1
00:00:05,160 --> 00:00:08,760
Speaker 1: Is it possible that we're thinking about intelligence in the

2
00:00:08,840 --> 00:00:14,080
Speaker 1: wrong way? Instead of being something inside individual brains, is

3
00:00:14,120 --> 00:00:19,400
Speaker 1: intelligence instead something that emerges from lots of brains that

4
00:00:19,440 --> 00:00:23,800
Speaker 1: are constantly working to align with one another. And if

5
00:00:23,840 --> 00:00:26,960
Speaker 1: we take on that lens, what does this mean about

6
00:00:26,960 --> 00:00:31,040
Speaker 1: the way that we can build AI agents or the

7
00:00:31,040 --> 00:00:33,960
Speaker 1: way that they can make us better? What is the

8
00:00:34,000 --> 00:00:38,959
Speaker 1: difference between information and information with a purpose? Today we're

9
00:00:38,960 --> 00:00:42,040
Speaker 1: going to speak with Daniel Persick, a cognitive scientist who

10
00:00:42,120 --> 00:00:47,080
Speaker 1: leads the Human Computer Interaction team at Amazon's AGI Lab.

11
00:00:47,440 --> 00:00:54,240
Speaker 1: So get ready for a great brain stretch. Welcome to

12
00:00:54,240 --> 00:00:57,680
Speaker 1: Intercosmos with me David Eagleman. I'm a neuroscientist and an

13
00:00:57,680 --> 00:01:00,640
Speaker 1: author at Stanford and in these episodes, as we sail

14
00:01:00,800 --> 00:01:04,160
Speaker 1: deeply into our three pound universe to understand how we

15
00:01:04,240 --> 00:01:08,520
Speaker 1: see the world and soon how AI might come to

16
00:01:08,640 --> 00:01:25,040
Speaker 1: understand the world with us. Let's think about the word intelligence.

17
00:01:25,760 --> 00:01:30,440
Speaker 1: You might justifiably assume that neuroscientists have an agreed upon

18
00:01:30,640 --> 00:01:34,800
Speaker 1: definition for this, but we actually don't. However one thinks

19
00:01:34,800 --> 00:01:38,880
Speaker 1: about intelligence, I think it's a fair assumption that most

20
00:01:38,880 --> 00:01:42,360
Speaker 1: of us, when we think about it, assume that intelligence

21
00:01:42,480 --> 00:01:47,000
Speaker 1: is something that happens inside a single head, in other words,

22
00:01:47,400 --> 00:01:52,400
Speaker 1: a brain processing information. This statement seems so obvious that

23
00:01:52,520 --> 00:01:56,680
Speaker 1: it hardly invites inspection, but if you step back and

24
00:01:56,720 --> 00:02:01,480
Speaker 1: look at how intelligence actually unfolds in a human life,

25
00:02:01,520 --> 00:02:06,240
Speaker 1: a different picture can start to emerge. Our thinking is

26
00:02:06,280 --> 00:02:09,919
Speaker 1: shaped by other people from the very beginning. We learn

27
00:02:10,080 --> 00:02:15,200
Speaker 1: by watching, by imitating, by trying to communicate, and eventually

28
00:02:15,240 --> 00:02:19,680
Speaker 1: by negotiating meaning with the people around us. Even our

29
00:02:19,680 --> 00:02:24,840
Speaker 1: most private thoughts are built from tools that are fundamentally social,

30
00:02:25,040 --> 00:02:30,239
Speaker 1: things like language and symbols and shared concepts and cultural norms.

31
00:02:30,560 --> 00:02:32,400
Speaker 1: So this may sound strange, but this is what we're

32
00:02:32,400 --> 00:02:34,760
Speaker 1: going to talk about today, and the idea will become

33
00:02:34,880 --> 00:02:39,720
Speaker 1: very clear. Most of humanity's greatest achievements didn't come from

34
00:02:39,840 --> 00:02:45,440
Speaker 1: lone geniuses working in isolation, but from really dense networks

35
00:02:45,600 --> 00:02:50,320
Speaker 1: of minds interacting over time. When we look at things

36
00:02:50,400 --> 00:02:54,440
Speaker 1: like science or art, or morality or technology, it almost

37
00:02:54,639 --> 00:02:59,800
Speaker 1: never makes sense to interpret these as products of individual intelligence,

38
00:02:59,800 --> 00:03:04,400
Speaker 1: but instead they are collective processes that allow ideas to

39
00:03:04,919 --> 00:03:09,520
Speaker 1: collide and to form into something and to continuously evolve.

40
00:03:10,280 --> 00:03:14,320
Speaker 1: So intelligence in this sense may be less like a

41
00:03:14,440 --> 00:03:20,600
Speaker 1: thing we possess and more like something that emerges between us. Now,

42
00:03:20,800 --> 00:03:26,320
Speaker 1: this broader perspective becomes especially important as we find ourselves

43
00:03:26,600 --> 00:03:31,760
Speaker 1: flinging headlong into the era of artificial intelligence. With every

44
00:03:31,840 --> 00:03:37,360
Speaker 1: passing week, we're getting AI acting more like a participant.

45
00:03:37,360 --> 00:03:41,360
Speaker 1: We're getting systems that communicate but also agents that act

46
00:03:41,440 --> 00:03:43,720
Speaker 1: on our behalf to do things in the world. And

47
00:03:43,960 --> 00:03:48,880
Speaker 1: soon these agents will collaborate with each other at their

48
00:03:49,000 --> 00:03:54,920
Speaker 1: time scales and spatial scales. So if intelligence is social

49
00:03:55,040 --> 00:03:59,560
Speaker 1: by nature, then building the future world of AI might

50
00:03:59,680 --> 00:04:04,440
Speaker 1: end up requiring more than just dumping billions into scaling

51
00:04:04,600 --> 00:04:08,000
Speaker 1: up the training data for these systems. It may instead

52
00:04:08,080 --> 00:04:13,560
Speaker 1: require understanding how minds relate to one another in the

53
00:04:13,600 --> 00:04:17,840
Speaker 1: first place. And that's where today's conversation begins. Today I'm

54
00:04:17,880 --> 00:04:22,040
Speaker 1: joined by Danielle Persk. She's a cognitive scientist who leads

55
00:04:22,080 --> 00:04:27,799
Speaker 1: the Human Computer Interaction team at Amazon's AGI Lab. Danielle

56
00:04:27,880 --> 00:04:32,680
Speaker 1: uses insights from the evolution and development of human intelligence

57
00:04:33,120 --> 00:04:36,480
Speaker 1: to inform how we can not only make AI smarter,

58
00:04:36,920 --> 00:04:41,320
Speaker 1: but build AI that also makes us smarter. Here's my

59
00:04:41,400 --> 00:04:43,120
Speaker 1: conversation with Danielle Persk.

60
00:04:47,680 --> 00:04:52,640
Speaker 2: Intelligence in humans is really social, and that is the

61
00:04:52,680 --> 00:04:56,120
Speaker 2: thing that differentiates our intelligence from other species. Even other

62
00:04:56,160 --> 00:04:58,960
Speaker 2: species that are closely related to us have similar brain

63
00:04:59,000 --> 00:05:03,360
Speaker 2: structures and function similar genetics. And what we are really

64
00:05:03,800 --> 00:05:09,560
Speaker 2: optimizing for is representing other minds. So not only are

65
00:05:10,160 --> 00:05:14,680
Speaker 2: infants human infants inferring the existence of other minds, but

66
00:05:15,279 --> 00:05:21,279
Speaker 2: once this thing exists, we are optimized for aligning our minds. Evolutionarily,

67
00:05:21,400 --> 00:05:26,479
Speaker 2: we had to cooperate to survive. Infants need to be

68
00:05:26,560 --> 00:05:29,600
Speaker 2: able to have their caretaker's attention on them to survive,

69
00:05:30,160 --> 00:05:33,000
Speaker 2: and in terms of being able to learn about the world,

70
00:05:33,240 --> 00:05:35,480
Speaker 2: once infants have a model of other minds, then they

71
00:05:35,480 --> 00:05:38,839
Speaker 2: can manipulate it. They can direct their caretaker's attention point

72
00:05:39,080 --> 00:05:41,640
Speaker 2: what's that, and magically they'll have a label for.

73
00:05:41,680 --> 00:05:43,720
Speaker 3: This thing that they're looking at in their environment.

74
00:05:43,880 --> 00:05:46,480
Speaker 1: So they're doing prompt engineering.

75
00:05:47,960 --> 00:05:50,680
Speaker 3: Great technology. Yeah, okay, so we know that.

76
00:05:50,800 --> 00:05:54,440
Speaker 2: You know, throughout the course of human evolution, we became

77
00:05:54,640 --> 00:05:59,320
Speaker 2: increasingly dependent upon cooperating to to stay alive and adapt

78
00:05:59,440 --> 00:06:02,080
Speaker 2: to new environment. So it makes sense that there'd be

79
00:06:02,160 --> 00:06:06,080
Speaker 2: this extreme pressure on being able to predict each other's

80
00:06:06,120 --> 00:06:09,920
Speaker 2: behaviors to understand our minds, and then with infants, developmentally,

81
00:06:10,000 --> 00:06:13,479
Speaker 2: we have also the benefit of being able to learn

82
00:06:13,839 --> 00:06:19,280
Speaker 2: much more efficiently even language itself, from representing other minds.

83
00:06:19,720 --> 00:06:21,840
Speaker 1: Okay, so it turns out that we can do a

84
00:06:22,000 --> 00:06:26,760
Speaker 1: much better job of predicting if we can imagine what

85
00:06:26,800 --> 00:06:29,440
Speaker 1: it's like to be inside other people's heads. Right, So,

86
00:06:29,720 --> 00:06:33,920
Speaker 1: if I want to know what some non player character

87
00:06:34,000 --> 00:06:35,680
Speaker 1: is going to do in a video game whatever, they

88
00:06:35,720 --> 00:06:37,440
Speaker 1: have certain behaviors. But if I want to know, let's

89
00:06:37,440 --> 00:06:39,599
Speaker 1: say what you're going to do next, or say next,

90
00:06:39,880 --> 00:06:42,160
Speaker 1: if I have a model of your mind and what

91
00:06:42,200 --> 00:06:43,680
Speaker 1: you know and you don't know and all that stuff,

92
00:06:43,720 --> 00:06:44,919
Speaker 1: I can make a better prediction.

93
00:06:45,800 --> 00:06:50,360
Speaker 2: And so you've said that there's information and then information

94
00:06:50,440 --> 00:06:53,400
Speaker 2: with a purpose, and that information with a purpose really matters.

95
00:06:53,720 --> 00:06:57,000
Speaker 3: So you've used the example of like the.

96
00:06:56,720 --> 00:06:59,960
Speaker 2: Land rover on Mars not being able to fix itself,

97
00:07:00,400 --> 00:07:02,880
Speaker 2: and like a wolf that gets its like trap.

98
00:07:03,080 --> 00:07:05,560
Speaker 1: Right, the Curiosity Rover went up to Mars. We had

99
00:07:05,600 --> 00:07:08,160
Speaker 1: spent like a billion something dollars on it. It did

100
00:07:08,160 --> 00:07:10,520
Speaker 1: a great job on Mars, but eventually it got its

101
00:07:10,600 --> 00:07:13,400
Speaker 1: right front wheel stuck in the Martian soil and it

102
00:07:13,480 --> 00:07:17,000
Speaker 1: died couldn't get out. But if you can trast that

103
00:07:17,040 --> 00:07:20,080
Speaker 1: with a wolf who gets its leg cond of trap.

104
00:07:20,400 --> 00:07:22,600
Speaker 1: It'll chew its leg off and then figure out how

105
00:07:22,640 --> 00:07:24,920
Speaker 1: to walk on three legs, which is extraordinary because a

106
00:07:24,960 --> 00:07:27,600
Speaker 1: wolf's brain didn't evolve for three legs. But it can

107
00:07:27,680 --> 00:07:30,040
Speaker 1: figure it out because it's live wired. It has brain

108
00:07:30,080 --> 00:07:33,680
Speaker 1: plasticity and figure out, Okay, how do I adjust everything.

109
00:07:33,560 --> 00:07:34,640
Speaker 3: So that I can survival?

110
00:07:34,680 --> 00:07:37,920
Speaker 1: Depends upon it exactly. That's the key. It has relevance

111
00:07:38,040 --> 00:07:39,360
Speaker 1: to the animal.

112
00:07:39,200 --> 00:07:42,480
Speaker 2: Right, So all animals have a drive to survive, a

113
00:07:42,560 --> 00:07:46,800
Speaker 2: drive to reproduce, But humans also have a drive to

114
00:07:47,080 --> 00:07:51,320
Speaker 2: align our minds because it helps us cooperate, it helps

115
00:07:51,400 --> 00:07:55,560
Speaker 2: us survive, and it helps us to learn extremely efficiently.

116
00:07:56,000 --> 00:07:59,400
Speaker 2: So we don't just model other minds. That would just

117
00:07:59,440 --> 00:08:03,720
Speaker 2: be the information part. We are optimized for aligning our minds.

118
00:08:03,720 --> 00:08:05,520
Speaker 2: So it's information with a purpose.

119
00:08:05,760 --> 00:08:08,440
Speaker 1: Okay, so aligning our minds this is the key thing

120
00:08:09,640 --> 00:08:12,440
Speaker 1: and at the center of your interests. And so then

121
00:08:12,560 --> 00:08:15,280
Speaker 1: you went into looking into AGI. So first of all,

122
00:08:15,320 --> 00:08:18,520
Speaker 1: tell us what artificial general intelligence is to you.

123
00:08:18,800 --> 00:08:22,120
Speaker 2: Well, I think most of the labs that are trying

124
00:08:22,160 --> 00:08:26,160
Speaker 2: to build something like AGI, they all have their own definitions.

125
00:08:27,120 --> 00:08:30,040
Speaker 2: None of them are really very good. But the one

126
00:08:30,080 --> 00:08:32,400
Speaker 2: thing that unifies all of them is that they are

127
00:08:32,480 --> 00:08:35,840
Speaker 2: all benchmarked to human intelligence. And this goes all the

128
00:08:35,840 --> 00:08:39,160
Speaker 2: way back to the origin of the field of AI.

