 --> 00:01:35,679
Speaker 1: of us.

32
00:01:36,240 --> 00:01:37,560
Speaker 2: I think there's probably no lesson.

33
00:01:40,040 --> 00:01:43,560
Speaker 1: When you started open ai in twenty fifteen, did you

34
00:01:43,640 --> 00:01:46,959
Speaker 1: imagine that within just a few years you would almost

35
00:01:46,959 --> 00:01:50,520
Speaker 1: find necessity you have to get on a plane fly

36
00:01:50,600 --> 00:01:54,120
Speaker 1: around the world to listen to people from every continent.

37
00:01:54,360 --> 00:01:56,600
Speaker 2: I've always tried to do this when I was running

38
00:01:56,600 --> 00:01:58,320
Speaker 2: my combinator, I would try to fly around and meet

39
00:01:58,320 --> 00:02:00,600
Speaker 2: people a lot. I think it's really important. I also

40
00:02:00,760 --> 00:02:04,640
Speaker 2: like I think most of my important insights come while

41
00:02:04,680 --> 00:02:07,720
Speaker 2: traveling in some way, and you get very different perspectives.

42
00:02:07,840 --> 00:02:10,120
Speaker 2: And certainly when we started opening I thought it probably

43
00:02:10,160 --> 00:02:12,520
Speaker 2: wasn't going to work. But if it did work, then

44
00:02:12,560 --> 00:02:15,400
Speaker 2: I thought like it would be an impactful technology and

45
00:02:15,800 --> 00:02:18,359
Speaker 2: getting input from the world would be a really important thing.

46
00:02:18,360 --> 00:02:18,600
Speaker 2: To do.

47
00:02:18,760 --> 00:02:22,639
Speaker 1: You're in an unprecedented position right now. In many cases,

48
00:02:23,040 --> 00:02:25,760
Speaker 1: in the Silicon Valley model that the founder of a

49
00:02:25,800 --> 00:02:28,080
Speaker 1: business like this owns a lot of equity, takes a

50
00:02:28,120 --> 00:02:31,120
Speaker 1: salary as well, has a financial upsided. You don't have

51
00:02:31,160 --> 00:02:33,800
Speaker 1: any of that. You just draw enough for your health insurance.

52
00:02:34,240 --> 00:02:38,760
Speaker 1: So what is the inner drive for you given the challenge,

53
00:02:38,800 --> 00:02:41,880
Speaker 1: given that the demands are on your time, your energy.

54
00:02:42,160 --> 00:02:44,480
Speaker 2: I can't think of any more exciting thing. I mean,

55
00:02:44,480 --> 00:02:46,480
Speaker 2: I hope this is like self explanatory, but I can't

56
00:02:46,520 --> 00:02:48,760
Speaker 2: think of anything more exciting to work on. I feel

57
00:02:48,919 --> 00:02:51,600
Speaker 2: like extremely privileged to be at this moment in history,

58
00:02:52,000 --> 00:02:54,120
Speaker 2: and more than that, like working with this particular team.

59
00:02:54,280 --> 00:02:56,560
Speaker 2: I don't know, like how I would possibly rather spend

60
00:02:56,600 --> 00:02:58,600
Speaker 2: the days. I was very fortunate. I made a bunch

61
00:02:58,600 --> 00:03:00,639
Speaker 2: of money very early in my career, so I don't

62
00:03:00,639 --> 00:03:02,880
Speaker 2: think it's some like great noble thing.

63
00:03:03,200 --> 00:03:05,400
Speaker 1: I was thinking about what does it mean to face

64
00:03:05,440 --> 00:03:08,000
Speaker 1: these types of exciting challenges? And some of them are

65
00:03:08,280 --> 00:03:11,800
Speaker 1: January sort of intellectually and exciting. Some are really hard,

66
00:03:11,960 --> 00:03:16,480
Speaker 1: thorny problems. And you are an unprecedented position. Do you

67
00:03:16,520 --> 00:03:18,640
Speaker 1: have mentors? Are there people you're learning from?

68
00:03:19,120 --> 00:03:22,440
Speaker 2: I feel like super fortunate to have had great mentors.

69
00:03:22,880 --> 00:03:25,359
Speaker 2: But I also think it's important not to try to

70
00:03:25,960 --> 00:03:27,680
Speaker 2: learn too much from other people and sort of do

71
00:03:27,760 --> 00:03:30,200
Speaker 2: things your own way. One of the magical things about

72
00:03:30,440 --> 00:03:35,120
Speaker 2: Silicon Valley is how much people care about mentorship and teaching,

73
00:03:35,240 --> 00:03:37,240
Speaker 2: and I've gotten way more of my first shore there.

74
00:03:37,320 --> 00:03:39,440
Speaker 1: If we pick out one or two lessons from the

75
00:03:39,480 --> 00:03:41,320
Speaker 1: great mentals you've had, what would they be.

