1
00:00:05,160 --> 00:00:09,040
Speaker 1: Is AI an intelligent agent or is there a totally

2
00:00:09,039 --> 00:00:12,080
Speaker 1: different way that we should be thinking about this. Perhaps

3
00:00:12,080 --> 00:00:16,520
Speaker 1: it's more like a piece of cultural technology. What in

4
00:00:16,560 --> 00:00:21,360
Speaker 1: the world is cultural technology? And how would rethinking this

5
00:00:21,960 --> 00:00:25,160
Speaker 1: change the way we approach what to do next? And

6
00:00:25,200 --> 00:00:26,880
Speaker 1: what does any of this have to do with the

7
00:00:26,920 --> 00:00:31,479
Speaker 1: myth of the golom or Socrates, or the printing press

8
00:00:31,720 --> 00:00:39,000
Speaker 1: or Martin Luther or the story of stone soup. Welcome

9
00:00:39,000 --> 00:00:42,199
Speaker 1: to Intercosmos with me David Eagleman. I'm a neuroscientist and

10
00:00:42,240 --> 00:00:45,919
Speaker 1: an author at Stanford and in these episodes we examine

11
00:00:46,000 --> 00:00:49,720
Speaker 1: brains and the world around us to understand who we

12
00:00:49,840 --> 00:01:09,319
Speaker 1: are and where we're going. So let's start with the

13
00:01:09,520 --> 00:01:12,280
Speaker 1: appreciation that we are smack in the middle of the

14
00:01:12,280 --> 00:01:17,160
Speaker 1: most dramatic technological shift in human history. Every few weeks,

15
00:01:17,200 --> 00:01:21,119
Speaker 1: a new AI system is released that can answer questions

16
00:01:21,160 --> 00:01:25,479
Speaker 1: of increasing complexity. As we've all seen, it can write

17
00:01:25,520 --> 00:01:29,800
Speaker 1: beautiful prose. It can punch out incredibly good code for software.

18
00:01:30,319 --> 00:01:34,720
Speaker 1: It composes music, It mimics voices. It produces images so

19
00:01:34,880 --> 00:01:38,160
Speaker 1: realistic that we've long ago lost the ability to tell

20
00:01:38,200 --> 00:01:39,679
Speaker 1: if a photo is real or not.

21
00:01:40,160 --> 00:01:41,760
Speaker 2: And this increasingly applies.

22
00:01:41,440 --> 00:01:45,560
Speaker 1: To video as well, So with every leap forward, the

23
00:01:45,600 --> 00:01:51,440
Speaker 1: same questions become louder. Is this an intelligent agent? Is

24
00:01:51,480 --> 00:01:55,520
Speaker 1: it conscious? Will it one day surpass us? And if so,

25
00:01:55,600 --> 00:01:59,200
Speaker 1: what happens to us? Today's episode is about how to

26
00:01:59,280 --> 00:02:01,720
Speaker 1: think about this from an angle that.

27
00:02:01,640 --> 00:02:03,080
Speaker 2: Will probably surprise you.

28
00:02:03,440 --> 00:02:06,880
Speaker 1: So as it stands now, in the public conversation about AI,

29
00:02:07,200 --> 00:02:10,680
Speaker 1: we really have just one metaphor, which is that AI

30
00:02:10,960 --> 00:02:16,040
Speaker 1: is an intelligent agent and of increasing intelligence. We all

31
00:02:16,080 --> 00:02:20,320
Speaker 1: talk about these systems as digital minds that can reason

32
00:02:20,360 --> 00:02:24,240
Speaker 1: and plan and act and perhaps even at some point desire.

33
00:02:25,000 --> 00:02:28,320
Speaker 1: This narrative of an AI with its own mind has

34
00:02:28,360 --> 00:02:30,880
Speaker 1: always been with us in science fiction, of course, but

35
00:02:30,919 --> 00:02:34,440
Speaker 1: today we hear it constantly in policy conversations and in

36
00:02:34,520 --> 00:02:38,680
Speaker 1: media headlines. And whether the tone is optimistic or anxious,

37
00:02:39,040 --> 00:02:43,600
Speaker 1: the underlying premise is the same that these are minds

38
00:02:43,760 --> 00:02:46,079
Speaker 1: in the making, that we are witnessing the birth of

39
00:02:46,120 --> 00:02:50,799
Speaker 1: a new kind of intelligence. But what if that metaphor

40
00:02:51,400 --> 00:02:55,680
Speaker 1: is misleading so much so that it's sending our conversations,

41
00:02:56,000 --> 00:03:01,400
Speaker 1: our policy, our research priorities off course. So today's episode

42
00:03:01,639 --> 00:03:06,720
Speaker 1: is about reframing what large AI models really are and

43
00:03:06,760 --> 00:03:10,720
Speaker 1: what they aren't My guest today is Alison Gopnik. She's

44
00:03:10,760 --> 00:03:13,880
Speaker 1: a professor of psychology at Berkeley, very well known in

45
00:03:13,919 --> 00:03:18,040
Speaker 1: the areas of cognitive and language development. She studies infants

46
00:03:18,080 --> 00:03:21,760
Speaker 1: and young children to understand how learning takes place. And

47
00:03:21,800 --> 00:03:23,799
Speaker 1: she was just by the way, elected to the National

48
00:03:23,800 --> 00:03:26,919
Speaker 1: Academy of Sciences. But I'm talking with her today about

49
00:03:26,919 --> 00:03:29,560
Speaker 1: a new paper she co authored in the journal Science

50
00:03:29,840 --> 00:03:35,040
Speaker 1: about AI with colleagues Henry Ferrell, Cosmishalsi, and James Evans.

51
00:03:35,400 --> 00:03:37,120
Speaker 2: The paper argues that.

52
00:03:37,120 --> 00:03:40,920
Speaker 1: We should stop thinking of large models as intelligent agents

53
00:03:41,480 --> 00:03:45,480
Speaker 1: and instead see them as a new kind of cultural

54
00:03:45,560 --> 00:03:47,160
Speaker 1: and social technology.

55
00:03:47,520 --> 00:03:48,440
Speaker 2: Now what does that mean.

56
00:03:48,960 --> 00:03:50,560
Speaker 1: Well, I'll give you a quick preview and then we'll

57
00:03:50,600 --> 00:03:54,520
Speaker 1: jump into the interview. All throughout history, humans have built

58
00:03:54,520 --> 00:03:59,480
Speaker 1: tools to organize information and transmit it. Think of spoken

59
00:03:59,600 --> 00:04:06,480
Speaker 1: language and then writing, and then printing, libraries, television, the Internet.

60
00:04:06,720 --> 00:04:11,280
Speaker 1: Each of these systems reshaped human culture, not because they

61
00:04:11,280 --> 00:04:16,120
Speaker 1: were intelligent in themselves, but because they allowed information to

62
00:04:16,279 --> 00:04:21,279
Speaker 1: be shared and transformed and coordinated in new ways. Just

63
00:04:21,279 --> 00:04:25,280
Speaker 1: think of how the printing press amplified voices, or how

64
00:04:25,320 --> 00:04:30,440
Speaker 1: something like markets distill the messy complexity of economies into

65
00:04:30,760 --> 00:04:32,440
Speaker 1: a single price.

66
00:04:32,279 --> 00:04:34,760
Speaker 2: Signal, how much does this thing cost?

67
00:04:35,520 --> 00:04:39,400
Speaker 1: Or how bureaucracies take the chaos of signals and sort

68
00:04:39,440 --> 00:04:43,840
Speaker 1: it into categories. These are not minds, but they are

69
00:04:43,880 --> 00:04:48,880
Speaker 1: powerful technologies of culture, technologies that change how we all

70
00:04:49,000 --> 00:04:50,120
Speaker 1: think and how we.

71
00:04:50,040 --> 00:04:50,960
Speaker 2: Act and how we live.

72
00:04:51,240 --> 00:04:54,080
Speaker 1: So the argument we'll hear today is that large AI

73
00:04:54,200 --> 00:04:59,680
Speaker 1: models are best understood in this lineage. They don't think,

74
00:05:00,120 --> 00:05:05,160
Speaker 1: but they process the vast collective output of human thought.

75
00:05:05,760 --> 00:05:09,480
Speaker 1: They are trained on millions of texts and images and voices,

76
00:05:09,839 --> 00:05:13,839
Speaker 1: everything from Shakespeare to Reddit threads to government paperwork, and

77
00:05:13,880 --> 00:05:19,159
Speaker 1: they summarize and reorganize and remix that cultural data. And

78
00:05:19,200 --> 00:05:22,400
Speaker 1: they also surface patterns in the data that maybe hadn't

79
00:05:22,440 --> 00:05:25,479
Speaker 1: been seen before. And so when you're interacting with such

80
00:05:25,480 --> 00:05:28,320
Speaker 1: a piece of technology, let's say, asking it to write

81
00:05:28,360 --> 00:05:31,600
Speaker 1: you a poem or explain a concept, you're not talking

82
00:05:31,640 --> 00:05:36,520
Speaker 1: to a mind. You're participating with a kind of cultural

83
00:05:36,640 --> 00:05:40,520
Speaker 1: compression and recombination machine. So see what you think of

84
00:05:40,560 --> 00:05:44,000
Speaker 1: the perspective that you hear today, because it can change

85
00:05:44,080 --> 00:05:48,479
Speaker 1: our concerns and our eventual legislative approaches if we stop

86
00:05:48,560 --> 00:05:52,040
Speaker 1: assuming that these are minds and instead treat them as

87
00:05:52,600 --> 00:05:58,240
Speaker 1: cultural infrastructures like search engines or even democratic institutions, then

88
00:05:58,279 --> 00:06:00,839
Speaker 1: we can start asking the questions.

89
00:06:01,200 --> 00:06:03,240
Speaker 2: One quick thing before we jump into the interview.

90
00:06:03,760 --> 00:06:07,840
Speaker 1: You've heard of large language models llms, and more recently

91
00:06:08,000 --> 00:06:11,400
Speaker 1: large multimodal models that are trained on words and images

92
00:06:11,800 --> 00:06:14,880
Speaker 1: and increasingly other data as well. So nowadays we just

93
00:06:14,960 --> 00:06:18,640
Speaker 1: refer to these as large models. So here's my interview

94
00:06:18,640 --> 00:06:25,560
Speaker 1: with Alison Gothnik. So, Alison, before we get started talking

95
00:06:25,560 --> 00:06:29,120
Speaker 1: about AI, you have built a very wonderful career studying

96
00:06:29,720 --> 00:06:32,640
Speaker 1: scientists who are unusually small and spend most of their

97
00:06:32,680 --> 00:06:34,320
Speaker 1: time lying down, So tell us about that.

98
00:06:35,080 --> 00:06:39,000
Speaker 3: So what I've always been most interested in is how

99
00:06:39,080 --> 00:06:41,440
Speaker 3: is it that people can figure out the world around them?

100
00:06:41,480 --> 00:06:44,039
Speaker 3: How is it that human beings with just a bunch

101
00:06:44,120 --> 00:06:47,440
Speaker 3: of photons hitting our eyes and little disturbances of air

102
00:06:47,560 --> 00:06:50,000
Speaker 3: at our ears, nevertheless, we know about a world of

103
00:06:50,040 --> 00:06:54,360
Speaker 3: people and objects and ultimately quarks and distant planets. How

104
00:06:54,360 --> 00:06:56,000
Speaker 3: could we ever do that? How could we ever learn

105
00:06:56,040 --> 00:06:58,359
Speaker 3: so much from so little? And of course the people

106
00:06:58,360 --> 00:07:01,400
Speaker 3: who are doing that more than anyone else are little children.

107
00:07:01,760 --> 00:07:04,920
Speaker 3: So for the past forty years, what I've been doing

108
00:07:04,960 --> 00:07:07,240
Speaker 3: is trying to figure out how is it that even

109
00:07:07,279 --> 00:07:10,480
Speaker 3: little children can learn so much so quickly from such

110
00:07:10,920 --> 00:07:14,600
Speaker 3: little information. And one of the questions is what kinds

111
00:07:14,600 --> 00:07:17,080
Speaker 3: of computations, what's going on in their brains? What are

112
00:07:17,120 --> 00:07:21,400
Speaker 3: their brains and minds doing that lets them solve these

113
00:07:21,520 --> 00:07:25,119
Speaker 3: really deep problems so quickly and so effectively. And that's

114
00:07:25,160 --> 00:07:28,239
Speaker 3: been the central idea in my career. And it's turned

115
00:07:28,240 --> 00:07:31,679
Speaker 3: out that by looking at kids empirically, by actually studying

116
00:07:31,680 --> 00:07:34,960
Speaker 3: them as scientists, we've discovered that they both know more

117
00:07:35,000 --> 00:07:37,160
Speaker 3: and learn more than we ever would have thought before.

118
00:07:37,200 --> 00:07:38,480
Speaker 3: They're the best learners.

119
00:07:38,120 --> 00:07:39,320
Speaker 4: That we know of in the universe.

120
00:07:39,600 --> 00:07:40,080
Speaker 2: Amazing.

121
00:07:40,400 --> 00:07:44,520
Speaker 1: So we're in this quite remarkable time where for both

122
00:07:44,520 --> 00:07:48,360
Speaker 1: of us we've been doing, you know, the same research

123
00:07:48,440 --> 00:07:51,640
Speaker 1: that we've been doing for many decades, and one might

124
00:07:51,680 --> 00:07:54,200
Speaker 1: have thought five years ago, okay, we'll probably be doing

125
00:07:54,240 --> 00:07:57,080
Speaker 1: that in twenty twenty five, But suddenly the world has

126
00:07:57,120 --> 00:08:02,120
Speaker 1: really changed around us because of A and so you

127
00:08:02,160 --> 00:08:03,800
Speaker 1: and I both are spending a lot of our time

128
00:08:03,920 --> 00:08:07,080
Speaker 1: writing about that and thinking about how to position AI,

129
00:08:07,160 --> 00:08:09,680
Speaker 1: how to understand what it does and does not mean.

130
00:08:10,440 --> 00:08:13,040
Speaker 1: So a lot of people, of course, are concerned about

131
00:08:14,360 --> 00:08:17,560
Speaker 1: super intelligence and the alignment problem, and so on. But

132
00:08:17,640 --> 00:08:20,880
Speaker 1: you and your colleagues have a quite different take that

133
00:08:20,920 --> 00:08:24,520
Speaker 1: you just wrote up in the journal Science in March,

134
00:08:25,040 --> 00:08:27,200
Speaker 1: and I thought it was a really lovely paper. So

135
00:08:27,240 --> 00:08:29,480
Speaker 1: that's what I want to ask you about. So you

136
00:08:29,760 --> 00:08:33,000
Speaker 1: are talking about the right way to look at large

137
00:08:33,040 --> 00:08:36,480
Speaker 1: models is as a social and cultural technology.

138
00:08:36,559 --> 00:08:38,840
Speaker 2: So let's unpack that, right.

139
00:08:38,960 --> 00:08:41,400
Speaker 3: So, as I said, you know, my career has been

140
00:08:41,440 --> 00:08:43,040
Speaker 3: about how could we learn as much as we do?

141
00:08:43,160 --> 00:08:44,880
Speaker 3: And how do children learn as much as they do?

142
00:08:45,320 --> 00:08:48,400
Speaker 3: And part of that has always been if we wanted

143
00:08:48,400 --> 00:08:51,040
Speaker 3: to design a computer or design an artificial system that

144
00:08:51,040 --> 00:08:53,360
Speaker 3: could learn the way children do, what would that system

145
00:08:53,440 --> 00:08:55,720
Speaker 3: look like, what could we put in, what kinds of

146
00:08:55,760 --> 00:08:59,440
Speaker 3: things would it have to do? So for twenty years

147
00:08:59,520 --> 00:09:02,560
Speaker 3: I've been laborating with computer scientists about what would that

148
00:09:02,640 --> 00:09:05,360
Speaker 3: kind of artificial system look like? But as you say,

149
00:09:05,760 --> 00:09:07,920
Speaker 3: even though this has been a long project, in the

150
00:09:08,000 --> 00:09:11,880
Speaker 3: last five years or so, these advances in AI have

151
00:09:12,000 --> 00:09:14,480
Speaker 3: really made us think about that in a different way.

152
00:09:14,720 --> 00:09:17,480
Speaker 3: And one of the interesting things is the big advances

153
00:09:17,520 --> 00:09:20,559
Speaker 3: have been in machine learning. They've actually been in designing

154
00:09:20,640 --> 00:09:23,960
Speaker 3: systems that don't just know things, but can learn things.

155
00:09:24,040 --> 00:09:26,160
Speaker 3: And as I say, children are the best learners we

156
00:09:26,160 --> 00:09:28,480
Speaker 3: know of in the universe. So there's been a really

157
00:09:28,520 --> 00:09:31,120
Speaker 3: interesting development, which is a lot of the people in

158
00:09:31,160 --> 00:09:35,720
Speaker 3: AI have been turning to developmental psychologists like me to say, look,

159
00:09:35,760 --> 00:09:38,199
Speaker 3: could we get some clues from how children are learning

160
00:09:38,440 --> 00:09:41,640
Speaker 3: to design systems that could learn that could learn in

161
00:09:41,640 --> 00:09:45,400
Speaker 3: the same way. Now, the interesting thing is that what

162
00:09:45,679 --> 00:09:49,079
Speaker 3: actually has happened in AI, specifically in the last five

163
00:09:49,160 --> 00:09:51,760
Speaker 3: years or so are these large models, these large language

164
00:09:51,800 --> 00:09:55,120
Speaker 3: models and more recently large language and vision models, and

165
00:09:55,160 --> 00:09:57,880
Speaker 3: they are the things that have really revolutionized our everyday

166
00:09:57,920 --> 00:10:02,000
Speaker 3: interactions with AI. It's important to say those are really

167
00:10:02,160 --> 00:10:05,480
Speaker 3: really different from children, and in fact, they're different from humans.