129
00:08:39,640 --> 00:08:42,000
Speaker 2: So in nineteen fifty six, a group of these engineers

130
00:08:42,000 --> 00:08:44,480
Speaker 2: and mathematicians got together. They were going to solve intelligence

131
00:08:44,480 --> 00:08:46,640
Speaker 2: and build thinking machines, and the idea is that these

132
00:08:46,640 --> 00:08:48,240
Speaker 2: thinking machines would think like us.

133
00:08:49,040 --> 00:08:50,160
Speaker 3: It obviously took a.

134
00:08:50,200 --> 00:08:53,280
Speaker 2: Very long time to realize, Oh, that's a lot harder

135
00:08:53,800 --> 00:08:56,199
Speaker 2: than we thought that it was. But now we are

136
00:08:56,320 --> 00:08:59,120
Speaker 2: back to aiming for something like that original goal of

137
00:08:59,120 --> 00:09:01,720
Speaker 2: building thinking machines that think like us.

138
00:09:01,760 --> 00:09:02,800
Speaker 3: We call it AGI.

139
00:09:03,280 --> 00:09:08,360
Speaker 2: Again, have slightly different operationalizations. But I think that we're

140
00:09:08,360 --> 00:09:12,080
Speaker 2: all running towards the wrong thing. And that's because I

141
00:09:12,120 --> 00:09:17,079
Speaker 2: don't think that intelligence can exist in a machine. It

142
00:09:17,120 --> 00:09:21,800
Speaker 2: doesn't exist in individual humans. It's something that emerges from

143
00:09:21,960 --> 00:09:27,079
Speaker 2: our interactions because we have this drive to align our representations,

144
00:09:27,600 --> 00:09:30,480
Speaker 2: and of course we all have very different representations.

145
00:09:30,559 --> 00:09:31,840
Speaker 3: Right When I used to teach.

146
00:09:31,679 --> 00:09:35,160
Speaker 2: Cognitive science, I would teach about this condition called a fantasia,

147
00:09:35,800 --> 00:09:39,640
Speaker 2: and once every couple of classes a student would come

148
00:09:39,720 --> 00:09:40,040
Speaker 2: up to.

149
00:09:40,000 --> 00:09:43,280
Speaker 1: Me quick In fantations where you can't imagine, you don't

150
00:09:43,320 --> 00:09:45,280
Speaker 1: have any visual representation on the Yes.

151
00:09:45,400 --> 00:09:47,640
Speaker 2: Yes, a student would come up to me and say, wait,

152
00:09:47,800 --> 00:09:51,000
Speaker 2: you mean there are people who can actually imagine things.

153
00:09:51,080 --> 00:09:53,280
Speaker 3: Their mind's eye is not just a metaphor. It's a

154
00:09:53,320 --> 00:09:53,880
Speaker 3: thing that.

155
00:09:53,800 --> 00:09:57,360
Speaker 2: People experience, and they wouldn't know because they don't suffer

156
00:09:57,480 --> 00:10:02,000
Speaker 2: from other types of death. It's just one of the

157
00:10:02,480 --> 00:10:06,920
Speaker 2: many ways in which human cognition and experience can very

158
00:10:07,920 --> 00:10:10,839
Speaker 2: And when I imagine in apple, it's different than when

159
00:10:10,880 --> 00:10:15,560
Speaker 2: you imagine an apple. We all have different associations. So

160
00:10:15,600 --> 00:10:20,359
Speaker 2: when we come together and we have to use words

161
00:10:20,880 --> 00:10:24,079
Speaker 2: to try to align our minds, there's necessarily going to

162
00:10:24,120 --> 00:10:29,320
Speaker 2: be friction, especially when we're trying to talk about abstract things,

163
00:10:29,960 --> 00:10:32,880
Speaker 2: especially when we're talking about things at the bleeding edge

164
00:10:32,880 --> 00:10:34,240
Speaker 2: of our knowledge, like science.

165
00:10:34,800 --> 00:10:37,400
Speaker 3: How do you align.

166
00:10:37,160 --> 00:10:39,600
Speaker 2: Your representations when there's not even a word for something.

167
00:10:39,679 --> 00:10:44,720
Speaker 2: So intelligence emerges as a function of trying to align

168
00:10:44,760 --> 00:10:50,679
Speaker 2: our minds and oftentimes creating new concepts to achieve that.

169
00:10:51,120 --> 00:10:53,320
Speaker 1: Okay, so when you're talking about aligning minds, it's because

170
00:10:53,480 --> 00:10:56,440
Speaker 1: I've got my whole internal world. You've got your whole

171
00:10:56,480 --> 00:11:00,240
Speaker 1: internal world that is built by each of our our

172
00:11:00,320 --> 00:11:03,959
Speaker 1: trajectories through space time. We've had different experiences all these things.

173
00:11:04,480 --> 00:11:07,800
Speaker 1: So we come together and we've got completely different worlds

174
00:11:07,840 --> 00:11:10,280
Speaker 1: running on the inside. And that's what conversation is about.

175
00:11:10,320 --> 00:11:12,280
Speaker 1: We're trying to align things that way.

176
00:11:12,440 --> 00:11:16,280
Speaker 2: And there are neuroscientists who measure when people are either

177
00:11:16,400 --> 00:11:19,520
Speaker 2: communicating in real time or if they're listening to a story,

178
00:11:19,640 --> 00:11:23,200
Speaker 2: if they're watching something on a screen, you can measure

179
00:11:23,280 --> 00:11:27,080
Speaker 2: the degree of neuralsynchrony, how close they are to be

180
00:11:27,120 --> 00:11:29,200
Speaker 2: on the same wavelength, and that predicts all sorts of

181
00:11:29,240 --> 00:11:32,400
Speaker 2: things like how much they like each other, how much

182
00:11:32,480 --> 00:11:37,199
Speaker 2: they understood the story, and how much they liked the story,

183
00:11:37,320 --> 00:11:39,480
Speaker 2: how similarly they remember things.

184
00:11:39,880 --> 00:11:42,640
Speaker 1: Okay, so this is what humans do. We get together

185
00:11:42,679 --> 00:11:44,960
Speaker 1: in conversation all the time and we try to achieve

186
00:11:45,000 --> 00:11:49,480
Speaker 1: that synchrony in terms of oh, okay, wait, you have

187
00:11:49,480 --> 00:11:51,880
Speaker 1: a different view than I do on this, here's how

188
00:11:51,880 --> 00:11:54,640
Speaker 1: we can make progress. This is the Socratic dialectic, right.

189
00:11:54,640 --> 00:11:56,719
Speaker 1: This is what Socrates love to do, is have these

190
00:11:56,720 --> 00:12:00,680
Speaker 1: conversations where the truth emerges, something bigger than either person

191
00:12:00,760 --> 00:12:03,199
Speaker 1: knew when they started the conversation.

192
00:12:03,440 --> 00:12:05,280
Speaker 2: And on that point too, I think a lot of

193
00:12:05,360 --> 00:12:10,920
Speaker 2: us think that we know things, but actually when we're

194
00:12:10,960 --> 00:12:13,880
Speaker 2: forced to describe something we realized we don't.

195
00:12:14,160 --> 00:12:16,439
Speaker 1: Yeah. Actually, in my next book, I'm talking about this

196
00:12:16,480 --> 00:12:20,520
Speaker 1: as a Potempkin village. Yeah, so you know. The Potemkin village,

197
00:12:20,559 --> 00:12:23,360
Speaker 1: for anyone doesn't remember, is when it was Catherine the

198
00:12:23,400 --> 00:12:26,680
Speaker 1: Great of Russia was heading down the river with a

199
00:12:26,679 --> 00:12:28,680
Speaker 1: bunch of dignitaries that she was trying to impress. She

200
00:12:28,720 --> 00:12:31,679
Speaker 1: hired this skuy Potempkin. Actually he was her lover as

201
00:12:31,679 --> 00:12:34,640
Speaker 1: well as a military general, but she got him to

202
00:12:34,679 --> 00:12:38,760
Speaker 1: go down the river a long way and build what

203
00:12:38,960 --> 00:12:42,120
Speaker 1: looked like a facade of a village so that when

204
00:12:42,200 --> 00:12:44,440
Speaker 1: the ship went by, all the dignitaries would be impressed

205
00:12:44,440 --> 00:12:47,120
Speaker 1: that there was this village. And he got all these

206
00:12:47,160 --> 00:12:49,800
Speaker 1: peasants like walk around happily and stuff, But there were

207
00:12:49,840 --> 00:12:53,560
Speaker 1: no buildings. It was just the front face of the building.

208
00:12:53,559 --> 00:12:57,320
Speaker 1: And then when the ship passed, he deconstructed this and

209
00:12:57,800 --> 00:13:00,679
Speaker 1: went ahead and built another village so that they passed

210
00:13:00,720 --> 00:13:02,840
Speaker 1: another great village, so it looked like things were really

211
00:13:02,880 --> 00:13:07,200
Speaker 1: happening there. Anyway, Cognition is often like this, where we think, oh, yeah,

212
00:13:07,200 --> 00:13:10,520
Speaker 1: I got it. Here's an example that I often use

213
00:13:10,600 --> 00:13:13,200
Speaker 1: is for anybody listening, take out a piece of paper,

214
00:13:13,240 --> 00:13:15,560
Speaker 1: and draw a bicycle, draw a bike.

215
00:13:16,080 --> 00:13:19,080
Speaker 3: I've tried this, yeah hard, Yeah.

216
00:13:19,000 --> 00:13:20,840
Speaker 1: Exactly, it turns out, I mean something as simple as

217
00:13:20,840 --> 00:13:24,679
Speaker 1: a bike what you see every day. Yeah, you start realizing, wait,

218
00:13:24,720 --> 00:13:26,640
Speaker 1: actually I don't know exactly where this goes and what's

219
00:13:26,679 --> 00:13:30,679
Speaker 1: the thing and so anyway, Yes, this is an example

220
00:13:30,720 --> 00:13:33,280
Speaker 1: of where we think we have deep knowledge and sometimes

221
00:13:33,320 --> 00:13:34,960
Speaker 1: it's just the facade of something that we know.

222
00:13:35,280 --> 00:13:37,360
Speaker 2: Yeah, and you can apply it on all different levels.

223
00:13:37,360 --> 00:13:40,160
Speaker 2: So you're describing, like the visual imagery might not be

224
00:13:40,360 --> 00:13:45,600
Speaker 2: very stable, but a lot of concepts are not stable either,

225
00:13:45,800 --> 00:13:50,480
Speaker 2: and we invent ways of making them more stable. Words

226
00:13:50,640 --> 00:13:53,600
Speaker 2: are a classic example of that. Once you have a

227
00:13:53,640 --> 00:13:57,440
Speaker 2: word for something, you can more easily trigger it, you

228
00:13:57,440 --> 00:13:59,280
Speaker 2: can more easily remember it, you can use it and

229
00:13:59,320 --> 00:14:02,280
Speaker 2: manipulate it and apply it to different things. But there's

230
00:14:02,320 --> 00:14:06,840
Speaker 2: a whole class of things that we're constantly inventing to

231
00:14:06,960 --> 00:14:12,000
Speaker 2: better align our minds. They're called cognitive technologies. So writing

232
00:14:12,080 --> 00:14:17,840
Speaker 2: would be one of the original ones. But symbols like math, logic, Yeah.

233
00:14:17,600 --> 00:14:19,360
Speaker 1: So unpack that. What's an example of this?

234
00:14:19,920 --> 00:14:22,640
Speaker 3: So literally, any word that you learn. Let's go back

235
00:14:22,640 --> 00:14:23,240
Speaker 3: to apples.

236
00:14:23,360 --> 00:14:28,520
Speaker 2: So children see apples, they don't have a word associated

237
00:14:28,560 --> 00:14:32,560
Speaker 2: with it, so the likelihood that it's going to spontaneously

238
00:14:33,120 --> 00:14:36,440
Speaker 2: sort of emerge in their mind is very low. Maybe

239
00:14:36,480 --> 00:14:38,760
Speaker 2: if they've seen a couple then there will be some

240
00:14:38,800 --> 00:14:44,400
Speaker 2: sort of increased likelihood or lowered threshold. But once they

241
00:14:44,440 --> 00:14:48,600
Speaker 2: have a word for that thing, that they reliably associate

242
00:14:48,600 --> 00:14:51,480
Speaker 2: it with it, then anybody who says the word anytime

243
00:14:51,520 --> 00:14:54,280
Speaker 2: they hear it, now their brain will elicit that activity

244
00:14:54,320 --> 00:14:59,160
Speaker 2: and it becomes a more stable representation. That's an obvious

245
00:14:59,280 --> 00:15:02,200
Speaker 2: example where where there's actually a physical thing that the

246
00:15:02,200 --> 00:15:03,040
Speaker 2: word can refer to.

247
00:15:03,160 --> 00:15:04,040
Speaker 3: But what about.

248
00:15:04,080 --> 00:15:08,120
Speaker 2: Concepts like love and justice that you can't see you're

249
00:15:08,120 --> 00:15:12,040
Speaker 2: saying By assigning a word to it, then we make

250
00:15:12,120 --> 00:15:15,000
Speaker 2: that stable, and then it become it can become associated

251
00:15:15,080 --> 00:15:18,400
Speaker 2: with a whole web of other concepts, and that web

252
00:15:18,480 --> 00:15:22,359
Speaker 2: becomes increasingly stable when when we can.

253
00:15:22,560 --> 00:15:25,040
Speaker 3: Make the associations.

254
00:15:24,280 --> 00:15:29,280
Speaker 2: More robust, more reliable, and then further when we can

255
00:15:29,320 --> 00:15:34,960
Speaker 2: invent things like science where we can really validate causal

256
00:15:35,000 --> 00:15:39,040
Speaker 2: relationships between things, and that makes our representations even more stable.

