1
00:00:05,120 --> 00:00:07,440
Speaker 1: AI seems like it burst out of the gate a

2
00:00:07,480 --> 00:00:10,920
Speaker 1: few years ago, But is it actually the latest chapter

3
00:00:11,119 --> 00:00:16,040
Speaker 1: in a three hundred year trajectory to turn thought into math?

4
00:00:16,680 --> 00:00:19,400
Speaker 2: Can the mind be captured with equations?

5
00:00:19,960 --> 00:00:23,400
Speaker 1: Why do current AI models need petabytes of data but

6
00:00:23,480 --> 00:00:26,759
Speaker 1: a child can learn from just a few examples. Why

7
00:00:26,760 --> 00:00:31,440
Speaker 1: does AI have jagged intelligence, meaning it looks brilliant in

8
00:00:31,480 --> 00:00:36,080
Speaker 1: one moment and then it does something totally nonsensical. In physics,

9
00:00:36,440 --> 00:00:39,479
Speaker 1: we have various laws, like the law of gravity or

10
00:00:39,520 --> 00:00:43,320
Speaker 1: the laws of motion, And today we're joined by cognitive

11
00:00:43,360 --> 00:00:47,080
Speaker 1: scientist Tom Griffiths from Princeton to talk about whether we

12
00:00:47,159 --> 00:00:55,400
Speaker 1: are moving towards nailing down laws of thought. Welcome to

13
00:00:55,400 --> 00:00:58,880
Speaker 1: Inner Cosmos with me David Eagleman. I'm a neuroscientist and

14
00:00:58,920 --> 00:01:02,400
Speaker 1: author at Stanford, and in these episodes we sail deeply

15
00:01:02,440 --> 00:01:05,960
Speaker 1: into our three pound universe to understand why and how.

16
00:01:05,800 --> 00:01:07,880
Speaker 2: Our lives look the way they do.

17
00:01:23,920 --> 00:01:27,000
Speaker 1: One thing that distinguishes Homo sapiens from all our cousins

18
00:01:27,040 --> 00:01:30,119
Speaker 1: in the animal kingdom is that we watch the world

19
00:01:30,120 --> 00:01:34,160
Speaker 1: around us and we try to abstract patterns from it.

20
00:01:34,600 --> 00:01:37,679
Speaker 1: For example, you might watch the way that a stone

21
00:01:37,760 --> 00:01:41,039
Speaker 1: falls to the ground and maybe you see a tree

22
00:01:41,120 --> 00:01:44,200
Speaker 1: branch fall, and maybe you see a glacier and one

23
00:01:44,240 --> 00:01:46,959
Speaker 1: day a huge wall of ice falls off it, and

24
00:01:47,040 --> 00:01:51,000
Speaker 1: pretty soon you start seeing an underlying similarity to the

25
00:01:51,040 --> 00:01:54,600
Speaker 1: way that things move. And eventually someone very very smart

26
00:01:54,680 --> 00:01:58,360
Speaker 1: comes along, like Isaac Newton and summarizes all this in

27
00:01:58,600 --> 00:02:02,080
Speaker 1: the law of gravity. And then the same smart guy,

28
00:02:02,160 --> 00:02:05,600
Speaker 1: Newton comes up with the three laws of motion. And

29
00:02:05,640 --> 00:02:08,560
Speaker 1: then another smart person is Einstein. He figures out the

30
00:02:08,919 --> 00:02:12,840
Speaker 1: conservation of mass and energy, which seems to be another

31
00:02:13,040 --> 00:02:16,919
Speaker 1: ironclad law, and then we have the laws of thermodynamics

32
00:02:16,960 --> 00:02:20,320
Speaker 1: and electrostatic laws, and all of this speaks to the

33
00:02:20,360 --> 00:02:24,079
Speaker 1: great success that we've had as the species in figuring

34
00:02:24,120 --> 00:02:28,160
Speaker 1: out the lowest level of code that's running in the universe.

35
00:02:28,880 --> 00:02:32,120
Speaker 1: But for most of human history, the concept of a

36
00:02:32,600 --> 00:02:35,840
Speaker 1: thought has felt like the most intimate thing we experience

37
00:02:35,960 --> 00:02:38,560
Speaker 1: and the least tractable thing to study.

38
00:02:39,200 --> 00:02:41,840
Speaker 2: What a thought is and how it occurs.

39
00:02:42,280 --> 00:02:46,200
Speaker 1: That seems to live in a different category of mystery

40
00:02:46,240 --> 00:02:49,560
Speaker 1: from how an object falls. Why, Well, it's because the

41
00:02:49,639 --> 00:02:53,600
Speaker 1: thought pops into your head and somehow it carries memory

42
00:02:53,680 --> 00:02:57,760
Speaker 1: and expectation and language and often a feeling. But it

43
00:02:57,840 --> 00:03:02,440
Speaker 1: feels vaporous and private. It feels like the one thing

44
00:03:02,960 --> 00:03:07,880
Speaker 1: that will forever escape formal description. But what's interesting is

45
00:03:07,919 --> 00:03:11,680
Speaker 1: that for centuries people have tried, there's always been a

46
00:03:11,800 --> 00:03:13,399
Speaker 1: deep human urge.

47
00:03:13,120 --> 00:03:16,480
Speaker 2: To ask whether thought has laws to it?

48
00:03:16,560 --> 00:03:19,720
Speaker 1: In other words, does the mind have principles that you

49
00:03:19,760 --> 00:03:23,760
Speaker 1: can write down? Does reasoning have a grammar to it?

50
00:03:24,120 --> 00:03:27,960
Speaker 1: Can you describe intelligence in a language that's precise enough

51
00:03:28,360 --> 00:03:31,440
Speaker 1: that once you understand the rules, you can begin to

52
00:03:31,480 --> 00:03:35,800
Speaker 1: build with them, like build artificial intelligence. Most of us

53
00:03:35,840 --> 00:03:38,400
Speaker 1: are old enough to remember that this question of AI

54
00:03:38,960 --> 00:03:42,840
Speaker 1: once lived in philosophy seminars and math departments, but now

55
00:03:42,880 --> 00:03:45,320
Speaker 1: it's sitting at the center of our economy.

56
00:03:46,280 --> 00:03:48,200
Speaker 2: Okay, So what is thought?

57
00:03:48,440 --> 00:03:52,640
Speaker 1: Can we capture it in formal systems like laws or equations?

58
00:03:53,040 --> 00:03:57,600
Speaker 1: Do different parts of intelligence come from logic, from learning,

59
00:03:57,720 --> 00:04:03,840
Speaker 1: from uncertainty, from memory, from prior knowledge, from living inside bodies,

60
00:04:03,920 --> 00:04:06,120
Speaker 1: from living inside our cultures?

61
00:04:06,640 --> 00:04:08,040
Speaker 2: From the particular.

62
00:04:07,600 --> 00:04:11,120
Speaker 1: Constraints of being a human animal with a short lifespan

63
00:04:11,520 --> 00:04:15,320
Speaker 1: and limited bandwidth. Our guest today is someone who lives

64
00:04:15,400 --> 00:04:19,240
Speaker 1: right at the intersection of all these questions. Tom Griffiths

65
00:04:19,320 --> 00:04:22,880
Speaker 1: is a professor at Princeton, where He directs the Computational

66
00:04:22,960 --> 00:04:27,680
Speaker 1: Cognitive Science Lab and the Princeton Laboratory for Artificial Intelligence.

67
00:04:28,160 --> 00:04:31,719
Speaker 1: He has spent years asking how minds work through the

68
00:04:31,839 --> 00:04:35,760
Speaker 1: different lenses of math and computation and learning. And he's

69
00:04:35,760 --> 00:04:38,919
Speaker 1: the author of a wonderful new book called The Laws

70
00:04:38,960 --> 00:04:43,719
Speaker 1: of Thought, which traces the long history of thinkers asking

71
00:04:44,080 --> 00:04:46,719
Speaker 1: are their rules to this? Can we understand what human

72
00:04:46,760 --> 00:04:50,839
Speaker 1: thinking is? In his book we get the lengthy arc

73
00:04:51,040 --> 00:04:55,839
Speaker 1: of minds trying to understand mind. This begins millennia ago

74
00:04:55,920 --> 00:05:00,520
Speaker 1: with Aristotle, who wondered whether logic itself could be math matized,

75
00:05:00,960 --> 00:05:05,520
Speaker 1: and Tom follows the trail through the architects of symbolic reasoning,

76
00:05:05,960 --> 00:05:09,240
Speaker 1: through the birth of computation, through the rise of neural networks,

77
00:05:09,640 --> 00:05:13,680
Speaker 1: through the realization that probability theory might serve as a

78
00:05:13,800 --> 00:05:17,719
Speaker 1: language for our beliefs about things. Along the way, in

79
00:05:17,760 --> 00:05:20,880
Speaker 1: his book, a picture emerges that there may not be

80
00:05:21,320 --> 00:05:24,440
Speaker 1: just a single tool for capturing their mind, but instead

81
00:05:24,440 --> 00:05:27,760
Speaker 1: there are different ways of trying to tackle the problem,

82
00:05:28,120 --> 00:05:32,520
Speaker 1: and each one sheds light on a different aspect of cognition.

83
00:05:33,240 --> 00:05:36,360
Speaker 1: So we're going to talk about ourselves human minds, and

84
00:05:36,440 --> 00:05:40,160
Speaker 1: we'll talk about AI what kind of intelligence is this

85
00:05:40,440 --> 00:05:47,480
Speaker 1: and what is missing? Here's my interview with Tom Griffiths.