168
00:10:05,520 --> 00:10:08,440
Speaker 3: They're doing something that I think is really really different

169
00:10:08,480 --> 00:10:15,280
Speaker 3: from human intelligence. And it's very natural for people to think, oh, okay,

170
00:10:15,320 --> 00:10:18,000
Speaker 3: look I talked to chatchept and I get an answer back,

171
00:10:18,320 --> 00:10:20,720
Speaker 3: it must have the same kind of intelligence that my

172
00:10:20,800 --> 00:10:23,320
Speaker 3: friend does or my child does. And it turns out

173
00:10:23,400 --> 00:10:26,480
Speaker 3: that that's not true. Those systems are really really different.

174
00:10:26,760 --> 00:10:28,600
Speaker 1: So before we go on, let's unpack that a little

175
00:10:28,640 --> 00:10:30,080
Speaker 1: bit in what ways are they different?

176
00:10:30,320 --> 00:10:34,920
Speaker 3: So a very common kind of model for how a

177
00:10:34,920 --> 00:10:37,640
Speaker 3: AI works is to think of it as if something

178
00:10:37,679 --> 00:10:40,880
Speaker 3: like chatcheept is an agent, an intelligent agent in the world,

179
00:10:41,000 --> 00:10:43,880
Speaker 3: like a person or even an animal that you know about.

180
00:10:45,160 --> 00:10:48,960
Speaker 3: But that's actually, I think, an illusion. That's what we've argued.

181
00:10:49,480 --> 00:10:51,720
Speaker 3: A better way to think about it is that as

182
00:10:51,800 --> 00:10:55,600
Speaker 3: long as we've been human, we've learned from other people,

183
00:10:55,800 --> 00:10:59,120
Speaker 3: and we've had great technological advances that have helped us

184
00:10:59,440 --> 00:11:02,400
Speaker 3: to learn more effectively for more and more people. So

185
00:11:02,440 --> 00:11:07,000
Speaker 3: if you think about language itself, or writing or print,

186
00:11:07,360 --> 00:11:11,000
Speaker 3: those are all examples of technological changes that made us

187
00:11:11,080 --> 00:11:13,880
Speaker 3: able to get more information from others. And what the

188
00:11:13,960 --> 00:11:16,600
Speaker 3: large models do is not go out into the world

189
00:11:16,600 --> 00:11:19,560
Speaker 3: and learn and think the way that babies do. What

190
00:11:19,600 --> 00:11:24,960
Speaker 3: they do is summarize information that human beings have actually

191
00:11:25,000 --> 00:11:28,000
Speaker 3: already discovered. So what they do is take all the

192
00:11:28,040 --> 00:11:30,440
Speaker 3: information and knowledge that human beings have put out on

193
00:11:30,840 --> 00:11:34,800
Speaker 3: the web essentially, and then summarize that in a way

194
00:11:34,840 --> 00:11:38,440
Speaker 3: that lets other people access it more efficiently. So it's

195
00:11:38,600 --> 00:11:42,520
Speaker 3: much more the technological development is much more like something

196
00:11:42,600 --> 00:11:45,400
Speaker 3: like writing that lets you find out what other people

197
00:11:45,400 --> 00:11:48,760
Speaker 3: are thinking than it is creating a system that could

198
00:11:48,800 --> 00:11:49,800
Speaker 3: learn and think itself.

199
00:11:50,160 --> 00:11:50,439
Speaker 2: Yeah.

200
00:11:50,720 --> 00:11:53,200
Speaker 1: In a previous episode of Intercosmos. They did a calculation

201
00:11:53,280 --> 00:11:56,920
Speaker 1: showing that the amount of information that one of these

202
00:11:57,000 --> 00:12:01,440
Speaker 1: dllms consumes would take you one thousand lifetimes.

203
00:12:00,880 --> 00:12:01,679
Speaker 2: For you to read.

204
00:12:02,600 --> 00:12:07,080
Speaker 1: And so it's consumed more than you could ever imagine.

205
00:12:07,280 --> 00:12:11,280
Speaker 1: And what it's doing fundamentally is when you ask it

206
00:12:11,320 --> 00:12:13,959
Speaker 1: a question, it's giving you an echo of the human

207
00:12:14,000 --> 00:12:16,200
Speaker 1: intelligence that's already in there. So I call this the

208
00:12:16,240 --> 00:12:19,320
Speaker 1: echo intelligence solution, where we feel like, wow, that thing's

209
00:12:19,360 --> 00:12:23,600
Speaker 1: really smart. But it's not smart. It's not smart in

210
00:12:23,640 --> 00:12:27,640
Speaker 1: the same way that a human is. It's taking advantage

211
00:12:27,679 --> 00:12:29,480
Speaker 1: of all the things that are already out there.

212
00:12:29,520 --> 00:12:31,760
Speaker 3: So here's here's a way I like to I like

213
00:12:31,840 --> 00:12:35,440
Speaker 3: to convey this. I think you know, storytelling, as you know,

214
00:12:35,600 --> 00:12:38,720
Speaker 3: is really important. So here's two stories you could tell

215
00:12:39,040 --> 00:12:42,840
Speaker 3: about how current ani work. So one story is sort

216
00:12:42,880 --> 00:12:45,800
Speaker 3: of the story of the Gollum, right, the Rabbi of

217
00:12:45,880 --> 00:12:49,319
Speaker 3: Progue and the Gollum. You create this artificial system and

218
00:12:49,960 --> 00:12:53,000
Speaker 3: it's magical and you put special magic in it, and

219
00:12:53,040 --> 00:12:55,080
Speaker 3: then it turns into something that's almost alive.

220
00:12:55,200 --> 00:12:57,839
Speaker 1: And it's interesting for anyone who doesn't know about the

221
00:12:57,840 --> 00:13:00,320
Speaker 1: story of the Gallum. That was a figure made of clay.

222
00:13:01,240 --> 00:13:04,959
Speaker 1: He was brought to life and defended the community.

223
00:13:05,760 --> 00:13:09,440
Speaker 3: But then, well, it turns out that these stories about

224
00:13:09,480 --> 00:13:12,319
Speaker 3: what would happen if you had something that wasn't human,

225
00:13:12,360 --> 00:13:15,120
Speaker 3: that was artificial that you brought to life are really

226
00:13:15,160 --> 00:13:18,760
Speaker 3: ancient there, way before even the Industrial Revolution. And I

227
00:13:18,760 --> 00:13:21,200
Speaker 3: can tell you right now it never ends well, the

228
00:13:21,320 --> 00:13:26,040
Speaker 3: story always variably. The end of the story is that

229
00:13:26,280 --> 00:13:29,719
Speaker 3: some terrible thing happens and the column goes mad and

230
00:13:30,640 --> 00:13:32,240
Speaker 3: causes trouble and chaos, and.

231
00:13:32,200 --> 00:13:36,200
Speaker 1: That inspired Frankenstein some hundreds of years later. They don't

232
00:13:36,200 --> 00:13:36,960
Speaker 1: have the same character.

233
00:13:37,200 --> 00:13:41,400
Speaker 3: So there's this very basic human fear about what would

234
00:13:41,440 --> 00:13:44,240
Speaker 3: it be like if there was something that wasn't actually

235
00:13:44,360 --> 00:13:46,920
Speaker 3: living that you treated as if it was living, as

236
00:13:47,000 --> 00:13:47,760
Speaker 3: you treated as.

237
00:13:47,640 --> 00:13:48,400
Speaker 4: If it was an agent.

238
00:13:48,440 --> 00:13:51,920
Speaker 3: And I think that basic picture, that's the sci fi picture,

239
00:13:52,080 --> 00:13:54,920
Speaker 3: that's the picture that a lot of people, including people

240
00:13:54,920 --> 00:13:57,680
Speaker 3: in the AI world themselves, have about what's happened in Ai.

241
00:13:58,320 --> 00:14:01,560
Speaker 3: Here's a really different story, also, a different ancient story.

242
00:14:01,600 --> 00:14:04,240
Speaker 3: This is the story of stone Soup. So what's the

243
00:14:04,280 --> 00:14:06,280
Speaker 3: story of stone soup. The story of stone Soup is

244
00:14:06,280 --> 00:14:09,960
Speaker 3: there's visitors who come to a village and they say,

245
00:14:10,000 --> 00:14:12,280
Speaker 3: we'd like some food and the villagers say, no, we

246
00:14:12,320 --> 00:14:14,360
Speaker 3: don't have any extra food. And they say, it's okay,

247
00:14:14,360 --> 00:14:16,319
Speaker 3: we're going to make stone soup. And they take out

248
00:14:16,320 --> 00:14:18,880
Speaker 3: a big pot. They put a couple of stones in it,

249
00:14:19,160 --> 00:14:21,360
Speaker 3: they put some water in, they start to boil it

250
00:14:21,480 --> 00:14:23,240
Speaker 3: up and they say, this will be delicious. We're going

251
00:14:23,280 --> 00:14:25,240
Speaker 3: to make stone soup just with these stones, and the

252
00:14:25,280 --> 00:14:28,760
Speaker 3: villagers say really. They say, yeah, it would be even

253
00:14:28,840 --> 00:14:30,960
Speaker 3: better if we had an onion and a carrot in it,

254
00:14:31,000 --> 00:14:32,360
Speaker 3: but if we don't, we don't.

255
00:14:32,400 --> 00:14:33,520
Speaker 4: And the villager says.

256
00:14:33,440 --> 00:14:35,360
Speaker 3: I think I have an onion and a carrot somewhere,

257
00:14:35,440 --> 00:14:38,360
Speaker 3: and they go and put it in and then they say,

258
00:14:38,480 --> 00:14:40,320
Speaker 3: you know, when we made this for the rich people,

259
00:14:40,400 --> 00:14:42,800
Speaker 3: we put barley and buttermilk in it, which makes it

260
00:14:42,840 --> 00:14:45,800
Speaker 3: even better. But it's okay, it'll still be good stone soup.

261
00:14:45,840 --> 00:14:48,360
Speaker 3: And another villager goes and gets the barley and buttermilk,

262
00:14:48,400 --> 00:14:49,360
Speaker 3: and you can imagine.

263
00:14:49,080 --> 00:14:49,640
Speaker 4: How this goes.

264
00:14:49,680 --> 00:14:52,320
Speaker 3: And they say, the king said that we should put

265
00:14:52,320 --> 00:14:54,480
Speaker 3: a chicken in it, which would make it really royal,

266
00:14:54,960 --> 00:14:57,760
Speaker 3: but we don't have any chicken. So another villager goes

267
00:14:57,760 --> 00:14:59,760
Speaker 3: and gets the chicken from the back, and by the

268
00:14:59,800 --> 00:15:02,480
Speaker 3: time they're done of course, they have this really wonderful

269
00:15:02,520 --> 00:15:05,440
Speaker 3: soup with all the contributions from all the villagers, and

270
00:15:05,600 --> 00:15:07,720
Speaker 3: they go to eat it, and the villagers say, this

271
00:15:07,880 --> 00:15:10,680
Speaker 3: is amazing. There's this wonderful soup and it was just

272
00:15:10,800 --> 00:15:13,600
Speaker 3: made from stones. Okay, here's the modern version of this.

273
00:15:14,000 --> 00:15:17,280
Speaker 3: There's a bunch of tech guys and they go to

274
00:15:17,320 --> 00:15:20,600
Speaker 3: the village of computer users and they say, we're going

275
00:15:20,680 --> 00:15:25,360
Speaker 3: to make artificial general intelligence just from gradient descent and

276
00:15:25,600 --> 00:15:30,280
Speaker 3: transformers and a few algorithms. And the computer used to say,

277
00:15:30,320 --> 00:15:33,120
Speaker 3: that sounds great. We're gonna have artificial general intelligence. And

278
00:15:33,160 --> 00:15:34,920
Speaker 3: they say, yeah, but it would be better if we

279
00:15:34,960 --> 00:15:37,000
Speaker 3: had more data. What we need for this, as you

280
00:15:37,120 --> 00:15:39,360
Speaker 3: just said, David, is lots and lots of data.

281
00:15:39,440 --> 00:15:41,520
Speaker 4: Could you guys put all of.

282
00:15:41,440 --> 00:15:44,120
Speaker 3: Your texts and pictures on the internet for us and

283
00:15:44,160 --> 00:15:45,840
Speaker 3: then let us use them to.

284
00:15:45,800 --> 00:15:48,240
Speaker 4: Train our systems, And the computer user to say, oh,

285
00:15:48,280 --> 00:15:49,000
Speaker 4: that sounds good.

286
00:15:49,080 --> 00:15:51,960
Speaker 3: We'll just keep putting more of our pictures and our

287
00:15:52,440 --> 00:15:56,680
Speaker 3: writings and our books on the internet, and I guess

288
00:15:56,720 --> 00:15:59,360
Speaker 3: you can just use them all for free, and then

289
00:15:59,480 --> 00:16:02,840
Speaker 3: the then the tech person, oh, this is really good.

290
00:16:02,840 --> 00:16:04,600
Speaker 3: This is getting to be more intelligent. But you know,

291
00:16:04,640 --> 00:16:06,840
Speaker 3: it still says really stupid things. A lot of the

292
00:16:06,840 --> 00:16:10,760
Speaker 3: time it says weird things. So what we could do

293
00:16:10,880 --> 00:16:13,360
Speaker 3: is reinforcement learning from human feedback, which is actually a

294
00:16:13,440 --> 00:16:16,360
Speaker 3: really important part of these systems. What we'll do is

295
00:16:16,400 --> 00:16:19,760
Speaker 3: we'll give them to humans and then people can say

296
00:16:19,800 --> 00:16:21,880
Speaker 3: whether what they're saying is good or not, and then

297
00:16:21,920 --> 00:16:23,280
Speaker 3: we'll use that for the training.

298
00:16:24,120 --> 00:16:27,000
Speaker 4: The computer used to say, oh, okay, we're happy to

299
00:16:27,080 --> 00:16:27,440
Speaker 4: do that.

300
00:16:27,560 --> 00:16:29,640
Speaker 3: We'll actually go out and say whether this is good

301
00:16:29,760 --> 00:16:30,960
Speaker 3: or not. There's a whole and this.

302
00:16:31,000 --> 00:16:31,840
Speaker 4: Is literally true.

303
00:16:31,880 --> 00:16:34,920
Speaker 3: There are whole villages in Kenya that we'll do this

304
00:16:35,080 --> 00:16:36,880
Speaker 3: for very small amounts of money.

305
00:16:37,840 --> 00:16:39,560
Speaker 4: And the tech pro said, oh.

306
00:16:39,440 --> 00:16:42,760
Speaker 3: Look see it's even smarter, but it's still saying really

307
00:16:42,760 --> 00:16:46,440
Speaker 3: stupid things sometimes. How about if you did prompt engineering.

308
00:16:46,520 --> 00:16:49,640
Speaker 3: So think really hard about exactly how to ask it

309
00:16:49,680 --> 00:16:52,720
Speaker 3: the right questions so that you can get the right answers,

310
00:16:52,720 --> 00:16:54,760
Speaker 3: because otherwise it's going to say stupid things. And the

311
00:16:54,880 --> 00:16:57,320
Speaker 3: users say, oh, okay, we'll do that. We'll sit down

312
00:16:57,320 --> 00:17:00,120
Speaker 3: and we'll figure out how to do prompt engineering. At

313
00:17:00,160 --> 00:17:04,160
Speaker 3: the end of this process, the tech bro say, seeing,

314
00:17:04,280 --> 00:17:07,320
Speaker 3: we told you we made artificial general intelligence and it

315
00:17:07,400 --> 00:17:10,439
Speaker 3: was just from a few algorithms and the computer you

316
00:17:10,560 --> 00:17:13,439
Speaker 3: to say that's amazing, that's amazing, We're going to have

317
00:17:13,560 --> 00:17:17,240
Speaker 3: artificial intelligence, and it's just you, brilliant tech guys who

318
00:17:17,400 --> 00:17:20,080
Speaker 3: invented it. So, of course, the point of this is

319
00:17:20,119 --> 00:17:23,040
Speaker 3: that it's a sort of debunking story, but it's also

320
00:17:23,400 --> 00:17:26,239
Speaker 3: in both versions, a positive story, because the point is

321
00:17:26,280 --> 00:17:30,159
Speaker 3: when you have a combination of lots and lots of contributions,

322
00:17:30,200 --> 00:17:34,159
Speaker 3: of lots and lots of intelligent people, lots of humans

323
00:17:34,160 --> 00:17:36,760
Speaker 3: who we know are intelligent, both in terms of the

324
00:17:36,840 --> 00:17:39,280
Speaker 3: data they provide and in terms of things like reinforcement

325
00:17:39,359 --> 00:17:42,840
Speaker 3: learning and from human feedback and prompt engineering, you've got

326
00:17:42,920 --> 00:17:46,600
Speaker 3: something that that's bigger than any individual human could have.

327
00:17:47,080 --> 00:17:49,520
Speaker 3: But it's not that what you've got is a gall

328
00:17:49,760 --> 00:17:51,600
Speaker 3: it's not that what you've got is an agent that's

329
00:17:51,640 --> 00:17:54,479
Speaker 3: gone out and been intelligent itself. It's really just a

330
00:17:54,520 --> 00:17:58,080
Speaker 3: system for putting together the thoughts of other agents.