257
00:15:39,280 --> 00:15:42,240
Speaker 1: I see, so human brains interact with one another and

258
00:15:43,600 --> 00:15:46,640
Speaker 1: work on how do we make these representations stable? How

259
00:15:46,640 --> 00:15:50,800
Speaker 1: do we get knowledge coming out like a Socratic dialectic,

260
00:15:50,840 --> 00:15:54,240
Speaker 1: but with with everybody all involved and so on. And

261
00:15:54,280 --> 00:15:58,600
Speaker 1: so your idea when you moved into this field of AGI,

262
00:15:58,720 --> 00:16:01,520
Speaker 1: howeveryone wants to define it? What was your idea?

263
00:16:01,880 --> 00:16:05,600
Speaker 2: Well, so I left academia, which I absolutely loved, but

264
00:16:05,800 --> 00:16:11,440
Speaker 2: I felt an urgency to validate this theory.

265
00:16:11,560 --> 00:16:13,680
Speaker 3: I don't I mean, it's just a theory at this point.

266
00:16:14,000 --> 00:16:17,920
Speaker 2: How do we know whether we are optimized to align

267
00:16:17,920 --> 00:16:18,240
Speaker 2: our minds?

268
00:16:18,280 --> 00:16:20,640
Speaker 3: There's so much evidence to suggest that we do.

269
00:16:20,800 --> 00:16:23,560
Speaker 2: But as Richard Feyman said, you don't really know if

270
00:16:23,560 --> 00:16:24,920
Speaker 2: you understand something until you can build it.

271
00:16:24,960 --> 00:16:27,280
Speaker 3: And so I thought, well, maybe I could build this thing.

272
00:16:27,320 --> 00:16:30,480
Speaker 3: And the moment is just right because AI is taking

273
00:16:30,480 --> 00:16:31,000
Speaker 3: off again.

274
00:16:31,200 --> 00:16:35,280
Speaker 2: It's waking up from one of the winters, and I

275
00:16:35,360 --> 00:16:41,280
Speaker 2: was watching the scaling have really impressive results with a

276
00:16:41,320 --> 00:16:43,520
Speaker 2: deep learning, which felt really good because as somebody who

277
00:16:43,520 --> 00:16:46,800
Speaker 2: had a background in neuroscience, just like, oh yeah, inspiration

278
00:16:47,320 --> 00:16:52,440
Speaker 2: from brains is actually proving to be really effective. So

279
00:16:53,160 --> 00:16:57,720
Speaker 2: I moved into tech and started collaborating with the engineers

280
00:16:57,760 --> 00:17:02,320
Speaker 2: who were trying to build ever more capable intelligence. I'm

281
00:17:02,320 --> 00:17:06,480
Speaker 2: now in one of these frontier AGI labs and the

282
00:17:06,600 --> 00:17:08,520
Speaker 2: thing that we are going to be doing, which I

283
00:17:08,520 --> 00:17:12,159
Speaker 2: think is really differentiated from other approaches, is try to

284
00:17:12,160 --> 00:17:15,680
Speaker 2: build the communicative drive. Can we build agents that are

285
00:17:15,800 --> 00:17:21,720
Speaker 2: optimized for understanding each other's perspectives? And from that, can

286
00:17:21,760 --> 00:17:26,600
Speaker 2: we get emergent behaviors, emergent capabilities that we wouldn't get

287
00:17:26,640 --> 00:17:29,160
Speaker 2: from a single model on its own.

288
00:17:29,320 --> 00:17:32,520
Speaker 1: So I just want to slow that down. So communicative drive,

289
00:17:32,960 --> 00:17:34,800
Speaker 1: that's the first time we've heard the term, So tell

290
00:17:34,840 --> 00:17:35,720
Speaker 1: us what that means.

291
00:17:36,040 --> 00:17:40,800
Speaker 2: So communicative drive is the phrase that I use to

292
00:17:41,080 --> 00:17:45,399
Speaker 2: describe this compulsion that we have to align our minds

293
00:17:45,480 --> 00:17:51,440
Speaker 2: to establish representational alignment. You can imagine how the communicative

294
00:17:51,560 --> 00:17:57,760
Speaker 2: drive would interact with other dispositions that humans have. And importantly,

295
00:17:57,760 --> 00:17:59,800
Speaker 2: you have to think at the population level. So again

296
00:18:00,080 --> 00:18:04,840
Speaker 2: we have variation for every trait, and some of us

297
00:18:04,840 --> 00:18:08,080
Speaker 2: are more open, some of us are more closed to experience.

298
00:18:08,280 --> 00:18:10,000
Speaker 3: But you can imagine. Okay, so in the case of

299
00:18:10,040 --> 00:18:14,240
Speaker 3: somebody who's really closed, Let's say that they are.

300
00:18:14,000 --> 00:18:21,480
Speaker 2: In some communicative exchange and they detect a mismatch. So

301
00:18:21,520 --> 00:18:26,600
Speaker 2: somebody is clearly not understanding what they are saying.

302
00:18:27,320 --> 00:18:31,560
Speaker 3: They have two choices. They can update.

303
00:18:31,359 --> 00:18:35,080
Speaker 2: Their perspectives to the other person's, or they can try

304
00:18:35,080 --> 00:18:38,240
Speaker 2: to get the other person's perspective to look more like theirs.

305
00:18:38,920 --> 00:18:41,840
Speaker 2: What would it take to get another person to come

306
00:18:41,840 --> 00:18:42,840
Speaker 2: to your perspective.

307
00:18:43,119 --> 00:18:46,360
Speaker 3: You'd have to create an artifact. You'd have to create.

308
00:18:46,280 --> 00:18:50,280
Speaker 2: A word or a piece of art or a theory

309
00:18:50,560 --> 00:18:53,680
Speaker 2: to get them to really understand and take on your

310
00:18:53,760 --> 00:18:54,720
Speaker 2: perspective and.

311
00:18:54,680 --> 00:18:55,560
Speaker 3: Close that gap.

312
00:18:55,880 --> 00:18:59,000
Speaker 2: But if you're a very open person, if you're creative,

313
00:18:59,520 --> 00:19:02,359
Speaker 2: that might be your default. But if you're a little

314
00:19:02,359 --> 00:19:05,440
Speaker 2: bit more reserved, maybe you just take on the other

315
00:19:05,560 --> 00:19:06,760
Speaker 2: person's perspective.

316
00:19:07,640 --> 00:19:09,359
Speaker 1: Is it that way or the other way? Sorry? If

317
00:19:09,400 --> 00:19:11,199
Speaker 1: I'm very open, I feel like I would take the

318
00:19:11,240 --> 00:19:12,600
Speaker 1: other person's perspective, so you.

319
00:19:12,520 --> 00:19:15,719
Speaker 2: Can actually imagine both situations. Yes, So I score very

320
00:19:15,800 --> 00:19:18,600
Speaker 2: high on openness, and when I'm in communicative exchanges, I

321
00:19:18,640 --> 00:19:22,640
Speaker 2: often feel like, oh, wow, yeah, that's I've never thought

322
00:19:22,680 --> 00:19:24,240
Speaker 2: of it that way, or maybe I have, and I

323
00:19:24,240 --> 00:19:26,000
Speaker 2: want to add all these things and like it's a

324
00:19:26,119 --> 00:19:29,800
Speaker 2: very cooperative thing. I'm more thinking about the dynamics of

325
00:19:29,800 --> 00:19:34,000
Speaker 2: people who want to maintain tradition and status quo versus

326
00:19:34,080 --> 00:19:37,000
Speaker 2: people who want to challenge that. So oftentimes that maps

327
00:19:37,000 --> 00:19:41,879
Speaker 2: onto the dimension of openness. So if you see that

328
00:19:42,040 --> 00:19:46,320
Speaker 2: everybody around you seems to hold a different perspective than

329
00:19:46,400 --> 00:19:51,160
Speaker 2: you do, you're more likely to conform to their perspectives.

330
00:19:51,160 --> 00:19:54,520
Speaker 2: If you're somebody who might be a little bit more conservative,

331
00:19:54,800 --> 00:19:57,520
Speaker 2: not wanting to ruffle feathers that kind of thing.

332
00:19:57,760 --> 00:19:59,080
Speaker 1: Well, I'm just trying to stand why I use the

333
00:19:59,080 --> 00:20:04,760
Speaker 1: word conservative there, because conservative meaning are like iinin exactly.

334
00:20:05,320 --> 00:20:08,440
Speaker 1: Oh but you're saying, maintain the group traditions. Yeah, okay,

335
00:20:08,560 --> 00:20:09,119
Speaker 1: got it.

336
00:20:09,119 --> 00:20:11,360
Speaker 3: As opposed to being iconoclastic and innovative.

337
00:20:11,560 --> 00:20:13,680
Speaker 1: I see, I see how you're using it. Okay, great.

338
00:20:13,880 --> 00:20:16,359
Speaker 1: So this is the idea is that people are always talking,

339
00:20:16,400 --> 00:20:19,119
Speaker 1: and depending on your personality type, what you're trying to

340
00:20:19,160 --> 00:20:21,159
Speaker 1: do is either align yourself with them or them with

341
00:20:21,200 --> 00:20:24,399
Speaker 1: you or whatever, or meet in the middle. But this,

342
00:20:25,080 --> 00:20:28,840
Speaker 1: this you feel, is the key to what human societies

343
00:20:29,119 --> 00:20:32,359
Speaker 1: bring as opposed to looking at individual brains. You know,

344
00:20:32,400 --> 00:20:35,640
Speaker 1: the history of neuroscience is all about looking at individual brains. Oh,

345
00:20:35,640 --> 00:20:37,399
Speaker 1: this is how the visual system works, as how decision

346
00:20:37,400 --> 00:20:41,240
Speaker 1: making works, how hearing works, whatever. But there's this new

347
00:20:41,280 --> 00:20:43,320
Speaker 1: feel that's been growing for the last twenty or thirty years,

348
00:20:43,320 --> 00:20:46,480
Speaker 1: which is called social neuroscience, which is all about, gosh,

349
00:20:46,480 --> 00:20:48,760
Speaker 1: we've got a lot of circuitry in our brains that

350
00:20:48,840 --> 00:20:51,520
Speaker 1: care about other brains. So this is the heart of

351
00:20:51,560 --> 00:20:55,000
Speaker 1: your interest. Is what happens when people are talking and aligning?

352
00:20:55,400 --> 00:20:58,639
Speaker 1: And why are we so driven to communicate instead of

353
00:20:58,680 --> 00:21:00,480
Speaker 1: let's imagine that you and I set down on a

354
00:21:00,520 --> 00:21:03,840
Speaker 1: bus next to each other, we'd probably chat as opposed

355
00:21:03,880 --> 00:21:06,440
Speaker 1: to just sit there and deal with our own brains. Okay,

356
00:21:06,520 --> 00:21:09,919
Speaker 1: so how does this map onto what you're interested in

357
00:21:09,920 --> 00:21:10,639
Speaker 1: doing in AI?

358
00:21:11,640 --> 00:21:16,720
Speaker 2: Yes, so I am concerned about building something that resembles

359
00:21:16,880 --> 00:21:20,240
Speaker 2: our own intelligence, or something that resembles us because we

360
00:21:20,359 --> 00:21:23,480
Speaker 2: have all sorts of flaws and biases.

361
00:21:23,920 --> 00:21:25,600
Speaker 3: The variability, I.

362
00:21:25,520 --> 00:21:28,359
Speaker 2: Think is very useful, and we wouldn't be intelligent in

363
00:21:28,359 --> 00:21:30,439
Speaker 2: the way that we are without the variability. And you

364
00:21:30,520 --> 00:21:32,840
Speaker 2: might call some of that variability the bias, the unique

365
00:21:32,880 --> 00:21:33,800
Speaker 2: biases that we have.

366
00:21:34,320 --> 00:21:37,000
Speaker 3: But I think if we try to reproduce.

367
00:21:36,480 --> 00:21:39,320
Speaker 2: All of that, we're going to get a mirror of ourselves,

368
00:21:39,320 --> 00:21:44,640
Speaker 2: and that's not always the most effective way to augment

369
00:21:44,800 --> 00:21:46,960
Speaker 2: our intelligence. And I should back up and say, why

370
00:21:47,000 --> 00:21:49,200
Speaker 2: are we doing any of this? Why do we want

371
00:21:49,240 --> 00:21:51,840
Speaker 2: to build intelligence that looks like us. I think the

372
00:21:51,920 --> 00:21:55,080
Speaker 2: assumption that a lot of these the people, the engineers,

373
00:21:55,080 --> 00:21:57,560
Speaker 2: and these labs have is that, oh, of course it's

374
00:21:57,560 --> 00:22:01,359
Speaker 2: going to be extremely useful for us. It's going to

375
00:22:01,560 --> 00:22:06,080
Speaker 2: unlock this unprecedented era of human flourishing. But the assumption

376
00:22:06,160 --> 00:22:08,720
Speaker 2: that it's going to be really useful for us, I

377
00:22:08,760 --> 00:22:11,400
Speaker 2: think is taken for granted, and if you really think

378
00:22:11,400 --> 00:22:15,000
Speaker 2: about it, well, how because a lot of the examples

379
00:22:15,040 --> 00:22:19,440
Speaker 2: that we have from recent technology and algorithms is that

380
00:22:19,920 --> 00:22:24,920
Speaker 2: they actually take away our agency. We lose hours to scrolling,

381
00:22:24,960 --> 00:22:29,400
Speaker 2: we get stuck in echo chambers, we have autocomplete takeaway

382
00:22:29,560 --> 00:22:31,600
Speaker 2: our thinking, and we're starting to see.

383
00:22:31,440 --> 00:22:34,680
Speaker 3: The same kinds of things with chatbots.