76
00:03:41,960 --> 00:03:44,760
Speaker 2: Paul Graham, who ran my commentrator before it started, ran

77
00:03:44,800 --> 00:03:47,320
Speaker 2: it before I did, I think did more to teach

78
00:03:47,360 --> 00:03:51,560
Speaker 2: people about how startups work, very heavily from what it

79
00:03:51,600 --> 00:03:54,720
Speaker 2: takes to make a high functioning org and then traps

80
00:03:54,720 --> 00:03:57,560
Speaker 2: you want to avoid. There. Certainly learning from Elon about

81
00:03:57,640 --> 00:04:00,880
Speaker 2: what is possible to do and that you don't need

82
00:04:00,920 --> 00:04:03,240
Speaker 2: to accept that part technology is not something to ignore.

83
00:04:03,560 --> 00:04:04,720
Speaker 2: That's been super valuable.

84
00:04:05,520 --> 00:04:08,080
Speaker 1: And I can see both of those lessons in open

85
00:04:08,120 --> 00:04:11,160
Speaker 1: AI and in what you have shipped and have been

86
00:04:11,200 --> 00:04:14,800
Speaker 1: shipping for a few years. When we last spoke a

87
00:04:14,840 --> 00:04:17,600
Speaker 1: couple of years ago, you were talking about these large

88
00:04:17,640 --> 00:04:20,720
Speaker 1: language models, and we're currently on GPT four, but back

89
00:04:20,720 --> 00:04:23,720
Speaker 1: then the state of the art was GPT three, and

90
00:04:23,760 --> 00:04:26,320
Speaker 1: you said to me that the gap between gpt E

91
00:04:26,520 --> 00:04:30,039
Speaker 1: two and GBT three was just a baby step on

92
00:04:30,080 --> 00:04:33,039
Speaker 1: the continue. It's just a little baby step. When you

93
00:04:33,080 --> 00:04:35,880
Speaker 1: now look at GPT four, would you say that's another

94
00:04:35,920 --> 00:04:36,560
Speaker 1: baby step.

95
00:04:37,320 --> 00:04:39,560
Speaker 2: It'll look like that in retrospect. I think.

96
00:04:42,240 --> 00:04:45,280
Speaker 1: That's a nature of the exponential in retrospect. But when

97
00:04:45,320 --> 00:04:46,440
Speaker 1: we're living through it, it.

98
00:04:46,400 --> 00:04:48,880
Speaker 2: Felt like a big jump for a little while, and

99
00:04:48,920 --> 00:04:51,080
Speaker 2: now people are it's very much like, what have you've

100
00:04:51,080 --> 00:04:54,800
Speaker 2: done for me lately? Where's GPT five? And that's fine.

101
00:04:54,839 --> 00:04:56,480
Speaker 2: That's the way of the world. That's how it's supposed

102
00:04:56,520 --> 00:04:59,280
Speaker 2: to go. People get used to anything. We established new

103
00:04:59,279 --> 00:05:00,000
Speaker 2: baselines very quick.

104
00:05:00,800 --> 00:05:03,919
Speaker 1: I'm curious what were the insights that you gained in

105
00:05:04,040 --> 00:05:07,640
Speaker 1: developing GPT for and in the months following its release

106
00:05:07,680 --> 00:05:10,679
Speaker 1: that were different to the ones from the previous models

107
00:05:10,680 --> 00:05:11,359
Speaker 1: that you released.

108
00:05:11,720 --> 00:05:15,200
Speaker 2: We finished training GPT for like eight ish months before

109
00:05:15,200 --> 00:05:17,600
Speaker 2: you released it, I think, and that was by far

110
00:05:17,680 --> 00:05:19,880
Speaker 2: the longest we've ever spent on a model pre release.

111
00:05:20,440 --> 00:05:22,080
Speaker 2: One of the things that we had learned with GPD

112
00:05:22,160 --> 00:05:24,520
Speaker 2: three was all of the ways these things break down

113
00:05:24,600 --> 00:05:26,560
Speaker 2: once you put them out in the wild. We think

114
00:05:26,560 --> 00:05:30,200
Speaker 2: it is really important to deploy models incrementally, to give

115
00:05:30,240 --> 00:05:33,280
Speaker 2: the world time to adapt, understand what we think is

116
00:05:33,320 --> 00:05:35,719
Speaker 2: going to happen, what might happen, to give people time

117
00:05:35,720 --> 00:05:38,480
Speaker 2: to figure out what the risks are, what the benefits are,

118
00:05:38,680 --> 00:05:40,480
Speaker 2: what the rules should be. We don't want to put

119
00:05:40,480 --> 00:05:42,080
Speaker 2: out a model that we know has a bunch of problems.

120
00:05:42,120 --> 00:05:45,520
Speaker 2: So we spent more time applying the lessons from the

121
00:05:45,560 --> 00:05:48,400
Speaker 2: earlier versions of GPT three to this one, and it's

122
00:05:48,440 --> 00:05:51,479
Speaker 2: been nice to see it's behaving as advertised most of

123
00:05:51,520 --> 00:05:53,200
Speaker 2: the time, much more of a time than before. So

124
00:05:53,240 --> 00:05:54,719
Speaker 2: that was a lesson, which is that if we really

125
00:05:54,760 --> 00:05:58,000
Speaker 2: spent a lot of time on alignment, auditing, testing our

126
00:05:58,040 --> 00:05:59,880
Speaker 2: whole safety system, we can make a lot of progress.