331
00:17:58,320 --> 00:18:01,800
Speaker 1: So the lesson that surfaces here is that although we

332
00:18:01,960 --> 00:18:07,280
Speaker 1: humans love to anthromorphize things and love to make inanimate

333
00:18:07,320 --> 00:18:12,000
Speaker 1: objects into agents in our minds, it's probably not the

334
00:18:12,080 --> 00:18:15,199
Speaker 1: right way to think about these language models. Or let

335
00:18:15,240 --> 00:18:19,040
Speaker 1: me say, these large models, language, vision, multimodal. So what

336
00:18:19,160 --> 00:18:21,280
Speaker 1: is the right way to think about it? So let's

337
00:18:21,320 --> 00:18:24,080
Speaker 1: really unpack this issue about what a social or cultural

338
00:18:24,119 --> 00:18:25,080
Speaker 1: technology is.

339
00:18:25,400 --> 00:18:31,800
Speaker 3: Yeah, so, ever since we've been human, we've made progress

340
00:18:32,040 --> 00:18:36,040
Speaker 3: by learning from other humans, and a number of people

341
00:18:36,640 --> 00:18:39,320
Speaker 3: like Joseph Henrick, for example, and Brob Boyd have argued

342
00:18:39,359 --> 00:18:41,760
Speaker 3: that that's kind of our human our great human gift,

343
00:18:42,680 --> 00:18:45,520
Speaker 3: that's really our secret sauce is not so much that

344
00:18:45,560 --> 00:18:49,520
Speaker 3: we can individually learn things that other creatures can't, although

345
00:18:49,520 --> 00:18:51,400
Speaker 3: I think that's part of it, but that we can

346
00:18:51,440 --> 00:18:54,119
Speaker 3: take advantage of all the things that other humans have

347
00:18:54,240 --> 00:18:57,440
Speaker 3: done over many, many, many generations. I like to think

348
00:18:57,440 --> 00:19:00,960
Speaker 3: of this in terms of the postmenopausal grandmother. I think,

349
00:19:01,200 --> 00:19:03,040
Speaker 3: you know, one of the distinctive human things is that

350
00:19:03,040 --> 00:19:05,320
Speaker 3: we have these postmenopausal grandmothers, and a lot of what

351
00:19:05,359 --> 00:19:08,040
Speaker 3: they do is tell us about the things that they've

352
00:19:08,400 --> 00:19:12,440
Speaker 3: learned in their long, wise lives. And by taking advantage

353
00:19:12,480 --> 00:19:15,719
Speaker 3: of what granny says, you can make progress, even if

354
00:19:15,760 --> 00:19:18,000
Speaker 3: what you're doing is now finding new things that you

355
00:19:18,000 --> 00:19:22,560
Speaker 3: will tell your grandchildren. And that capacity is really the

356
00:19:22,600 --> 00:19:26,600
Speaker 3: capacity that makes us special. And we've had special technologies

357
00:19:26,680 --> 00:19:30,159
Speaker 3: ever since we evolved, that tuned up that capacity, that

358
00:19:30,200 --> 00:19:33,720
Speaker 3: made it more powerful. So language itself, of course, which

359
00:19:33,760 --> 00:19:36,520
Speaker 3: is one of our distinctive things about humans, lets us

360
00:19:36,600 --> 00:19:39,280
Speaker 3: learn from others. But even more if you think about

361
00:19:39,280 --> 00:19:42,679
Speaker 3: something like the invention of writing that enabled us to

362
00:19:42,760 --> 00:19:45,040
Speaker 3: learn from you know, not just our own granny, but

363
00:19:45,080 --> 00:19:47,679
Speaker 3: granny's who were far away in space and in time,

364
00:19:48,080 --> 00:19:52,080
Speaker 3: and it's fascinating. Socrates famously has a whole section about

365
00:19:52,080 --> 00:19:55,800
Speaker 3: why he thinks writing is a terrible idea because exactly

366
00:19:55,880 --> 00:19:58,480
Speaker 3: because he thinks, people will read something in a book

367
00:19:58,520 --> 00:20:00,879
Speaker 3: and they'll think it's actually a person. They'll think that

368
00:20:00,960 --> 00:20:04,520
Speaker 3: it's a person who said this, and it's not. It's

369
00:20:04,640 --> 00:20:06,600
Speaker 3: just something that's written in a book. You won't be

370
00:20:06,640 --> 00:20:09,560
Speaker 3: able to have Socratic dialogues with something that's written in

371
00:20:09,600 --> 00:20:13,120
Speaker 3: a book. It's not really a person, but because it's language,

372
00:20:13,440 --> 00:20:16,720
Speaker 3: will treat it as if it's a person. So writing

373
00:20:16,880 --> 00:20:20,000
Speaker 3: is a good example of something that even though the

374
00:20:20,040 --> 00:20:24,040
Speaker 3: books aren't intelligent, the books in some sense don't know things.

375
00:20:24,160 --> 00:20:28,000
Speaker 3: In another sense, we know things because of books a way.

376
00:20:28,040 --> 00:20:30,119
Speaker 3: I think if this, sometimes they suppose someone asks you,

377
00:20:30,760 --> 00:20:34,680
Speaker 3: who knows more me or the UC Berkeley library. Well,

378
00:20:35,000 --> 00:20:37,600
Speaker 3: the library has much more knowledge in it. It's got

379
00:20:37,800 --> 00:20:40,159
Speaker 3: vast amounts of knowledge that I could never actually have

380
00:20:40,280 --> 00:20:43,160
Speaker 3: in my head. But it's not the sort of thing

381
00:20:43,200 --> 00:20:45,639
Speaker 3: that knows. I'm the sort of person who knows and

382
00:20:45,880 --> 00:20:48,920
Speaker 3: I know things because I can do things like consult

383
00:20:48,960 --> 00:20:52,000
Speaker 3: the library. And then you have print, which is even

384
00:20:52,119 --> 00:20:55,680
Speaker 3: more powerful and has even more powerful effects. You have

385
00:20:56,800 --> 00:21:00,600
Speaker 3: video and film. You have pictures, which I think are

386
00:21:00,640 --> 00:21:03,560
Speaker 3: a really important medium that we don't pay enough attention to.

387
00:21:03,920 --> 00:21:07,040
Speaker 3: So when we talk about vision models, for example, they're

388
00:21:07,040 --> 00:21:10,000
Speaker 3: not actually using vision. What they're using is all the

389
00:21:10,000 --> 00:21:12,199
Speaker 3: pictures that we put on the Internet, and pictures are

390
00:21:12,240 --> 00:21:16,400
Speaker 3: a really important source of communication to So to say

391
00:21:16,440 --> 00:21:19,119
Speaker 3: that it's a cultural technology, to say it's one of

392
00:21:19,160 --> 00:21:22,199
Speaker 3: these technologies that lets humans learn from other humans is

393
00:21:22,240 --> 00:21:24,840
Speaker 3: not at all to dismiss it. Those cultural technologies are

394
00:21:24,840 --> 00:21:27,480
Speaker 3: the things that have led, for better or for worse,

395
00:21:27,600 --> 00:21:30,000
Speaker 3: to the world that we have now. But it's just

396
00:21:30,160 --> 00:21:32,920
Speaker 3: a really different thing. It's what philosophers would call a

397
00:21:32,960 --> 00:21:36,800
Speaker 3: category mistake to think that it's like an intelligent agent,

398
00:21:36,800 --> 00:21:38,359
Speaker 3: which is not to say that at some point in

399
00:21:38,400 --> 00:21:42,560
Speaker 3: the future AI might not develop intelligent agents, but that's

400
00:21:42,600 --> 00:21:44,600
Speaker 3: not what the large models are doing, and that's a

401
00:21:44,680 --> 00:21:47,880
Speaker 3: much much harder lift, something we're much further away from

402
00:21:48,000 --> 00:21:55,800
Speaker 3: than this fantastic, interesting, powerful new cultural technology that we've invented.

403
00:22:11,920 --> 00:22:14,280
Speaker 1: So when we're thinking about cultural technologies, those are all

404
00:22:14,440 --> 00:22:17,680
Speaker 1: amazing examples that you gave about the invention of writing

405
00:22:18,800 --> 00:22:21,680
Speaker 1: and the printing press and then the Internet and so on.

406
00:22:22,640 --> 00:22:25,480
Speaker 1: In the paper that you wrote with your colleagues, you

407
00:22:25,560 --> 00:22:31,000
Speaker 1: mentioned other things also, like markets and bureaucracies and representative democracies.

408
00:22:31,080 --> 00:22:33,560
Speaker 1: Just give us a sense of how those are our

409
00:22:33,640 --> 00:22:34,960
Speaker 1: cultural technologies as well.

410
00:22:35,119 --> 00:22:39,040
Speaker 3: Yeah, so this paper, it was a wonderful collaboration between

411
00:22:39,080 --> 00:22:42,600
Speaker 3: me and Henry Farrell, who's a really distinguished political scientist,

412
00:22:42,880 --> 00:22:46,680
Speaker 3: James Evans, who's a fantastic sociologist, and Cosmo Shalizi, who's

413
00:22:46,720 --> 00:22:48,840
Speaker 3: the statistician. So it was kind of like everybody in

414
00:22:48,880 --> 00:22:52,520
Speaker 3: the social sciences. And what Henry and James have pointed

415
00:22:52,560 --> 00:22:55,720
Speaker 3: out is that we have things like writing in print,

416
00:22:55,760 --> 00:22:58,399
Speaker 3: but if you think about things like a market, this

417
00:22:58,480 --> 00:23:02,399
Speaker 3: is an old observation in economics, what a market really

418
00:23:02,440 --> 00:23:06,399
Speaker 3: does is to summarize information from lots and lots of

419
00:23:06,400 --> 00:23:09,399
Speaker 3: individual people. Imagine that you're in a forager culture and

420
00:23:09,440 --> 00:23:12,000
Speaker 3: you want to exchange you know, I have a two

421
00:23:12,840 --> 00:23:15,600
Speaker 3: turnips and I want to exchange them for some beads. Right,

422
00:23:16,000 --> 00:23:18,880
Speaker 3: I have to work that out for each individual group

423
00:23:18,920 --> 00:23:20,919
Speaker 3: of people. And what a market does is let you

424
00:23:20,960 --> 00:23:23,399
Speaker 3: do that literally on a planetary scale, lets you do

425
00:23:23,440 --> 00:23:26,040
Speaker 3: it for billions of people. And the price is a

426
00:23:26,119 --> 00:23:29,600
Speaker 3: kind of summary of here's all the desires and all

427
00:23:29,600 --> 00:23:32,680
Speaker 3: the goals and all the preferences of all of these

428
00:23:32,680 --> 00:23:35,960
Speaker 3: people just summarized in this one, you know number that

429
00:23:36,000 --> 00:23:39,280
Speaker 3: you find when you look on an Amazon. Now, of

430
00:23:39,280 --> 00:23:43,159
Speaker 3: course that's interesting because it didn't even rely on you know,

431
00:23:43,280 --> 00:23:47,480
Speaker 3: markets start in the industrial age, but they don't rely

432
00:23:47,600 --> 00:23:50,760
Speaker 3: on computations. They they're way before we have computers. They're

433
00:23:50,800 --> 00:23:53,679
Speaker 3: way before we even have calculators. But the invention of

434
00:23:53,760 --> 00:23:56,879
Speaker 3: markets was a kind of information processing invention that let

435
00:23:56,960 --> 00:24:01,399
Speaker 3: us take individual desires and put them together. And democratic

436
00:24:01,440 --> 00:24:04,080
Speaker 3: elections are like that too, where we have all of

437
00:24:04,119 --> 00:24:08,320
Speaker 3: these people who have different preferences, and the democratic process

438
00:24:08,440 --> 00:24:11,600
Speaker 3: lets us figure out a way of combining them all.

439
00:24:11,640 --> 00:24:13,960
Speaker 3: And that's another thing that these large models can do,

440
00:24:14,000 --> 00:24:16,400
Speaker 3: that can take lots of information from lots of different people,

441
00:24:16,440 --> 00:24:18,480
Speaker 3: put it together in a single format. And again this

442
00:24:18,600 --> 00:24:20,280
Speaker 3: is for good or for ill, and maybe we could

443
00:24:20,320 --> 00:24:23,440
Speaker 3: talk about both of those sides in a bit.

444
00:24:23,680 --> 00:24:25,240
Speaker 2: Great. Well, actually let's go there now.

445
00:24:25,359 --> 00:24:28,840
Speaker 1: So all previous social and cultural technologies always come with

446
00:24:28,880 --> 00:24:29,400
Speaker 1: good and bad.

447
00:24:29,520 --> 00:24:31,560
Speaker 2: So what do you see.

448
00:24:31,280 --> 00:24:34,480
Speaker 1: As far as large models go with our current moment

449
00:24:34,520 --> 00:24:37,640
Speaker 1: of AI. What do you see as the potential good

450
00:24:37,680 --> 00:24:38,080
Speaker 1: and bad?

451
00:24:38,440 --> 00:24:38,720
Speaker 4: Yeah?

452
00:24:39,119 --> 00:24:41,680
Speaker 3: So, one thing that I think is worth pointing out

453
00:24:41,880 --> 00:24:45,359
Speaker 3: is that actually the big technological change was not the

454
00:24:45,560 --> 00:24:48,639
Speaker 3: change in large models. It was the fact that around

455
00:24:48,680 --> 00:24:52,000
Speaker 3: the year two thousand there was this remarkable thing happened

456
00:24:52,040 --> 00:24:54,840
Speaker 3: that nobody really noticed or paid attention to, which is

457
00:24:54,880 --> 00:24:58,280
Speaker 3: that all the previous media got turned into bits. So

458
00:24:58,920 --> 00:25:05,040
Speaker 3: it's fascinating. Around two thousand, you get the first computerized movies,

459
00:25:05,080 --> 00:25:08,880
Speaker 3: You get things like Toy Story and Pixar. You get PDFs.

460
00:25:09,000 --> 00:25:11,520
Speaker 3: PDFs are taking print and turning them into bits. You

461
00:25:11,560 --> 00:25:17,440
Speaker 3: get HDTV, so you get TV that's now digital, And suddenly,

462
00:25:17,880 --> 00:25:20,920
Speaker 3: in a very short space of time, the only analog

463
00:25:21,040 --> 00:25:21,720
Speaker 3: media are.

464
00:25:21,600 --> 00:25:23,080
Speaker 4: In museums in kindergartens.

465
00:25:23,080 --> 00:25:25,920
Speaker 3: You know, everything else is everything else is digital, which

466
00:25:25,960 --> 00:25:28,840
Speaker 3: means that now not only do you have information, but

467
00:25:28,920 --> 00:25:33,800
Speaker 3: it's infinitely reproducible, and it's instantaneously transmissible because it's in

468
00:25:33,840 --> 00:25:36,959
Speaker 3: the format of bits. And once that happened, then it

469
00:25:37,000 --> 00:25:39,080
Speaker 3: was just a matter of time before we found new

470
00:25:39,160 --> 00:25:44,040
Speaker 3: ways of accessing and summarizing and organizing that information. And

471
00:25:44,240 --> 00:25:47,760
Speaker 3: large models, I think are the are the result of that. Okay,

472
00:25:47,840 --> 00:25:51,960
Speaker 3: so what do we know about these past changes in technologies?

473
00:25:52,960 --> 00:25:53,919
Speaker 4: You go back to writing.

474
00:25:53,960 --> 00:25:57,320
Speaker 3: As I mentioned, Socrates was very dubious about whether writing

475
00:25:57,480 --> 00:26:00,280
Speaker 3: was going to be a good thing or not, because

476
00:26:00,320 --> 00:26:03,320
Speaker 3: he pointed out that you don't have Socratic dialogues when

477
00:26:03,320 --> 00:26:06,600
Speaker 3: you have books, when you have writing, and you don't

478
00:26:07,040 --> 00:26:09,560
Speaker 3: memorize all of Homer when you have books, and he

479
00:26:09,680 --> 00:26:12,560
Speaker 3: was right, we don't memorize all of Homer anymore. Well,

480
00:26:12,560 --> 00:26:14,359
Speaker 3: we do tend to think that things that are written

481
00:26:14,359 --> 00:26:16,320
Speaker 3: down are truer than they actually are.

482
00:26:16,600 --> 00:26:17,199
Speaker 2: And by the way, in.

483
00:26:17,200 --> 00:26:19,560
Speaker 1: The fifteenth century, once the printing press was invented, there

484
00:26:19,600 --> 00:26:21,960
Speaker 1: were a lot of these same complaints that surfaced, where

485
00:26:21,960 --> 00:26:25,080
Speaker 1: people said, look, you know, if you ask a kid

486
00:26:25,080 --> 00:26:26,600
Speaker 1: a question, now they're just going to go to the

487
00:26:26,640 --> 00:26:29,119
Speaker 1: shelf and pull it right off and there's the answer.

488
00:26:29,160 --> 00:26:30,159
Speaker 2: They don't have to think about it.

489
00:26:30,280 --> 00:26:35,280
Speaker 3: That's exactly right, and so there was misinformation. You know,

490
00:26:35,320 --> 00:26:37,600
Speaker 3: when you can pass on information, one of the first

491
00:26:37,640 --> 00:26:40,480
Speaker 3: things that happens is you can also pass on misinformation.

492
00:26:41,640 --> 00:26:44,760
Speaker 3: An example that I really like to give is, so

493
00:26:44,800 --> 00:26:48,400
Speaker 3: you were mentioning about the fifteenth century when printing starts.