86
00:05:48,160 --> 00:05:50,320
Speaker 3: As soon as you turn thought into math, it becomes

87
00:05:50,320 --> 00:05:52,480
Speaker 3: something that machines would be able to do. And so

88
00:05:52,600 --> 00:05:57,160
Speaker 3: our modern AI systems are really a consequence of, you know,

89
00:05:57,200 --> 00:06:00,520
Speaker 3: that thought that people were having hundreds of years ago,

90
00:06:01,080 --> 00:06:04,160
Speaker 3: of being able to turn thought into something that can

91
00:06:04,160 --> 00:06:05,800
Speaker 3: be expressed in mathematical terms.

92
00:06:05,920 --> 00:06:07,440
Speaker 1: And so one of the things that I loved about

93
00:06:07,440 --> 00:06:09,400
Speaker 1: your book, by the way, is that you really tell

94
00:06:09,560 --> 00:06:11,640
Speaker 1: stories of all the thinkers.

95
00:06:12,240 --> 00:06:14,880
Speaker 2: You dive into the lives, you tell them with real color.

96
00:06:15,160 --> 00:06:17,440
Speaker 1: If you were going to start with one thinker that

97
00:06:17,480 --> 00:06:19,160
Speaker 1: you think is the most important, who would that be.

98
00:06:19,400 --> 00:06:21,160
Speaker 3: There are a couple of people who have this sort

99
00:06:21,160 --> 00:06:23,960
Speaker 3: of enduring influence throughout the book. One of them is Leibnitz,

100
00:06:24,040 --> 00:06:26,680
Speaker 3: who kind of started this enterprise in some sense. He

101
00:06:26,760 --> 00:06:30,360
Speaker 3: was really trying to take the idea of logic as

102
00:06:30,400 --> 00:06:33,280
Speaker 3: expressed by Aristotle and turn it into math, but ultimately

103
00:06:33,279 --> 00:06:35,960
Speaker 3: failed in doing that. But along the way he also

104
00:06:36,120 --> 00:06:38,799
Speaker 3: discovered the calculus, which turned out to be really important

105
00:06:38,800 --> 00:06:41,440
Speaker 3: when people wanted to make neural networks that could learn

106
00:06:41,520 --> 00:06:44,480
Speaker 3: from data. It turns out that the trick for doing

107
00:06:44,560 --> 00:06:46,840
Speaker 3: that is actually a trick that lad had figured out

108
00:06:46,880 --> 00:06:50,599
Speaker 3: all that time ago. And then another key figure here,

109
00:06:50,760 --> 00:06:53,320
Speaker 3: as might be suggested by the title of the book,

110
00:06:53,400 --> 00:06:57,800
Speaker 3: is George Bull, who was a nineteenth century mathematician. He

111
00:06:57,920 --> 00:06:59,600
Speaker 3: was a school teacher for most of his life and

112
00:06:59,680 --> 00:07:01,600
Speaker 3: did a lot of like serious math on the side

113
00:07:01,640 --> 00:07:03,680
Speaker 3: instead of you know, had a big effect on the

114
00:07:03,720 --> 00:07:08,080
Speaker 3: history of mathematics. But he was really the person who

115
00:07:08,160 --> 00:07:11,560
Speaker 3: then first solved that problem that Leibnitz had posed. And

116
00:07:11,600 --> 00:07:15,480
Speaker 3: in addition to the impact that that work had, he's

117
00:07:15,560 --> 00:07:19,000
Speaker 3: also the great grandfather of Jeff Hinton, who was one

118
00:07:19,000 --> 00:07:21,400
Speaker 3: of the people who played an important role in developing

119
00:07:21,440 --> 00:07:24,040
Speaker 3: these algorithms for learning from your own networks. And so

120
00:07:24,200 --> 00:07:26,240
Speaker 3: you could make an argument that without Boole we would

121
00:07:26,240 --> 00:07:29,480
Speaker 3: be a fair way back from where we are today.

122
00:07:29,920 --> 00:07:30,080
Speaker 2: You know.

123
00:07:30,200 --> 00:07:33,160
Speaker 1: Interestingly, when most people think about Boole, they only know

124
00:07:33,320 --> 00:07:38,360
Speaker 1: about Boolean numbers. They know about zero and one binary numbers,

125
00:07:38,680 --> 00:07:41,520
Speaker 1: and that's essentially the extent of the think. But he

126
00:07:41,560 --> 00:07:43,760
Speaker 1: was quite celebrated in his life right even though he

127
00:07:43,840 --> 00:07:48,440
Speaker 1: was a headmaster and not formally involved as a professor.

128
00:07:48,560 --> 00:07:51,000
Speaker 1: Am I correct about this? He nonetheless was quite recognized

129
00:07:51,000 --> 00:07:51,800
Speaker 1: as a mathematician.

130
00:07:52,400 --> 00:07:55,080
Speaker 3: Yeah, he became a university professor later in his life,

131
00:07:55,120 --> 00:07:57,080
Speaker 3: but spent most of his life as a teacher and

132
00:07:57,120 --> 00:08:01,080
Speaker 3: a head master. But yeah, he won a gold medal

133
00:08:01,080 --> 00:08:04,080
Speaker 3: in mathematics from the Royal Society. Was a very prestigious award,

134
00:08:05,080 --> 00:08:08,960
Speaker 3: and you know, was this amazing person who was having

135
00:08:09,000 --> 00:08:11,880
Speaker 3: these high level correspondences with the leading mathematicians of the

136
00:08:11,960 --> 00:08:16,360
Speaker 3: day while holding down his job running a small school.

137
00:08:16,960 --> 00:08:17,640
Speaker 2: Yeah.

138
00:08:17,800 --> 00:08:22,840
Speaker 1: Now, in the book, you essentially use three different frameworks.

139
00:08:22,920 --> 00:08:26,440
Speaker 1: What phenomenon does each framework explain?

140
00:08:26,560 --> 00:08:27,960
Speaker 2: Unusually Well, the.

141
00:08:27,840 --> 00:08:29,880
Speaker 3: Three frameworks I talk about in the book are what

142
00:08:29,920 --> 00:08:32,360
Speaker 3: I call rules and symbols, which is what we've been

143
00:08:32,400 --> 00:08:34,800
Speaker 3: talking about. This kind of like approach that stems out

144
00:08:34,840 --> 00:08:37,120
Speaker 3: of logic, where the idea is that you're going to

145
00:08:37,200 --> 00:08:39,719
Speaker 3: be able to write down some rules that characterize the

146
00:08:39,720 --> 00:08:41,960
Speaker 3: structure of thought, and by following those rules, you end

147
00:08:42,000 --> 00:08:47,319
Speaker 3: up with interesting consequences. The second approach is networks, features

148
00:08:47,320 --> 00:08:47,840
Speaker 3: and spaces.

149
00:08:47,920 --> 00:08:48,079
Speaker 4: Right.

150
00:08:48,120 --> 00:08:50,640
Speaker 3: This is neural networks, which you can kind of think

151
00:08:50,640 --> 00:08:54,160
Speaker 3: about as a system for doing computation when you start

152
00:08:54,200 --> 00:08:56,839
Speaker 3: representing things as points in a space. Right, So if

153
00:08:56,840 --> 00:09:01,600
Speaker 3: you start to think about you know, every object that

154
00:09:01,640 --> 00:09:03,360
Speaker 3: you could see in the world is not being something

155
00:09:03,400 --> 00:09:05,920
Speaker 3: that's described by rules, but being described by a location

156
00:09:06,040 --> 00:09:08,560
Speaker 3: along some dimensions. You need to have a way of

157
00:09:08,600 --> 00:09:11,400
Speaker 3: talking about how to map between those spaces and your

158
00:09:11,440 --> 00:09:14,120
Speaker 3: all network solve that problem. And then the third is

159
00:09:14,960 --> 00:09:20,439
Speaker 3: probability and statistics. And probability theory is really powerful because

160
00:09:20,720 --> 00:09:24,120
Speaker 3: it is the complement to logic, where logic tells us

161
00:09:24,160 --> 00:09:26,120
Speaker 3: how to go from things that we know to be

162
00:09:26,200 --> 00:09:28,959
Speaker 3: true to other things that we're equally certain or true.

163
00:09:29,360 --> 00:09:31,680
Speaker 3: Probability theory tells us what to do when we're uncertain.

164
00:09:32,160 --> 00:09:34,920
Speaker 3: So if we get some information we want to draw

165
00:09:34,920 --> 00:09:37,319
Speaker 3: a conclusion, but we're not able to draw that conclusion

166
00:09:37,360 --> 00:09:40,559
Speaker 3: with perfect certainty, Probability theory tells us how to do that,

167
00:09:40,920 --> 00:09:44,000
Speaker 3: and it tells us how to combine our sort of

168
00:09:44,400 --> 00:09:49,320
Speaker 3: background beliefs, the other sources of information we have our

169
00:09:49,360 --> 00:09:52,120
Speaker 3: biases in with the data that we see in a

170
00:09:52,120 --> 00:09:54,160
Speaker 3: way that helps us to explain how it's possible to

171
00:09:54,240 --> 00:09:56,440
Speaker 3: learn from small amounts of data. And that's one thing

172
00:09:56,480 --> 00:09:59,199
Speaker 3: which is still something that discriminates human learning from the

173
00:09:59,280 --> 00:10:01,120
Speaker 3: learning that's done by AI systems today.

174
00:10:01,520 --> 00:10:03,560
Speaker 1: Okay, great, so we're going to dive into each of

175
00:10:03,600 --> 00:10:06,560
Speaker 1: these three lenses. But just before we do, do you

176
00:10:06,640 --> 00:10:11,680
Speaker 1: see the AI conversation today over indexing on one of

177
00:10:11,720 --> 00:10:13,160
Speaker 1: these lenses over the others.