127
00:06:00,120 --> 00:06:04,400
Speaker 1: So you build this model, it's an incredibly complicated machine.

128
00:06:04,560 --> 00:06:07,760
Speaker 1: GPT three, the precursor, had one hundred and seventy five

129
00:06:07,960 --> 00:06:11,040
Speaker 1: billion parameters, which I think of as sliders on a

130
00:06:11,080 --> 00:06:14,800
Speaker 1: graphic equalizer. It's a lot of configuration, and GPT four

131
00:06:15,360 --> 00:06:19,000
Speaker 1: is larger still, although you haven't formally said how much larger.

132
00:06:19,360 --> 00:06:24,080
Speaker 1: How do you take that machine and get it to

133
00:06:24,240 --> 00:06:27,320
Speaker 1: do what we want it to do and not do

134
00:06:27,400 --> 00:06:30,039
Speaker 1: what we don't want to do. That's the alignment problem,

135
00:06:30,200 --> 00:06:31,960
Speaker 1: and that's where you've spent this eight months.

136
00:06:32,080 --> 00:06:35,279
Speaker 2: Yeah, so I want to be clear on this. Just

137
00:06:35,320 --> 00:06:38,520
Speaker 2: because we're able to align GPT four does not mean

138
00:06:38,560 --> 00:06:40,800
Speaker 2: we're like out of the woods, not even close, as

139
00:06:40,839 --> 00:06:42,960
Speaker 2: I hope is obvious. We have a huge amount of

140
00:06:42,960 --> 00:06:44,839
Speaker 2: work to do to figure out how to we're going

141
00:06:44,880 --> 00:06:48,599
Speaker 2: to align superintelligence and much more powerful systems than what

142
00:06:48,600 --> 00:06:51,280
Speaker 2: we have now. And I worry that when we say, hey,

143
00:06:51,279 --> 00:06:53,479
Speaker 2: we can align GPT four pretty well, people think we

144
00:06:53,520 --> 00:06:55,680
Speaker 2: think we've solved the problem. We don't. But it is

145
00:06:55,880 --> 00:06:58,599
Speaker 2: I think remarkable that we can take the base model

146
00:06:58,600 --> 00:07:01,039
Speaker 2: of GPT four, which if you use it, you'd be like,

147
00:07:01,160 --> 00:07:03,719
Speaker 2: this is not very impressive, or it's the least extremely

148
00:07:03,720 --> 00:07:06,800
Speaker 2: difficult to use, and with so little effort we can

149
00:07:06,920 --> 00:07:09,440
Speaker 2: do URLHF and get the model to be so usable

150
00:07:09,480 --> 00:07:10,240
Speaker 2: and so aligned.

151
00:07:10,320 --> 00:07:13,480
Speaker 1: And URLHF is reinforcement learning with human feedback, which I

152
00:07:13,480 --> 00:07:16,440
Speaker 1: think is the way that you get people to answer

153
00:07:16,520 --> 00:07:18,680
Speaker 1: questions from GPT four and tell it when it's been

154
00:07:18,880 --> 00:07:22,600
Speaker 1: good and when it's not met expectations, and.

155
00:07:22,640 --> 00:07:25,360
Speaker 2: It turns and it's very tiny amounts of feedback, and

156
00:07:25,440 --> 00:07:28,400
Speaker 2: it's very unsophisticated too. It's really just like thumbs up

157
00:07:28,440 --> 00:07:30,560
Speaker 2: thumbs down, and the fact that this works, I think

158
00:07:30,640 --> 00:07:31,640
Speaker 2: is quite remarkable.

159
00:07:31,760 --> 00:07:35,840
Speaker 1: You've said you're not training GPT five right now, and

160
00:07:35,920 --> 00:07:38,640
Speaker 1: I was curious about why that was. Was it that

161
00:07:39,080 --> 00:07:42,520
Speaker 1: there's not enough data. Was it work that there aren't

162
00:07:42,600 --> 00:07:45,200
Speaker 1: enough computer chips to train it on. Was it that

163
00:07:45,240 --> 00:07:48,320
Speaker 1: you saw things going on when you were making GPT

164
00:07:48,480 --> 00:07:51,000
Speaker 1: four happen that you thought you need to figure out

165
00:07:51,280 --> 00:07:53,440
Speaker 1: how to tackle these before we build the next.

166
00:07:53,440 --> 00:07:55,560
Speaker 2: These models are very difficult to build, Like the time

167
00:07:55,560 --> 00:07:58,760
Speaker 2: between GPT three and four was almost three years. It

168
00:07:58,840 --> 00:08:00,480
Speaker 2: just takes a while. There's a lot of research to

169
00:08:00,520 --> 00:08:02,960
Speaker 2: go do. There's also a lot of other stuff we

170
00:08:03,000 --> 00:08:05,240
Speaker 2: want to do with GPT four, not that it's done.

171
00:08:05,320 --> 00:08:07,640
Speaker 2: We want to study post training a lot, We want

172
00:08:07,680 --> 00:08:09,640
Speaker 2: to expand it in all sorts of ways. The fact

173
00:08:09,680 --> 00:08:11,760
Speaker 2: that they can ship an iPhone every year is incredible to me.