494
00:26:49,000 --> 00:26:50,880
Speaker 3: Something that I just found out recently that I think

495
00:26:50,960 --> 00:26:53,359
Speaker 3: is really fascinating is, you know how we all have

496
00:26:53,480 --> 00:26:57,920
Speaker 3: this mythology about Luther nailed the articles to the door,

497
00:26:58,320 --> 00:27:00,440
Speaker 3: and that was, you know, the big defiance thing. Well,

498
00:27:00,480 --> 00:27:02,760
Speaker 3: it turns out nailing things to the door was like

499
00:27:02,840 --> 00:27:05,400
Speaker 3: having post it notes, Like everybody was always nailing things

500
00:27:05,440 --> 00:27:06,960
Speaker 3: to the door. That was just like the way that

501
00:27:07,000 --> 00:27:11,040
Speaker 3: you distributed it. What Luther did was print his ideas,

502
00:27:11,520 --> 00:27:15,160
Speaker 3: and when they were printed, then they could be distributed

503
00:27:15,200 --> 00:27:19,399
Speaker 3: to everybody, including like the common people. And that was

504
00:27:19,440 --> 00:27:24,560
Speaker 3: the revolutionary thing. That was the disruptive That was the disruptive.

505
00:27:24,080 --> 00:27:27,439
Speaker 1: Thing, analogous to let's say, off Hillary using radio in

506
00:27:27,480 --> 00:27:30,320
Speaker 1: the thirties, reaching a much wider audience.

507
00:27:30,560 --> 00:27:34,080
Speaker 3: But my favorite example is in the eighteenth century, the

508
00:27:34,240 --> 00:27:37,840
Speaker 3: late eighteenth century, there were further technological changes in printing,

509
00:27:37,880 --> 00:27:41,520
Speaker 3: which meant that it became extremely I mean Essentially, anybody

510
00:27:41,520 --> 00:27:43,879
Speaker 3: could go and find a printing press and print a

511
00:27:43,920 --> 00:27:45,399
Speaker 3: pamphlet and distributed.

512
00:27:45,920 --> 00:27:47,760
Speaker 4: And there's a very good.

513
00:27:47,640 --> 00:27:51,359
Speaker 3: Argument that this was responsible for the Enlightenment, that you know,

514
00:27:51,440 --> 00:27:55,520
Speaker 3: something like it's not a coincidence that Ben Franklin was

515
00:27:55,560 --> 00:27:58,720
Speaker 3: a printer. A lot of the source of the American

516
00:27:58,720 --> 00:28:03,879
Speaker 3: Revolution was these pamphlets that were spreading new ideas from

517
00:28:03,920 --> 00:28:07,399
Speaker 3: the Enlightenment, things like Tom Payne's common Sense, the idea

518
00:28:07,480 --> 00:28:11,240
Speaker 3: that people could have a democratic state. Those were all

519
00:28:11,280 --> 00:28:14,040
Speaker 3: things that were distributed through printing. So we all think

520
00:28:14,480 --> 00:28:18,080
Speaker 3: that's great, like, that's fantastic, we could get things really quickly.

521
00:28:18,600 --> 00:28:22,960
Speaker 3: But at the same time, the scholar Robert Darten pointed

522
00:28:23,000 --> 00:28:25,640
Speaker 3: this out a long time ago. If you actually look

523
00:28:25,720 --> 00:28:29,360
Speaker 3: and read all of the pamphlets that were produced in France,

524
00:28:29,480 --> 00:28:31,520
Speaker 3: so they were happening in America, they were also happening

525
00:28:31,560 --> 00:28:32,040
Speaker 3: in France.

526
00:28:32,760 --> 00:28:34,399
Speaker 4: You will be amazed to hear this, David.

527
00:28:34,480 --> 00:28:39,520
Speaker 3: But most of them were libel lies, misinformation, and a

528
00:28:39,560 --> 00:28:42,760
Speaker 3: lot of soft core porn. That's the first thing when

529
00:28:42,920 --> 00:28:45,600
Speaker 3: people have a new cultural technology, that's the first thing

530
00:28:45,640 --> 00:28:48,000
Speaker 3: they use it for. And a lot of the French

531
00:28:48,040 --> 00:28:52,240
Speaker 3: Revolution came because, for example, Marie Antoinette saying let them

532
00:28:52,280 --> 00:28:52,680
Speaker 3: eat cake.

533
00:28:53,040 --> 00:28:55,560
Speaker 4: That was a meme. That wasn't something that she said.

534
00:28:55,640 --> 00:28:58,880
Speaker 3: That was something that came out of this underground of

535
00:28:59,360 --> 00:29:03,200
Speaker 3: people just beating pamphlets. So you get the same benefits

536
00:29:03,200 --> 00:29:07,240
Speaker 3: and drawbacks. You get information being distributed really quickly to

537
00:29:07,320 --> 00:29:09,480
Speaker 3: many more people in a way that means that new

538
00:29:09,520 --> 00:29:13,360
Speaker 3: ideas can spread. But it also means that misinformation and

539
00:29:15,680 --> 00:29:18,719
Speaker 3: libel and other kinds of things you don't want can spread.

540
00:29:18,760 --> 00:29:21,240
Speaker 3: And I think we see this happening with the current

541
00:29:21,600 --> 00:29:26,640
Speaker 3: systems as well. So lms are representing the humans you know,

542
00:29:26,680 --> 00:29:29,959
Speaker 3: who are putting You know that the soup just tastes

543
00:29:30,080 --> 00:29:32,040
Speaker 3: like what all the people have put in the soup,

544
00:29:32,560 --> 00:29:35,280
Speaker 3: And that means that if people are putting in things

545
00:29:35,320 --> 00:29:42,400
Speaker 3: that are racist or sexist, or just wrong or misinformation

546
00:29:42,840 --> 00:29:44,880
Speaker 3: or outrage, that's.

547
00:29:44,680 --> 00:29:47,040
Speaker 4: What's going to show up from the llms.

548
00:29:47,160 --> 00:29:50,160
Speaker 1: And of course that's no different than the Berkeley Library

549
00:29:50,200 --> 00:29:52,200
Speaker 1: in the sense that if people are writing books and

550
00:29:52,240 --> 00:29:56,600
Speaker 1: they understand something about planetary motion or dark energy or whatever,

551
00:29:56,640 --> 00:29:58,640
Speaker 1: and they write books on it, that's that's all we

552
00:29:58,720 --> 00:30:01,000
Speaker 1: have to draw from R two things.

553
00:30:01,080 --> 00:30:02,920
Speaker 4: What is it that they can't do well? What they

554
00:30:03,000 --> 00:30:05,320
Speaker 4: can't do is go out and find something new.

555
00:30:05,800 --> 00:30:07,920
Speaker 3: And that's the great thing that human beings can do,

556
00:30:07,960 --> 00:30:11,040
Speaker 3: and especially actually human children can do, is go out

557
00:30:11,080 --> 00:30:13,959
Speaker 3: into the world say this is what everyone's told me

558
00:30:14,080 --> 00:30:16,400
Speaker 3: is true, but you know what, I'm not sure it

559
00:30:16,440 --> 00:30:18,760
Speaker 3: is true. Let me go and find out something new.

560
00:30:19,000 --> 00:30:22,040
Speaker 3: That's what kids do, that's what teenagers do. That's what

561
00:30:22,120 --> 00:30:26,480
Speaker 3: scientists do. Go out in the world, revise, change, think

562
00:30:26,520 --> 00:30:29,040
Speaker 3: about things in new ways, find out something new about

563
00:30:29,040 --> 00:30:29,400
Speaker 3: the world.

564
00:30:29,600 --> 00:30:30,600
Speaker 4: And that's exactly the.

565
00:30:30,520 --> 00:30:32,880
Speaker 3: Thing that ellms can't do. What elms can do is

566
00:30:32,880 --> 00:30:35,440
Speaker 3: summarize what all the other humans have said. They can't

567
00:30:35,480 --> 00:30:40,320
Speaker 3: go out and find something that's not just not just

568
00:30:40,320 --> 00:30:43,000
Speaker 3: an extrapolation from what people have already done.

569
00:30:44,560 --> 00:30:47,560
Speaker 1: So there's something really interesting about this, because what elms

570
00:30:47,600 --> 00:30:50,600
Speaker 1: are great at doing, of course, is interpolating between different

571
00:30:50,680 --> 00:30:54,840
Speaker 1: things that have been said, and sometimes those are lucune,

572
00:30:54,880 --> 00:30:58,320
Speaker 1: those are holes that humans haven't explored in before for

573
00:30:58,360 --> 00:31:01,880
Speaker 1: whatever reason. So what I and some of my colleagues

574
00:31:01,880 --> 00:31:04,000
Speaker 1: has been doing for a while is, you know, pitching

575
00:31:04,040 --> 00:31:08,720
Speaker 1: these really strange scientific questions at it and saying, you know,

576
00:31:09,080 --> 00:31:11,080
Speaker 1: what's your hypothesis, and it comes up with something and

577
00:31:11,080 --> 00:31:15,040
Speaker 1: we say, okay, you think you know. More broadly, just

578
00:31:15,760 --> 00:31:18,400
Speaker 1: come up with something that could explain this, and you know,

579
00:31:18,560 --> 00:31:20,360
Speaker 1: and you keep pushing it and it comes up with

580
00:31:20,440 --> 00:31:23,680
Speaker 1: pretty creative things. And they're creative in the sense that

581
00:31:23,720 --> 00:31:27,080
Speaker 1: they are remixes of what it's taken in before. And

582
00:31:27,120 --> 00:31:31,320
Speaker 1: it can do something, you know, interpolating between points that

583
00:31:31,360 --> 00:31:32,120
Speaker 1: are already known.

584
00:31:32,600 --> 00:31:35,040
Speaker 2: Now I totally agree with you. Of course, what it

585
00:31:35,120 --> 00:31:35,840
Speaker 2: can't do.

586
00:31:36,280 --> 00:31:39,400
Speaker 1: Is think about something outside of the sphere of human knowledge,

587
00:31:39,440 --> 00:31:42,320
Speaker 1: which is what let's say, you know Einstein does when

588
00:31:42,320 --> 00:31:44,200
Speaker 1: he says, what if I'm writing on a photon and

589
00:31:44,240 --> 00:31:45,480
Speaker 1: what would things look like?

590
00:31:45,560 --> 00:31:45,840
Speaker 2: And so on.

591
00:31:45,960 --> 00:31:49,480
Speaker 1: It comes up with a special theory of relativity that is,

592
00:31:50,720 --> 00:31:53,560
Speaker 1: as you say, taking all the stuff that's coming forward

593
00:31:53,560 --> 00:31:55,600
Speaker 1: and saying, hey, maybe that's not right and there's a.

594
00:31:55,600 --> 00:31:57,400
Speaker 2: Completely different way to look at it.

595
00:31:57,920 --> 00:32:02,520
Speaker 1: And of course with that, Record Buyers is the ability

596
00:32:02,560 --> 00:32:04,960
Speaker 1: to not only come up with a new idea, but

597
00:32:05,000 --> 00:32:09,080
Speaker 1: then simulate that idea to its conclusions and say, oh,

598
00:32:09,120 --> 00:32:12,280
Speaker 1: you know, that explains things better than what we currently have.

599
00:32:12,480 --> 00:32:15,120
Speaker 3: And a very important piece of that is it also

600
00:32:15,240 --> 00:32:18,200
Speaker 3: involves going out and testing it, so you know, it

601
00:32:18,240 --> 00:32:21,760
Speaker 3: wasn't just Einstein, it was people going out and actually,

602
00:32:21,840 --> 00:32:25,760
Speaker 3: you know, measuring the eclipses. That meant that that theory

603
00:32:25,880 --> 00:32:29,320
Speaker 3: was confirmed. So one of the things that even little

604
00:32:29,360 --> 00:32:33,280
Speaker 3: kids do is test things, go out experiment. When kids

605
00:32:33,320 --> 00:32:35,680
Speaker 3: do it, we call it getting into everything. But we've

606
00:32:35,720 --> 00:32:38,520
Speaker 3: been recently doing a lot of work showing that even

607
00:32:38,560 --> 00:32:41,040
Speaker 3: when little kids are getting into everything, what they're actually

608
00:32:41,080 --> 00:32:44,040
Speaker 3: doing is trying to get data that they can use

609
00:32:44,120 --> 00:32:46,840
Speaker 3: to change what they think, to revise what they think.

610
00:32:46,920 --> 00:32:50,400
Speaker 3: And that's a very human kind of intelligence. The reason

611
00:32:50,480 --> 00:32:55,280
Speaker 3: why the llms hallucinate is what people often call it.

612
00:32:55,040 --> 00:32:57,600
Speaker 3: It's not that they're hallucinating, it's that they just don't

613
00:32:58,000 --> 00:33:00,920
Speaker 3: They're not designed to know the difference between truth and falsehood.

614
00:33:01,160 --> 00:33:04,240
Speaker 3: They're not their objective function, as they say. The thing

615
00:33:04,240 --> 00:33:06,720
Speaker 3: they're trying to do is not to get the truth.

616
00:33:06,880 --> 00:33:09,200
Speaker 3: It's not going out and doing an experiment. It's not

617
00:33:09,320 --> 00:33:12,560
Speaker 3: changing their minds. It's trying to get the best summary

618
00:33:12,640 --> 00:33:15,760
Speaker 3: of the things that they already have heard from they

619
00:33:15,800 --> 00:33:17,520
Speaker 3: already have heard from other human beings.

620
00:33:17,760 --> 00:33:20,640
Speaker 1: Now, just for clarification, everything that you and I are

621
00:33:20,680 --> 00:33:25,000
Speaker 1: talking about right now is with current large models. But

622
00:33:25,520 --> 00:33:27,680
Speaker 1: what a lot of people are talking about now is

623
00:33:27,800 --> 00:33:30,840
Speaker 1: the third wave. The next wave of AI is probably

624
00:33:30,880 --> 00:33:34,560
Speaker 1: going to be agents who are experimenting in the world,

625
00:33:34,600 --> 00:33:36,800
Speaker 1: who are doing things in the world and gathering data

626
00:33:36,800 --> 00:33:39,960
Speaker 1: that way, because clearly we've already done the common crawl,

627
00:33:40,320 --> 00:33:43,480
Speaker 1: where these AI agents have crawled everything that has been

628
00:33:43,520 --> 00:33:46,800
Speaker 1: written by humans and there's no more data to be had.

629
00:33:47,240 --> 00:33:48,560
Speaker 2: So the next step.

630
00:33:48,320 --> 00:33:52,720
Speaker 1: Is run experiments in the real world, try out hypotheses,

631
00:33:53,240 --> 00:33:55,800
Speaker 1: and so that's going to change everything again.

632
00:33:56,800 --> 00:33:59,440
Speaker 3: Yeah, I mean that's where the kids are like a

633
00:33:59,440 --> 00:34:02,800
Speaker 3: wonderful because that's what kids are doing. And you know,

634
00:34:02,840 --> 00:34:05,320
Speaker 3: you mentioned something that we think about something like Einstein

635
00:34:05,360 --> 00:34:07,960
Speaker 3: as being this big change. But think about I think

636
00:34:07,960 --> 00:34:11,839
Speaker 3: about this with my uh my grandchildren. For example, when

637
00:34:11,880 --> 00:34:15,920
Speaker 3: I look at a phone, right or a computer, what

638
00:34:16,000 --> 00:34:18,120
Speaker 3: I say to myself is okay, well you use.

639
00:34:18,520 --> 00:34:20,239
Speaker 4: You use a keyboard.

640
00:34:20,480 --> 00:34:22,600
Speaker 3: But then there's these other gimmicks about, like I could

641
00:34:22,640 --> 00:34:24,200
Speaker 3: talk to it or I could touch it.

642
00:34:24,200 --> 00:34:24,799
Speaker 4: It would work.

643
00:34:25,280 --> 00:34:31,279
Speaker 3: My grandchildren, even my eighteen month old grandchildren, think, oh no,

644
00:34:31,640 --> 00:34:33,440
Speaker 3: this is a system that you talk to and you

645
00:34:33,520 --> 00:34:35,680
Speaker 3: touch and then things happen when you talk. You know,

646
00:34:35,719 --> 00:34:37,920
Speaker 3: if you actually have this little physical object and you

647
00:34:37,960 --> 00:34:40,800
Speaker 3: talk to it and touch it, you're going to get effects.

648
00:34:41,080 --> 00:34:42,960
Speaker 3: And they think that the you know, the keyboard is like,

649
00:34:43,040 --> 00:34:47,000
Speaker 3: what is this? This is this weird, strange, awkward thing

650
00:34:47,239 --> 00:34:49,040
Speaker 3: that you know, peripheral device.

651
00:34:49,080 --> 00:34:50,440
Speaker 4: We don't want to have anything to do with that.

652
00:34:50,960 --> 00:34:54,760
Speaker 3: Eighteen month olds, right aren't reading, let alone using a keyboard,

653
00:34:55,040 --> 00:34:57,840
Speaker 3: so they're even just you know, the difference between my

654
00:34:58,320 --> 00:35:01,440
Speaker 3: vision of what a computer is and what my eighteen

655
00:35:01,440 --> 00:35:08,120
Speaker 3: month old grandson's vision is is already a really radically

656
00:35:08,120 --> 00:35:08,840
Speaker 3: different vision.

657
00:35:09,200 --> 00:35:09,680
Speaker 2: That's right.