178
00:10:14,760 --> 00:10:17,520
Speaker 3: I think there's a lot of emphasis on neural networks,

179
00:10:17,559 --> 00:10:21,240
Speaker 3: which are fundamentally the sort of engineering technology which is

180
00:10:21,280 --> 00:10:25,360
Speaker 3: making possible the creation of our chatbots and the other

181
00:10:25,520 --> 00:10:29,600
Speaker 3: sort of big AI systems that are deployed. I think

182
00:10:30,200 --> 00:10:34,360
Speaker 3: that potentially misses out the importance of these other threads

183
00:10:34,559 --> 00:10:37,679
Speaker 3: right where. One thing that's important to remember is that

184
00:10:37,720 --> 00:10:40,680
Speaker 3: those neural networks are being trained on what is essentially

185
00:10:40,720 --> 00:10:43,200
Speaker 3: a system of rules and symbols. They're being trained on

186
00:10:43,720 --> 00:10:47,480
Speaker 3: human language, which is symbolic and rule like in various ways,

187
00:10:48,160 --> 00:10:50,680
Speaker 3: and they're being trained on code, which is even more

188
00:10:50,720 --> 00:10:54,240
Speaker 3: symbolic and even more rule like, And those things together

189
00:10:54,280 --> 00:10:56,559
Speaker 3: provide some of the substrate for developing the kind of

190
00:10:56,600 --> 00:10:59,280
Speaker 3: intelligence that they demonstrate. And then the way that they're

191
00:10:59,320 --> 00:11:03,079
Speaker 3: trained is by learning to predict the next token, right,

192
00:11:03,120 --> 00:11:05,080
Speaker 3: the next word or part of word, based on what

193
00:11:05,080 --> 00:11:07,760
Speaker 3: they've seen so far. And that way of training them

194
00:11:07,800 --> 00:11:12,040
Speaker 3: is actually using probability theory. So that's a probabilistic problem

195
00:11:12,040 --> 00:11:14,079
Speaker 3: because you're making a guess about what the next thing

196
00:11:14,120 --> 00:11:15,920
Speaker 3: is going to be based on the things that you see,

197
00:11:16,200 --> 00:11:18,520
Speaker 3: and so that's an important ingredient in their success as well,

198
00:11:18,559 --> 00:11:22,000
Speaker 3: is that they're essentially learning to approximate a big probability distribution.

199
00:11:22,360 --> 00:11:25,079
Speaker 1: So let's dive into the first one, rules and symbols.

200
00:11:25,280 --> 00:11:28,600
Speaker 1: So take us back to the original urge. Why did

201
00:11:28,679 --> 00:11:33,559
Speaker 1: early thinkers believe that this could be used to explain thinking.

202
00:11:35,200 --> 00:11:38,240
Speaker 3: I think a lot of the draw of rules and

203
00:11:38,280 --> 00:11:41,760
Speaker 3: symbols was that that really was, in some way what

204
00:11:42,120 --> 00:11:45,400
Speaker 3: mathematics was to people, right, So Leibniz, part of the

205
00:11:45,400 --> 00:11:47,400
Speaker 3: reason why he wasn't able to solve this problem of

206
00:11:47,440 --> 00:11:50,160
Speaker 3: figuring out how to turn thought into math is that

207
00:11:50,679 --> 00:11:53,560
Speaker 3: what he thought math was, or the kind of math

208
00:11:53,600 --> 00:11:55,440
Speaker 3: that he was trying to use to solve that problem,

209
00:11:56,000 --> 00:11:59,080
Speaker 3: was really arithmetic, right, And arithmetic was kind of like

210
00:11:59,120 --> 00:12:01,560
Speaker 3: the model that they had for a mathematical system. So

211
00:12:01,600 --> 00:12:04,120
Speaker 3: you can think about ideas being added together or subtracting

212
00:12:04,120 --> 00:12:07,080
Speaker 3: one idea from another, and really thinking about the operators

213
00:12:07,120 --> 00:12:09,160
Speaker 3: that you're using as being the things that are sort

214
00:12:09,160 --> 00:12:11,280
Speaker 3: of coming from this familiar mathematical language.

215
00:12:11,360 --> 00:12:13,640
Speaker 4: And so I think part of.

216
00:12:13,640 --> 00:12:16,240
Speaker 3: The reason that we end up with that approach is

217
00:12:16,280 --> 00:12:19,120
Speaker 3: because of the kind of math that has been successful

218
00:12:19,120 --> 00:12:22,920
Speaker 3: in other settings, right where we need to do arithmetic

219
00:12:23,040 --> 00:12:25,320
Speaker 3: to you know, that's a good description of certain.

220
00:12:25,160 --> 00:12:26,320
Speaker 4: Kinds of things that human minds do.

221
00:12:27,720 --> 00:12:30,120
Speaker 3: Google had the insight that you needed a different kind

222
00:12:30,120 --> 00:12:32,640
Speaker 3: of algebra in order to describe thought, and then that's

223
00:12:32,679 --> 00:12:36,200
Speaker 3: what leads to modern mathematical logic. But it's still in

224
00:12:36,240 --> 00:12:39,560
Speaker 3: this kind of symbolic language, although Gooole also talked about

225
00:12:39,559 --> 00:12:42,840
Speaker 3: probability theory as being important for capturing languages as well.

226
00:12:42,920 --> 00:12:45,360
Speaker 3: So I think it's really more about what are the

227
00:12:45,440 --> 00:12:48,640
Speaker 3: kinds of mathematical systems that it was sort of straightforward

228
00:12:48,640 --> 00:12:51,199
Speaker 3: to formalize, and that gave us something that we could

229
00:12:51,240 --> 00:12:53,600
Speaker 3: try to map thought onto. And that's what we do

230
00:12:53,679 --> 00:12:58,320
Speaker 3: as scientists is often taking mathematical systems that mathematicians have

231
00:12:58,360 --> 00:13:00,840
Speaker 3: defined for us and then saying, oh, I think this

232
00:13:01,000 --> 00:13:03,800
Speaker 3: mathematical system maps onto the thing that I want to understand,

233
00:13:04,200 --> 00:13:06,400
Speaker 3: and so trying to establish that correspondence and not just

234
00:13:06,520 --> 00:13:08,199
Speaker 3: then allow us to derive its consequences.

235
00:13:09,160 --> 00:13:12,840
Speaker 1: So speaking of rules and symbols, So thinkers like Newl

236
00:13:12,920 --> 00:13:17,040
Speaker 1: and Simon, they popularize this idea of goals and sub goals.

237
00:13:17,480 --> 00:13:21,920
Speaker 1: What did that viewpoint get exactly right about human problem solving.

238
00:13:23,440 --> 00:13:26,760
Speaker 3: So now we're fast forwarding a bit right from we

239
00:13:26,840 --> 00:13:30,800
Speaker 3: have Boule figuring out the structure of logic. That turns

240
00:13:30,840 --> 00:13:32,480
Speaker 3: into you know, lots of people then sort of turn

241
00:13:32,520 --> 00:13:34,840
Speaker 3: that into a sort of mature theory of logic. You

242
00:13:34,920 --> 00:13:39,160
Speaker 3: get aalenteering kind of turning this into a theory of computation,

243
00:13:39,480 --> 00:13:41,960
Speaker 3: thinking about what an abstract mathematician is doing when they're

244
00:13:42,000 --> 00:13:43,880
Speaker 3: doing something like logic, and thinking about how you can

245
00:13:43,920 --> 00:13:48,240
Speaker 3: make a machine do that. And then we have people

246
00:13:48,520 --> 00:13:51,560
Speaker 3: starting to realize that, you know, as digital computers are

247
00:13:51,559 --> 00:13:55,560
Speaker 3: being developed, maybe those provide a good model for how

248
00:13:55,600 --> 00:13:59,880
Speaker 3: thinking works in general, and then trying to use a

249
00:14:00,000 --> 00:14:03,200
Speaker 3: computer as a sort of foundation for you know, thinking

250
00:14:03,200 --> 00:14:05,280
Speaker 3: about things like how people might solve problems. And so

251
00:14:06,040 --> 00:14:09,800
Speaker 3: Alan Ewele and Herbert Simon were influential cognitive scientists who

252
00:14:10,600 --> 00:14:14,679
Speaker 3: did exactly that. They had this idea that maybe there

253
00:14:14,760 --> 00:14:17,400
Speaker 3: is a way that you could make computers smarter by

254
00:14:17,440 --> 00:14:20,720
Speaker 3: using insights from human cognition, but also get a better

255
00:14:20,800 --> 00:14:23,160
Speaker 3: understanding of what humans are doing when they're solving problems

256
00:14:23,200 --> 00:14:25,400
Speaker 3: by using the sort of ideas that come from things

257
00:14:25,440 --> 00:14:29,360
Speaker 3: like computer programming, and so they set up you know this,

258
00:14:29,640 --> 00:14:31,240
Speaker 3: you know, when we're trying to solve a problem or

259
00:14:31,240 --> 00:14:33,360
Speaker 3: prove a mathematical theorem or play a game of chess,

260
00:14:33,720 --> 00:14:35,520
Speaker 3: they set this up as a problem of searching through

261
00:14:35,560 --> 00:14:40,600
Speaker 3: a tree of possibilities, where what you're doing is making choices,

262
00:14:40,880 --> 00:14:42,480
Speaker 3: and then each of those choices gives you a new

263
00:14:42,520 --> 00:14:44,240
Speaker 3: set of choices, and each of those choices gives you

264
00:14:44,240 --> 00:14:46,360
Speaker 3: a new set of choices, and the hard thing is

265
00:14:46,880 --> 00:14:49,160
Speaker 3: finding a path through those choices that leads to the

266
00:14:49,160 --> 00:14:51,160
Speaker 3: point that you want to end up at. And so

267
00:14:51,720 --> 00:14:54,120
Speaker 3: that's something where you can take inspiration from how human

268
00:14:54,120 --> 00:14:57,720
Speaker 3: mathematicians solve problems. You can take inspiration from the kind

269
00:14:57,760 --> 00:14:59,960
Speaker 3: of you know, tricks like working backwards from the end

270
00:15:00,080 --> 00:15:03,320
Speaker 3: towards the start. Right, Those were principles that they were

271
00:15:03,360 --> 00:15:05,280
Speaker 3: able to use to try and explain these aspects of

272
00:15:05,360 --> 00:15:07,800
Speaker 3: human cognition as well as making the machines work better.