174
00:08:12,120 --> 00:08:13,480
Speaker 2: But we're just going to be on a longer than

175
00:08:13,480 --> 00:08:14,200
Speaker 2: one year cadence.

176
00:08:14,320 --> 00:08:16,200
Speaker 1: You said that there's more research to be done, and

177
00:08:16,240 --> 00:08:19,560
Speaker 1: there are a number of very well storied AI researchers

178
00:08:19,640 --> 00:08:24,920
Speaker 1: who have said that large language models are limited. They

179
00:08:24,920 --> 00:08:28,480
Speaker 1: will not get us to the next performance increase that

180
00:08:28,520 --> 00:08:32,959
Speaker 1: you can't build artificial general intelligence with llms. Do you

181
00:08:33,040 --> 00:08:33,600
Speaker 1: agree with that?

182
00:08:33,760 --> 00:08:36,440
Speaker 2: I mean, first of all, I think most of those

183
00:08:36,960 --> 00:08:40,720
Speaker 2: commentators have been horribly wrong about what lllms are going

184
00:08:40,760 --> 00:08:43,120
Speaker 2: to be able to do, and a lot of them

185
00:08:43,160 --> 00:08:45,680
Speaker 2: have now switched to saying, well, it's not that lllms

186
00:08:45,679 --> 00:08:47,199
Speaker 2: aren't going to work, it's that they work too well

187
00:08:47,240 --> 00:08:49,840
Speaker 2: and they're too dangerous and we've got to stop them.

188
00:08:50,200 --> 00:08:52,439
Speaker 2: Or others have just said, well, you know, it's all

189
00:08:52,480 --> 00:08:54,440
Speaker 2: still like a parlor trick and this is not any

190
00:08:54,440 --> 00:08:58,079
Speaker 2: real learning. Some of the more sophisticated ones say, Okay,

191
00:08:58,600 --> 00:09:00,760
Speaker 2: lllms work better than expect, but they're not going to

192
00:09:00,760 --> 00:09:02,840
Speaker 2: get all the way to AGI in the current paradigm,

193
00:09:02,920 --> 00:09:05,760
Speaker 2: and that we agree with. So I think we absolutely

194
00:09:05,760 --> 00:09:08,440
Speaker 2: should push as far as we can in the current paradigm,

195
00:09:08,559 --> 00:09:10,120
Speaker 2: but we're hard at work trying to figure out the

196
00:09:10,120 --> 00:09:13,560
Speaker 2: next paradigm. The thing I'm personally most excited about, maybe

197
00:09:13,640 --> 00:09:17,720
Speaker 2: of the whole AGI world, is that these models at

198
00:09:17,720 --> 00:09:19,720
Speaker 2: some point are going to help us discover new science

199
00:09:20,920 --> 00:09:23,880
Speaker 2: fast and in really meaningful ways. But I think the

200
00:09:23,920 --> 00:09:25,959
Speaker 2: fastest way to get there is to go beyond the

201
00:09:26,000 --> 00:09:29,599
Speaker 2: GPT paradigm models that can generate new knowledge, models that

202
00:09:29,679 --> 00:09:31,760
Speaker 2: can come up with new ideas, models that can sort

203
00:09:31,800 --> 00:09:33,640
Speaker 2: of just figure things out that they haven't seen before,

204
00:09:33,720 --> 00:09:35,120
Speaker 2: and that's going to require new work.

205
00:09:35,440 --> 00:09:39,160
Speaker 1: I've been using GPT four obsessively. I'm happy to hear

206
00:09:39,160 --> 00:09:41,800
Speaker 1: the last few months. It's quite something, and I do

207
00:09:41,840 --> 00:09:44,040
Speaker 1: feel that it's sometimes coming up with new knowledge. I

208
00:09:44,040 --> 00:09:46,040
Speaker 1: haven't done a robust test that I'm sitting here as

209
00:09:46,080 --> 00:09:48,840
Speaker 1: somebody who works in research, and I'm thinking I have

210
00:09:48,960 --> 00:09:50,840
Speaker 1: learned something new here. So what's going on?

211
00:09:51,360 --> 00:09:53,600
Speaker 2: Yeah, I mean there's like glimpses of it, right, and

212
00:09:53,679 --> 00:09:57,160
Speaker 2: it can do small things, but it can't self correct

213
00:09:57,200 --> 00:09:59,160
Speaker 2: and stay on the rails enough where you can just say, hey,

214
00:09:59,160 --> 00:10:03,240
Speaker 2: GPT four, please go cure cancer. That's not going to happen, right,

215
00:10:03,320 --> 00:10:05,160
Speaker 2: but it would be nice if we had a system

216
00:10:05,200 --> 00:10:05,800
Speaker 2: that could do that.