658
00:35:09,800 --> 00:35:13,319
Speaker 1: So the long arc of moral progress, as well as

659
00:35:13,400 --> 00:35:18,200
Speaker 1: the arc of human knowledge is all about is all

660
00:35:18,239 --> 00:35:20,640
Speaker 1: about us passing on to the next generation. Hey, here's

661
00:35:20,680 --> 00:35:23,000
Speaker 1: the things we've learned exactly, and then they pick it

662
00:35:23,080 --> 00:35:25,080
Speaker 1: up and they springboard right off of that.

663
00:35:25,480 --> 00:35:27,399
Speaker 3: So that's going to be you know, if you look

664
00:35:27,440 --> 00:35:29,600
Speaker 3: at the difference between how much progress has been made

665
00:35:29,600 --> 00:35:33,120
Speaker 3: with the LMS and something like robotics, where there's progress,

666
00:35:33,120 --> 00:35:36,880
Speaker 3: but it's much slower and still very very far removed

667
00:35:36,920 --> 00:35:41,719
Speaker 3: from what every baby's doing. You know, I do think

668
00:35:42,040 --> 00:35:44,719
Speaker 3: and we've been thinking about whether you could design a

669
00:35:44,760 --> 00:35:50,160
Speaker 3: system that, for instance, let me describe this one of

670
00:35:50,200 --> 00:35:54,840
Speaker 3: the kinds of AI systems, not an LM. But a

671
00:35:54,880 --> 00:35:58,520
Speaker 3: really different approach is what's called reinforcement learning. And reinforcement

672
00:35:58,600 --> 00:36:00,720
Speaker 3: learning is when a system does something and then learns

673
00:36:00,719 --> 00:36:03,360
Speaker 3: from the outcomes of what it does. But that's still

674
00:36:03,640 --> 00:36:09,000
Speaker 3: very very labor and computation intensive. It's hard, it's hard

675
00:36:09,040 --> 00:36:12,319
Speaker 3: to do efficiently, and it has a big problem, which

676
00:36:12,360 --> 00:36:15,120
Speaker 3: is in reinforcement learning, what those systems are doing is

677
00:36:15,160 --> 00:36:17,439
Speaker 3: just trying to, say, get a big higher score in game.

678
00:36:17,520 --> 00:36:20,320
Speaker 3: So reinforcement learning is how they solved go and chess,

679
00:36:20,320 --> 00:36:22,719
Speaker 3: So you can say, okay, you want to win the game.

680
00:36:22,719 --> 00:36:24,080
Speaker 3: What do you have to do to win the game.

681
00:36:24,480 --> 00:36:26,480
Speaker 3: But of course a lot of what children are doing

682
00:36:26,600 --> 00:36:27,880
Speaker 3: is just trying.

683
00:36:27,600 --> 00:36:29,040
Speaker 4: To figure out how the world works.

684
00:36:29,120 --> 00:36:31,360
Speaker 3: It doesn't really matter whether you're winning or losing, or

685
00:36:31,920 --> 00:36:35,400
Speaker 3: you're getting things or not. It's you just want to

686
00:36:35,440 --> 00:36:37,719
Speaker 3: figure out how the world works. So one thing we've

687
00:36:37,760 --> 00:36:39,640
Speaker 3: been doing is saying, what would happen if you had

688
00:36:39,640 --> 00:36:42,480
Speaker 3: a reinforcement learning system and instead of trying to get

689
00:36:42,480 --> 00:36:44,920
Speaker 3: a high score, it was trying to get more information,

690
00:36:45,280 --> 00:36:47,840
Speaker 3: or it was trying to figure out how to be

691
00:36:47,920 --> 00:36:50,640
Speaker 3: more effective or figure out how cause and effect worked.

692
00:36:50,920 --> 00:36:53,520
Speaker 3: Would that be a better way of describing that certainly

693
00:36:53,560 --> 00:36:55,560
Speaker 3: is much closer to what the kids are doing. But

694
00:36:55,680 --> 00:36:58,160
Speaker 3: part of the problem is that in the systems that

695
00:36:58,200 --> 00:37:00,480
Speaker 3: we have now you know, you were mentioning, well, you

696
00:37:00,520 --> 00:37:03,680
Speaker 3: could go to a chat subutee and say, okay, give

697
00:37:03,719 --> 00:37:06,960
Speaker 3: me five different ways of answering this question. But the

698
00:37:07,000 --> 00:37:11,920
Speaker 3: way that they currently are generating that kind of variability

699
00:37:11,920 --> 00:37:14,560
Speaker 3: and novelty is basically just by being random, just by

700
00:37:15,440 --> 00:37:18,080
Speaker 3: what they sometimes call turning up the temperature, just doing

701
00:37:18,120 --> 00:37:21,480
Speaker 3: things that are more different from one another. They're not

702
00:37:21,960 --> 00:37:26,160
Speaker 3: able to say, Okay, this is a plausible answer, this

703
00:37:26,239 --> 00:37:28,160
Speaker 3: could be the right answer, and this is just you know,

704
00:37:28,719 --> 00:37:31,920
Speaker 3: completely irrelevant. That's part of the reason why they hallucinate

705
00:37:32,080 --> 00:37:35,600
Speaker 3: is because they'll generate something and it seems like it

706
00:37:35,640 --> 00:37:38,480
Speaker 3: could be potentially be an answer, and they're not evaluating

707
00:37:38,600 --> 00:37:40,960
Speaker 3: does this make sense or not? And that's something that

708
00:37:41,400 --> 00:37:43,680
Speaker 3: we know that kids are doing, and we don't have

709
00:37:43,719 --> 00:37:45,160
Speaker 3: a very good account of.

710
00:37:45,120 --> 00:37:45,680
Speaker 4: How they do it.

711
00:37:45,800 --> 00:37:48,680
Speaker 3: So, you know, this is the old This shows how

712
00:37:48,760 --> 00:37:51,799
Speaker 3: old I am. There used to be a show about

713
00:37:51,840 --> 00:37:54,360
Speaker 3: you know, kids say the darkness things, and there's.

714
00:37:54,320 --> 00:37:58,400
Speaker 4: Slightly more expletive latent things like that on the internet.

715
00:37:58,719 --> 00:38:02,000
Speaker 3: Kids are always saying the creative strange things, but they're

716
00:38:02,000 --> 00:38:05,480
Speaker 3: not random, like they're not. It's not that they're just

717
00:38:06,040 --> 00:38:10,360
Speaker 3: randomly saying saying things that make no sense. They're really

718
00:38:10,480 --> 00:38:14,000
Speaker 3: different from what a grown up would ever say, but

719
00:38:14,040 --> 00:38:16,720
Speaker 3: they're not, but they kind of make sense. A lovely

720
00:38:16,760 --> 00:38:19,880
Speaker 3: example recently is one of a post doc was taking

721
00:38:19,920 --> 00:38:22,319
Speaker 3: her child out for her four year old out for

722
00:38:22,320 --> 00:38:24,920
Speaker 3: a walk on the Berkeley campus and the campus has

723
00:38:24,960 --> 00:38:27,279
Speaker 3: a companyli it's a bell tower that's, you know, a

724
00:38:27,360 --> 00:38:29,600
Speaker 3: clock that's very high up on top of a tower.

725
00:38:30,200 --> 00:38:32,399
Speaker 3: And a little boy looked at it and said, there's

726
00:38:32,440 --> 00:38:33,640
Speaker 3: a clock up there.

727
00:38:34,040 --> 00:38:36,600
Speaker 4: And then he thought to himself and he said, why

728
00:38:36,640 --> 00:38:39,680
Speaker 4: did they put the clock up there? And he said

729
00:38:40,520 --> 00:38:43,800
Speaker 4: it must be so the students and the children couldn't

730
00:38:43,800 --> 00:38:47,600
Speaker 4: break it, which is of course a wonderful explanation, right,

731
00:38:47,600 --> 00:38:48,880
Speaker 4: like you put it up really.

732
00:38:48,719 --> 00:38:52,280
Speaker 3: High so it'll be out of reach of the students

733
00:38:52,280 --> 00:38:55,200
Speaker 3: and they can't go and break the valuable clock. Not

734
00:38:55,320 --> 00:38:57,960
Speaker 3: something that a grown up would think of, but something

735
00:38:58,000 --> 00:39:02,239
Speaker 3: that's kind of plausible. And it's that kind of capacity

736
00:39:02,320 --> 00:39:05,799
Speaker 3: to generate something that isn't random but makes sense that

737
00:39:05,920 --> 00:39:08,160
Speaker 3: kids are really really good at doing. And we've done

738
00:39:08,239 --> 00:39:10,279
Speaker 3: a bunch of experiments at show. In some cases they're

739
00:39:10,320 --> 00:39:14,120
Speaker 3: better at doing that than grown ups are. But that's

740
00:39:14,120 --> 00:39:17,759
Speaker 3: something that's still very much not part of what's available

741
00:39:17,800 --> 00:39:22,240
Speaker 3: in AI. And there's an old observation in AI, sometimes

742
00:39:22,320 --> 00:39:27,000
Speaker 3: called Mirovich's paradox haunts Morovit was originally the person who

743
00:39:27,080 --> 00:39:29,480
Speaker 3: noticed it, which is that a lot of the things

744
00:39:29,480 --> 00:39:32,759
Speaker 3: that we as humans think are really really hard and

745
00:39:32,800 --> 00:39:36,680
Speaker 3: require a lot of intelligence, like playing chess, are actually

746
00:39:37,080 --> 00:39:40,040
Speaker 3: not that hard for AI systems. Things that we think

747
00:39:40,080 --> 00:39:42,800
Speaker 3: of is just taking for granted, like picking up a

748
00:39:42,920 --> 00:39:45,319
Speaker 3: chess piece and putting it on the board, turn out

749
00:39:45,320 --> 00:39:49,080
Speaker 3: to be a lot harder than playing chess. So the

750
00:39:49,120 --> 00:39:51,520
Speaker 3: thing that the kids can do, which is take a

751
00:39:51,560 --> 00:39:53,439
Speaker 3: bunch of you know, take a box full of mixed

752
00:39:53,480 --> 00:39:55,719
Speaker 3: up chess pieces and pull out the right ones and

753
00:39:55,719 --> 00:39:59,240
Speaker 3: put them in the right place, or have someone say,

754
00:39:59,760 --> 00:40:01,359
Speaker 3: are right, now, we're going to play chess, but we're

755
00:40:01,360 --> 00:40:03,560
Speaker 3: going to have a different role. You can move the

756
00:40:03,640 --> 00:40:07,360
Speaker 3: ponds as many spaces as you want. Those kinds of

757
00:40:07,440 --> 00:40:10,040
Speaker 3: abilities are exactly the ones that even the great chess

758
00:40:10,040 --> 00:40:12,680
Speaker 3: playing programs are going to have a hard time of doing.

759
00:40:28,280 --> 00:40:31,360
Speaker 1: That's right, and this is what I was mentioning before

760
00:40:31,440 --> 00:40:33,839
Speaker 1: about this third wave. So the first wave of AI

761
00:40:34,040 --> 00:40:38,239
Speaker 1: was really about reinforcement learning entirely. It was let's nail chess,

762
00:40:38,320 --> 00:40:41,120
Speaker 1: let's nail go, and it's all about let's play millions

763
00:40:41,160 --> 00:40:43,959
Speaker 1: of games new reinforcement. Then the second wave of AI

764
00:40:44,080 --> 00:40:46,480
Speaker 1: was totally different. It was, Hey, let's just absorb everything

765
00:40:46,520 --> 00:40:49,080
Speaker 1: that's out there. And so the third wave is really

766
00:40:49,280 --> 00:40:52,040
Speaker 1: something like becoming closer to a child than interacting with

767
00:40:52,080 --> 00:40:56,160
Speaker 1: the world and getting that feedback, right, And that's the future.

768
00:40:56,160 --> 00:40:58,279
Speaker 1: I mean, we don't have that yet in twenty twenty five,

769
00:40:58,360 --> 00:41:00,640
Speaker 1: but it'll be great to revi is that this part

770
00:41:00,680 --> 00:41:04,440
Speaker 1: of the conversation in twenty twenty eight and see where

771
00:41:04,480 --> 00:41:05,719
Speaker 1: things are and what they look like.

772
00:41:06,120 --> 00:41:08,880
Speaker 3: Yeah, I mean it maybe that we're headed for another

773
00:41:08,960 --> 00:41:12,400
Speaker 3: you know, the famous AI springs at AI Winters. My

774
00:41:13,000 --> 00:41:16,120
Speaker 3: intuition is that we're getting to the limits of what

775
00:41:16,239 --> 00:41:19,120
Speaker 3: the LLM cultural technology piece can do now that will

776
00:41:19,200 --> 00:41:21,920
Speaker 3: change the world, There's no question about it. And in

777
00:41:21,960 --> 00:41:25,120
Speaker 3: a way, you know the fact that kids can get

778
00:41:25,120 --> 00:41:27,759
Speaker 3: on the internet and find out kids anyone can pick

779
00:41:27,840 --> 00:41:30,040
Speaker 3: up something that's in their pocket and find out all

780
00:41:30,080 --> 00:41:32,200
Speaker 3: of this information about the world that is going to

781
00:41:32,600 --> 00:41:36,680
Speaker 3: change the world. But if we're thinking about intelligence, then

782
00:41:37,600 --> 00:41:39,279
Speaker 3: I think we still have a long way to go

783
00:41:39,320 --> 00:41:41,799
Speaker 3: to have something that has the kind of intelligence that

784
00:41:41,840 --> 00:41:42,920
Speaker 3: every child has.

785
00:41:42,719 --> 00:41:46,359
Speaker 1: Has in the AI being an intelligence being exactly right,

786
00:41:47,000 --> 00:41:47,640
Speaker 1: exactly right.

787
00:41:47,719 --> 00:41:48,960
Speaker 2: I'm curious what you think about this.

788
00:41:49,120 --> 00:41:51,400
Speaker 1: I've been sort of campaigning on this point for a

789
00:41:51,400 --> 00:41:53,799
Speaker 1: long time that I think the next generation is going

790
00:41:53,880 --> 00:41:57,279
Speaker 1: to be smarter than we are simply because of their

791
00:41:57,800 --> 00:42:02,120
Speaker 1: access to the world's knowledge. With the rectangle in their pocket,

792
00:42:01,480 --> 00:42:04,319
Speaker 1: the moment they're curious about a question and have the

793
00:42:04,360 --> 00:42:09,960
Speaker 1: right neurotransmitters present, they get the answer, and the exposure

794
00:42:10,680 --> 00:42:14,680
Speaker 1: to everything that's known is so incredible as far as

795
00:42:15,280 --> 00:42:17,759
Speaker 1: filling their warehouse so they can do remixes and come

796
00:42:17,840 --> 00:42:18,760
Speaker 1: up with new ideas.

797
00:42:18,920 --> 00:42:22,799
Speaker 2: What's your take hoting children or children growing up in

798
00:42:22,840 --> 00:42:23,960
Speaker 2: the digital children.

799
00:42:23,680 --> 00:42:24,640
Speaker 4: Growing up in the digital age.

800
00:42:24,680 --> 00:42:27,800
Speaker 3: Yeah, so I think that's a really good That's obviously

801
00:42:27,800 --> 00:42:31,000
Speaker 3: a really good, profound question. And I think again, if

802
00:42:31,040 --> 00:42:36,320
Speaker 3: you think about those past examples, like what did writing

803
00:42:36,600 --> 00:42:42,520
Speaker 3: and pictures and print and then the Internet itself, the

804
00:42:42,560 --> 00:42:45,760
Speaker 3: ways that that really changed in some ways that really

805
00:42:45,920 --> 00:42:51,520
Speaker 3: really beefed up human intelligence in really important significant ways.

806
00:42:51,600 --> 00:42:54,200
Speaker 3: The amount of things that's just the number of things

807
00:42:54,239 --> 00:42:56,680
Speaker 3: we know, right, the number of things we could know,

808
00:42:56,800 --> 00:42:58,280
Speaker 3: the number of things we could access.

809
00:42:58,360 --> 00:42:59,120
Speaker 4: All of those things.

810
00:42:59,280 --> 00:43:02,279
Speaker 3: Those culture technologies have really changed, and I think there's

811
00:43:02,320 --> 00:43:04,400
Speaker 3: every reason to believe that that will happen with the

812
00:43:04,440 --> 00:43:07,479
Speaker 3: new technologies as well. Now. At the same time, there's

813
00:43:07,520 --> 00:43:11,080
Speaker 3: always been these trade offs where other kinds of intelligence

814
00:43:11,120 --> 00:43:13,920
Speaker 3: that we had before, like the kind of intelligence that

815
00:43:13,960 --> 00:43:16,200
Speaker 3: you need to have to be a hunter gatherer, for example,

816
00:43:16,280 --> 00:43:18,719
Speaker 3: being able to be out in the world, take in

817
00:43:18,760 --> 00:43:22,680
Speaker 3: lots of information, go out and act, those kinds of

818
00:43:22,760 --> 00:43:27,200
Speaker 3: intelligences might suffer and probably will suffer as a result

819
00:43:27,280 --> 00:43:31,840
Speaker 3: of that. So, you know, being able to being able

820
00:43:31,880 --> 00:43:35,080
Speaker 3: to build a house, right, which is a really useful

821
00:43:35,080 --> 00:43:37,200
Speaker 3: thing to be able to do, is not something that

822
00:43:37,239 --> 00:43:40,680
Speaker 3: you can do just based on YouTube videos, although maybe

823
00:43:40,719 --> 00:43:45,000
Speaker 3: YouTube videos can maybe YouTube videos can help. My suspicion

824
00:43:45,120 --> 00:43:47,760
Speaker 3: is that what tends to happen with these technological changes

825
00:43:47,880 --> 00:43:51,080
Speaker 3: is they make all the difference in the world and

826
00:43:51,120 --> 00:43:53,640
Speaker 3: nobody realizes it because by the time they've made all

827
00:43:53,680 --> 00:43:56,000
Speaker 3: the difference, you just sort of take it for granted.