273
00:15:08,040 --> 00:15:09,960
Speaker 1: Okay, but then one of the things that happened is

274
00:15:10,000 --> 00:15:13,120
Speaker 1: that at least one of these attempts had ballooned into

275
00:15:13,200 --> 00:15:17,320
Speaker 1: twenty five million rules. And so what does that teach

276
00:15:17,400 --> 00:15:20,520
Speaker 1: us about the shape of human intelligence.

277
00:15:21,960 --> 00:15:23,800
Speaker 4: This rules and symbols enterprise.

278
00:15:24,040 --> 00:15:24,160
Speaker 2: Right.

279
00:15:24,240 --> 00:15:27,000
Speaker 4: The sort of appeal that this had was that maybe

280
00:15:27,040 --> 00:15:27,520
Speaker 4: one day.

281
00:15:27,400 --> 00:15:29,560
Speaker 3: You could just write down all of the rules that

282
00:15:29,560 --> 00:15:31,720
Speaker 3: you need to write down, and then you've characterized how

283
00:15:31,760 --> 00:15:34,320
Speaker 3: intelligence works. Right, So it's just a matter of getting

284
00:15:34,400 --> 00:15:37,440
Speaker 3: enough rules in a way that's very reminiscent today, right

285
00:15:37,520 --> 00:15:40,920
Speaker 3: of you know, the way that our modern AI systems

286
00:15:40,920 --> 00:15:43,520
Speaker 3: are being made is by training them on more and

287
00:15:43,600 --> 00:15:46,640
Speaker 3: more data, right, feeding in more and more language. There

288
00:15:46,680 --> 00:15:48,520
Speaker 3: was a hope that you could just like, yeah, like

289
00:15:48,800 --> 00:15:50,760
Speaker 3: document all of the rules that you need to capture

290
00:15:51,080 --> 00:15:54,360
Speaker 3: the structure of human knowledge. And so that led to

291
00:15:54,960 --> 00:15:57,240
Speaker 3: you know, companies being started to try and engage in

292
00:15:57,280 --> 00:16:01,880
Speaker 3: that enterprise, ultimately I would say, unsuccessfully, but giving us

293
00:16:01,920 --> 00:16:06,440
Speaker 3: some kind of characterization of like particular subsets of human knowledge.

294
00:16:06,520 --> 00:16:08,640
Speaker 3: And so I think the thing that came out of

295
00:16:08,920 --> 00:16:12,280
Speaker 3: that enterprise was revealing that maybe you need something more

296
00:16:12,360 --> 00:16:16,600
Speaker 3: than just rules, right, that maybe thinking about logic as

297
00:16:16,640 --> 00:16:19,480
Speaker 3: a basis for our model of intelligence was missing something.

298
00:16:20,040 --> 00:16:22,240
Speaker 3: It's an approach that worked really well for certain kinds

299
00:16:22,280 --> 00:16:26,200
Speaker 3: of problems like doing arithmetic, playing games or chess, but

300
00:16:26,240 --> 00:16:28,520
Speaker 3: it didn't work very well for other kinds of problems

301
00:16:28,640 --> 00:16:31,520
Speaker 3: like figuring out what you're seeing in the world, or

302
00:16:31,880 --> 00:16:34,360
Speaker 3: actually learning language or these other kinds of things.

303
00:16:34,400 --> 00:16:37,080
Speaker 1: And so this is what leads to your second lens,

304
00:16:37,240 --> 00:16:40,800
Speaker 1: which is neural networks. And you talk about these as

305
00:16:40,880 --> 00:16:44,080
Speaker 1: having you know, a boom and bust history. So, first,

306
00:16:44,120 --> 00:16:46,720
Speaker 1: what happened in the last decade that allowed them to

307
00:16:46,760 --> 00:16:48,880
Speaker 1: turn into the dominant paradigm.

308
00:16:49,320 --> 00:16:53,440
Speaker 3: The big breakthrough in the last decade was really about

309
00:16:53,760 --> 00:16:56,680
Speaker 3: being able to make bigger in neural networks that could

310
00:16:56,720 --> 00:17:01,840
Speaker 3: be trained on more data in a way that could scale, right,

311
00:17:01,880 --> 00:17:05,159
Speaker 3: and so bigger here means what these are. An artificial

312
00:17:05,240 --> 00:17:08,840
Speaker 3: neural network is a set of units that are communicating

313
00:17:08,840 --> 00:17:13,120
Speaker 3: with one another. They're communicating along weighted connections, a sort

314
00:17:13,160 --> 00:17:15,639
Speaker 3: of you know, imagine like how neurons are connected in

315
00:17:15,680 --> 00:17:17,840
Speaker 3: your brain, and those neurons are connected to one another

316
00:17:17,880 --> 00:17:20,359
Speaker 3: and sending each other signals. An artificial neural network is

317
00:17:20,400 --> 00:17:23,719
Speaker 3: basically simulating that kind of structure inside a computer. And

318
00:17:23,760 --> 00:17:26,679
Speaker 3: so for a long time, the sort of the history

319
00:17:26,680 --> 00:17:30,040
Speaker 3: of neural networks has been one of people figuring out

320
00:17:30,119 --> 00:17:33,880
Speaker 3: how to make bigger neural networks work. So the very

321
00:17:33,920 --> 00:17:37,800
Speaker 3: first you know, learning neural networks. They had a learning

322
00:17:37,840 --> 00:17:40,719
Speaker 3: algorithm that worked for one layer of weights, and then

323
00:17:40,760 --> 00:17:42,560
Speaker 3: there was a breakthrough in the nineteen eighties that meant,

324
00:17:42,600 --> 00:17:44,280
Speaker 3: now you had a learning algorithm that could work for

325
00:17:44,359 --> 00:17:46,600
Speaker 3: multiple layers of weights, but it didn't work for very

326
00:17:46,760 --> 00:17:48,920
Speaker 3: deep neural networks with lots of layers of net weights

327
00:17:48,960 --> 00:17:52,160
Speaker 3: because it I can sort of explain the technical reasons

328
00:17:52,200 --> 00:17:54,159
Speaker 3: behind it, but you know, sort of like the basic

329
00:17:54,160 --> 00:17:57,760
Speaker 3: algorithm didn't quite work. And so the big breakthroughs of

330
00:17:57,840 --> 00:18:00,000
Speaker 3: the last you know, ten to fifteen years have been

331
00:18:00,080 --> 00:18:03,480
Speaker 3: about coming up with ways to take those algorithms and

332
00:18:03,520 --> 00:18:05,440
Speaker 3: actually make them work for neural networks that are bigger

333
00:18:05,480 --> 00:18:07,679
Speaker 3: and bigger and deeper and deeper, that are able to

334
00:18:07,840 --> 00:18:11,840
Speaker 3: easily learn more complex functions and can do so from

335
00:18:12,160 --> 00:18:14,879
Speaker 3: massive amounts of data in a way that means that

336
00:18:14,920 --> 00:18:18,000
Speaker 3: they're able to discover sort of complex relationships between things

337
00:18:18,000 --> 00:18:19,960
Speaker 3: that are necessary to produce intelligent behavior.

338
00:18:20,280 --> 00:18:24,080
Speaker 1: And so, what are these neural networks capture about cognition

339
00:18:24,760 --> 00:18:29,200
Speaker 1: that symbols missed, especially in terms of things like similarity

340
00:18:29,240 --> 00:18:32,040
Speaker 1: and fuzziness and graded concepts.

341
00:18:32,560 --> 00:18:35,000
Speaker 3: Fuzziness is a really good way of describing it. It's

342
00:18:35,040 --> 00:18:39,080
Speaker 3: that you know, if you ask somebody, you know, whether

343
00:18:39,119 --> 00:18:41,800
Speaker 3: something is a piece of furniture, they're going to say,

344
00:18:42,119 --> 00:18:43,840
Speaker 3: you know, if you show them a chair, they'll say, yes,

345
00:18:43,920 --> 00:18:46,879
Speaker 3: definitely a piece of furniture. If you show them a rug,

346
00:18:47,440 --> 00:18:51,119
Speaker 3: they'll say, yeah, maybe a piece of furniture. Right, it

347
00:18:51,119 --> 00:18:52,960
Speaker 3: doesn't sort of fit with our you know, week sort

348
00:18:52,960 --> 00:18:56,240
Speaker 3: of have a prototypical idea of what furniture is, which

349
00:18:56,240 --> 00:18:58,479
Speaker 3: contains things like chairs and tables and ottomans and these

350
00:18:58,480 --> 00:19:02,119
Speaker 3: other kinds of things, and then rugs and treadmills, and you.

351
00:19:02,160 --> 00:19:04,040
Speaker 4: Know, like these are things that maybe.

352
00:19:03,760 --> 00:19:06,320
Speaker 3: You're in this category, but maybe an't right. And so

353
00:19:07,160 --> 00:19:09,520
Speaker 3: we need to have a way of thinking about concepts

354
00:19:09,520 --> 00:19:11,760
Speaker 3: that's not just the sort of yes or no, true

355
00:19:11,800 --> 00:19:14,760
Speaker 3: or false one or zero that logic would give us.

356
00:19:14,800 --> 00:19:16,840
Speaker 3: We need to have something which has that fuzziness in it.