217
00:10:11,720 --> 00:10:15,840
Speaker 1: Obviously, we're talking about how powerful these technologies are and

218
00:10:15,880 --> 00:10:19,000
Speaker 1: there will also be downsides, and let's start with one

219
00:10:19,440 --> 00:10:23,280
Speaker 1: that is quite approximate today. So GPT for these other

220
00:10:23,320 --> 00:10:26,400
Speaker 1: large language models are very very good at producing text

221
00:10:26,600 --> 00:10:30,199
Speaker 1: human sounding text, and so it opens up that risk

222
00:10:30,280 --> 00:10:33,840
Speaker 1: of misinformation and disinformation in particular, as we head in

223
00:10:33,880 --> 00:10:37,679
Speaker 1: towards important elections in the United States. How serious a

224
00:10:37,760 --> 00:10:40,000
Speaker 1: risk do you see that.

225
00:10:40,400 --> 00:10:44,680
Speaker 2: I do think disinformation is becoming a bigger challenge in

226
00:10:44,720 --> 00:10:50,199
Speaker 2: the world. And also I think it's a somewhat fraught category.

227
00:10:50,280 --> 00:10:52,640
Speaker 2: You know, we've labeled things as disinformation as a society

228
00:10:52,640 --> 00:10:54,520
Speaker 2: that turned out to be true, right, We've kicked people

229
00:10:54,520 --> 00:10:56,640
Speaker 2: off platforms for saying things that turned out to be true.

230
00:10:57,240 --> 00:10:59,839
Speaker 2: And we're going to have to find a balance where

231
00:10:59,840 --> 00:11:03,840
Speaker 2: we preserve the ability to be wrong in exchange for

232
00:11:03,960 --> 00:11:08,880
Speaker 2: sometimes exposing important information and without saying everything is intentional

233
00:11:08,880 --> 00:11:13,720
Speaker 2: disinformation used to manipulate, but people that are intentionally being

234
00:11:13,920 --> 00:11:16,840
Speaker 2: wrong in order to manipulate, I think is a real problem,

235
00:11:16,880 --> 00:11:19,520
Speaker 2: and we've seen more of that with technology.

236
00:11:19,640 --> 00:11:22,760
Speaker 1: I mean three point five in particular is really quite good.

237
00:11:23,120 --> 00:11:25,280
Speaker 1: So if there was going to be a disinformation wave,

238
00:11:25,320 --> 00:11:26,120
Speaker 1: wouldn't it have come?

239
00:11:26,640 --> 00:11:29,199
Speaker 2: So I was going to get there. I think humans

240
00:11:29,240 --> 00:11:32,320
Speaker 2: are already good at making disinformation, and maybe the GPT

241
00:11:32,440 --> 00:11:34,800
Speaker 2: models make it easier, but that's not the thing I'm

242
00:11:34,840 --> 00:11:37,560
Speaker 2: afraid of. Also, I think it's tempting to compare AI

243
00:11:37,559 --> 00:11:39,920
Speaker 2: and social media, but they're super different. Like you can

244
00:11:39,960 --> 00:11:42,920
Speaker 2: generate all the disinformation you want with GPT four, but

245
00:11:43,040 --> 00:11:44,959
Speaker 2: if it's just for yourself and it's not being spread,

246
00:11:44,960 --> 00:11:46,560
Speaker 2: it's not going to do much. I think what is

247
00:11:46,559 --> 00:11:48,959
Speaker 2: worth considering is what's going to be different with AI

248
00:11:49,080 --> 00:11:50,920
Speaker 2: and where is it going to plug into channels that

249
00:11:50,920 --> 00:11:53,520
Speaker 2: could help it spread. And I think one thing that

250
00:11:53,559 --> 00:11:58,320
Speaker 2: will be different is the interactive, personalized persuasive ability of

251
00:11:58,320 --> 00:11:59,000
Speaker 2: these systems.

252
00:12:00,320 --> 00:12:03,560
Speaker 1: That I might get a robocall on my phone, I

253
00:12:03,640 --> 00:12:06,440
Speaker 1: pick it up, and then the messaging in there is

254
00:12:06,559 --> 00:12:10,720
Speaker 1: really attuned to me, so it's emotionally resonant, really realistic,

255
00:12:10,760 --> 00:12:11,840
Speaker 1: and read out by machinery.

256
00:12:12,120 --> 00:12:14,880
Speaker 2: That's what I think the new challenge will be, and

257
00:12:15,080 --> 00:12:17,360
Speaker 2: there's a lot to do there. We can build refusals

258
00:12:17,360 --> 00:12:19,880
Speaker 2: into the models, we can build monitoring systems, so people

259
00:12:19,880 --> 00:12:22,719
Speaker 2: can't do that at scale, But we're going to have

260
00:12:22,920 --> 00:12:26,000
Speaker 2: powerful open source models out in the world, and those

261
00:12:26,480 --> 00:12:28,440
Speaker 2: the open eye techniques of what we can do on

262
00:12:28,440 --> 00:12:31,079
Speaker 2: our own systems won't work the same.

263
00:12:31,120 --> 00:12:33,320
Speaker 1: Right, So just to clarify that point, right, because with

264
00:12:33,360 --> 00:12:36,880
Speaker 1: open AI, you have an API and you have a

265
00:12:36,960 --> 00:12:39,280
Speaker 1: named customer, so if you see bad behavior, you can

266
00:12:39,320 --> 00:12:42,600
Speaker 1: turn that person off, whereas an open source model could

267
00:12:42,600 --> 00:12:45,640
Speaker 1: be run by anyone on their desktop computer at some point,

268
00:12:45,679 --> 00:12:49,720
Speaker 1: and it's actually much harder. There's a proliferation problem. Yeah,

269
00:12:49,760 --> 00:12:53,480
Speaker 1: but solving this can't just be open eyes reams, right.