828
00:43:56,239 --> 00:43:58,600
Speaker 3: So an example that I like to give is as

829
00:43:58,640 --> 00:44:02,600
Speaker 3: I'm walking down the street. Now, I'm spending a lot

830
00:44:02,600 --> 00:44:06,280
Speaker 3: of time decoding text, So as I walk down the street,

831
00:44:06,600 --> 00:44:11,080
Speaker 3: I'm completely surrounded by these signs. You know, go call

832
00:44:11,160 --> 00:44:16,080
Speaker 3: our lawyer for your accident in case you're in an accident,

833
00:44:16,160 --> 00:44:19,080
Speaker 3: right or whatever, all the signs that are on the street.

834
00:44:19,520 --> 00:44:24,400
Speaker 3: I don't think to myself, God, it's exhausting every minute.

835
00:44:24,680 --> 00:44:26,560
Speaker 3: I never get to just walk down the street. I'm

836
00:44:26,600 --> 00:44:29,080
Speaker 3: always putting all this energy into trying to take all

837
00:44:29,080 --> 00:44:31,239
Speaker 3: this text and read it. And if you think about it,

838
00:44:31,280 --> 00:44:33,359
Speaker 3: if you were a pre literate person, it probably would

839
00:44:33,440 --> 00:44:35,280
Speaker 3: be exhausting if you had to sit down and figure

840
00:44:35,280 --> 00:44:37,080
Speaker 3: out what is it that each one of these signs

841
00:44:37,160 --> 00:44:40,240
Speaker 3: is saying. But of course we don't even I mean literally,

842
00:44:40,239 --> 00:44:43,000
Speaker 3: we're not even conscious of it. It's just part of

843
00:44:43,080 --> 00:44:45,839
Speaker 3: what goes on in the background. And we know from

844
00:44:46,080 --> 00:44:48,960
Speaker 3: neuroscience that in fact, parts of our brain have been

845
00:44:49,040 --> 00:44:52,040
Speaker 3: adapted to just do this kind of processing really quickly,

846
00:44:52,080 --> 00:44:54,080
Speaker 3: so we don't even notice that we're doing it. And

847
00:44:54,160 --> 00:44:56,320
Speaker 3: my suspicion is that that's what's going to happen with

848
00:44:57,760 --> 00:45:02,000
Speaker 3: the next generation and the the internet and information. We'll

849
00:45:02,040 --> 00:45:04,640
Speaker 3: be doing these things we won't even think about it.

850
00:45:04,680 --> 00:45:07,239
Speaker 3: But there's a really important caveat to that, which is

851
00:45:07,239 --> 00:45:10,120
Speaker 3: that if you go back to those examples of print

852
00:45:10,239 --> 00:45:16,000
Speaker 3: and writing, why is it that we didn't just have

853
00:45:16,239 --> 00:45:21,080
Speaker 3: all the evil misinformation and libel of the French Revolution indefinitely. Well,

854
00:45:21,120 --> 00:45:23,359
Speaker 3: what's happened is every time we've had one of these

855
00:45:23,440 --> 00:45:27,680
Speaker 3: new cultural technologies, we've also developed systems for keeping them

856
00:45:27,760 --> 00:45:32,040
Speaker 3: under control. So newspapers, you know, again we just sort

857
00:45:32,040 --> 00:45:35,040
Speaker 3: of take newspapers for granted, and newspapers are sort of disappearing,

858
00:45:35,320 --> 00:45:39,040
Speaker 3: But newspapers were away of taking that printed information and

859
00:45:39,120 --> 00:45:42,600
Speaker 3: curating it and having editors who said this is true,

860
00:45:42,640 --> 00:45:44,560
Speaker 3: and this isn't true, and this is something we want

861
00:45:44,600 --> 00:45:46,120
Speaker 3: to tell our readers, and this is something we don't

862
00:45:46,120 --> 00:45:48,799
Speaker 3: want to tell our readers. And you had norms that

863
00:45:48,920 --> 00:45:53,680
Speaker 3: developed things like journalism or journalism school or libel laws

864
00:45:53,719 --> 00:45:56,480
Speaker 3: that said, no, here's a way that we can take

865
00:45:56,520 --> 00:46:01,040
Speaker 3: this new culture and make control it in a way

866
00:46:01,080 --> 00:46:03,359
Speaker 3: that will let the good parts come out and keep

867
00:46:03,400 --> 00:46:06,239
Speaker 3: the bad parts under control. And every time there's a

868
00:46:06,239 --> 00:46:10,560
Speaker 3: new cultural technology, we've had new laws, we've had new norms,

869
00:46:10,600 --> 00:46:13,480
Speaker 3: we've had new institutions, we've had new kinds of people

870
00:46:13,840 --> 00:46:18,080
Speaker 3: who were there to try to make these institutions work

871
00:46:18,120 --> 00:46:19,839
Speaker 3: for and the same thing, by the way, it's true

872
00:46:19,840 --> 00:46:23,040
Speaker 3: for markets. You know, think about the way that markets

873
00:46:23,080 --> 00:46:25,960
Speaker 3: are great for coordinating people, but we all know all

874
00:46:26,000 --> 00:46:28,319
Speaker 3: the terrible things that can happen with markets do so

875
00:46:28,360 --> 00:46:30,920
Speaker 3: you need to have laws, you need to have institutions,

876
00:46:30,920 --> 00:46:33,239
Speaker 3: you need to have ways to regulate them. And I

877
00:46:33,239 --> 00:46:35,200
Speaker 3: think the same thing's going to be true with AI.

878
00:46:35,400 --> 00:46:38,040
Speaker 3: If we're going to succeed, we're both going to have

879
00:46:38,120 --> 00:46:38,560
Speaker 3: to have.

880
00:46:38,760 --> 00:46:41,440
Speaker 4: Internal you know, internal norms.

881
00:46:41,560 --> 00:46:43,960
Speaker 3: I think you see some of that happening with kids,

882
00:46:43,960 --> 00:46:47,279
Speaker 3: where kids will say things like, oh no, I don't

883
00:46:47,440 --> 00:46:49,520
Speaker 3: you know this is the wrong thing to pay attention to.

884
00:46:49,920 --> 00:46:52,080
Speaker 3: If you go to that YouTube site, it's full of

885
00:46:52,520 --> 00:46:54,960
Speaker 3: you know, it's full of nonsense. This one is actually

886
00:46:55,400 --> 00:46:59,239
Speaker 3: this one is actually better. And also we're going to

887
00:46:59,280 --> 00:47:02,440
Speaker 3: have to have all those boring things like legislation and

888
00:47:02,640 --> 00:47:08,759
Speaker 3: code and regulatory agencies that's already sort of starting that

889
00:47:08,800 --> 00:47:10,879
Speaker 3: will make sure that things are good rather than bad.

890
00:47:11,160 --> 00:47:14,319
Speaker 3: Another example, I like to give a great technology, technological

891
00:47:14,400 --> 00:47:17,240
Speaker 3: change that we don't think about very much. Think about

892
00:47:17,360 --> 00:47:19,879
Speaker 3: it's nineteen hundred and people are saying.

893
00:47:20,200 --> 00:47:21,000
Speaker 4: You know what we should do.

894
00:47:21,040 --> 00:47:23,080
Speaker 3: We should take all our wooden houses and we should

895
00:47:23,080 --> 00:47:25,880
Speaker 3: put electricity in them. So we should put wires that

896
00:47:25,920 --> 00:47:28,880
Speaker 3: have electricity, which we know burns things, and we should

897
00:47:28,880 --> 00:47:32,520
Speaker 3: put them in everybody's house, including everybody's wooden.

898
00:47:32,200 --> 00:47:35,200
Speaker 4: House, and it'll be fine, right, it'll be great.

899
00:47:35,760 --> 00:47:39,319
Speaker 3: Well. The only reason why we can do that, and

900
00:47:39,400 --> 00:47:41,279
Speaker 3: we did have a lot of houses burned down to

901
00:47:41,320 --> 00:47:44,600
Speaker 3: begin with, is because we have, as anyone who's you know,

902
00:47:44,640 --> 00:47:47,680
Speaker 3: anyone who's a contractor or has said a reno, Nos,

903
00:47:47,800 --> 00:47:50,040
Speaker 3: we have this thing called code, which is this book

904
00:47:50,160 --> 00:47:53,840
Speaker 3: that's like this thick that has all the rules about

905
00:47:53,960 --> 00:47:56,360
Speaker 3: here's what you have to do, here's how the wiring

906
00:47:56,400 --> 00:47:57,960
Speaker 3: has to work, here's all the things that you have

907
00:47:58,000 --> 00:48:01,520
Speaker 3: to do to make electricity work. And nobody thinks about

908
00:48:01,520 --> 00:48:05,440
Speaker 3: that big book of code as being a fantastic human invention.

909
00:48:05,560 --> 00:48:07,319
Speaker 3: We mostly think of it as, oh, God, that's why

910
00:48:07,360 --> 00:48:10,480
Speaker 3: my contractor is charging me so much. But that invention,

911
00:48:10,719 --> 00:48:12,840
Speaker 3: in a way, is just as important as the invention

912
00:48:12,880 --> 00:48:16,560
Speaker 3: of electricity itself. The invention of electricity itself was a

913
00:48:16,600 --> 00:48:21,600
Speaker 3: big scientific advance, but you couldn't use that unless you

914
00:48:21,600 --> 00:48:23,880
Speaker 3: had this other kind of advance, which was the code,

915
00:48:24,239 --> 00:48:27,640
Speaker 3: the laws, the regulations, the legislation. And I think that's

916
00:48:27,680 --> 00:48:30,680
Speaker 3: something that a lot of people in AI are realizing.

917
00:48:32,640 --> 00:48:35,400
Speaker 1: So when you're thinking about what will be the legislation

918
00:48:35,480 --> 00:48:37,919
Speaker 1: and the norms that are coming down the pike, when

919
00:48:37,920 --> 00:48:40,000
Speaker 1: you squint into the future, can you see what that's

920
00:48:40,040 --> 00:48:40,560
Speaker 1: going to look like?

921
00:48:41,360 --> 00:48:44,120
Speaker 3: I think that's a good question, and it will have

922
00:48:44,200 --> 00:48:46,680
Speaker 3: to be somewhat different. You know, each one of the

923
00:48:46,719 --> 00:48:48,920
Speaker 3: way that we deal with language, which is, you know,

924
00:48:49,000 --> 00:48:51,319
Speaker 3: sort of having a moral principle that you shouldn't lie,

925
00:48:51,360 --> 00:48:53,200
Speaker 3: for example, is different from what we have to do

926
00:48:53,239 --> 00:48:55,240
Speaker 3: with writing, is different from what we do with print,

927
00:48:55,680 --> 00:48:58,280
Speaker 3: is different from what we had to do with the Internet.

928
00:48:58,760 --> 00:48:59,719
Speaker 4: One of the points that.

929
00:48:59,680 --> 00:49:02,799
Speaker 3: We make in are in that paper that I think

930
00:49:02,920 --> 00:49:06,880
Speaker 3: is really important is that there are real dangers in

931
00:49:06,920 --> 00:49:10,960
Speaker 3: the fact that this technology has been monopolized by a

932
00:49:11,000 --> 00:49:14,040
Speaker 3: few big agencies, for example, a few big companies.

933
00:49:14,400 --> 00:49:15,200
Speaker 4: So one thing is.

934
00:49:15,160 --> 00:49:18,200
Speaker 3: How are we going to make sure that that it's democratized,

935
00:49:18,320 --> 00:49:22,520
Speaker 3: that in fact, people can use it outside of the

936
00:49:22,560 --> 00:49:25,120
Speaker 3: control of just a few big companies.

937
00:49:25,160 --> 00:49:28,040
Speaker 1: And because of just a quick interjection, I actually I

938
00:49:28,080 --> 00:49:29,960
Speaker 1: have no worries about that, and I'll tell you why.

939
00:49:30,000 --> 00:49:31,680
Speaker 2: It's because everything starts this way.

940
00:49:32,320 --> 00:49:35,080
Speaker 1: And you know, we saw when deep Sea came out

941
00:49:35,120 --> 00:49:39,000
Speaker 1: recently they had reportedly done it for six million bucks

942
00:49:39,040 --> 00:49:43,840
Speaker 1: instead of billions. And it's only going to get cheaper

943
00:49:43,920 --> 00:49:45,120
Speaker 1: and easier to do this.

944
00:49:46,280 --> 00:49:46,880
Speaker 2: So I think it.

945
00:49:46,880 --> 00:49:50,399
Speaker 1: Will quickly become democratized, just like every other example we've

946
00:49:50,440 --> 00:49:53,359
Speaker 1: had of things like this leg printing. I mean, now

947
00:49:53,400 --> 00:49:55,880
Speaker 1: we all have a printer on our desk at home, right,

948
00:49:56,080 --> 00:49:58,040
Speaker 1: whereas a printing press was a big.

949
00:49:57,880 --> 00:49:59,239
Speaker 2: Deal that only a few people had.

950
00:49:59,520 --> 00:50:01,760
Speaker 4: Yeah, well, I think that's partly true.

951
00:50:01,800 --> 00:50:04,919
Speaker 3: But on the other hand, the fact that these all

952
00:50:04,960 --> 00:50:09,480
Speaker 3: depend on these pre trained systems that do take an

953
00:50:09,600 --> 00:50:13,279
Speaker 3: enormous amount of money to get started, and even with

954
00:50:13,400 --> 00:50:15,920
Speaker 3: deep Seak, it seems like part of what Deepseek was

955
00:50:15,920 --> 00:50:20,160
Speaker 3: doing was taking advantage of the fact that this pre

956
00:50:20,280 --> 00:50:22,880
Speaker 3: training had already been done by other systems.

957
00:50:22,920 --> 00:50:24,799
Speaker 4: So I think that will happen.

958
00:50:24,880 --> 00:50:28,360
Speaker 3: But I think the question about who will have control

959
00:50:28,840 --> 00:50:33,360
Speaker 3: is a really important political and economic question. Here's another

960
00:50:33,560 --> 00:50:37,680
Speaker 3: interesting point that my colleague Henry Ferrell made in that

961
00:50:38,080 --> 00:50:41,200
Speaker 3: has made and we made in that science piece. If

962
00:50:41,200 --> 00:50:47,560
Speaker 3: you think about cultural technologies they intrinsically always involve attention

963
00:50:48,200 --> 00:50:52,560
Speaker 3: between the creators and the distributors. So the ideas cultural

964
00:50:52,600 --> 00:50:55,839
Speaker 3: technology is I can get information from other people. That's

965
00:50:55,920 --> 00:50:59,799
Speaker 3: essentially the idea, But that means someone has to make

966
00:50:59,840 --> 00:51:02,319
Speaker 3: up that information, someone has to generate it, someone has

967
00:51:02,360 --> 00:51:04,440
Speaker 3: to find out what's true, for example.

968
00:51:05,200 --> 00:51:09,960
Speaker 4: And there's this kind of paradoxical synergy between the people.

969
00:51:09,680 --> 00:51:12,200
Speaker 3: Who are out there creating things that are new and

970
00:51:12,239 --> 00:51:14,480
Speaker 3: then the people who are distributing them. So there's no

971
00:51:14,600 --> 00:51:17,399
Speaker 3: point in writing, as you and I both know as writers, right,

972
00:51:17,400 --> 00:51:19,880
Speaker 3: there's no point. Well, or maybe there's some point, but

973
00:51:19,880 --> 00:51:21,759
Speaker 3: there's not much point in writing a book unless you

974
00:51:21,760 --> 00:51:24,279
Speaker 3: have a publisher who can actually get it out into

975
00:51:24,280 --> 00:51:24,640
Speaker 3: the world.

976
00:51:24,840 --> 00:51:25,000
Speaker 4: Right.

977
00:51:26,160 --> 00:51:29,120
Speaker 3: But for the publisher, there's no point in being a

978
00:51:29,160 --> 00:51:32,239
Speaker 3: publisher unless you can get people to write books for you,

979
00:51:32,360 --> 00:51:35,120
Speaker 3: unless you can get new content, new ideas.

980
00:51:36,280 --> 00:51:37,800
Speaker 4: But that also means that there's.

981
00:51:37,560 --> 00:51:41,879
Speaker 3: This intrinsic tension, economic tension between who's going to pay right,

982
00:51:42,000 --> 00:51:44,319
Speaker 3: So for the distributors, it's always going to be in

983
00:51:44,360 --> 00:51:46,480
Speaker 3: their interest for the creators.

984
00:51:46,000 --> 00:51:48,840
Speaker 4: To get paid as little as possible.

985
00:51:48,680 --> 00:51:50,400
Speaker 3: And for the creators it's always going to be in

986
00:51:50,440 --> 00:51:51,920
Speaker 3: their interests for the distributors to.

987
00:51:51,920 --> 00:51:53,640
Speaker 4: Get paid as little as possible.

988
00:51:53,960 --> 00:51:56,440
Speaker 3: So if you're thinking about that from an economic perspective,

989
00:51:56,440 --> 00:51:59,280
Speaker 3: there's always going to be this tension between the people

990
00:51:59,280 --> 00:52:02,160
Speaker 3: who are creating and the people who are distributing. And

991
00:52:02,200 --> 00:52:04,960
Speaker 3: you can already see that in what's happened in journalism,

992
00:52:04,960 --> 00:52:08,880
Speaker 3: for instance, or what's happened to the death of local newspapers.