357
00:19:17,160 --> 00:19:19,399
Speaker 3: One way of getting fuzziness is by thinking about a

358
00:19:19,480 --> 00:19:22,760
Speaker 3: concept in terms of points in space, right where you

359
00:19:22,760 --> 00:19:26,399
Speaker 3: could think chairs are here in one location, rugs are

360
00:19:26,440 --> 00:19:29,040
Speaker 3: here in another location, and maybe what it is to

361
00:19:29,040 --> 00:19:30,439
Speaker 3: be a piece of furniture is to just be in

362
00:19:30,440 --> 00:19:32,560
Speaker 3: some part of that space, and how close you are

363
00:19:32,600 --> 00:19:34,280
Speaker 3: to that part of the space is like how good

364
00:19:34,320 --> 00:19:36,560
Speaker 3: you are as an example of that kind of furniture.

365
00:19:37,359 --> 00:19:39,800
Speaker 3: And so as soon as you think in those terms,

366
00:19:39,800 --> 00:19:42,160
Speaker 3: you have a new problem, which is with our rules

367
00:19:42,160 --> 00:19:44,400
Speaker 3: and symbols. We knew how to do computation, we knew

368
00:19:44,400 --> 00:19:47,000
Speaker 3: how to describe thinking. Thinking was a matter of applying

369
00:19:47,040 --> 00:19:49,240
Speaker 3: the rules and seat of you know, repeating that process.

370
00:19:50,040 --> 00:19:52,640
Speaker 3: But we don't have a way of doing computation in spaces.

371
00:19:52,720 --> 00:19:54,439
Speaker 3: And that's what youral networks give us. So you can

372
00:19:54,520 --> 00:19:58,760
Speaker 3: kind of think about a space corresponding to the activation

373
00:19:58,920 --> 00:20:00,639
Speaker 3: of the units inside this neural network.

374
00:20:00,680 --> 00:20:03,040
Speaker 4: How much you know, how much input.

375
00:20:02,720 --> 00:20:05,600
Speaker 3: Each neural unit in that neural network is receiving, and

376
00:20:04,920 --> 00:20:08,600
Speaker 3: how much of a response it's making that characterizes some

377
00:20:08,680 --> 00:20:10,639
Speaker 3: kind of space. And then neural network gives us a

378
00:20:10,640 --> 00:20:13,359
Speaker 3: way of mapping from the inputs that it's getting to

379
00:20:13,440 --> 00:20:14,080
Speaker 3: some output.

380
00:20:14,200 --> 00:20:16,040
Speaker 4: So you could put in you know.

381
00:20:16,000 --> 00:20:18,239
Speaker 3: Your picture of a chair, and it maps that to

382
00:20:18,240 --> 00:20:20,560
Speaker 3: some point in space, and then it put sort of

383
00:20:20,560 --> 00:20:23,440
Speaker 3: produces out an output the corresponds to, yes, this is

384
00:20:23,480 --> 00:20:26,080
Speaker 3: a piece of furniture. And because those outputs can now

385
00:20:26,119 --> 00:20:29,199
Speaker 3: be continuous values, you can capture the fuzziness and other

386
00:20:29,280 --> 00:20:31,360
Speaker 3: kinds of things that you want for your concepts.

387
00:20:31,640 --> 00:20:34,960
Speaker 1: And so, in what sense are these modern systems, these

388
00:20:35,040 --> 00:20:38,360
Speaker 1: artificial neural networks learning, and in what sense are they

389
00:20:38,400 --> 00:20:42,720
Speaker 1: doing something that's maybe categorically different from how children learn.

390
00:20:44,520 --> 00:20:48,320
Speaker 3: This is a fundamental question, right, That's the kind of

391
00:20:48,320 --> 00:20:50,440
Speaker 3: thing that we cognitive scientists think about a lot, and

392
00:20:50,800 --> 00:20:53,080
Speaker 3: I think that AI researchers are starting to care about

393
00:20:53,119 --> 00:20:55,359
Speaker 3: a lot too, which is, you know, what are these

394
00:20:55,440 --> 00:20:59,720
Speaker 3: sort of meaningful differences between human minds, human brains and

395
00:20:59,720 --> 00:21:01,680
Speaker 3: what we building in these AI systems or these sort

396
00:21:01,680 --> 00:21:07,240
Speaker 3: of artificial brains. I think one very salient difference is

397
00:21:07,359 --> 00:21:10,399
Speaker 3: the amount of data which is needed for a human

398
00:21:10,640 --> 00:21:12,639
Speaker 3: to learn language compared to the amount of data you

399
00:21:12,640 --> 00:21:15,159
Speaker 3: need to put into on neural network. So if you

400
00:21:15,200 --> 00:21:18,640
Speaker 3: take a system like chat GPT, right, one of these chatbots,

401
00:21:19,000 --> 00:21:21,639
Speaker 3: those systems are trained on the equivalent of something like

402
00:21:21,720 --> 00:21:24,920
Speaker 3: five thousand to fifty thousand years of continuous speech. There's

403
00:21:24,920 --> 00:21:26,800
Speaker 3: sort of massive amounts of data that are going into

404
00:21:26,800 --> 00:21:28,919
Speaker 3: that system. So it's like on the order of a

405
00:21:28,960 --> 00:21:31,159
Speaker 3: thousand or ten thousand times as much data as a

406
00:21:31,240 --> 00:21:33,879
Speaker 3: human child might get in order to learn language. And

407
00:21:33,960 --> 00:21:37,280
Speaker 3: the reason is that those artificial neural networks are really

408
00:21:37,359 --> 00:21:40,399
Speaker 3: kind of like undifferentiated learning machines. You can take that

409
00:21:40,440 --> 00:21:41,960
Speaker 3: same kind of neural network, you can get it to

410
00:21:42,040 --> 00:21:44,400
Speaker 3: learn all sorts of different kinds of things. It works

411
00:21:44,400 --> 00:21:46,119
Speaker 3: really well for learning language, but you can use it

412
00:21:46,160 --> 00:21:47,920
Speaker 3: to learn something about vision or something. You know, you

413
00:21:47,960 --> 00:21:49,280
Speaker 3: can sort of take all sorts of problems and give

414
00:21:49,280 --> 00:21:50,520
Speaker 3: it to them and it can learn how to do that.

415
00:21:51,040 --> 00:21:56,000
Speaker 3: And so as a consequence, they have what we call

416
00:21:56,000 --> 00:21:59,760
Speaker 3: in cognitive science machine learning inductive biases. They're not biased

417
00:21:59,800 --> 00:22:04,120
Speaker 3: to towards any particular solution to the learning problem, and

418
00:22:04,400 --> 00:22:09,399
Speaker 3: human brains have stronger inductive biases for things like learning language. Right,

419
00:22:09,440 --> 00:22:12,200
Speaker 3: we're sort of disposed towards certain kinds of things, which

420
00:22:12,200 --> 00:22:14,679
Speaker 3: are human languages. The things that we call human languages

421
00:22:14,680 --> 00:22:16,880
Speaker 3: are the things that we're disposed to learn. And as

422
00:22:16,880 --> 00:22:19,480
Speaker 3: a consequence, you know, we're able to sort of narrow

423
00:22:19,560 --> 00:22:21,320
Speaker 3: down the space of possibilities in a way that means

424
00:22:21,320 --> 00:22:22,680
Speaker 3: that we're able to learn from less data.

425
00:22:36,960 --> 00:22:39,399
Speaker 1: Okay, this makes a great segue to the third lens.

426
00:22:39,440 --> 00:22:42,880
Speaker 1: So you talked about rules and symbols, and you talked

427
00:22:42,920 --> 00:22:45,560
Speaker 1: about artificial neural networks. The third part of your book

428
00:22:45,640 --> 00:22:51,200
Speaker 1: is about probabilities and statistics. So why did probability become

429
00:22:51,240 --> 00:22:56,000
Speaker 1: an attractive candidate language for thinking about cognition?

430
00:22:56,920 --> 00:22:59,520
Speaker 3: Probability there is a good way of answering certain kinds

431
00:22:59,520 --> 00:23:02,840
Speaker 3: of why questions that we might have, right so, and

432
00:23:02,880 --> 00:23:05,520
Speaker 3: the reason is that it's a way of characterizing how

433
00:23:05,560 --> 00:23:08,960
Speaker 3: a rational agent should make an inference. So all the

434
00:23:08,960 --> 00:23:13,159
Speaker 3: way back in the eighteenth century, British nonconformist minister, the

435
00:23:13,160 --> 00:23:16,440
Speaker 3: Reverend Thomas Bays had this radical idea that you could

436
00:23:17,119 --> 00:23:19,800
Speaker 3: talk about, you know, again, like take a mathematical system

437
00:23:19,880 --> 00:23:22,399
Speaker 3: probability theory, which we would use for describing what happens

438
00:23:22,440 --> 00:23:24,639
Speaker 3: when you roll dice or flip coins right, sort of

439
00:23:24,720 --> 00:23:27,400
Speaker 3: you know, sort of language of gambling and saying, oh,

440
00:23:27,480 --> 00:23:30,240
Speaker 3: in fact, that mathematical system might also be a really

441
00:23:30,280 --> 00:23:33,560
Speaker 3: good system for describing how beliefs work. And so what

442
00:23:33,640 --> 00:23:35,399
Speaker 3: he was interested in was if you think about a

443
00:23:35,400 --> 00:23:38,280
Speaker 3: belief as you know, a degree of belief, right, you

444
00:23:38,320 --> 00:23:40,160
Speaker 3: can say, oh, I think it's going to rain tomorrow,

445
00:23:40,600 --> 00:23:42,800
Speaker 3: and I'll put put that on a scale which goes

446
00:23:42,800 --> 00:23:45,479
Speaker 3: from zero to one, where zero is you know, not

447
00:23:45,520 --> 00:23:47,359
Speaker 3: going to rain, and one is one hundred percent it's

448
00:23:47,359 --> 00:23:47,760
Speaker 3: going to rain.