270
00:12:53,559 --> 00:12:54,760
Speaker 1: You must be asking for help.

271
00:12:55,040 --> 00:12:58,080
Speaker 2: There's regulatory things that we can do that will help some.

272
00:12:58,320 --> 00:13:01,839
Speaker 2: The real solution here is to you educate people about

273
00:13:01,840 --> 00:13:05,040
Speaker 2: what's happening. We've been through this before. When photoshop first

274
00:13:05,240 --> 00:13:07,520
Speaker 2: became popular. There was a brief period of time where

275
00:13:07,559 --> 00:13:09,560
Speaker 2: people like seeing as believing it's got to be real,

276
00:13:09,640 --> 00:13:11,760
Speaker 2: and then people learn quickly that it's not. And some

277
00:13:11,760 --> 00:13:13,760
Speaker 2: people still fall for this stuff. But on the whole,

278
00:13:13,800 --> 00:13:15,520
Speaker 2: if you've see an image, you know it might be

279
00:13:15,600 --> 00:13:18,760
Speaker 2: digitally manipulated. Well understood. The same thing will happen with

280
00:13:18,840 --> 00:13:21,880
Speaker 2: these new technologies. But the sooner we can educate people

281
00:13:21,920 --> 00:13:24,360
Speaker 2: about it, because the emotional resonance is going to be

282
00:13:24,440 --> 00:13:26,240
Speaker 2: so much higher, I think the better.

283
00:13:33,160 --> 00:13:36,560
Speaker 1: Let's turn to education. We're at a global university here,

284
00:13:36,920 --> 00:13:40,320
Speaker 1: and of course education is closely connected to the job market.

285
00:13:41,080 --> 00:13:45,400
Speaker 1: When we previous times, we've seen powerful new technologies emerge.

286
00:13:45,800 --> 00:13:50,359
Speaker 1: They have really impacted power dynamics between workers and employers.

287
00:13:50,400 --> 00:13:53,040
Speaker 1: I think back to the late eighteenth century there was

288
00:13:53,120 --> 00:13:57,240
Speaker 1: Engels Pause, the point in time in England where GDP

289
00:13:57,440 --> 00:14:00,680
Speaker 1: went up the worker wages were stagnant. Looking at Ai,

290
00:14:01,000 --> 00:14:03,880
Speaker 1: we might see something similar, and neither you nor I,

291
00:14:03,880 --> 00:14:06,359
Speaker 1: I think want historians of the future to be describing

292
00:14:06,720 --> 00:14:11,400
Speaker 1: Altman's pause, when wages suffered under a point of wage

293
00:14:11,400 --> 00:14:15,079
Speaker 1: pressure because of the new technology. What are the interventions

294
00:14:15,120 --> 00:14:17,400
Speaker 1: that are needed to make sure that there is a

295
00:14:17,720 --> 00:14:20,800
Speaker 1: equitable sharing of the gains from the technology.

296
00:14:20,840 --> 00:14:23,280
Speaker 2: Well, first of all, we just need gains. We need growth.

297
00:14:23,320 --> 00:14:26,160
Speaker 2: I think one of the problems in the developed the

298
00:14:26,160 --> 00:14:28,560
Speaker 2: world right now is we don't have enough sustainable growth,

299
00:14:28,920 --> 00:14:31,320
Speaker 2: and that's causing all sorts of problems. So I'm excited

300
00:14:31,360 --> 00:14:34,720
Speaker 2: that this technology can bring the missing productivity gains. In

301
00:14:34,760 --> 00:14:39,520
Speaker 2: the last few decades back, some technologies are reduced inequality

302
00:14:39,520 --> 00:14:42,040
Speaker 2: by nature, and some enhance it. I'm not totally sure

303
00:14:42,120 --> 00:14:43,560
Speaker 2: this one's going to go, but I think this is

304
00:14:43,600 --> 00:14:47,440
Speaker 2: a technology that the shape of which is to reduce inequality.

305
00:14:47,880 --> 00:14:50,320
Speaker 2: My basic model of the world is that the cost

306
00:14:50,360 --> 00:14:52,600
Speaker 2: of intelligence and the cost of energy are the two

307
00:14:53,400 --> 00:14:57,920
Speaker 2: limiting inputs, and if you can make those dramatically cheaper,

308
00:14:57,960 --> 00:15:01,320
Speaker 2: dramatically more accessible, that does more to help poor people

309
00:15:01,320 --> 00:15:04,120
Speaker 2: than rich people. Frankly, all thought to help everyone a lot.