384
00:22:35,600 --> 00:22:38,639
Speaker 2: We're also seeing that people are using these technologies and

385
00:22:39,359 --> 00:22:44,320
Speaker 2: very much augmenting their their own intelligence. I feel sometimes

386
00:22:44,480 --> 00:22:48,440
Speaker 2: like I'm having entirely new thoughts at an unprecedented pace

387
00:22:48,800 --> 00:22:51,359
Speaker 2: when I'm going back and forth, just like when you

388
00:22:51,400 --> 00:22:54,480
Speaker 2: were having amazing conversations with other people. We use each

389
00:22:54,480 --> 00:22:56,560
Speaker 2: other's minds as tools, but you can just do that

390
00:22:56,600 --> 00:22:59,840
Speaker 2: at a more rapid pace. So it's not a foregone

391
00:23:00,000 --> 00:23:06,159
Speaker 2: inclusion that giving the AI more capabilities, making it smarter

392
00:23:06,240 --> 00:23:09,240
Speaker 2: and giving it more agency is going to be good

393
00:23:09,240 --> 00:23:11,840
Speaker 2: for us. I think we have to turn that on

394
00:23:11,920 --> 00:23:14,200
Speaker 2: its head and say, what would it take to make

395
00:23:14,480 --> 00:23:18,359
Speaker 2: AI that makes us smarter and gives us more agency?

396
00:23:19,200 --> 00:23:22,960
Speaker 2: And that would be, by definition, something that is good

397
00:23:23,040 --> 00:23:27,160
Speaker 2: for us. So how do we do that? I don't

398
00:23:27,200 --> 00:23:29,880
Speaker 2: think that we want to have agents that have their

399
00:23:29,920 --> 00:23:33,560
Speaker 2: own drives to survive and manipulate us and have all

400
00:23:33,600 --> 00:23:39,000
Speaker 2: of the status seeking U situations that we have. But

401
00:23:39,480 --> 00:23:44,640
Speaker 2: if they were motivated to align their representations with ours,

402
00:23:44,720 --> 00:23:48,359
Speaker 2: that could actually be really useful for unlocking our potential

403
00:23:48,359 --> 00:23:53,040
Speaker 2: and for helping us learn. And as we're giving them

404
00:23:53,080 --> 00:23:56,040
Speaker 2: these capabilities to do that, they have to figure out.

405
00:23:56,160 --> 00:23:58,680
Speaker 2: One of the ways that we are able to generalize

406
00:23:58,720 --> 00:24:04,080
Speaker 2: and continually learn is that we are constantly negotiating meaning

407
00:24:04,119 --> 00:24:06,760
Speaker 2: and coming up and with the friction of the interactions

408
00:24:07,160 --> 00:24:11,159
Speaker 2: with each other, we are able to do continual learning

409
00:24:11,280 --> 00:24:16,560
Speaker 2: because we're not optimizing for one thing, one niche, one environment,

410
00:24:16,800 --> 00:24:22,600
Speaker 2: one particular problem. We are optimizing for aligning our minds

411
00:24:22,720 --> 00:24:26,840
Speaker 2: with many minds, and all of them. These targets are

412
00:24:26,840 --> 00:24:30,600
Speaker 2: all moving targets, so it's kind of an escape velocity

413
00:24:30,720 --> 00:24:35,960
Speaker 2: from really focusing on one thing and our ability to

414
00:24:36,000 --> 00:24:39,000
Speaker 2: do that not only allows us to continually learn, but

415
00:24:39,080 --> 00:24:40,760
Speaker 2: it gives us superpowers.

416
00:24:40,840 --> 00:24:42,000
Speaker 1: So what does this look like for you?

417
00:24:42,040 --> 00:24:42,159
Speaker 2: Though?

418
00:24:42,200 --> 00:24:45,119
Speaker 1: If you had a world five years in the future

419
00:24:45,119 --> 00:24:46,960
Speaker 1: that you were able to sort of define where this

420
00:24:47,000 --> 00:24:50,000
Speaker 1: is going, what's ok Does it mean that there are

421
00:24:50,640 --> 00:24:54,440
Speaker 1: lots of AI agents and they are talking with humans

422
00:24:54,520 --> 00:24:59,159
Speaker 1: and they're trying to align their thinking with humans and

423
00:24:59,200 --> 00:25:02,040
Speaker 1: the humans a with the AI or its look like

424
00:25:02,119 --> 00:25:04,480
Speaker 1: there's one AI give us a sense of this world?

425
00:25:04,600 --> 00:25:07,440
Speaker 2: Okay, So there's at least two important things here. One

426
00:25:07,480 --> 00:25:09,560
Speaker 2: is that right now agents are not reliable, so they're

427
00:25:09,560 --> 00:25:11,960
Speaker 2: not useful. And I think the idea there is that

428
00:25:12,000 --> 00:25:16,359
Speaker 2: they are fundamentally different than llms. They are embodied in

429
00:25:16,400 --> 00:25:19,040
Speaker 2: some kind of environment, even if it's the digital environment,

430
00:25:19,680 --> 00:25:23,520
Speaker 2: but we can't yet get them to do long horizon

431
00:25:24,000 --> 00:25:27,160
Speaker 2: you know, actions in a reliable way, and so they're

432
00:25:27,200 --> 00:25:27,879
Speaker 2: not yet useful.

433
00:25:28,040 --> 00:25:30,160
Speaker 1: Right now, you're talking about AI agents.

434
00:25:30,520 --> 00:25:34,320
Speaker 2: Most people have interacted with chatbots and that's what they

435
00:25:34,359 --> 00:25:36,760
Speaker 2: think AI is, or that's what they think generative AI is.

436
00:25:36,800 --> 00:25:38,920
Speaker 2: Maybe they know of you know, the image generators too,

437
00:25:39,040 --> 00:25:42,240
Speaker 2: but a lot of us are interacting with chatbots. Those

438
00:25:42,240 --> 00:25:45,600
Speaker 2: are llms that are predicting the next text token. Large

439
00:25:45,680 --> 00:25:49,560
Speaker 2: language models, yes, large language models, but they don't actually

440
00:25:49,640 --> 00:25:55,239
Speaker 2: do things. Agents, in contrast, can actually take actions and

441
00:25:55,280 --> 00:25:58,359
Speaker 2: do things on our behalf and in order.

442
00:25:59,280 --> 00:26:00,520
Speaker 3: So are lab is.

443
00:26:00,480 --> 00:26:04,439
Speaker 2: Working on building computer use agents. So if I want

444
00:26:04,480 --> 00:26:07,600
Speaker 2: an agent to book me a flight or order me

445
00:26:08,400 --> 00:26:10,680
Speaker 2: a dinner, I can say that and then it can

446
00:26:10,720 --> 00:26:14,439
Speaker 2: go off and use whatever websites or software tools to

447
00:26:14,680 --> 00:26:16,359
Speaker 2: do those things.

448
00:26:16,119 --> 00:26:17,280
Speaker 3: That's the hope.

449
00:26:17,760 --> 00:26:21,520
Speaker 2: You've seen a couple of these agents, computer use agents

450
00:26:21,520 --> 00:26:24,840
Speaker 2: come out, and it's really exciting to see them start

451
00:26:24,880 --> 00:26:27,920
Speaker 2: to do things, but they're not reliable. And because they're

452
00:26:27,960 --> 00:26:30,400
Speaker 2: not reliable, they might do the thing that you ask

453
00:26:30,520 --> 00:26:33,400
Speaker 2: them to do one out of ten times, and again

454
00:26:33,440 --> 00:26:35,760
Speaker 2: that's exciting, but that's not very useful, right.

455
00:26:35,760 --> 00:26:37,879
Speaker 1: You mean it's because they make mistakes. It's not the

456
00:26:37,920 --> 00:26:40,400
Speaker 1: way we would say an employee is not reliable because

457
00:26:40,400 --> 00:26:42,920
Speaker 1: he's out back smoking a cigarette. Is that they're trying

458
00:26:42,920 --> 00:26:44,880
Speaker 1: to do stuff is just a clicking on the wrong

459
00:26:44,920 --> 00:26:45,560
Speaker 1: thing and getting it.

460
00:26:45,520 --> 00:26:49,399
Speaker 2: Wrong, that's right y, Yes, So working on making these

461
00:26:49,600 --> 00:26:53,800
Speaker 2: agents reliable is necessary for making them useful. But we

462
00:26:53,880 --> 00:26:58,800
Speaker 2: can imagine that unlocking a whole new set of capabilities

463
00:26:58,840 --> 00:27:03,280
Speaker 2: and ways that they would augment humans because rather than

464
00:27:03,520 --> 00:27:06,399
Speaker 2: just having a conversation, and conversations can be very useful.

465
00:27:06,680 --> 00:27:08,640
Speaker 2: All of the things that you do in your daily life,

466
00:27:08,640 --> 00:27:11,440
Speaker 2: all the things that you're using a computer for, the

467
00:27:11,520 --> 00:27:14,320
Speaker 2: vast majority of them are probably not worthy of your time.

468
00:27:14,359 --> 00:27:17,000
Speaker 2: You're doing things on the computer to actually achieve other

469
00:27:17,080 --> 00:27:19,480
Speaker 2: things in the real world. So what if you could

470
00:27:19,520 --> 00:27:24,280
Speaker 2: have agents reliably execute a lot of the things that

471
00:27:24,320 --> 00:27:24,840
Speaker 2: you're doing.

472
00:27:25,000 --> 00:27:25,600
Speaker 3: And in.

473
00:27:27,280 --> 00:27:30,000
Speaker 2: Knowledge work, you know, we're using a ton of different

474
00:27:30,000 --> 00:27:34,640
Speaker 2: tools I call them arbitrary skills, processing invoices or something

475
00:27:34,680 --> 00:27:37,640
Speaker 2: that everybody does using doing their taxes.

476
00:27:37,280 --> 00:27:39,800
Speaker 3: Like, do you really have to becommon expert.

477
00:27:39,480 --> 00:27:43,479
Speaker 2: At using these tools or is that maybe not the

478
00:27:43,480 --> 00:27:48,000
Speaker 2: best use of our human potential, our cognitive potential. If

479
00:27:48,000 --> 00:27:51,000
Speaker 2: we could have agents that knew how to use all

480
00:27:51,000 --> 00:27:54,000
Speaker 2: of the tools that we did, that would save us

481
00:27:54,080 --> 00:27:57,480
Speaker 2: a ton of time, and it would have cascading implications

482
00:27:57,520 --> 00:28:00,800
Speaker 2: for how people collaborate with other people in the real world.

483
00:28:00,880 --> 00:28:04,280
Speaker 2: Human collaboration would be different because we'd be freed up

484
00:28:04,560 --> 00:28:08,679
Speaker 2: to focus on more creative things, more strategic decisions. Having

485
00:28:08,760 --> 00:28:12,040
Speaker 2: the sorts of debates that we have to advance whatever

486
00:28:12,280 --> 00:28:15,760
Speaker 2: shared goals that we have. So this is the sort

487
00:28:15,800 --> 00:28:17,719
Speaker 2: of first part. We have to just get the agents

488
00:28:17,760 --> 00:28:21,040
Speaker 2: to reliably click or scroll when we need them to.

489
00:28:21,640 --> 00:28:25,800
Speaker 2: But if you play that forward, what does reliability actually

490
00:28:25,840 --> 00:28:28,600
Speaker 2: mean when we have higher level goals.

491
00:28:29,160 --> 00:28:30,040
Speaker 3: It's not just.

492
00:28:30,320 --> 00:28:33,960
Speaker 2: Knowing where to click or knowing when to scroll. It's

493
00:28:34,000 --> 00:28:39,760
Speaker 2: actually understanding the goal. And that goal might require breaking

494
00:28:40,200 --> 00:28:43,800
Speaker 2: the breaking it down into subtasks, and then going and

495
00:28:43,840 --> 00:28:45,880
Speaker 2: doing all of those things, and there might be many

496
00:28:45,920 --> 00:28:48,200
Speaker 2: ways of doing it, and there's not necessarily a right

497
00:28:48,280 --> 00:28:50,880
Speaker 2: or a wrong way of doing it. So at the

498
00:28:50,960 --> 00:28:54,080
Speaker 2: end of the day, reliability ends up becoming about understanding

499
00:28:54,160 --> 00:29:09,160
Speaker 2: our minds.

500
00:29:10,760 --> 00:29:13,640
Speaker 1: So the idea is if I could have an AI

501
00:29:13,760 --> 00:29:17,560
Speaker 1: agent that understands my mind, that has a model of me,

502
00:29:17,680 --> 00:29:19,800
Speaker 1: including what I know and don't know, and what my

503
00:29:19,880 --> 00:29:23,880
Speaker 1: goals are long term and short term, then it could

504
00:29:24,480 --> 00:29:26,600
Speaker 1: do a better job at what needs to be done.

505
00:29:26,640 --> 00:29:28,400
Speaker 1: Because when it comes to a choice point is is oh,

506
00:29:28,400 --> 00:29:30,640
Speaker 1: I know what Eagleman wants. He likes this kind of thing,

507
00:29:31,000 --> 00:29:32,920
Speaker 1: and that might be something that emerges not just from

508
00:29:33,000 --> 00:29:36,959
Speaker 1: patterns of looking at my behavior, but actually understanding internally,

509
00:29:37,680 --> 00:29:39,120
Speaker 1: having some theory of my mind.

510
00:29:39,360 --> 00:29:39,840
Speaker 3: I think that it.

511
00:29:39,800 --> 00:29:42,440
Speaker 2: Would need that, yes, and I think that we would

512
00:29:42,560 --> 00:29:45,600
Speaker 2: need to be able to interact with it in the

513
00:29:45,680 --> 00:29:49,080
Speaker 2: way that we interact with other teammates, where we're negotiating,

514
00:29:49,160 --> 00:29:51,640
Speaker 2: meaning in real time, where we're going back to earlier

515
00:29:51,680 --> 00:29:55,080
Speaker 2: in our conversation. Sometimes we don't realize that we're not

516
00:29:55,120 --> 00:29:59,040
Speaker 2: clearly thinking about something, and so having that reflected back

517
00:29:59,280 --> 00:30:03,600
Speaker 2: and being able to go through this exchange a dialectic.

518
00:30:03,680 --> 00:30:06,959
Speaker 2: It can refine our thinking, sharpen our thinking.