449
00:23:47,800 --> 00:23:48,200
Speaker 4: Tomorrow.

450
00:23:48,280 --> 00:23:52,040
Speaker 3: Right, That is a belief that you've expressed in the

451
00:23:52,080 --> 00:23:54,960
Speaker 3: form of a probability. And now if you, you know,

452
00:23:55,119 --> 00:23:56,639
Speaker 3: wake up in the morning and look out the window

453
00:23:56,640 --> 00:23:59,239
Speaker 3: and you see gray storm clouds, you've got a new

454
00:23:59,240 --> 00:24:02,480
Speaker 3: piece of data. You need to revise your beliefs, and

455
00:24:02,560 --> 00:24:05,119
Speaker 3: probability theory actually tells you how to do that. It says,

456
00:24:06,320 --> 00:24:09,200
Speaker 3: you know, for each hypothesis, right, so our hypotheses here

457
00:24:09,240 --> 00:24:11,040
Speaker 3: are it's going.

458
00:24:10,960 --> 00:24:12,080
Speaker 4: To rain or it's not going to rain.

459
00:24:12,440 --> 00:24:15,360
Speaker 3: Right, You're going to modify that belief based on how

460
00:24:15,440 --> 00:24:18,639
Speaker 3: likely the data is that you saw if that hypothesis

461
00:24:18,680 --> 00:24:21,359
Speaker 3: were true. So, because gray storm clouds are more likely

462
00:24:21,440 --> 00:24:23,920
Speaker 3: if it's going to rain that day, we should increase

463
00:24:23,960 --> 00:24:27,600
Speaker 3: our belief that it's going to rain. And as a consequence, well,

464
00:24:27,680 --> 00:24:29,040
Speaker 3: we'll end up with a number that's a little bit

465
00:24:29,119 --> 00:24:30,359
Speaker 3: higher than the number we had before.

466
00:24:30,600 --> 00:24:32,280
Speaker 4: And probability theory tells us how to do that.

467
00:24:32,520 --> 00:24:35,240
Speaker 3: There's a principle of probability theory called Bays rule after

468
00:24:35,240 --> 00:24:37,680
Speaker 3: the Reverend Thomas Beys that tells you how to take

469
00:24:37,760 --> 00:24:40,960
Speaker 3: your original beliefs and then turn them into the beliefs

470
00:24:40,960 --> 00:24:43,680
Speaker 3: that you get after seeing data. And that turns out

471
00:24:43,720 --> 00:24:45,600
Speaker 3: to be exactly the tool that you need to answer

472
00:24:45,640 --> 00:24:49,560
Speaker 3: these kinds of questions about how inductive biases work. Right, So,

473
00:24:50,160 --> 00:24:53,320
Speaker 3: how is it that children are able to learn from

474
00:24:53,600 --> 00:24:56,880
Speaker 3: less data than anural networks. Well, it's a consequence of

475
00:24:57,440 --> 00:25:00,000
Speaker 3: you know, these things that we can describe using different

476
00:25:00,160 --> 00:25:03,399
Speaker 3: probabilities being assigned to different hypotheses, whether hypotheses correspond to

477
00:25:03,400 --> 00:25:05,080
Speaker 3: the structure of the languages that are being learned.

478
00:25:05,320 --> 00:25:10,320
Speaker 1: And when people call humans irrational, what changes if we

479
00:25:10,440 --> 00:25:16,120
Speaker 1: look at mistakes as resource limited inferences.

480
00:25:17,640 --> 00:25:20,280
Speaker 3: This is one of the things that I explore in

481
00:25:19,760 --> 00:25:23,399
Speaker 3: my own research is this question of how we should

482
00:25:23,400 --> 00:25:26,919
Speaker 3: actually think about rationality for real agents, right, And this

483
00:25:27,000 --> 00:25:28,919
Speaker 3: is relevant if you're building an AI system or if

484
00:25:28,920 --> 00:25:31,760
Speaker 3: you're just trying to understand human behavior. So I think,

485
00:25:33,040 --> 00:25:35,720
Speaker 3: like I said, probability theory gives us a characterization of

486
00:25:35,800 --> 00:25:38,679
Speaker 3: what you should do as an ideal rational agent, But

487
00:25:38,760 --> 00:25:42,640
Speaker 3: that assumes that you have infinite computational resources. We mirror,

488
00:25:42,760 --> 00:25:46,440
Speaker 3: humans don't have infinite computational resources, nor do our AI systems,

489
00:25:46,560 --> 00:25:49,280
Speaker 3: And so you can ask what should a rational agent

490
00:25:49,320 --> 00:25:52,280
Speaker 3: do if they don't have all of the computational resources

491
00:25:52,280 --> 00:25:54,359
Speaker 3: that you might need and then out of that you

492
00:25:54,400 --> 00:25:57,160
Speaker 3: get the answer is that you know, you should follow

493
00:25:57,480 --> 00:26:02,320
Speaker 3: an algorithm, follow us strategy makes sense given the resources

494
00:26:02,359 --> 00:26:04,280
Speaker 3: that you have. That's what it means to be rational

495
00:26:04,280 --> 00:26:06,840
Speaker 3: in those circumstances where you're sort of doing the best

496
00:26:06,920 --> 00:26:12,159
Speaker 3: job you can of approximating probabilistic inference given those resource constraints.

497
00:26:12,200 --> 00:26:14,479
Speaker 3: And so some of the things that people do when

498
00:26:14,520 --> 00:26:16,400
Speaker 3: people do weird things, and we do lots of weird things,

499
00:26:16,440 --> 00:26:19,000
Speaker 3: and we don't always follow probability theory, things that we

500
00:26:19,000 --> 00:26:22,320
Speaker 3: can understand as us, you know, running into those resource

501
00:26:22,359 --> 00:26:25,200
Speaker 3: limitations and then coming up with, you know, reasonable strategies

502
00:26:25,200 --> 00:26:26,719
Speaker 3: for trying to approximate the right answer.

503
00:26:28,040 --> 00:26:29,080
Speaker 2: So, if we look.

504
00:26:28,960 --> 00:26:33,800
Speaker 1: At probability being the grammar of uncertainty, one thing we

505
00:26:33,880 --> 00:26:36,840
Speaker 1: know is that our prior expectations matter. And one of

506
00:26:36,880 --> 00:26:39,280
Speaker 1: the things I've been obsessed with and doing a lot

507
00:26:39,320 --> 00:26:41,000
Speaker 1: of research on and talking a lot about on the

508
00:26:41,000 --> 00:26:44,200
Speaker 1: podcast is the way that all of us drop into

509
00:26:44,200 --> 00:26:48,520
Speaker 1: the world and our cultures influence us and our language

510
00:26:48,560 --> 00:26:51,160
Speaker 1: and our moment in time and our neighborhood, and this

511
00:26:51,320 --> 00:26:55,760
Speaker 1: leads to people being quite different on the inside. Is

512
00:26:55,800 --> 00:26:58,159
Speaker 1: this something that you think about sometimes about how we

513
00:26:58,240 --> 00:27:03,760
Speaker 1: develop our priors differently based on you know what where

514
00:27:03,760 --> 00:27:04,360
Speaker 1: we grow up.

515
00:27:05,600 --> 00:27:08,679
Speaker 3: Yeah, So priors is that Daysian language, right, for talking

516
00:27:08,680 --> 00:27:11,240
Speaker 3: about the beliefs that you have before you get data

517
00:27:11,440 --> 00:27:14,840
Speaker 3: that you then update into what we call posterior probabilities

518
00:27:14,840 --> 00:27:17,440
Speaker 3: that are informed by those data. And so yeah, I

519
00:27:18,359 --> 00:27:20,720
Speaker 3: spend a lot of my research time thinking about these

520
00:27:20,800 --> 00:27:23,920
Speaker 3: kinds of questions of you know what are these sort

521
00:27:23,960 --> 00:27:28,880
Speaker 3: of prior distributions for humans? How do we acquire good

522
00:27:28,960 --> 00:27:31,959
Speaker 3: prior distributions for solving different kinds of problems. One thing

523
00:27:32,000 --> 00:27:34,520
Speaker 3: there is that calling it a prior makes it sound

524
00:27:34,560 --> 00:27:36,840
Speaker 3: like maybe it's something you're born with, but in fact,

525
00:27:36,880 --> 00:27:37,680
Speaker 3: it just means.

526
00:27:37,720 --> 00:27:38,960
Speaker 4: It's before you get data.

527
00:27:39,040 --> 00:27:41,439
Speaker 3: Right, So when you're seeing that storm cloud in the morning,

528
00:27:42,040 --> 00:27:44,520
Speaker 3: you had a prior probability from last night, and then

529
00:27:44,640 --> 00:27:47,560
Speaker 3: that prior probability was informed by everything else that you know, Right,

530
00:27:47,680 --> 00:27:49,920
Speaker 3: The priors are all of the biases and knowledge that

531
00:27:49,960 --> 00:27:51,880
Speaker 3: we bring to bear when we're trying to make an difference.

532
00:27:51,960 --> 00:27:55,080
Speaker 3: And so yeah, I think I think understanding that is

533
00:27:55,119 --> 00:27:57,280
Speaker 3: a big part of the project of understanding human cognition.

534
00:27:57,600 --> 00:27:59,120
Speaker 2: So let's zoom the camera out.

535
00:27:59,200 --> 00:28:01,639
Speaker 1: We've talked about the three lenses that you describe in

536
00:28:01,640 --> 00:28:04,119
Speaker 1: the book. Now, you also point out that we have

537
00:28:04,160 --> 00:28:08,199
Speaker 1: a lot of constraints like finite lives in limited compute,

538
00:28:08,200 --> 00:28:11,760
Speaker 1: and limited bandwidth, and so how do these constraints sculpt

539
00:28:12,040 --> 00:28:13,320
Speaker 1: human intelligence?