993
00:52:09,239 --> 00:52:13,160
Speaker 3: We had this kind of strange equilibrium where you know,

994
00:52:13,360 --> 00:52:16,400
Speaker 3: the want ads and the weather reports paid for the

995
00:52:16,440 --> 00:52:22,480
Speaker 3: investigative journalism and the arts criticism. And that's once that

996
00:52:22,560 --> 00:52:25,400
Speaker 3: two thousand digital convergence happened, that wasn't going to be

997
00:52:25,400 --> 00:52:26,200
Speaker 3: there anymore.

998
00:52:26,320 --> 00:52:27,840
Speaker 4: And now it's.

999
00:52:27,960 --> 00:52:30,879
Speaker 3: Really sort of up for grabs about who is how

1000
00:52:30,920 --> 00:52:34,319
Speaker 3: people are going to get compensated, and who is going

1001
00:52:34,360 --> 00:52:37,480
Speaker 3: to have the power of the distributors or the creators,

1002
00:52:37,520 --> 00:52:39,400
Speaker 3: And I think that's going to be a really important

1003
00:52:39,440 --> 00:52:41,239
Speaker 3: issue as we're going forward with this too.

1004
00:52:41,560 --> 00:52:44,080
Speaker 1: One idea that people have been floating recently is to

1005
00:52:44,200 --> 00:52:47,960
Speaker 1: have digital dividends come back to the creators, so that

1006
00:52:48,000 --> 00:52:51,480
Speaker 1: the people whose work has been used to train the

1007
00:52:51,520 --> 00:52:55,720
Speaker 1: models and maybe provided point one percent of the answer

1008
00:52:55,800 --> 00:52:58,600
Speaker 1: that that model just gave gets a few pennies back

1009
00:52:59,000 --> 00:53:04,320
Speaker 1: each time. That's idea how those economics models will evolve

1010
00:53:05,000 --> 00:53:05,920
Speaker 1: over the coming year.

1011
00:53:06,280 --> 00:53:07,920
Speaker 3: I mean, if you think about it, like what we

1012
00:53:07,960 --> 00:53:11,200
Speaker 3: did for print was we invented copyright law, so we

1013
00:53:11,239 --> 00:53:14,120
Speaker 3: have a whole lot of laws about Look, you can't,

1014
00:53:14,560 --> 00:53:16,959
Speaker 3: you know, I can't just take a David Eagleman book

1015
00:53:17,000 --> 00:53:19,640
Speaker 3: and copy it and say that it's mine. But of

1016
00:53:19,680 --> 00:53:24,319
Speaker 3: course that involves the actual text. So copyright depends on

1017
00:53:24,360 --> 00:53:26,799
Speaker 3: the fact that you're taking word for words something that

1018
00:53:26,800 --> 00:53:30,279
Speaker 3: someone else has written. How about if Google decides to

1019
00:53:30,360 --> 00:53:33,440
Speaker 3: use David Eagleman's book to train its large language model.

1020
00:53:33,800 --> 00:53:37,080
Speaker 3: It's not taking every single word, but the reason why

1021
00:53:37,160 --> 00:53:40,239
Speaker 3: it's giving you good answers about neuroscience is because it's

1022
00:53:40,239 --> 00:53:42,560
Speaker 3: been trained on those books. And we know as a

1023
00:53:42,560 --> 00:53:45,840
Speaker 3: matter of fact that those large models have been trained

1024
00:53:45,920 --> 00:53:49,239
Speaker 3: on my books and your books and all the other

1025
00:53:49,239 --> 00:53:50,200
Speaker 3: books that are out there.

1026
00:53:50,280 --> 00:53:50,440
Speaker 4: Now.

1027
00:53:50,440 --> 00:53:54,759
Speaker 3: They're not copying it down exactly, but the fact that

1028
00:53:54,800 --> 00:53:56,799
Speaker 3: you could go to chat ChiPT and it could tell

1029
00:53:56,800 --> 00:54:01,520
Speaker 3: you here's what Alison Govnik would say about about children's learning,

1030
00:54:01,840 --> 00:54:04,719
Speaker 3: comes because it's gone out there and used all that

1031
00:54:04,800 --> 00:54:06,280
Speaker 3: information in all of those books.

1032
00:54:06,320 --> 00:54:07,880
Speaker 4: So it's really tricky.

1033
00:54:08,120 --> 00:54:08,600
Speaker 2: Yeah.

1034
00:54:08,960 --> 00:54:10,920
Speaker 1: What I think is really tricky about this is the

1035
00:54:11,000 --> 00:54:12,960
Speaker 1: reason you and I are able to write books is

1036
00:54:12,960 --> 00:54:16,920
Speaker 1: because we've sat in thousands of lectures, we've heard ideas

1037
00:54:16,960 --> 00:54:18,480
Speaker 1: from people, and we've put things.

1038
00:54:18,239 --> 00:54:20,280
Speaker 2: Together in our own way and our own voice.

1039
00:54:20,960 --> 00:54:23,920
Speaker 1: And in a sense, the lllm's doing what humans do

1040
00:54:24,000 --> 00:54:25,120
Speaker 1: in terms of creativity.

1041
00:54:25,160 --> 00:54:27,080
Speaker 2: We're remixing all of our inputs.

1042
00:54:27,320 --> 00:54:29,960
Speaker 4: Yeah, but of course we wouldn't.

1043
00:54:30,840 --> 00:54:33,239
Speaker 3: Our books wouldn't be worth reading if they were just

1044
00:54:33,320 --> 00:54:35,880
Speaker 3: summaries of everything that had gone before. So part of

1045
00:54:35,880 --> 00:54:38,680
Speaker 3: what we want is to have exactly this.

1046
00:54:39,360 --> 00:54:39,799
Speaker 4: People in.

1047
00:54:41,600 --> 00:54:45,160
Speaker 3: A field called cultural evolution actually study how is it

1048
00:54:45,200 --> 00:54:48,320
Speaker 3: that information comes from one generation to another.

1049
00:54:48,719 --> 00:54:52,080
Speaker 4: And we've studied this in children. It's really fascinating.

1050
00:54:52,200 --> 00:54:55,600
Speaker 3: In The Gardener and the Carpenter, which is my book

1051
00:54:55,640 --> 00:54:59,399
Speaker 3: about parents and children, looking at how children learn from

1052
00:54:59,400 --> 00:55:03,000
Speaker 3: other people, there's lots of examples in there about how

1053
00:55:03,120 --> 00:55:07,120
Speaker 3: children can take information from their parents, for example, but

1054
00:55:07,160 --> 00:55:10,080
Speaker 3: they don't just swallow it hole they think about it

1055
00:55:10,120 --> 00:55:14,279
Speaker 3: in new ways, they remix it. In cultural evolution they

1056
00:55:14,280 --> 00:55:17,239
Speaker 3: talk about this as the balance between imitation and innovation.

1057
00:55:17,719 --> 00:55:18,520
Speaker 4: So you need both.

1058
00:55:18,560 --> 00:55:20,440
Speaker 3: You need to be able to imitate the other people

1059
00:55:20,480 --> 00:55:23,000
Speaker 3: around you to make progress, but then you also want

1060
00:55:23,040 --> 00:55:25,359
Speaker 3: to do something that's not just what the other people

1061
00:55:25,400 --> 00:55:27,000
Speaker 3: around you have done. I mean, there'd be no point

1062
00:55:27,040 --> 00:55:30,320
Speaker 3: in imitating if somebody hadn't innovated.

1063
00:55:29,920 --> 00:55:31,160
Speaker 4: At some point in the past.

1064
00:55:31,600 --> 00:55:35,560
Speaker 3: And it's a really hard scientific problem about how do

1065
00:55:35,600 --> 00:55:36,800
Speaker 3: you get that balance.

1066
00:55:36,880 --> 00:55:38,759
Speaker 4: And what you see with children is that.

1067
00:55:38,719 --> 00:55:43,760
Speaker 3: If you give children information, they will They're very likely,

1068
00:55:43,880 --> 00:55:48,000
Speaker 3: especially little children, will kind of imitate you literally. You know,

1069
00:55:48,400 --> 00:55:51,359
Speaker 3: you produce a gesture and the kids will imitate it.

1070
00:55:51,840 --> 00:55:54,560
Speaker 3: But then they'll also change it depending on whether they

1071
00:55:54,600 --> 00:55:55,640
Speaker 3: think it makes sense or not.

1072
00:55:56,120 --> 00:56:00,399
Speaker 1: Yeah, And as you know, in the animal kingdom, see

1073
00:56:00,440 --> 00:56:04,919
Speaker 1: that all animals have this trade off between exploration and exploitation.

1074
00:56:05,080 --> 00:56:07,759
Speaker 1: So they'll spend eighty percent of a time exploiting the

1075
00:56:07,840 --> 00:56:11,120
Speaker 1: things that they know, as in under that rock, there

1076
00:56:11,120 --> 00:56:13,719
Speaker 1: are these grubs that I eat, and then they'll spend

1077
00:56:13,719 --> 00:56:16,960
Speaker 1: twenty percent of their time exploring other places that they

1078
00:56:16,960 --> 00:56:19,840
Speaker 1: haven't seen before. Across the animal kingdom, we see this

1079
00:56:19,880 --> 00:56:22,000
Speaker 1: exploration exploitation trade off, and it strikes me that the

1080
00:56:23,120 --> 00:56:28,160
Speaker 1: imitation innovation trade off is essentially the cognitive equivalent analogy

1081
00:56:28,200 --> 00:56:28,480
Speaker 1: of that.

1082
00:56:28,920 --> 00:56:31,760
Speaker 4: No I think that's exactly right. And what I've argued.

1083
00:56:31,560 --> 00:56:35,560
Speaker 3: Is that if you think about childhood right, childhood is

1084
00:56:35,680 --> 00:56:38,759
Speaker 3: very evolutionarily paradoxical, like why would we have this long

1085
00:56:38,800 --> 00:56:41,120
Speaker 3: time when not only are we not producing anything, but

1086
00:56:41,160 --> 00:56:47,520
Speaker 3: we're extracting resources from the other people around us. And

1087
00:56:47,560 --> 00:56:51,520
Speaker 3: there's a really interesting biological generalization which is that the

1088
00:56:51,600 --> 00:56:55,120
Speaker 3: smarter the adultonimal, the longer childhood it has. And that's

1089
00:56:55,360 --> 00:57:00,319
Speaker 3: very very general. It's mammals, birds, marsupials, even insects see

1090
00:57:00,320 --> 00:57:04,440
Speaker 3: this relationship between the length of childhood and the intelligency

1091
00:57:04,440 --> 00:57:08,560
Speaker 3: in the anthropomorphic sense of the adult, how good the

1092
00:57:08,600 --> 00:57:11,840
Speaker 3: adult is learning and figuring out the world. And I've

1093
00:57:11,880 --> 00:57:14,840
Speaker 3: made the argument that this is really an example of

1094
00:57:14,880 --> 00:57:18,800
Speaker 3: this explore exploit trade off. So childhood is evolution's way

1095
00:57:19,040 --> 00:57:22,840
Speaker 3: of resolving the explore exploit trade off, because what it

1096
00:57:22,880 --> 00:57:25,720
Speaker 3: does is it gives you this period of childhood.

1097
00:57:25,880 --> 00:57:26,560
Speaker 4: Well you don't have.

1098
00:57:26,520 --> 00:57:28,880
Speaker 3: To worry about exploiting, you don't actually have to worry

1099
00:57:28,920 --> 00:57:30,960
Speaker 3: about taking care of yourself and going out in the

1100
00:57:30,960 --> 00:57:34,360
Speaker 3: world and getting resources. As I always say, babies and

1101
00:57:34,360 --> 00:57:38,080
Speaker 3: young children have exactly one utility function, as the economists say,

1102
00:57:38,080 --> 00:57:40,320
Speaker 3: which is be as cute as you possibly can be,

1103
00:57:40,600 --> 00:57:43,160
Speaker 3: and they're very very very good at maximizing that.

1104
00:57:43,440 --> 00:57:45,280
Speaker 4: And as long as you're as cute as you possibly

1105
00:57:45,280 --> 00:57:47,480
Speaker 4: can be, you don't have to worry about anything.

1106
00:57:47,240 --> 00:57:49,640
Speaker 3: Else, right, you don't have to worry about getting fed

1107
00:57:49,680 --> 00:57:50,560
Speaker 3: and taken care of.

1108
00:57:51,040 --> 00:57:52,800
Speaker 4: But what that means is that that.

1109
00:57:54,600 --> 00:57:57,440
Speaker 3: Frees you up to do the things that babies and

1110
00:57:57,480 --> 00:58:02,000
Speaker 3: young children do, which is play and explore and experiment

1111
00:58:02,200 --> 00:58:06,600
Speaker 3: and have weird, crazy, imaginative pretend and do all these

1112
00:58:06,680 --> 00:58:09,480
Speaker 3: kinds of explore functions. And then because the children are

1113
00:58:09,520 --> 00:58:12,240
Speaker 3: doing that, then the adults can take advantage of all

1114
00:58:12,360 --> 00:58:16,680
Speaker 3: the novelty and innovation that the children are producing. So

1115
00:58:17,800 --> 00:58:20,280
Speaker 3: and it's interesting that in the computer science literature they

1116
00:58:20,280 --> 00:58:22,720
Speaker 3: also talk a lot about this explore exploit trade off,

1117
00:58:23,080 --> 00:58:26,640
Speaker 3: and there's literally proofs that you can't. It's a trade off.

1118
00:58:26,640 --> 00:58:28,760
Speaker 3: You can't have both of them at the same time.

1119
00:58:29,240 --> 00:58:33,000
Speaker 3: And often the best solution is, and you know, I

1120
00:58:33,040 --> 00:58:36,000
Speaker 3: think people can feel this intuitively, is start out exploring

1121
00:58:36,360 --> 00:58:39,240
Speaker 3: when you're trying to solve.

1122
00:58:38,200 --> 00:58:38,840
Speaker 4: A new problem.

1123
00:58:39,200 --> 00:58:43,720
Speaker 3: Start out just brainstorming, having crazy ideas, not worrying too

1124
00:58:43,760 --> 00:58:46,680
Speaker 3: much about whether you're getting there or not. And then

1125
00:58:47,000 --> 00:58:49,960
Speaker 3: narrow in and say, Okay, now I have a solution.

1126
00:58:50,040 --> 00:58:51,680
Speaker 3: Now I want to figure out how do I find

1127
00:58:51,720 --> 00:58:56,440
Speaker 3: tune that solution. And I think childhood is nature's way

1128
00:58:56,520 --> 00:59:01,520
Speaker 3: of implementing that idea. Explore first, explore later, and again,

1129
00:59:01,640 --> 00:59:04,960
Speaker 3: if you're thinking about these artificial intelligence systems, it's a

1130
00:59:05,000 --> 00:59:08,520
Speaker 3: really deep problem about how could you get design a

1131
00:59:08,560 --> 00:59:12,760
Speaker 3: system that can trade off those two capacities in an

1132
00:59:12,800 --> 00:59:13,720
Speaker 3: intelligent way?

1133
00:59:14,040 --> 00:59:16,959
Speaker 1: And you know, I think that's that's the third wave

1134
00:59:17,040 --> 00:59:19,640
Speaker 1: that's coming. Yeah, that's the part that we don't have currently,

1135
00:59:19,760 --> 00:59:21,120
Speaker 1: But that's what's saying next.

1136
00:59:21,280 --> 00:59:25,000
Speaker 3: What's interesting is in humans, the way that we do

1137
00:59:25,080 --> 00:59:28,920
Speaker 3: it is because we have a particular what biologists call

1138
00:59:29,000 --> 00:59:33,120
Speaker 3: life history, We have a particular developmental history. We start

1139
00:59:33,120 --> 00:59:36,960
Speaker 3: out being children, we become adults. Then for humans, we

1140
00:59:37,040 --> 00:59:40,320
Speaker 3: become elders, we become post mental pausal grandmothers. Now we

1141
00:59:40,440 --> 00:59:44,280
Speaker 3: tend to think, and you hear a lot of this

1142
00:59:44,600 --> 00:59:49,560
Speaker 3: from AI guys, with guys being the relevant term that

1143
00:59:49,680 --> 00:59:53,680
Speaker 3: somehow the thirty five year old, the thirty five year

1144
00:59:53,680 --> 00:59:57,880
Speaker 3: old adult male is like the peak of the psychologist

1145
00:59:58,000 --> 01:00:00,439
Speaker 3: or the philosopher or the AI guy is the peak

1146
01:00:00,520 --> 01:00:03,960
Speaker 3: of human intelligence. And then there's this thing called intelligence

1147
01:00:03,960 --> 01:00:05,520
Speaker 3: that you can have a little of or a lot of,

1148
01:00:05,560 --> 01:00:08,280
Speaker 3: and the thirty five year olds or maybe even the

1149
01:00:08,320 --> 01:00:10,400
Speaker 3: twenty five year olds like they have the most of it,

1150
01:00:10,480 --> 01:00:13,000
Speaker 3: and then childhood is just building up to it, and

1151
01:00:13,040 --> 01:00:15,840
Speaker 3: then elderhood is just falling off from it. But that

1152
01:00:15,880 --> 01:00:19,880
Speaker 3: doesn't make much sense from an evolutionary perspective. In fact,

1153
01:00:20,040 --> 01:00:22,160
Speaker 3: what seems to be true is that we have these

1154
01:00:22,280 --> 01:00:27,840
Speaker 3: different functions, really different, radically different kinds of intelligence that

1155
01:00:27,960 --> 01:00:30,720
Speaker 3: trade off. They're not just different, they trade off against

1156
01:00:30,800 --> 01:00:36,919
Speaker 3: one another, like exploration exploitation, and our developmental trajectory from

1157
01:00:37,000 --> 01:00:40,920
Speaker 3: childhood to adulthood to elderhood is a way that we

1158
01:00:41,040 --> 01:00:43,520
Speaker 3: managed to deal with those kinds of trade offs. So

1159
01:00:43,520 --> 01:00:47,520
Speaker 3: we have this childhood that lets us explore and lets

1160
01:00:47,600 --> 01:00:50,840
Speaker 3: us get information from other people. We have this adulthood

1161
01:00:50,880 --> 01:00:54,400
Speaker 3: where we can go out and use that information to

1162
01:00:54,480 --> 01:00:56,920
Speaker 3: do all the things that Gronim's do, like find mates

1163
01:00:56,960 --> 01:01:00,200
Speaker 3: and resources and find our way in the hierarchy. And

1164
01:01:00,240 --> 01:01:03,920
Speaker 3: then we have this elderhood where now we're motivated to

1165
01:01:04,960 --> 01:01:07,800
Speaker 3: use our resources to help the next generation, to pass

1166
01:01:07,880 --> 01:01:09,880
Speaker 3: on the wisdom and information that we've got to the

1167
01:01:09,920 --> 01:01:10,640
Speaker 3: next generation.