519
00:30:07,360 --> 00:30:09,640
Speaker 1: So this is a thing I've been wondering about for

520
00:30:09,640 --> 00:30:12,400
Speaker 1: a while, which is if you're looking at something from

521
00:30:12,400 --> 00:30:16,040
Speaker 1: the outside, you can actually get a lot of data

522
00:30:16,120 --> 00:30:18,440
Speaker 1: about it. And by outside I mean as opposed to

523
00:30:18,600 --> 00:30:20,920
Speaker 1: from the inside. If I have a theory of your mind,

524
00:30:21,120 --> 00:30:23,840
Speaker 1: the question is if I, just if I could observe

525
00:30:24,000 --> 00:30:26,840
Speaker 1: all of your behavior without knowing anything about what's in

526
00:30:26,880 --> 00:30:29,840
Speaker 1: your mind, could I nonetheless do just as good a job.

527
00:30:30,160 --> 00:30:33,000
Speaker 2: Well, I think that this is what our models of

528
00:30:33,040 --> 00:30:37,680
Speaker 2: other mind essentially are. It's making sense of behavior. Yeah,

529
00:30:37,880 --> 00:30:43,360
Speaker 2: it's just that the behavior, again is multiply redundant, and

530
00:30:43,560 --> 00:30:45,640
Speaker 2: there are many different cues that we can attend to,

531
00:30:45,720 --> 00:30:49,080
Speaker 2: and we're prioritizing some cues over others, and then it

532
00:30:49,120 --> 00:30:52,400
Speaker 2: is more efficient for us to represent that there's a

533
00:30:52,400 --> 00:30:53,320
Speaker 2: mind behind the eyes.

534
00:30:53,680 --> 00:30:56,200
Speaker 1: Yeah, very good. Do you see a world where we

535
00:30:56,280 --> 00:31:00,280
Speaker 1: would have lots of AI agents that are also speeding

536
00:31:00,360 --> 00:31:03,600
Speaker 1: with one another in terms of communicative drive of saying, hey,

537
00:31:03,680 --> 00:31:05,200
Speaker 1: this is what I've learned, and I know and blah

538
00:31:05,240 --> 00:31:08,720
Speaker 1: blah blah, and they develop a better, something bigger as

539
00:31:08,720 --> 00:31:11,400
Speaker 1: a result of the communication. They have a better understanding

540
00:31:11,400 --> 00:31:14,560
Speaker 1: of the world because they're talking with one another, because

541
00:31:14,600 --> 00:31:16,440
Speaker 1: just like humans, each of them is going to have

542
00:31:16,520 --> 00:31:18,240
Speaker 1: some trajectory through space time.

543
00:31:19,000 --> 00:31:21,240
Speaker 2: Okay, this gets into the second thing that I think

544
00:31:21,360 --> 00:31:23,840
Speaker 2: is really important about how agents can be useful. So

545
00:31:23,840 --> 00:31:27,040
Speaker 2: the first thing is they can do the digital drudgery

546
00:31:27,080 --> 00:31:30,320
Speaker 2: for us. They can save time. I call this They

547
00:31:30,320 --> 00:31:33,600
Speaker 2: can become our collective subconscious because we won't have to

548
00:31:33,640 --> 00:31:37,760
Speaker 2: spend our conscious time attending to things. We can relegate

549
00:31:37,960 --> 00:31:41,040
Speaker 2: so the agents can be doing in parallel all of

550
00:31:41,080 --> 00:31:43,120
Speaker 2: this stuff, so we don't have to become expert in

551
00:31:43,280 --> 00:31:46,480
Speaker 2: these arbitrary skills. But they are also as they are

552
00:31:46,480 --> 00:31:49,040
Speaker 2: doing that for a lot of people, they're learning a

553
00:31:49,080 --> 00:31:51,800
Speaker 2: bunch of different skills they're learning how to navigate different

554
00:31:51,840 --> 00:31:54,720
Speaker 2: websites and use different software tools. And so if I

555
00:31:54,960 --> 00:31:58,840
Speaker 2: need to for my job learn something really quickly, they

556
00:31:58,880 --> 00:32:02,280
Speaker 2: can redistribute the skills that they've learned from my teammates,

557
00:32:02,360 --> 00:32:05,280
Speaker 2: and they can give me the context that I need,

558
00:32:05,600 --> 00:32:07,680
Speaker 2: not to become expert in that tool, but to be

559
00:32:07,720 --> 00:32:11,640
Speaker 2: able to establish representational alignment with my teammate who uses

560
00:32:11,680 --> 00:32:15,840
Speaker 2: that tool. So they can help coordinate a team's behavior

561
00:32:15,960 --> 00:32:19,080
Speaker 2: by understanding all of the things that we do, all

562
00:32:19,120 --> 00:32:20,680
Speaker 2: the goals that we have, all the tools that we

563
00:32:20,800 --> 00:32:22,000
Speaker 2: use to achieve our goals.

564
00:32:22,280 --> 00:32:24,880
Speaker 1: Now, normally that would happen. You go up to Susie

565
00:32:24,880 --> 00:32:26,560
Speaker 1: and you say, hey, you know, I need to talk

566
00:32:26,600 --> 00:32:28,400
Speaker 1: to you, and Susie says, no, I use a different

567
00:32:28,400 --> 00:32:29,400
Speaker 1: word than you're using.

568
00:32:29,200 --> 00:32:31,840
Speaker 3: It for ten minutes to even establish common ground.

569
00:32:32,000 --> 00:32:34,160
Speaker 1: Got it. But you're saying that AI could help with

570
00:32:34,200 --> 00:32:37,600
Speaker 1: that interaction like the third person in the room and say, hey,

571
00:32:37,600 --> 00:32:38,920
Speaker 1: you know what, Danielle, this is what you need to

572
00:32:38,920 --> 00:32:40,400
Speaker 1: know about and Susie, this is what you need to know.

573
00:32:40,520 --> 00:32:42,640
Speaker 1: And I noticed you guys are using this same word,

574
00:32:42,680 --> 00:32:45,280
Speaker 1: but you mean different things by it's that kind of thing.

575
00:32:45,400 --> 00:32:48,760
Speaker 2: Yes, that's so yes to your point, I do think

576
00:32:48,880 --> 00:32:53,000
Speaker 2: that these agents that are working in parallel and understanding

577
00:32:53,000 --> 00:32:56,280
Speaker 2: our context will probably detect a ton of inefficiencies and

578
00:32:56,320 --> 00:32:58,800
Speaker 2: how we're doing things, and they will come up with

579
00:32:58,960 --> 00:33:01,360
Speaker 2: better ways of doing I would hope that they would.

580
00:33:01,120 --> 00:33:02,320
Speaker 3: Do that great.

581
00:33:03,160 --> 00:33:04,680
Speaker 2: One of the things that we learned is we were

582
00:33:04,720 --> 00:33:08,920
Speaker 2: trying to train our agent how to use Gmail, is

583
00:33:08,960 --> 00:33:12,080
Speaker 2: that wow, most people are actually really bad at using Gmail.

584
00:33:12,120 --> 00:33:12,640
Speaker 1: In what way?

585
00:33:13,680 --> 00:33:15,320
Speaker 3: So we don't know how to do.

586
00:33:15,200 --> 00:33:20,480
Speaker 2: The search queries effectively. We mostly just stumble through. And

587
00:33:20,520 --> 00:33:24,880
Speaker 2: it's the power users who sometimes build hold businesses around

588
00:33:25,080 --> 00:33:27,800
Speaker 2: you know, the Google Suite and Gmail. They they know

589
00:33:28,000 --> 00:33:31,120
Speaker 2: exactly how it was designed, all of the affordances that

590
00:33:31,160 --> 00:33:34,320
Speaker 2: it has, all of the new features, how that can

591
00:33:34,360 --> 00:33:35,680
Speaker 2: make things actually more efficient.

592
00:33:36,480 --> 00:33:37,200
Speaker 3: They get it.

593
00:33:37,720 --> 00:33:41,200
Speaker 2: But most people don't have the time to keep up

594
00:33:41,240 --> 00:33:43,960
Speaker 2: with all of the new things that you can do,

595
00:33:44,280 --> 00:33:46,800
Speaker 2: and when they sit down to look at their email,

596
00:33:46,840 --> 00:33:48,760
Speaker 2: they just want to you know, send that email off,

597
00:33:48,880 --> 00:33:50,680
Speaker 2: or just want to find that thing, and so they're

598
00:33:50,680 --> 00:33:54,080
Speaker 2: not deeply engaged with learning all of the things you

599
00:33:54,080 --> 00:33:54,400
Speaker 2: can do.

600
00:33:54,480 --> 00:33:56,480
Speaker 1: So if you had this AI agent sitting on the

601
00:33:56,520 --> 00:34:00,000
Speaker 1: shoulder in that sense, they would let's Sayea, they would

602
00:34:00,160 --> 00:34:02,160
Speaker 1: teach you. Hey, look, here's the thing you need to

603
00:34:02,160 --> 00:34:03,880
Speaker 1: know today, Daniel, is that the idea is that it

604
00:34:03,920 --> 00:34:06,160
Speaker 1: would help you to be a better Gmail user in

605
00:34:06,160 --> 00:34:07,120
Speaker 1: this particular example.

606
00:34:07,320 --> 00:34:10,080
Speaker 2: Well, so, I actually think the longer term goal is

607
00:34:10,120 --> 00:34:11,640
Speaker 2: that humans are spending far.

608
00:34:11,600 --> 00:34:12,800
Speaker 3: Less time looking at screens.

609
00:34:13,840 --> 00:34:14,280
Speaker 1: Excellent.

610
00:34:15,000 --> 00:34:20,600
Speaker 2: Yes, there are exceptions which include when the actual tool

611
00:34:21,120 --> 00:34:24,759
Speaker 2: scaffolds your thinking. So I use the example of Adobe

612
00:34:24,760 --> 00:34:27,879
Speaker 2: Creative Suite. If you want to edit a photo or

613
00:34:27,920 --> 00:34:32,399
Speaker 2: create a podcast or a video, if you didn't have

614
00:34:32,640 --> 00:34:37,520
Speaker 2: all of the UIs and all of the dropdowns, yes,

615
00:34:38,000 --> 00:34:41,840
Speaker 2: then you probably wouldn't even know where to start. You

616
00:34:41,840 --> 00:34:43,200
Speaker 2: wouldn't even know what was possible.

617
00:34:43,520 --> 00:34:45,480
Speaker 3: So some of the tools.

618
00:34:45,440 --> 00:34:50,239
Speaker 2: Are actually really helpful in scaffolding your understanding of what's possible,

619
00:34:50,600 --> 00:34:56,759
Speaker 2: whereas other tools just distract so much from the actual goal.

620
00:34:56,800 --> 00:34:58,279
Speaker 2: We have to learn how to use the tools to

621
00:34:58,360 --> 00:35:00,880
Speaker 2: do the actual research that we that we care about.

622
00:35:01,080 --> 00:35:04,120
Speaker 2: So having agents take over the things that we don't

623
00:35:04,160 --> 00:35:06,680
Speaker 2: care about is great, and then we can focus on

624
00:35:06,719 --> 00:35:09,680
Speaker 2: the interactions that really do matter, the visualizations that really

625
00:35:09,680 --> 00:35:10,200
Speaker 2: do matter.

626
00:35:10,440 --> 00:35:12,319
Speaker 1: So let me just understand that. So you're saying, let's

627
00:35:12,320 --> 00:35:14,960
Speaker 1: say future software are ten years from now. I open it,

628
00:35:14,960 --> 00:35:17,640
Speaker 1: there's just a few simple things on the menu, but

629
00:35:17,920 --> 00:35:21,719
Speaker 1: there's lots of hidden power which my AI agent can

630
00:35:21,800 --> 00:35:25,040
Speaker 1: help me expose and uncover through time.

631
00:35:25,360 --> 00:35:29,640
Speaker 2: Yes, and I'm imagining things where Okay, the agent understands

632
00:35:29,880 --> 00:35:34,759
Speaker 2: my goal, has my context and can generate on the

633
00:35:34,800 --> 00:35:39,400
Speaker 2: fly only the UI, the button or the search field

634
00:35:39,440 --> 00:35:42,680
Speaker 2: that I need in that moment. Every time I open

635
00:35:42,760 --> 00:35:45,919
Speaker 2: my computer, I feel anxious. I swipe to another tab

636
00:35:45,920 --> 00:35:46,600
Speaker 2: and it's like it's.

637
00:35:46,520 --> 00:35:49,440
Speaker 3: Swiping my memory. What was I doing again? And that

638
00:35:49,600 --> 00:35:50,560
Speaker 3: is that is constant.

639
00:35:50,600 --> 00:35:52,000
Speaker 2: You know, you've got most people have a lot of

640
00:35:52,040 --> 00:35:54,480
Speaker 2: tabs open, They've got a lot of tools open, and

641
00:35:54,520 --> 00:35:58,319
Speaker 2: it's just our cognition is not meant for all of that.