540
00:28:13,800 --> 00:28:16,520
Speaker 3: I think this is really important to just thinking about

541
00:28:16,560 --> 00:28:18,399
Speaker 3: the moment that we're in where there's a lot of

542
00:28:18,400 --> 00:28:21,520
Speaker 3: anxiety around AI, right, And I think if you think

543
00:28:21,520 --> 00:28:24,439
Speaker 3: about intelligence as a kind of one dimensional quantity, you

544
00:28:24,440 --> 00:28:26,560
Speaker 3: can kind of imagine that you know, humans are somewhere,

545
00:28:26,920 --> 00:28:28,320
Speaker 3: our AI systems are somewhere.

546
00:28:28,359 --> 00:28:29,560
Speaker 4: It seems like they're approaching us.

547
00:28:29,600 --> 00:28:31,560
Speaker 3: Maybe they're going to overtake us, and then, oh my god,

548
00:28:31,560 --> 00:28:33,120
Speaker 3: what is going to happen when that happens, Right, We're

549
00:28:33,160 --> 00:28:34,600
Speaker 3: just going to become redundant. There's not going to be

550
00:28:34,600 --> 00:28:36,520
Speaker 3: any jobs. Everything is going to fall apart. And so

551
00:28:37,160 --> 00:28:39,520
Speaker 3: that's a consequence of having a particular conception of what

552
00:28:39,640 --> 00:28:41,880
Speaker 3: intelligence is, which is this kind of one dimensional way

553
00:28:41,880 --> 00:28:44,040
Speaker 3: of thinking about it. And I think there's a different

554
00:28:44,080 --> 00:28:46,200
Speaker 3: way of thinking about it which gives us a little

555
00:28:46,200 --> 00:28:48,000
Speaker 3: more flexibility and maybe a little more hope in the

556
00:28:48,040 --> 00:28:50,120
Speaker 3: way that we think about what's going to happen with AI,

557
00:28:50,560 --> 00:28:55,840
Speaker 3: and that is thinking about intelligence as being an adaptation

558
00:28:56,040 --> 00:28:59,240
Speaker 3: to the kinds of computational problems that a system has

559
00:28:59,360 --> 00:29:02,520
Speaker 3: either of or being trained to solve, right, And so

560
00:29:03,360 --> 00:29:06,240
Speaker 3: for human beings, those computational problems are shaped by the

561
00:29:06,280 --> 00:29:08,440
Speaker 3: constraints that we operate under. And a lot of those

562
00:29:08,440 --> 00:29:11,080
Speaker 3: constraints come from our biology, right that we, as you said,

563
00:29:11,480 --> 00:29:14,880
Speaker 3: have limited lives, have you know, limited compute resources what

564
00:29:14,880 --> 00:29:17,560
Speaker 3: we can carry around inside our heads, have limited bandwidth

565
00:29:17,600 --> 00:29:20,480
Speaker 3: for communication. We have to like make noises at each other,

566
00:29:20,800 --> 00:29:22,800
Speaker 3: or wiggle our fingers or you know, somehow use our

567
00:29:22,840 --> 00:29:26,280
Speaker 3: bodies to transfer data from one human mind to another

568
00:29:26,360 --> 00:29:30,440
Speaker 3: human mind. It's very inefficient. And so those constraints are

569
00:29:30,520 --> 00:29:34,200
Speaker 3: things that mean that human intelligence takes a particular shape,

570
00:29:34,440 --> 00:29:37,920
Speaker 3: which is we're able to learn from limited data because

571
00:29:37,920 --> 00:29:40,080
Speaker 3: we have to because we don't live that long. Right,

572
00:29:40,280 --> 00:29:44,120
Speaker 3: You can't rely on getting five thousand years of language

573
00:29:44,160 --> 00:29:47,600
Speaker 3: data or multiple human lifetimes of you know, chess playing

574
00:29:47,680 --> 00:29:51,400
Speaker 3: or whatever it is. Right, you have to be good

575
00:29:51,400 --> 00:29:53,960
Speaker 3: at using the resources that you have in ways that

576
00:29:53,960 --> 00:29:56,000
Speaker 3: are efficient. And so that's kind of like deciding what

577
00:29:56,080 --> 00:29:59,160
Speaker 3: to think about being able to recognize when a problem

578
00:29:59,200 --> 00:30:01,800
Speaker 3: has a structure that you I've seen before being able

579
00:30:01,840 --> 00:30:04,280
Speaker 3: to you know, sort of like become sort of automatic

580
00:30:04,320 --> 00:30:07,160
Speaker 3: in using certain kinds of patterns of thinking and strategies

581
00:30:07,160 --> 00:30:11,120
Speaker 3: for solving problems, really trying to make it as easy as.

582
00:30:11,000 --> 00:30:13,320
Speaker 4: Possible for us to use the resources that we have.

583
00:30:13,800 --> 00:30:18,760
Speaker 3: And then you need to develop capacities for trying to

584
00:30:18,800 --> 00:30:22,000
Speaker 3: circumvent those bandwidth constraints in order to be able to

585
00:30:22,040 --> 00:30:24,320
Speaker 3: do things that transcend what any individual human can do,

586
00:30:24,640 --> 00:30:30,600
Speaker 3: and that means developing things like language writing societies LLCs,

587
00:30:30,920 --> 00:30:34,920
Speaker 3: you know, all of the sort of libraries, right institutions,

588
00:30:35,240 --> 00:30:37,960
Speaker 3: all of the theory of mind, right for reasoning about

589
00:30:37,960 --> 00:30:40,360
Speaker 3: what someone else might be trying to communicate to you.

590
00:30:40,840 --> 00:30:43,120
Speaker 3: All of this stuff is actually sort of like human

591
00:30:43,160 --> 00:30:46,920
Speaker 3: stuff that's a consequence of these constraints. And so as

592
00:30:46,920 --> 00:30:50,800
Speaker 3: we make AI systems that are smarter, those AI systems

593
00:30:50,800 --> 00:30:53,400
Speaker 3: are in turn being shaped by what they're being trained

594
00:30:53,400 --> 00:30:56,280
Speaker 3: to do and what constraints they operate under. But those

595
00:30:56,320 --> 00:30:58,760
Speaker 3: constraints are different from the ones that humans have. They

596
00:30:58,800 --> 00:31:02,320
Speaker 3: can you know, get more data, they can get access

597
00:31:02,320 --> 00:31:04,520
Speaker 3: to more compute, they don't have bandwidth limitations.

598
00:31:04,520 --> 00:31:05,520
Speaker 4: You can just copy.

599
00:31:05,240 --> 00:31:09,280
Speaker 3: A you know, a state of an AI system across machines.

600
00:31:09,360 --> 00:31:12,440
Speaker 3: You can use the same data to train multiple AI systems,

601
00:31:13,000 --> 00:31:15,280
Speaker 3: and all of those things mean that I think, rather

602
00:31:15,320 --> 00:31:17,600
Speaker 3: than being sort of on one axis where we're sort

603
00:31:17,600 --> 00:31:20,800
Speaker 3: of thinking about better and worse, it's more that there

604
00:31:20,800 --> 00:31:23,560
Speaker 3: are many axes that we can think about intelligent systems

605
00:31:23,600 --> 00:31:25,680
Speaker 3: developing along, and we're just going to end up in

606
00:31:25,680 --> 00:31:28,080
Speaker 3: a state where we have human intelligence and we have AIS,

607
00:31:28,520 --> 00:31:30,640
Speaker 3: and they're going to be meaningfully different from one another,

608
00:31:31,160 --> 00:31:33,200
Speaker 3: rather than things that are sort of directly competing in

609
00:31:33,320 --> 00:31:34,600
Speaker 3: terms of the capacities that they have.

610
00:31:35,280 --> 00:31:36,640
Speaker 2: Yeah, I agree with you on that.

611
00:31:37,280 --> 00:31:41,480
Speaker 1: When you think about the way that humans beat machines

612
00:31:41,640 --> 00:31:44,400
Speaker 1: on data efficiency, what do you think that means is

613
00:31:44,520 --> 00:31:47,320
Speaker 1: missing architecturally from our AI systems.

614
00:31:48,880 --> 00:31:52,360
Speaker 3: I think it's actually it's a great question, and the

615
00:31:52,360 --> 00:31:55,040
Speaker 3: way I would express it is not in terms of architecture.