310
00:15:04,480 --> 00:15:06,960
Speaker 2: This technology will lift all of the world up. Most

311
00:15:07,000 --> 00:15:09,120
Speaker 2: people in this room, if they need some sort of

312
00:15:09,200 --> 00:15:12,840
Speaker 2: intellectual cognitive labor, they can afford it. Most people in

313
00:15:12,880 --> 00:15:16,800
Speaker 2: the world often can't. And if we can commoditize that,

314
00:15:17,640 --> 00:15:19,760
Speaker 2: I think that is an equalizing force and an important one.

315
00:15:19,800 --> 00:15:21,400
Speaker 2: Can I say one, YEA cool? I think there will

316
00:15:21,400 --> 00:15:22,960
Speaker 2: be way more jobs on the other side of this

317
00:15:23,000 --> 00:15:25,200
Speaker 2: technological revolution. I'm not a believer that this is the

318
00:15:25,280 --> 00:15:27,240
Speaker 2: end of work at all. I think like we will

319
00:15:27,280 --> 00:15:29,120
Speaker 2: look back at the mundane jobs many of us do

320
00:15:29,160 --> 00:15:30,840
Speaker 2: today and be like, that was really bad. This is

321
00:15:30,880 --> 00:15:33,080
Speaker 2: much better and more interesting now. I still think we'll

322
00:15:33,120 --> 00:15:37,760
Speaker 2: have to think about distribution of wealth differently than we

323
00:15:37,800 --> 00:15:40,280
Speaker 2: do today, and that's fine. We actually think about that

324
00:15:40,440 --> 00:15:44,880
Speaker 2: somewhat differently after every technological revolution. I also think, given

325
00:15:44,920 --> 00:15:48,000
Speaker 2: the shape of this particular one, the way that access

326
00:15:48,040 --> 00:15:50,960
Speaker 2: to these systems is distributed fairly is going to be

327
00:15:50,960 --> 00:15:52,680
Speaker 2: a very challenging question.

328
00:15:52,760 --> 00:15:54,920
Speaker 1: Right, And you know, I think access is so important.

329
00:15:54,920 --> 00:15:58,560
Speaker 1: And in those previous revolutions, the technology revolutions, the thing

330
00:15:58,640 --> 00:16:01,360
Speaker 1: that drew us together was political structures I mean it

331
00:16:01,440 --> 00:16:05,160
Speaker 1: was trade unionism and labor collectives in the late nineteenth century.

332
00:16:05,560 --> 00:16:07,960
Speaker 1: When we look at something like AI, can you imagine

333
00:16:07,960 --> 00:16:11,800
Speaker 1: the types of structures that would be needed for recognizing

334
00:16:11,840 --> 00:16:16,480
Speaker 1: and redistributing the gains from unpaid or low paid work

335
00:16:16,480 --> 00:16:19,480
Speaker 1: that's often not recognized, for example, the work that women

336
00:16:19,640 --> 00:16:20,600
Speaker 1: are doing around the world.

337
00:16:21,520 --> 00:16:25,480
Speaker 2: I think there will be an important and overdo shift

338
00:16:25,560 --> 00:16:29,200
Speaker 2: in the kinds of work that we value, and providing

339
00:16:29,600 --> 00:16:32,160
Speaker 2: human connection to people will all of a sudden be

340
00:16:32,320 --> 00:16:34,320
Speaker 2: as I think should be one of the most valued

341
00:16:34,320 --> 00:16:36,840
Speaker 2: types of work, happening all kinds of different ways.

342
00:16:37,080 --> 00:16:40,840
Speaker 1: So when you reflect on how AI has progressed to

343
00:16:41,000 --> 00:16:44,560
Speaker 1: this point, what lessons, if any, can we draw about

344
00:16:44,560 --> 00:16:49,960
Speaker 1: the journey towards artificial superintelligence and how that might emerge.

345
00:16:50,040 --> 00:16:53,240
Speaker 1: This is the idea of having an artificial intelligence that

346
00:16:53,360 --> 00:16:57,760
Speaker 1: is more capable than humans in every and all domains.

347
00:16:57,800 --> 00:16:59,320
Speaker 2: It's hard to give a short answer this question, but

348
00:16:59,520 --> 00:17:02,120
Speaker 2: you've got I think there's a lot of things that

349
00:17:02,160 --> 00:17:04,520
Speaker 2: we've learned so far, but one of them is that

350
00:17:06,240 --> 00:17:09,520
Speaker 2: a we have an algorithm that can genuinely truly learn

351
00:17:09,720 --> 00:17:13,840
Speaker 2: and be It gets predictably better with skill, and these

352
00:17:13,840 --> 00:17:17,560
Speaker 2: are two remarkable facts put together, and I think even

353
00:17:17,600 --> 00:17:19,359
Speaker 2: though we think about that every day, I suspect we

354
00:17:19,400 --> 00:17:24,520
Speaker 2: don't quite feel how important that is. One observation is

355
00:17:24,560 --> 00:17:28,280
Speaker 2: that it's just going to keep going. Another observation is

356
00:17:28,320 --> 00:17:32,639
Speaker 2: that we will have these discontinuous increases occasionally where we

357
00:17:32,640 --> 00:17:36,159
Speaker 2: figure out something new. And a third is that I

358
00:17:36,160 --> 00:17:38,600
Speaker 2: think the way that I used to think about heading