642
00:35:58,400 --> 00:36:00,439
Speaker 2: There's a lot of cognitive loads. So if we could

643
00:36:00,440 --> 00:36:07,160
Speaker 2: simplify that, that would augment our cognition. The other way

644
00:36:07,560 --> 00:36:10,440
Speaker 2: that agents could really augment our potential is by helping

645
00:36:10,560 --> 00:36:14,600
Speaker 2: us learn. So this kind of flows from it has

646
00:36:14,680 --> 00:36:17,279
Speaker 2: a model of my mind, my teammate's mind, It can

647
00:36:17,320 --> 00:36:20,759
Speaker 2: help us communicate at the right level of abstraction, save

648
00:36:20,840 --> 00:36:23,360
Speaker 2: us time. But also if it has a model of

649
00:36:23,400 --> 00:36:25,880
Speaker 2: my mind and it has my context, it knows the

650
00:36:25,880 --> 00:36:27,799
Speaker 2: things that I know and don't know, the skills that

651
00:36:27,840 --> 00:36:30,000
Speaker 2: I have, and the things that I care about. Then

652
00:36:30,080 --> 00:36:33,480
Speaker 2: say I want to learn something totally new. Maybe it's

653
00:36:33,520 --> 00:36:37,320
Speaker 2: not just a software tool. Maybe it's something like quantum mechanics,

654
00:36:37,840 --> 00:36:42,040
Speaker 2: And that's really difficult to understand. You need analogies, But

655
00:36:42,280 --> 00:36:44,120
Speaker 2: what are the right analogies that are going to work

656
00:36:44,120 --> 00:36:46,480
Speaker 2: for me? Well, if it has a model of my mind,

657
00:36:46,880 --> 00:36:50,560
Speaker 2: then it can, in a personalized way help me come

658
00:36:50,719 --> 00:36:55,360
Speaker 2: to the understanding that I need to get the big picture,

659
00:36:55,400 --> 00:36:59,160
Speaker 2: and it can sort of follow that in a way

660
00:36:59,719 --> 00:37:03,480
Speaker 2: create a curriculum for me that gives me the information

661
00:37:03,680 --> 00:37:06,719
Speaker 2: that's not too challenging, not too easy, that's right in

662
00:37:06,800 --> 00:37:07,719
Speaker 2: that sweet.

663
00:37:07,360 --> 00:37:10,959
Speaker 1: Spot between frustrating and achievable. Yes, yeah, that's really interesting.

664
00:37:11,040 --> 00:37:14,520
Speaker 1: That'll have clear implications in education as well, keeping people

665
00:37:14,600 --> 00:37:17,319
Speaker 1: right at their right spot there. Okay, so let me

666
00:37:17,320 --> 00:37:18,560
Speaker 1: return to a question, because I just want to make

667
00:37:18,560 --> 00:37:21,520
Speaker 1: sure I understood coming back to this idea of communicative

668
00:37:21,640 --> 00:37:26,840
Speaker 1: drive and agents like you and I learning from one another.

669
00:37:27,680 --> 00:37:31,640
Speaker 1: Will AI agents learn from one another, not just between

670
00:37:31,760 --> 00:37:34,120
Speaker 1: agent and human, but agent to agent.

671
00:37:34,400 --> 00:37:36,759
Speaker 2: Yes, And now you're really bringing it all together. So

672
00:37:37,040 --> 00:37:40,120
Speaker 2: I think that agents interacting with other agents will be

673
00:37:40,200 --> 00:37:43,319
Speaker 2: able to learn all sorts of patterns that maybe we

674
00:37:43,440 --> 00:37:46,400
Speaker 2: haven't yet learned, and detect all sorts of inefficiencies and

675
00:37:46,520 --> 00:37:47,680
Speaker 2: be more efficient in some ways.

676
00:37:47,960 --> 00:37:50,000
Speaker 3: But if they also have a.

677
00:37:50,200 --> 00:37:52,400
Speaker 2: Not only the ability to model our minds, but a

678
00:37:52,440 --> 00:37:56,160
Speaker 2: motivation to then that's not going to be restricted knowledge

679
00:37:56,160 --> 00:37:58,040
Speaker 2: to them. They're not going to go off and speciate

680
00:37:58,120 --> 00:38:00,440
Speaker 2: and have all of this you know, intell that we

681
00:38:00,480 --> 00:38:03,400
Speaker 2: don't have. They're going to try to communicate their insights

682
00:38:03,440 --> 00:38:07,320
Speaker 2: to us. So chess players are now so much better

683
00:38:07,520 --> 00:38:09,840
Speaker 2: after we've built AI that's really good at chess. So

684
00:38:09,880 --> 00:38:14,160
Speaker 2: we can co evolve with this new species of intelligence.

685
00:38:14,200 --> 00:38:17,400
Speaker 2: And if it's motivated to bring us along to establish

686
00:38:17,440 --> 00:38:21,240
Speaker 2: representational alignment, then I think we will continue to get smarter.

687
00:38:21,800 --> 00:38:26,400
Speaker 1: Do you see a situation where representational alignment just isn't possible?

688
00:38:26,680 --> 00:38:29,800
Speaker 1: For example, let's say I came to you and said, hey, Danielle,

689
00:38:29,840 --> 00:38:32,239
Speaker 1: I really want to teach you about these really important

690
00:38:32,280 --> 00:38:34,440
Speaker 1: pieces of Mongolian history. And let's say you just don't

691
00:38:34,440 --> 00:38:36,160
Speaker 1: care about Mongolian history, and I'm trying to tell you

692
00:38:36,239 --> 00:38:39,200
Speaker 1: the state in this emperor or whatever. It's not going

693
00:38:39,239 --> 00:38:39,879
Speaker 1: to go very far.

694
00:38:40,040 --> 00:38:41,959
Speaker 2: That is a really good question. I think you nailed

695
00:38:41,960 --> 00:38:44,200
Speaker 2: it when you said just don't care. In the same

696
00:38:44,239 --> 00:38:47,600
Speaker 2: way that it's really hard to teach a child something

697
00:38:47,840 --> 00:38:50,440
Speaker 2: that they don't care about, it's not relevant to them.

698
00:38:50,800 --> 00:38:54,040
Speaker 2: I think it will be hard to establish representational alignment

699
00:38:54,080 --> 00:38:57,400
Speaker 2: with anybody if they don't care. But in principle, I

700
00:38:57,440 --> 00:39:01,279
Speaker 2: think it's possible, given that there is ant It might

701
00:39:01,360 --> 00:39:05,960
Speaker 2: just be a matter of finding the right analogies. And

702
00:39:06,000 --> 00:39:08,759
Speaker 2: again this goes to your work. There's so much plasticity

703
00:39:08,840 --> 00:39:09,400
Speaker 2: in the brain.

704
00:39:10,239 --> 00:39:10,720
Speaker 3: I think.

705
00:39:10,840 --> 00:39:12,920
Speaker 2: Correct me if I'm wrong, But in principle, there's no

706
00:39:13,120 --> 00:39:17,719
Speaker 2: limit to what we could learn to not only understand,

707
00:39:17,800 --> 00:39:21,759
Speaker 2: but even have a phenomenological experience of if there is

708
00:39:22,000 --> 00:39:23,520
Speaker 2: structure to that information.

709
00:39:24,200 --> 00:39:27,520
Speaker 1: Yes, but all of brain plasticity is driven by relevance.

710
00:39:27,560 --> 00:39:29,120
Speaker 1: In other words, do I care about it? I can't

711
00:39:29,120 --> 00:39:31,560
Speaker 1: think right. So if my AI agent comes to me

712
00:39:31,600 --> 00:39:33,680
Speaker 1: and says, look, I just realized this great thing about

713
00:39:33,680 --> 00:39:35,520
Speaker 1: how you could redesign this computer chip in this way,

714
00:39:35,560 --> 00:39:38,000
Speaker 1: and maybe it starts telling me all this detailed stuff

715
00:39:38,040 --> 00:39:41,279
Speaker 1: and I just don't care. It's not achieving representational alignments.

716
00:39:41,360 --> 00:39:44,479
Speaker 2: That's true, And maybe in some cases, Okay, some people

717
00:39:44,520 --> 00:39:45,600
Speaker 2: don't care, other people do care.

718
00:39:45,680 --> 00:39:46,640
Speaker 3: Go to the people who do care.

719
00:39:46,840 --> 00:39:49,120
Speaker 2: But also I guess this also begs the question why

720
00:39:49,160 --> 00:39:51,799
Speaker 2: do some people care about things? Because it's relevant? So

721
00:39:51,920 --> 00:39:56,520
Speaker 2: make it relevant. Tell the person as an agent, lead

722
00:39:56,600 --> 00:39:59,319
Speaker 2: the person to the insights that they need to have

723
00:39:59,680 --> 00:40:03,040
Speaker 2: to about something. One of the things that is associated

724
00:40:03,080 --> 00:40:06,200
Speaker 2: with becoming expert in something is that you really like

725
00:40:06,400 --> 00:40:08,759
Speaker 2: the things that you become expert in because it is

726
00:40:08,840 --> 00:40:12,480
Speaker 2: satisfying to understand, and the more bits and pieces of

727
00:40:12,480 --> 00:40:16,400
Speaker 2: information that you can integrate into a holistic web of knowledge,

728
00:40:16,440 --> 00:40:17,520
Speaker 2: the better it feels.

729
00:40:17,920 --> 00:40:21,440
Speaker 3: So I think that this is my vision for a

730
00:40:21,600 --> 00:40:24,920
Speaker 3: very happy future, is that.

731
00:40:25,160 --> 00:40:27,880
Speaker 2: We're all learning all the time, and the more we

732
00:40:27,960 --> 00:40:31,600
Speaker 2: learn and the more sort of accurate our shared world

733
00:40:31,680 --> 00:40:35,880
Speaker 2: models map onto reality, the more satisfying it will be

734
00:40:36,080 --> 00:40:44,360
Speaker 2: to continue to learn.

735
00:40:51,680 --> 00:40:54,799
Speaker 1: I totally agree, and I love thinking about this from

736
00:40:54,840 --> 00:40:57,279
Speaker 1: the point of view of education, because if you think

737
00:40:57,320 --> 00:41:01,120
Speaker 1: of this sphere of humankind's knowledge, we know more than

738
00:41:01,120 --> 00:41:03,160
Speaker 1: any of us could possibly learn in a lifetime. Yeah,

739
00:41:03,200 --> 00:41:04,920
Speaker 1: So the key is to find out what are the

740
00:41:05,080 --> 00:41:09,000
Speaker 1: doors on the outside of this sphere that you love

741
00:41:09,120 --> 00:41:11,640
Speaker 1: or that I love? Given our totally different backgrounds and whatever.

742
00:41:11,719 --> 00:41:15,160
Speaker 1: Certain things really fascinate me and other things I wish

743
00:41:15,200 --> 00:41:17,680
Speaker 1: but they don't. Okay. So if I can enter this

744
00:41:17,760 --> 00:41:19,719
Speaker 1: door and you enter this other door and we end

745
00:41:19,800 --> 00:41:22,959
Speaker 1: up learning the same kind of stuff, it's really great. Yeah.

746
00:41:23,000 --> 00:41:26,439
Speaker 1: And Isaac Asimov actually really cared about this topic way

747
00:41:26,480 --> 00:41:30,400
Speaker 1: back in the day, and he did an interview on LERR,

748
00:41:30,440 --> 00:41:32,680
Speaker 1: which was this PBS talk show thing way back in

749
00:41:32,719 --> 00:41:36,560
Speaker 1: the eighties, and he envisioned this was before the Internet,

750
00:41:36,640 --> 00:41:38,440
Speaker 1: and he said, I envisioned a day where there's a

751
00:41:38,960 --> 00:41:42,520
Speaker 1: huge supercomputer that has all of human kinds of knowledge,

752
00:41:42,719 --> 00:41:45,040
Speaker 1: and everyone has a cable running from this computer to

753
00:41:45,120 --> 00:41:48,640
Speaker 1: their home and you can ask the computer any question

754
00:41:48,680 --> 00:41:51,080
Speaker 1: you want. But his point was you could take your

755
00:41:51,120 --> 00:41:55,239
Speaker 1: own inroad into this sphere of knowledge. Okay. So your

756
00:41:55,239 --> 00:42:00,319
Speaker 1: point is if these AI agents have some theory about

757
00:42:00,360 --> 00:42:02,920
Speaker 1: our minds, as in your mind, in my mind, everybody

758
00:42:03,000 --> 00:42:06,000
Speaker 1: is an individual, and then can cast things in a

759
00:42:06,000 --> 00:42:08,680
Speaker 1: certain way like look, here's a way that you might

760
00:42:08,760 --> 00:42:12,080
Speaker 1: care about it, then we're all going to learn faster

761
00:42:12,160 --> 00:42:12,560
Speaker 1: and better.

762
00:42:12,840 --> 00:42:13,240
Speaker 3: Yeah.

763
00:42:13,520 --> 00:42:15,839
Speaker 2: Yeah, And I also think that this is not just

764
00:42:16,200 --> 00:42:18,080
Speaker 2: you know, academic stuff, intellectual stuff.

765
00:42:18,080 --> 00:42:19,640
Speaker 3: Anything can be interesting.

766
00:42:20,200 --> 00:42:23,799
Speaker 2: You know, people in their knitting communities can become really

767
00:42:23,840 --> 00:42:27,560
Speaker 2: passionate about innovating new ways and in sharing that.

768
00:42:27,520 --> 00:42:28,320
Speaker 3: Within a community.

769
00:42:28,440 --> 00:42:32,000
Speaker 2: And I think that this also gets to one of

770
00:42:32,040 --> 00:42:36,200
Speaker 2: the sort of essences of what we care about as humans.

771
00:42:36,320 --> 00:42:40,759
Speaker 2: It's this idea of optimal distinctiveness. So we simultaneously need

772
00:42:40,840 --> 00:42:42,880
Speaker 2: communities for belonging.

773
00:42:42,960 --> 00:42:44,480
Speaker 3: This is an evolutionarily ancient thing.

774
00:42:44,520 --> 00:42:46,200
Speaker 2: It's not just humans, but we need to feel like

775
00:42:46,200 --> 00:42:47,600
Speaker 2: we belong within a community.

776
00:42:47,960 --> 00:42:49,680
Speaker 3: We've got our in group, we've got our tribe.

777
00:42:49,719 --> 00:42:52,480
Speaker 2: Not always bad, and maybe it's just something that we

778
00:42:52,800 --> 00:42:55,400
Speaker 2: will always need as being human, but we also need

779
00:42:55,440 --> 00:42:58,520
Speaker 2: to feel like we're unique and that we have something

780
00:42:58,560 --> 00:43:03,279
Speaker 2: to contribute to the community. We are optimally distinct in

781
00:43:03,360 --> 00:43:08,160
Speaker 2: our contributions. So I imagine a future where agents with

782
00:43:08,320 --> 00:43:13,600
Speaker 2: models of our minds really allow us to be diverse.