1168
01:01:11,160 --> 01:01:11,800
Speaker 4: So away.

1169
01:01:11,840 --> 01:01:15,920
Speaker 3: I put this perhaps a little meanly sometimes, is basically,

1170
01:01:15,960 --> 01:01:19,000
Speaker 3: we're humans up to puberty and after menopause, and in

1171
01:01:19,040 --> 01:01:22,000
Speaker 3: between we're sort of basically glorified primates. In between we're

1172
01:01:22,000 --> 01:01:25,640
Speaker 3: doing all those things like you know, mating and predating

1173
01:01:25,680 --> 01:01:30,480
Speaker 3: and finding resources and all the fun stuff is what

1174
01:01:30,520 --> 01:01:32,520
Speaker 3: the kids and the grandmom's fit to do, which is

1175
01:01:32,720 --> 01:01:38,080
Speaker 3: play and explore and tell stories and pass on recipes

1176
01:01:38,200 --> 01:01:40,880
Speaker 3: and do Broadway show tunes, all the things that I

1177
01:01:40,920 --> 01:01:44,800
Speaker 3: do as a grandmother with my with my grandchildren. But

1178
01:01:44,840 --> 01:01:46,880
Speaker 3: there is something that's deeper than that, which is that

1179
01:01:47,280 --> 01:01:51,640
Speaker 3: typically AI has not kind of had this developmental perspective.

1180
01:01:51,680 --> 01:01:54,880
Speaker 3: And I think that's another thing that developmental science can

1181
01:01:54,920 --> 01:01:58,280
Speaker 3: contribute to AI, is think not just about that there's

1182
01:01:58,320 --> 01:02:00,640
Speaker 3: this thing called intelligence that we want more of or

1183
01:02:00,720 --> 01:02:05,080
Speaker 3: less of, but how do you in a society and

1184
01:02:05,160 --> 01:02:08,760
Speaker 3: in an individual across time trade off these really different

1185
01:02:08,800 --> 01:02:10,000
Speaker 3: kinds of intelligences.

1186
01:02:12,200 --> 01:02:14,520
Speaker 1: So to zoom this back out to the big picture,

1187
01:02:14,560 --> 01:02:18,880
Speaker 1: then the idea of not looking at AI as an

1188
01:02:18,920 --> 01:02:22,439
Speaker 1: intelligent agent, but looking at it as a cultural technology.

1189
01:02:22,520 --> 01:02:24,480
Speaker 1: One of the things that suggests, and you say this

1190
01:02:24,520 --> 01:02:27,600
Speaker 1: in the paper, is that it's not just engineers who

1191
01:02:27,640 --> 01:02:30,360
Speaker 1: should be studying AI, but it's social scientists who should

1192
01:02:30,360 --> 01:02:31,240
Speaker 1: be commenting on it.

1193
01:02:31,560 --> 01:02:34,920
Speaker 4: Yeah, so just yeah, it's sort of fascinating.

1194
01:02:34,960 --> 01:02:37,680
Speaker 3: So if we wanted to say, let's try and figure

1195
01:02:37,680 --> 01:02:40,160
Speaker 3: out what's going to be good, what's going to be bad,

1196
01:02:40,280 --> 01:02:42,320
Speaker 3: how do we regulate it, how do we make.

1197
01:02:42,200 --> 01:02:43,000
Speaker 4: It good or bad?

1198
01:02:43,800 --> 01:02:46,240
Speaker 3: The people we should be talking to are people who

1199
01:02:46,280 --> 01:02:51,440
Speaker 3: know something about how print and writing and pictures affected society,

1200
01:02:51,640 --> 01:02:57,440
Speaker 3: or how developing markets and democracies and bureaucracies restructured society.

1201
01:02:57,520 --> 01:03:01,720
Speaker 3: And they all had enormous effects, right, I mean in

1202
01:03:01,760 --> 01:03:04,840
Speaker 3: a way like it's easy and fun to bring another

1203
01:03:04,880 --> 01:03:08,280
Speaker 3: intelligent agent into the universe. It's a lot more fun

1204
01:03:08,320 --> 01:03:10,400
Speaker 3: than coding. Actually, we know how to do that and

1205
01:03:10,400 --> 01:03:12,960
Speaker 3: we're doing it all the time, and it makes a

1206
01:03:12,960 --> 01:03:15,400
Speaker 3: little difference, but not an enormous difference. Bringing a new

1207
01:03:15,400 --> 01:03:20,880
Speaker 3: cultural technology into the universe like print, that's a giant change.

1208
01:03:21,080 --> 01:03:23,160
Speaker 3: But we've done it before. It's not as if this

1209
01:03:23,240 --> 01:03:25,600
Speaker 3: is a singularity. It's not as if this isn't something

1210
01:03:25,600 --> 01:03:28,240
Speaker 3: that humans have done before. The people who know about

1211
01:03:28,280 --> 01:03:32,360
Speaker 3: it are psychologists and political scientists and sociologists and historians

1212
01:03:32,360 --> 01:03:33,560
Speaker 3: of technology, and.

1213
01:03:33,560 --> 01:03:35,240
Speaker 4: One thing that I think would be really.

1214
01:03:35,000 --> 01:03:39,760
Speaker 3: Really helpful would be to get this historical perspective into

1215
01:03:39,880 --> 01:03:42,720
Speaker 3: the way that we think about AI, rather than having

1216
01:03:42,720 --> 01:03:46,400
Speaker 3: the sort of Gallum Frankenstein's story perspective of this is

1217
01:03:46,440 --> 01:03:48,760
Speaker 3: something that's unlike anything that's ever happened in.

1218
01:03:50,560 --> 01:03:51,560
Speaker 4: Humankind before.

1219
01:03:51,680 --> 01:03:56,040
Speaker 3: It's something that is important and really has had significant effects,

1220
01:03:56,040 --> 01:03:58,840
Speaker 3: but it is something that's happened in human life before,

1221
01:03:58,920 --> 01:04:01,400
Speaker 3: and it's something that comes out of the way that

1222
01:04:01,480 --> 01:04:04,880
Speaker 3: human beings work, not something that comes out of some

1223
01:04:05,080 --> 01:04:08,120
Speaker 3: strange alien inhuman kind of development.

1224
01:04:12,680 --> 01:04:15,800
Speaker 1: That was my interview with Alison Gopnik, professor at Berkeley.

1225
01:04:16,080 --> 01:04:18,240
Speaker 1: There are two points I want to return to. I've

1226
01:04:18,280 --> 01:04:21,880
Speaker 1: previously argued this point in Inner Cosmos in episode seventy two,

1227
01:04:22,320 --> 01:04:25,760
Speaker 1: that AI is best thought of as a processor of

1228
01:04:25,800 --> 01:04:30,720
Speaker 1: the intelligence of billions of humans. And that's because I

1229
01:04:30,760 --> 01:04:33,680
Speaker 1: was noticing, even back then, the number of people who

1230
01:04:33,720 --> 01:04:36,680
Speaker 1: typed in a sophisticated question and they got back what

1231
01:04:36,720 --> 01:04:39,720
Speaker 1: appeared to be a sophisticated answer, and they concluded, Wow,

1232
01:04:39,760 --> 01:04:44,200
Speaker 1: this thing is truly intelligent, but they were simply confusing

1233
01:04:44,280 --> 01:04:49,280
Speaker 1: that with the intellectual endeavors of humans before them. Maybe

1234
01:04:49,600 --> 01:04:52,360
Speaker 1: dozens of people had written about that topic, or maybe

1235
01:04:52,440 --> 01:04:55,920
Speaker 1: hundreds or thousands, but they simply didn't know that, and

1236
01:04:55,960 --> 01:05:00,400
Speaker 1: so they heard the echo and they misinterpreted that as

1237
01:05:00,480 --> 01:05:03,800
Speaker 1: the proud voice of AI. So I named this the

1238
01:05:04,040 --> 01:05:08,920
Speaker 1: intelligence echo illusion. Second point is, right after my conversation

1239
01:05:09,000 --> 01:05:11,600
Speaker 1: with Alison, I found myself thinking a lot about the question,

1240
01:05:12,200 --> 01:05:14,680
Speaker 1: when there's a giant machine that collects up data and

1241
01:05:14,720 --> 01:05:19,440
Speaker 1: processes it, why are we so quick to anthropomorphize it,

1242
01:05:19,520 --> 01:05:23,080
Speaker 1: to see it as a being. Well, it turns out

1243
01:05:23,120 --> 01:05:26,080
Speaker 1: that brains are very quick to do that. For example,

1244
01:05:26,480 --> 01:05:29,520
Speaker 1: imagine you hear some sound in your home at night.

1245
01:05:29,920 --> 01:05:33,840
Speaker 1: You assume it's a living, intentional creature, even though you

1246
01:05:33,920 --> 01:05:37,400
Speaker 1: might discover after some investigation with a flashlight in a

1247
01:05:37,440 --> 01:05:39,960
Speaker 1: baseball bat that it was just the wind blowing and

1248
01:05:40,040 --> 01:05:44,960
Speaker 1: knocking something around. Presumably this is a survival mechanism to

1249
01:05:45,400 --> 01:05:49,360
Speaker 1: assume everything is alive and has intention. But I think

1250
01:05:49,360 --> 01:05:52,040
Speaker 1: there's another aspect to it also, which is that we

1251
01:05:52,120 --> 01:05:55,880
Speaker 1: seem unable to think about complex systems and so we

1252
01:05:56,040 --> 01:06:00,800
Speaker 1: have to assign a central character to it. Examples of this,

1253
01:06:00,840 --> 01:06:02,840
Speaker 1: which I'm going to talk about in an episode in

1254
01:06:02,920 --> 01:06:05,240
Speaker 1: a few weeks from now, but for now, I'll just

1255
01:06:05,360 --> 01:06:08,439
Speaker 1: use as an example what's known is the Great Man

1256
01:06:08,680 --> 01:06:12,880
Speaker 1: view of history. We look to some historical figure, this

1257
01:06:12,880 --> 01:06:15,680
Speaker 1: could be a woman as well, and we attribute a

1258
01:06:15,920 --> 01:06:19,959
Speaker 1: historical outcome to that person. But this is a very

1259
01:06:20,000 --> 01:06:25,160
Speaker 1: misleading narrative tendency, because anything that happens historically is a

1260
01:06:25,360 --> 01:06:31,320
Speaker 1: hugely complex event, shaped by thousands of people and often institutions,

1261
01:06:31,360 --> 01:06:36,160
Speaker 1: and sometimes random contingencies. We say that Hitler led Germany

1262
01:06:36,200 --> 01:06:40,120
Speaker 1: into World War Two, but that glosses over the collaborators

1263
01:06:40,160 --> 01:06:44,440
Speaker 1: and the predecessors and the cultural scaffolding that made all

1264
01:06:44,480 --> 01:06:48,160
Speaker 1: these things possible. My assertion is that this isn't just

1265
01:06:48,200 --> 01:06:53,440
Speaker 1: a storytelling shortcut. It reflects something deep about human cognition.

1266
01:06:53,840 --> 01:06:57,280
Speaker 1: We're just not able to hold that level of complexity

1267
01:06:57,320 --> 01:07:02,240
Speaker 1: in our heads, so we become story tellers. We anthropomorphize

1268
01:07:02,360 --> 01:07:05,400
Speaker 1: to one or a few central characters. In other words,

1269
01:07:05,520 --> 01:07:09,640
Speaker 1: it's really hard for our brains to grasp sprawling, distributed

1270
01:07:09,680 --> 01:07:13,280
Speaker 1: networks of influence, so we turn a vast social tapestry

1271
01:07:13,640 --> 01:07:17,080
Speaker 1: into a single silhouette. Okay, so now let's zoom back

1272
01:07:17,080 --> 01:07:20,240
Speaker 1: out to the main picture. What happens when we stop

1273
01:07:20,400 --> 01:07:24,280
Speaker 1: asking whether a large model is intelligent and instead ask

1274
01:07:24,680 --> 01:07:28,640
Speaker 1: what kind of cultural machine it is. Today's conversation with

1275
01:07:28,720 --> 01:07:32,680
Speaker 1: Alison Gopnik took us through a reframing that I think

1276
01:07:32,800 --> 01:07:36,400
Speaker 1: is clarifying and possibly really important. When we think of

1277
01:07:36,600 --> 01:07:41,560
Speaker 1: large models as minds, as proto agents, it's easy to

1278
01:07:41,560 --> 01:07:43,920
Speaker 1: get swept up in a speculative drama.

1279
01:07:44,280 --> 01:07:45,080
Speaker 2: We ask if.

1280
01:07:44,920 --> 01:07:48,520
Speaker 1: They'll surpass us, if they'll enslave us. But when we

1281
01:07:48,960 --> 01:07:53,400
Speaker 1: shift the lens, when we see these systems as cultural technologies,

1282
01:07:53,760 --> 01:07:58,480
Speaker 1: more akin to libraries or markets, then our questions become

1283
01:07:59,000 --> 01:08:00,960
Speaker 1: more grounded and actionable.

1284
01:08:01,280 --> 01:08:03,080
Speaker 2: Large models are changing.

1285
01:08:02,800 --> 01:08:06,120
Speaker 1: Everything about our lives, how we write, how we seek information,

1286
01:08:06,240 --> 01:08:11,120
Speaker 1: how we work. They are reorganizing the structure of knowledge

1287
01:08:11,240 --> 01:08:15,520
Speaker 1: in real time, and like every major information technology before them,

1288
01:08:15,960 --> 01:08:20,240
Speaker 1: whether that's writing or printing or broadcast media, these tools

1289
01:08:20,240 --> 01:08:24,760
Speaker 1: are going to shape culture not because they think, but

1290
01:08:24,840 --> 01:08:28,800
Speaker 1: simply because they change the flow of ideas. So maybe

1291
01:08:28,840 --> 01:08:32,320
Speaker 1: the first step towards wise stewardship of this new era

1292
01:08:32,880 --> 01:08:38,400
Speaker 1: is to improve our metaphors less Golum and Frankenstein, more

1293
01:08:38,800 --> 01:08:44,840
Speaker 1: printing press and library, less digital brain, more public infrastructure,

1294
01:08:45,360 --> 01:08:47,560
Speaker 1: less will it become sentient?

1295
01:08:47,840 --> 01:08:51,800
Speaker 2: And more? What cultural shifts does this give rise to?

1296
01:08:52,200 --> 01:08:54,960
Speaker 1: There's a lot at stake and how we frame these systems,

1297
01:08:55,000 --> 01:08:59,320
Speaker 1: because our analogies and our metaphors are never neutral. If

1298
01:08:59,360 --> 01:09:03,599
Speaker 1: we imagine a large model as a proto, human will

1299
01:09:03,680 --> 01:09:06,400
Speaker 1: fear it and will regulate it accordingly. But if we

1300
01:09:06,479 --> 01:09:10,639
Speaker 1: see it as a new kind of social technology, something

1301
01:09:10,720 --> 01:09:14,599
Speaker 1: like a market or a library or a sprawling editorial system,

1302
01:09:15,080 --> 01:09:19,200
Speaker 1: we can draw on the history of how societies have

1303
01:09:19,360 --> 01:09:23,960
Speaker 1: dealt with the cultural impact of information systems. In other words,

1304
01:09:24,439 --> 01:09:29,000
Speaker 1: how we think about new technology will shape the world

1305
01:09:29,360 --> 01:09:35,920
Speaker 1: we're about to live into. Go to eagleman dot com

1306
01:09:35,920 --> 01:09:39,840
Speaker 1: slash podcast for more information and find further reading. Check

1307
01:09:39,840 --> 01:09:42,200
Speaker 1: out my newsletter on substack and be a part of

1308
01:09:42,240 --> 01:09:44,080
Speaker 1: the online conversation there.

1309
01:09:44,600 --> 01:09:45,880
Speaker 2: Finally, you can watch.

1310
01:09:45,680 --> 01:09:48,320
Speaker 1: Videos of Inner Cosmos on YouTube, where you can leave

1311
01:09:48,400 --> 01:09:54,960
Speaker 1: comments Until next time. I'm David Eagleman and this is

1312
01:09:55,080 --> 01:10:03,439
Speaker 1: inner Cosmos. You can help you to have to