359
00:17:38,680 --> 00:17:42,440
Speaker 2: towards superintelligence is we were going to build this one

360
00:17:42,840 --> 00:17:45,720
Speaker 2: extremely capable system, and there were a bunch of challenge

361
00:17:45,720 --> 00:17:47,800
Speaker 2: safety challenges with that, and it was sort of a

362
00:17:47,800 --> 00:17:51,280
Speaker 2: world that was going to feel quite unstable. But I

363
00:17:51,320 --> 00:17:54,840
Speaker 2: think we now see a path where we very much

364
00:17:55,080 --> 00:17:58,239
Speaker 2: build these tools, not creatures, tools that get more and

365
00:17:58,240 --> 00:18:02,119
Speaker 2: more powerful, and they're there's billions of copies, trillions of

366
00:18:02,119 --> 00:18:06,159
Speaker 2: copies being used in the world, helping individual people just

367
00:18:06,200 --> 00:18:08,600
Speaker 2: be way more effective, capable of doing way more The

368
00:18:08,600 --> 00:18:11,800
Speaker 2: amount of output that one person can have can dramatically increase.

369
00:18:12,119 --> 00:18:16,560
Speaker 2: And where the super intelligence emerges is not just the

370
00:18:16,600 --> 00:18:21,119
Speaker 2: capability of our biggest single neural network, but all of

371
00:18:21,119 --> 00:18:23,360
Speaker 2: the new science we're discovering, all of the new things

372
00:18:23,400 --> 00:18:24,399
Speaker 2: we're creating.

373
00:18:24,520 --> 00:18:27,960
Speaker 1: And the interactions between these billions and trillions of other systems.

374
00:18:28,040 --> 00:18:31,800
Speaker 2: The society we build up, which is AI assisted humans

375
00:18:32,520 --> 00:18:34,919
Speaker 2: using these tools to build up this society, and the

376
00:18:34,960 --> 00:18:37,560
Speaker 2: knowledge and the technology and the institutions and the norms

377
00:18:37,560 --> 00:18:41,080
Speaker 2: that we have, and that vision of living with superintelligence

378
00:18:41,440 --> 00:18:44,400
Speaker 2: seems to me way better all around and a way

379
00:18:44,440 --> 00:18:47,440
Speaker 2: more exciting future for me, for all of you. I hope,

380
00:18:47,520 --> 00:18:49,240
Speaker 2: hope you agree on this than that. Kind of like

381
00:18:49,320 --> 00:18:50,399
Speaker 2: one super briin, you.

382
00:18:50,480 --> 00:18:53,520
Speaker 1: Really brought a visionary picture to us today. Thank you

383
00:18:53,560 --> 00:19:02,320
Speaker 1: so much, thank you, thank you. Reflecting on this conversation

384
00:19:02,359 --> 00:19:04,159
Speaker 1: with Sam, I'm struck by how willing he is to

385
00:19:04,200 --> 00:19:07,679
Speaker 1: engage with the profound risks that AI could pose. Maybe

386
00:19:07,680 --> 00:19:10,119
Speaker 1: this is because the technology is evolving so quickly that

387
00:19:10,160 --> 00:19:13,119
Speaker 1: it's hard even for someone in his position to figure

388
00:19:13,160 --> 00:19:17,080
Speaker 1: out what comes next. One thing remains true. I believe

389
00:19:17,080 --> 00:19:19,120
Speaker 1: it isn't just down to the tech bosses to work

390
00:19:19,160 --> 00:19:22,320
Speaker 1: out how this technology can help us. Instead, this is

391
00:19:22,320 --> 00:19:32,159
Speaker 1: a process we should all have a saying. Thanks for

392
00:19:32,200 --> 00:19:35,520
Speaker 1: listening to the exponentially podcast. If you enjoy the show,

393
00:19:35,760 --> 00:19:38,720
Speaker 1: please leave a review or rating. It really does help

394
00:19:38,760 --> 00:19:42,359
Speaker 1: others find us. The exponentially podcast is presented by me

395
00:19:42,600 --> 00:19:45,840
Speaker 1: Azeem as are. The sound designer is Will Horricks. The

396
00:19:45,920 --> 00:19:48,760
Speaker 1: research was led by Chloe Ippaer and music composed by

397
00:19:48,840 --> 00:19:52,959
Speaker 1: Emily Green and John Zarcone. The show is produced by

398
00:19:53,000 --> 00:19:57,000
Speaker 1: Frederick Cassella, Maria Garrilov and me Azeem As are special

399
00:19:57,040 --> 00:20:00,399
Speaker 1: thanks to Sage Bauman, Jeff Grocott and Magnus Henrikson. The

400
00:20:00,480 --> 00:20:04,520
Speaker 1: executive producers are Andrew Barden, Adam Camiski and Kyle Kramer.

401
00:20:04,800 --> 00:20:08,520
Speaker 1: David Ravella is the managing editor. Exponentially was created by

402
00:20:08,520 --> 00:20:10,840
Speaker 1: Frederick Cassella and is an Eat the Pie I plus

403
00:20:10,840 --> 00:20:14,680
Speaker 1: one limited production in association with Bloomberg LLC.