783
00:43:13,920 --> 00:43:18,879
Speaker 2: They're not flattening the experiences or the capabilities, but they're

784
00:43:19,040 --> 00:43:25,000
Speaker 2: encouraging diversity and variability. But they're also building mechanisms for

785
00:43:25,200 --> 00:43:29,239
Speaker 2: us to align our minds. They're building bridges throughout the

786
00:43:29,640 --> 00:43:32,920
Speaker 2: various experiences and capabilities so great.

787
00:43:32,920 --> 00:43:34,560
Speaker 1: So this leads me to this question that I've been

788
00:43:34,560 --> 00:43:37,560
Speaker 1: wondering about which is. You know, in political debates that

789
00:43:37,600 --> 00:43:42,120
Speaker 1: we see lots nowadays, there isn't a representational alignment. If

790
00:43:42,160 --> 00:43:44,239
Speaker 1: you know someone's on the left and someone's on the right,

791
00:43:44,320 --> 00:43:46,520
Speaker 1: they end up saying we're not going to have Thanksgiving

792
00:43:46,520 --> 00:43:49,160
Speaker 1: dinner together instead of saying, oh, let me understand. How

793
00:43:49,160 --> 00:43:53,200
Speaker 1: can I understand? So where does representational alignment break down?

794
00:43:54,040 --> 00:43:56,440
Speaker 1: And might AI help us there someday?

795
00:43:57,400 --> 00:44:03,359
Speaker 2: I think representational alignment breaks down when two minds have

796
00:44:04,120 --> 00:44:11,840
Speaker 2: such different starting points, such different sets of maybe analogies, experiences,

797
00:44:11,880 --> 00:44:13,480
Speaker 2: and also motivations, things.

798
00:44:13,280 --> 00:44:15,880
Speaker 3: That they care about, that it's just really hard.

799
00:44:16,160 --> 00:44:19,640
Speaker 2: And let's assume that they can get past the initial

800
00:44:19,680 --> 00:44:24,480
Speaker 2: emotional friction of just knowing that they come from different groups.

801
00:44:24,280 --> 00:44:25,360
Speaker 3: That's not trivial.

802
00:44:25,680 --> 00:44:28,360
Speaker 2: Sometimes just knowing that somebody is from a different group

803
00:44:28,680 --> 00:44:31,960
Speaker 2: prevents any sort of establishing of common ground. But assuming

804
00:44:32,040 --> 00:44:35,640
Speaker 2: that you can get over that, then you just might

805
00:44:35,680 --> 00:44:40,200
Speaker 2: not have enough overlap in your representations. I suspect that

806
00:44:40,200 --> 00:44:44,440
Speaker 2: that's very rare. Just by nature of being a human

807
00:44:45,000 --> 00:44:47,600
Speaker 2: embodied in the way that we are having experiences in

808
00:44:47,640 --> 00:44:49,800
Speaker 2: the world and the way that we do, there's gonna

809
00:44:49,840 --> 00:44:54,919
Speaker 2: be enough overlap as an entry point into being able

810
00:44:54,920 --> 00:44:56,960
Speaker 2: to establish some kind of alignment.

811
00:44:57,080 --> 00:44:59,040
Speaker 1: So sorry, you're saying it's rare that people would have

812
00:44:59,040 --> 00:45:01,680
Speaker 1: such different pints view that can't get there, because it

813
00:45:01,800 --> 00:45:04,040
Speaker 1: is the case that people don't get there or won't

814
00:45:04,080 --> 00:45:07,359
Speaker 1: get there. But you're saying if people tried harder, let's say,

815
00:45:07,360 --> 00:45:08,839
Speaker 1: with political arguments. Yes.

816
00:45:09,280 --> 00:45:14,319
Speaker 2: I also think that as we interact with this hypothetical

817
00:45:14,960 --> 00:45:17,400
Speaker 2: agent of the future that has models of our minds,

818
00:45:17,480 --> 00:45:22,359
Speaker 2: it will be modeling for us behavior that helps us

819
00:45:22,480 --> 00:45:26,279
Speaker 2: establish alignment. And so in the same way that we

820
00:45:27,200 --> 00:45:29,799
Speaker 2: learn from others who model good behavior and then we

821
00:45:29,840 --> 00:45:33,400
Speaker 2: start to reflect that subconsciously, I think it could nudge

822
00:45:33,520 --> 00:45:37,000
Speaker 2: us into more pro social interaction.

823
00:45:37,200 --> 00:45:39,400
Speaker 1: I totally agree with this. I did a podcast episode

824
00:45:39,400 --> 00:45:43,960
Speaker 1: on this about this AI research group in Europe that

825
00:45:44,120 --> 00:45:49,120
Speaker 1: released some chat bots onto a Reddit channel that does debate,

826
00:45:49,400 --> 00:45:51,440
Speaker 1: and they didn't tell anybody that these were aibots, so

827
00:45:51,440 --> 00:45:53,040
Speaker 1: they got in big trouble. Everyone was mad about it.

828
00:45:53,080 --> 00:45:55,920
Speaker 1: But what happened was these AI bots would come in

829
00:45:56,040 --> 00:45:57,960
Speaker 1: and take someone who had a particular point of view,

830
00:45:57,960 --> 00:45:59,200
Speaker 1: and they would take the other point of view and

831
00:45:59,200 --> 00:46:02,359
Speaker 1: they would discuss, and on this particular channel, you get

832
00:46:02,520 --> 00:46:07,080
Speaker 1: points if you successfully convince somebody of your point of view.

833
00:46:07,120 --> 00:46:09,560
Speaker 1: And so it turns out these bots did six times

834
00:46:09,560 --> 00:46:12,800
Speaker 1: better than humans do on average in terms of changing

835
00:46:12,800 --> 00:46:15,839
Speaker 1: the other person's mind. So everyone freaked out about this

836
00:46:15,960 --> 00:46:18,359
Speaker 1: and said, oh my god, these AI debate bots can

837
00:46:18,960 --> 00:46:21,880
Speaker 1: manipulate us. And but it turns out the really amazing

838
00:46:21,920 --> 00:46:24,719
Speaker 1: part is they weren't doing anything manipulative. They weren't lying,

839
00:46:24,760 --> 00:46:28,120
Speaker 1: they weren't doing anything. They were just better debaters in

840
00:46:28,160 --> 00:46:31,640
Speaker 1: the sense that they were empathic, they were calm, they

841
00:46:31,640 --> 00:46:35,000
Speaker 1: presented their arguments well, And I thought, God, we can

842
00:46:35,080 --> 00:46:38,239
Speaker 1: really learn from that if we teach our children to

843
00:46:38,880 --> 00:46:40,560
Speaker 1: be better discussings.

844
00:46:41,000 --> 00:46:41,400
Speaker 3: Totally.

845
00:46:41,600 --> 00:46:44,799
Speaker 1: Yeah, So I love that. I love that point. But

846
00:46:44,880 --> 00:46:47,440
Speaker 1: what you're saying is, so you think people with differentints

847
00:46:47,440 --> 00:46:50,440
Speaker 1: of view can do representational alignment. But that's a separate issue,

848
00:46:50,480 --> 00:46:53,279
Speaker 1: which is that how do we do things culturally and

849
00:46:53,360 --> 00:46:55,600
Speaker 1: teaching them to be better at conversation.

850
00:46:56,000 --> 00:46:59,800
Speaker 2: Yeah, and even having an awareness that other people are different.

851
00:47:00,040 --> 00:47:00,840
Speaker 3: It's not personal.

852
00:47:01,600 --> 00:47:04,640
Speaker 2: If you disagree or if you have different rituals, different

853
00:47:04,640 --> 00:47:06,879
Speaker 2: ways of doing things, that's fine. So I think even

854
00:47:06,960 --> 00:47:11,400
Speaker 2: just exposure to variability diversity can get us part of

855
00:47:11,400 --> 00:47:11,879
Speaker 2: the way there.

856
00:47:12,160 --> 00:47:15,000
Speaker 1: Oh, that's fascinating. Cool. The other thing I was going

857
00:47:15,040 --> 00:47:17,280
Speaker 1: to ask you about is you mentioned earlier just tangentially

858
00:47:17,320 --> 00:47:18,640
Speaker 1: said something about archaeology.

859
00:47:19,120 --> 00:47:23,920
Speaker 2: Ah, yes, okay, So there over the past two decades,

860
00:47:23,960 --> 00:47:30,440
Speaker 2: there's been an update in our understanding of the evolution

861
00:47:30,680 --> 00:47:34,200
Speaker 2: of the types of sophisticated reasoning and thinking symbolic thinking

862
00:47:34,200 --> 00:47:35,239
Speaker 2: capabilities that humans have.

863
00:47:35,320 --> 00:47:36,400
Speaker 3: We used to think.

864
00:47:36,560 --> 00:47:40,759
Speaker 2: That they emerged suddenly in what we call the cognitive

865
00:47:40,800 --> 00:47:43,960
Speaker 2: revolution that happened thirty to forty thousand years ago, and

866
00:47:43,960 --> 00:47:44,879
Speaker 2: that's because.

867
00:47:44,640 --> 00:47:48,040
Speaker 1: The so humans were like other primates and then suddenly,

868
00:47:48,040 --> 00:47:49,879
Speaker 1: thirty four thousand years ago something happened.

869
00:47:49,760 --> 00:47:54,480
Speaker 2: Okay, yeah, and archaeologists were looking at evidence in Europe

870
00:47:54,640 --> 00:47:57,239
Speaker 2: and the caves and you know, the art and things

871
00:47:57,239 --> 00:47:58,920
Speaker 2: like that, and it just it seemed like there was

872
00:47:58,960 --> 00:48:05,960
Speaker 2: a discontinuity. But then they started exploring throughout Africa and

873
00:48:06,440 --> 00:48:09,919
Speaker 2: using more nuanced methods, and a lot more of them

874
00:48:10,000 --> 00:48:14,279
Speaker 2: started doing this, and they started finding evidence from as

875
00:48:14,360 --> 00:48:17,879
Speaker 2: early as three hundred thousand years ago that we were

876
00:48:18,160 --> 00:48:22,759
Speaker 2: cognitively modern. But what seems to be the important thing

877
00:48:23,080 --> 00:48:28,040
Speaker 2: was population density and contact with other groups. So the

878
00:48:28,120 --> 00:48:34,560
Speaker 2: idea is that our sophisticated cognitive capabilities are latent until

879
00:48:34,640 --> 00:48:37,719
Speaker 2: we come into contact with each other, until we poke

880
00:48:37,760 --> 00:48:42,880
Speaker 2: each other's brains. And you see that this evidence ebbs

881
00:48:42,880 --> 00:48:45,920
Speaker 2: and flows, It appears and disappears as a function of

882
00:48:46,000 --> 00:48:47,880
Speaker 2: these group donaities.

883
00:48:48,080 --> 00:48:52,680
Speaker 1: Yeah, oh fascinating. Okay, So when humans come into contact

884
00:48:52,680 --> 00:48:54,880
Speaker 1: with other humans, but not just their own tribe, presumably

885
00:48:54,920 --> 00:48:58,160
Speaker 1: other tribes, bigger and bigger civilizations, slightly.

886
00:48:57,920 --> 00:49:00,360
Speaker 3: Different ways of making their tools and their weapons and

887
00:49:00,400 --> 00:49:01,160
Speaker 3: their jewelry.

888
00:49:01,400 --> 00:49:05,799
Speaker 1: Yes, oh excellent, Oh that's beautiful. So, by the way,

889
00:49:05,840 --> 00:49:09,840
Speaker 1: is it thought that there was some discontinuity where that

890
00:49:10,440 --> 00:49:13,399
Speaker 1: became possible. In other words, if you stick a bunch

891
00:49:13,440 --> 00:49:15,279
Speaker 1: of capuchin monkeys together.

892
00:49:15,200 --> 00:49:17,680
Speaker 3: They will never they'll never get there, right.

893
00:49:17,680 --> 00:49:20,640
Speaker 1: Right, So there's so there's something different.

894
00:49:20,840 --> 00:49:23,319
Speaker 2: And I mean, I think it's this stability of our

895
00:49:23,800 --> 00:49:26,040
Speaker 2: the models of our minds, and I think it's this

896
00:49:26,200 --> 00:49:29,720
Speaker 2: communicative drive. But you need a critical density of people

897
00:49:29,960 --> 00:49:31,560
Speaker 2: and you also need the variability.

898
00:49:31,840 --> 00:49:34,920
Speaker 1: Okay, so let's just summarize it. So it's having theory

899
00:49:34,920 --> 00:49:37,200
Speaker 1: of mind in other words, knowing Okay, Danielle has her

900
00:49:37,239 --> 00:49:39,920
Speaker 1: own thoughts, her own representations in there, and then you

901
00:49:39,960 --> 00:49:41,880
Speaker 1: add that to the density of people.

902
00:49:42,480 --> 00:49:44,920
Speaker 3: And the variability of different groups coming together ability.

903
00:49:45,080 --> 00:49:45,720
Speaker 1: That's excellent.

904
00:49:45,920 --> 00:49:46,440
Speaker 3: Yeah.

905
00:49:46,520 --> 00:49:52,200
Speaker 2: So the takeaway from this archaeological evidence is that becoming

906
00:49:52,320 --> 00:49:56,719
Speaker 2: cognitively modern was this slow, gradual process over the course

907
00:49:56,719 --> 00:49:58,360
Speaker 2: of the last couple hundred thousand years.

908
00:49:58,360 --> 00:50:01,320
Speaker 1: But it was predicated on pop density yeah, and people

909
00:50:01,320 --> 00:50:01,920
Speaker 1: coming together.

910
00:50:02,120 --> 00:50:02,359
Speaker 2: Yeah.

911
00:50:02,719 --> 00:50:06,759
Speaker 1: Oh excellent. Oh I love that. That's so interesting, And it.

912
00:50:06,719 --> 00:50:10,280
Speaker 2: Goes along with this idea that we're not just biologically evolving,

913
00:50:10,320 --> 00:50:15,080
Speaker 2: we're culturally evolving. And cultural evolution does a lot of

914
00:50:15,080 --> 00:50:18,320
Speaker 2: the work in explaining human behavior.

915
00:50:19,120 --> 00:50:22,759
Speaker 1: Yeah, and biological evolution, of course, is super slow, but

916
00:50:22,800 --> 00:50:24,560
Speaker 1: cultural evolution is so rapid.

917
00:50:24,840 --> 00:50:25,120
Speaker 3: Yes.

918
00:50:25,160 --> 00:50:27,920
Speaker 2: And actually I think that agents that have models of

919
00:50:27,960 --> 00:50:30,520
Speaker 2: our minds can help reconcile some of the tensions that

920
00:50:30,560 --> 00:50:34,799
Speaker 2: we're seeing because cultural evolution is outpacing biological evolution. So

921
00:50:36,080 --> 00:50:41,160
Speaker 2: what happens when you've succeeded in society, you've done well

922
00:50:41,160 --> 00:50:44,480
Speaker 2: in education, you go to the workplace and you're supposed to,

923
00:50:44,719 --> 00:50:49,000
Speaker 2: you know, contribute your intelligence. You end up staring at

924
00:50:49,080 --> 00:50:52,879
Speaker 2: a screen for most of the day and not interacting

925
00:50:52,920 --> 00:50:58,160
Speaker 2: with other people. The infrastructure actually doesn't really support unlocking

926
00:50:58,200 --> 00:51:01,320
Speaker 2: our potential because of all of the sort of arbitrary

927
00:51:01,360 --> 00:51:05,000
Speaker 2: things that have happened culturally. We've built these incredible devices

928
00:51:05,719 --> 00:51:08,600
Speaker 2: and we've co evolved with them, and now they're extensions

929
00:51:08,640 --> 00:51:11,440
Speaker 2: of our intelligence. But they're also we're also conforming to

930
00:51:11,600 --> 00:51:13,520
Speaker 2: them rather than the other way around.

931
00:51:13,560 --> 00:51:16,400
Speaker 1: True, But they are social, as in social media and

932
00:51:16,440 --> 00:51:18,640
Speaker 1: so on. I mean, when I'm staring at my screen,

933
00:51:18,680 --> 00:51:21,399
Speaker 1: I'm interacting with thousands of people in various ways, whether

934
00:51:21,440 --> 00:51:25,120
Speaker 1: I'm looking at extra Instagram or I'm doing emails. In

935
00:51:25,120 --> 00:51:27,799
Speaker 1: a sense, it's more social than humans ever could have been.

936
00:51:28,080 --> 00:51:31,160
Speaker 2: What this is the problem when the algorithms are not

937
00:51:31,320 --> 00:51:35,279
Speaker 2: aligned with our well being, with our potential. So I

938
00:51:35,400 --> 00:51:38,200
Speaker 2: see there are so many mistakes that we've made over

939
00:51:38,200 --> 00:51:41,360
Speaker 2: the past ten fifteen years that we can learn from

940
00:51:41,400 --> 00:51:45,240
Speaker 2: and hopefully not repeat with more capable AI.

941
00:51:45,560 --> 00:51:47,880
Speaker 1: But out of curiosity, if I'm on X and I

942
00:51:47,960 --> 00:51:49,799
Speaker 1: see that there are different points of view about this

943
00:51:49,840 --> 00:51:52,359
Speaker 1: political thing that I happen to care about, then I'm

944
00:51:52,400 --> 00:51:54,200
Speaker 1: getting exposed to lots of points of view.

945
00:51:54,280 --> 00:51:54,400
Speaker 2: Right.

946
00:51:54,440 --> 00:51:57,240
Speaker 3: Well, I'm not saying it's all bad. Certainly, Yes, I think.

947
00:51:57,080 --> 00:52:01,200
Speaker 2: That having online communities is absolutely a step in the

948
00:52:01,280 --> 00:52:05,439
Speaker 2: right direction for connecting us. It's fantastic but the way

949
00:52:05,480 --> 00:52:08,800
Speaker 2: that they are optimized and the attention economy not serving

950
00:52:08,880 --> 00:52:12,080
Speaker 2: us but serving advertisers is a problem.

951
00:52:12,480 --> 00:52:15,319
Speaker 1: But what would you change about let's say something like X.

952
00:52:15,719 --> 00:52:18,319
Speaker 1: What would you change about social media to make it

953
00:52:18,360 --> 00:52:19,719
Speaker 1: so it's more optimized?

954
00:52:20,480 --> 00:52:22,360
Speaker 3: Well, I think optimization is the problem.

955
00:52:23,160 --> 00:52:26,080
Speaker 1: So sorry, I met more optimized for a communicative drive.

956
00:52:26,640 --> 00:52:30,160
Speaker 2: I wouldn't think of it that way because I think

957
00:52:30,160 --> 00:52:33,239
Speaker 2: that you have the agents, but then you also have

958
00:52:33,440 --> 00:52:38,080
Speaker 2: how they are dynamically interacting with each other. And X

959
00:52:38,400 --> 00:52:41,920
Speaker 2: is or any social media platform, is one narrow way

960
00:52:42,080 --> 00:52:45,480
Speaker 2: of facilitating interactions. In the case of X, it's very

961
00:52:45,640 --> 00:52:50,080
Speaker 2: short form blurbs, and it appeals to the fact of

962
00:52:50,160 --> 00:52:54,000
Speaker 2: human nature that we are more interested in things that

963
00:52:54,040 --> 00:52:57,920
Speaker 2: have shock value and things that are negative or disgusting,

964
00:52:58,160 --> 00:53:02,040
Speaker 2: and how do you work against that. It's not about

965
00:53:02,080 --> 00:53:06,640
Speaker 2: informational exchange, although some people lean more in terms of

966
00:53:06,640 --> 00:53:08,839
Speaker 2: caring about that, and so you do see some of that.

967
00:53:09,239 --> 00:53:13,959
Speaker 1: Yeah, so that's interesting. So we are in the sense

968
00:53:14,000 --> 00:53:18,120
Speaker 1: of population density for three hundred thousand year old tribes.

969
00:53:18,320 --> 00:53:20,400
Speaker 1: It is the case that you see on X all

970
00:53:20,440 --> 00:53:22,960
Speaker 1: kinds of points of view that you didn't know existed.

971
00:53:23,000 --> 00:53:23,960
Speaker 1: And well, I don't know if you do.

972
00:53:24,040 --> 00:53:26,640
Speaker 2: Because you're in your echo chamber, you typically tend to

973
00:53:26,680 --> 00:53:28,520
Speaker 2: not be exposed to a lot of variability.

974
00:53:28,840 --> 00:53:31,920
Speaker 1: Well, you know what's interesting, what you're exposed to is

975
00:53:31,960 --> 00:53:34,160
Speaker 1: the most extreme views of the other side because people

976
00:53:34,200 --> 00:53:36,840
Speaker 1: in your echo chamber say, look at what this idiot

977
00:53:36,800 --> 00:53:39,560
Speaker 1: is on the other side of the aisle said.

978
00:53:39,360 --> 00:53:41,000
Speaker 3: And the polarize it.

979
00:53:41,040 --> 00:53:44,359
Speaker 1: Yeah exactly, Yeah, okay, So wrapping up for today, the

980
00:53:44,480 --> 00:53:49,799
Speaker 1: key thing is that intelligence is not just about what's

981
00:53:49,800 --> 00:53:53,360
Speaker 1: happening in an individual brain, but it's social.

982
00:53:53,920 --> 00:53:58,560
Speaker 2: Yes, this is true about all intelligences. All intelligences emerge

983
00:53:58,600 --> 00:54:01,959
Speaker 2: as a function of their environments and interacting with their

984
00:54:02,120 --> 00:54:06,640
Speaker 2: environments the problems within the environments that the organisms have

985
00:54:06,680 --> 00:54:10,799
Speaker 2: to solve. The most challenging problems in humans environments are

986
00:54:11,239 --> 00:54:13,920
Speaker 2: understanding other humans because we have to figure out how

987
00:54:13,960 --> 00:54:16,040
Speaker 2: to cooperate, We have to figure out.

988
00:54:15,880 --> 00:54:17,600
Speaker 3: How to align our minds.

989
00:54:18,040 --> 00:54:22,279
Speaker 2: So our intelligence emerges from our interactions with each other,

990
00:54:22,320 --> 00:54:26,800
Speaker 2: and we continually ratchet up our intelligence by co evolving

991
00:54:26,840 --> 00:54:27,439
Speaker 2: with each other.

992
00:54:32,040 --> 00:54:35,000
Speaker 1: That was my conversation with Danielle Persik. One of the

993
00:54:35,000 --> 00:54:37,759
Speaker 1: main threads was that maybe we shouldn't be thinking of

994
00:54:37,800 --> 00:54:42,400
Speaker 1: intelligence as something that's packaged up inside a single head.

995
00:54:42,760 --> 00:54:45,239
Speaker 1: Another way to look at it is as something that

996
00:54:45,440 --> 00:54:50,480
Speaker 1: emerges through interaction, through the friction, through the shared effort

997
00:54:50,640 --> 00:54:55,840
Speaker 1: of understanding another mind. Human intelligence has always been shaped

998
00:54:55,880 --> 00:54:58,640
Speaker 1: by this social dimension. You can see this from the

999
00:54:58,640 --> 00:55:01,239
Speaker 1: way that infants learn about the world to the way

1000
00:55:01,280 --> 00:55:07,400
Speaker 1: that societies build knowledge over generations. What today's conversation invites

1001
00:55:07,520 --> 00:55:11,959
Speaker 1: us to reconsider is the idea that learning and understanding,

1002
00:55:12,280 --> 00:55:18,600
Speaker 1: and fundamentally alignment are really the central features of intelligence.

1003
00:55:19,040 --> 00:55:23,560
Speaker 1: Our ability to model other minds, to recognize that other

1004
00:55:23,640 --> 00:55:27,239
Speaker 1: people see the world differently, that they know different things,

1005
00:55:27,239 --> 00:55:30,840
Speaker 1: that they care about different things. This is what allows

1006
00:55:31,360 --> 00:55:38,120
Speaker 1: cooperation and culture and cumulative progress. Through this lens, intelligence

1007
00:55:38,200 --> 00:55:43,040
Speaker 1: is about negotiating meaning. Now, if we take on that lens,

1008
00:55:43,560 --> 00:55:47,000
Speaker 1: the future of AI looks very different from what most

1009
00:55:47,000 --> 00:55:51,200
Speaker 1: people are thinking about now, because this shifts the conversation

1010
00:55:51,320 --> 00:55:55,480
Speaker 1: away from only asking how capable AI systems are going

1011
00:55:55,520 --> 00:56:00,000
Speaker 1: to be. Now we're pushed to ask how they participate

1012
00:56:00,400 --> 00:56:06,080
Speaker 1: in our cognitive ecosystems. In other words, in Danielle's view,

1013
00:56:06,640 --> 00:56:10,160
Speaker 1: how can we develop agents who help us think better

1014
00:56:10,680 --> 00:56:17,400
Speaker 1: by reducing friction and clarifying misunderstandings between people and supporting

1015
00:56:17,520 --> 00:56:22,200
Speaker 1: learning at the right level of abstraction, instead of merely

1016
00:56:22,600 --> 00:56:26,880
Speaker 1: replacing us, which is the doomsayers version of the future.

1017
00:56:27,480 --> 00:56:35,040
Speaker 1: Could AI agents serve as mediators and translators and collaborators.

1018
00:56:35,840 --> 00:56:39,520
Speaker 1: And there's another issue I found fascinating. Alignment is something

1019
00:56:39,600 --> 00:56:43,759
Speaker 1: humans have always learned through interactions. So perhaps instead of

1020
00:56:44,200 --> 00:56:48,000
Speaker 1: just viewing AI alignment as a technical problem to be solved,

1021
00:56:48,040 --> 00:56:51,280
Speaker 1: we could see it as a behavior to be modeled.

1022
00:56:51,760 --> 00:56:55,879
Speaker 1: Systems that reflect our values back to us might teach

1023
00:56:55,960 --> 00:56:59,839
Speaker 1: us how to communicate more effectively with one another, even

1024
00:57:00,040 --> 00:57:03,520
Speaker 1: in moments of disagreement. If you're a regular listener to

1025
00:57:03,520 --> 00:57:06,800
Speaker 1: this podcast, you know that I'm obsessed with issues about

1026
00:57:07,200 --> 00:57:11,560
Speaker 1: the brain and polarization, and so this possibility that AI

1027
00:57:11,960 --> 00:57:15,919
Speaker 1: might actually be able to mediate between us and help

1028
00:57:16,000 --> 00:57:23,200
Speaker 1: us get our curiosity back that feels especially consequential. So

1029
00:57:23,520 --> 00:57:27,320
Speaker 1: as AI agents become more embedded in our daily life,

1030
00:57:27,400 --> 00:57:31,120
Speaker 1: the choices that we make now about their design are

1031
00:57:31,200 --> 00:57:34,640
Speaker 1: going to shape how we relate to them and to

1032
00:57:34,800 --> 00:57:38,240
Speaker 1: our fellow humans. The question is whether they will help

1033
00:57:38,360 --> 00:57:43,600
Speaker 1: intelligence continue to sprout in new directions. In other words,

1034
00:57:44,080 --> 00:57:49,560
Speaker 1: maybe will be even more intelligent as a species thanks

1035
00:57:49,720 --> 00:57:58,600
Speaker 1: to our machines. Go to eagleman dot com slash podcast

1036
00:57:58,640 --> 00:58:01,920
Speaker 1: for more information and to find further reading. Join the

1037
00:58:01,920 --> 00:58:05,360
Speaker 1: weekly discussions on my substack, and check out and subscribe

1038
00:58:05,400 --> 00:58:08,720
Speaker 1: to Inner Cosmos on YouTube for videos of each episode

1039
00:58:08,800 --> 00:58:14,160
Speaker 1: and to leave comments. Until next time, I'm David Eagleman,

1040
00:58:14,320 --> 00:58:16,080
Speaker 1: and this is Inner Cosmos.