616
00:31:55,720 --> 00:31:58,280
Speaker 3: So it's actually in terms of a different part of

617
00:31:58,320 --> 00:32:01,000
Speaker 3: a neural network. So when we think about this problem

618
00:32:01,000 --> 00:32:05,520
Speaker 3: of inductive bias, right, which is we're you know, what

619
00:32:06,160 --> 00:32:09,120
Speaker 3: a system is sort of disposed towards learning. As I said,

620
00:32:09,240 --> 00:32:11,840
Speaker 3: our neural networks, the way we normally set them up,

621
00:32:12,040 --> 00:32:14,040
Speaker 3: are pretty weak inductive biases. They can learn all sorts

622
00:32:14,040 --> 00:32:17,320
Speaker 3: of things. The inductive bias that a neural network has

623
00:32:17,560 --> 00:32:20,640
Speaker 3: it is constrained by its architecture, but it's also constrained

624
00:32:20,640 --> 00:32:23,080
Speaker 3: by where it starts out in the space of the

625
00:32:23,120 --> 00:32:26,680
Speaker 3: settings of all of those weights. And normally the default

626
00:32:26,720 --> 00:32:28,360
Speaker 3: is that you set up your neural networks so those

627
00:32:28,360 --> 00:32:31,000
Speaker 3: weights start out really small, close to zero, and then

628
00:32:31,000 --> 00:32:32,960
Speaker 3: they sort of grow away from that as it starts

629
00:32:33,000 --> 00:32:36,040
Speaker 3: to learn how to do things. We've had success in

630
00:32:36,120 --> 00:32:40,719
Speaker 3: taking neural networks that are architecturally identical but setting them

631
00:32:40,720 --> 00:32:43,720
Speaker 3: with different initial weights in order to create an inductive

632
00:32:43,760 --> 00:32:47,000
Speaker 3: bias that enables rapid learning. So we act to use

633
00:32:47,040 --> 00:32:49,720
Speaker 3: a technique called meta learning, which is a method from

634
00:32:49,720 --> 00:32:53,440
Speaker 3: machine learning where you take the same neural network architecture

635
00:32:53,600 --> 00:32:55,960
Speaker 3: and the same initial weights, and you use it to

636
00:32:56,040 --> 00:32:57,960
Speaker 3: learn to solve lots of different problems, like you can

637
00:32:58,040 --> 00:33:00,800
Speaker 3: use it to learn lots of different languages, say, from

638
00:33:00,840 --> 00:33:04,000
Speaker 3: limited data, and then you try and optimize the initial

639
00:33:04,040 --> 00:33:05,800
Speaker 3: weights of the neural network to make it so it

640
00:33:05,800 --> 00:33:09,000
Speaker 3: can learn all of those languages better using the same

641
00:33:09,080 --> 00:33:11,000
Speaker 3: kinds of algorithms we use for training the weights of

642
00:33:11,040 --> 00:33:12,880
Speaker 3: the neural network. When we have these giant data sets.

643
00:33:12,880 --> 00:33:15,560
Speaker 3: You can instead use those algorithms to train the initial

644
00:33:15,560 --> 00:33:17,680
Speaker 3: weights of a neural network for a small data set,

645
00:33:17,720 --> 00:33:20,760
Speaker 3: for lots of small data sets, And when you do that,

646
00:33:20,800 --> 00:33:22,320
Speaker 3: you end up with a neural network that has an

647
00:33:22,360 --> 00:33:25,400
Speaker 3: inductive bias that makes it possible to learn from small

648
00:33:25,400 --> 00:33:26,960
Speaker 3: amounts of data. And so that's the kind of thing

649
00:33:27,000 --> 00:33:29,920
Speaker 3: we've been exploring in my lab is can we find

650
00:33:29,960 --> 00:33:33,080
Speaker 3: a way of taking exactly these same neural network architectures

651
00:33:33,600 --> 00:33:36,000
Speaker 3: and just starting them out in a different place that

652
00:33:36,120 --> 00:33:39,120
Speaker 3: maybe aligns better with the kinds of things that humans do. Okay, well,

653
00:33:39,160 --> 00:33:40,880
Speaker 3: this is a really good segue to what I wanted

654
00:33:40,880 --> 00:33:43,600
Speaker 3: to ask you, which is, if you're looking at rules

655
00:33:43,640 --> 00:33:46,880
Speaker 3: and systems is one sort of math to describe the mind,

656
00:33:46,880 --> 00:33:49,320
Speaker 3: and artificial neural networks is another kind of math, and

657
00:33:49,360 --> 00:33:54,200
Speaker 3: probability is another. What does an optimal hybrid look like?

658
00:33:54,320 --> 00:33:58,920
Speaker 3: Given that no single, no one of these describes everything

659
00:33:58,960 --> 00:34:02,240
Speaker 3: about what's going on with minds, So what does the

660
00:34:02,360 --> 00:34:04,120
Speaker 3: hybrid for an aisystem look like?

661
00:34:04,200 --> 00:34:07,120
Speaker 2: In twenty twenty six, the place.

662
00:34:06,920 --> 00:34:09,759
Speaker 3: Where I end up in the book is saying that

663
00:34:10,080 --> 00:34:13,600
Speaker 3: these different kinds of math really do all fit together

664
00:34:13,640 --> 00:34:16,239
Speaker 3: in an interesting way, and in order to understand that

665
00:34:16,280 --> 00:34:18,560
Speaker 3: we can talk about different levels of analysis when we're

666
00:34:18,560 --> 00:34:21,040
Speaker 3: trying to make sense of an information processing system. So

667
00:34:21,920 --> 00:34:23,880
Speaker 3: the most abstract level, this is an idea that was

668
00:34:23,880 --> 00:34:27,640
Speaker 3: introduced by the computational neuroscientist David Marr. He said, the

669
00:34:27,680 --> 00:34:29,560
Speaker 3: most abstract level is just thinking about the problem that

670
00:34:29,560 --> 00:34:32,279
Speaker 3: the system is solving in its ideal solution, right. And

671
00:34:32,960 --> 00:34:36,279
Speaker 3: I think logic and symbolic systems and probability theory give

672
00:34:36,360 --> 00:34:39,239
Speaker 3: us a good way of characterizing the kinds of problems

673
00:34:39,280 --> 00:34:43,279
Speaker 3: that minds have to solve, right They you know, probabilistic

674
00:34:43,280 --> 00:34:45,960
Speaker 3: inference because we have to make these uncertain inferences. And

675
00:34:46,000 --> 00:34:49,400
Speaker 3: then logic as a way of characterizing the kinds of

676
00:34:49,440 --> 00:34:50,960
Speaker 3: things that are in the world that have this rich

677
00:34:51,000 --> 00:34:54,480
Speaker 3: structure of you know, like a sort of combinatorial structure

678
00:34:54,520 --> 00:34:57,400
Speaker 3: that you get from from having symbols and rules that

679
00:34:57,440 --> 00:35:00,160
Speaker 3: combine together with things like language and dance and all

680
00:35:00,160 --> 00:35:02,400
Speaker 3: of these you know, structured things. Even you know, if

681
00:35:02,400 --> 00:35:03,960
Speaker 3: you look at trees, you can see they have like

682
00:35:04,120 --> 00:35:06,759
Speaker 3: recursive structures that are expressed in them. Right, So these

683
00:35:06,840 --> 00:35:08,600
Speaker 3: kind of occur in nature and are important to be

684
00:35:08,600 --> 00:35:11,640
Speaker 3: able to understand. And then at the level below that

685
00:35:11,840 --> 00:35:17,160
Speaker 3: you have how the system solves those problems, right, like

686
00:35:17,480 --> 00:35:20,279
Speaker 3: what algorithms it might use, what representations it might use,

687
00:35:20,400 --> 00:35:22,920
Speaker 3: And then below that it's you know, how that's actually

688
00:35:22,920 --> 00:35:25,960
Speaker 3: implemented in some kind of physical system, right, and artificial

689
00:35:25,960 --> 00:35:27,799
Speaker 3: neural networks give us a kind of story at those

690
00:35:27,880 --> 00:35:31,720
Speaker 3: levels where we can think about them as being a

691
00:35:31,760 --> 00:35:35,439
Speaker 3: good general purpose system for learning to approximate the things

692
00:35:35,440 --> 00:35:37,799
Speaker 3: that probabilistic in front tells you to do, and learning

693
00:35:37,840 --> 00:35:40,560
Speaker 3: to approximate the structure that's contained within those symbolic systems.

694
00:35:40,960 --> 00:35:45,160
Speaker 3: So I actually think, you know, the kind of story

695
00:35:45,200 --> 00:35:47,319
Speaker 3: that we have right now that's emerged out of these

696
00:35:47,400 --> 00:35:50,120
Speaker 3: advances in AI is actually a pretty good story for

697
00:35:50,200 --> 00:35:52,759
Speaker 3: how we could think about human minds working. The thing

698
00:35:52,800 --> 00:35:55,279
Speaker 3: that's missing, most important thing that's missing is this kind

699
00:35:55,320 --> 00:35:59,279
Speaker 3: of aspect of inductive bias, right where we haven't been

700
00:35:59,320 --> 00:36:01,440
Speaker 3: able to capture what human inductive biases are like in

701
00:36:01,520 --> 00:36:03,799
Speaker 3: machines and so that you have these meaningful differences that come.

702
00:36:03,680 --> 00:36:04,080
Speaker 4: Out of that.

703
00:36:05,440 --> 00:36:08,600
Speaker 3: But it's not a bad place for thinking about how

704
00:36:08,640 --> 00:36:11,080
Speaker 3: these pieces might fit together to give us an explanation

705
00:36:11,120 --> 00:36:12,279
Speaker 3: for how it is that mind's work.

706
00:36:12,680 --> 00:36:18,399
Speaker 1: So along these lines, which AI benchmarks feel to you misleading?

707
00:36:18,520 --> 00:36:21,080
Speaker 2: And how would you make better benchmarks?

708
00:36:22,640 --> 00:36:26,239
Speaker 3: So, in general, I'm not a huge fan of benchmarks,

709
00:36:26,360 --> 00:36:29,279
Speaker 3: because I think benchmarks are useful as an engineering tool,

710
00:36:29,880 --> 00:36:33,560
Speaker 3: but I, as a cognitive scientist, don't just want to

711
00:36:33,600 --> 00:36:36,120
Speaker 3: know how well something is doing something. I want to

712
00:36:36,120 --> 00:36:38,960
Speaker 3: know how it's doing that thing right and how it

713
00:36:39,040 --> 00:36:41,719
Speaker 3: might be sort of messing that up right. So when

714
00:36:41,800 --> 00:36:46,160
Speaker 3: we are designing experiments as cognitive scientists, we don't just say, oh,

715
00:36:46,200 --> 00:36:48,120
Speaker 3: here's one hundred math problems. Go do with one hundred

716
00:36:48,160 --> 00:36:50,720
Speaker 3: math problems and we'll get a score. We say, let's

717
00:36:50,840 --> 00:36:54,480
Speaker 3: choose a set of math problems so that which answers

718
00:36:54,640 --> 00:36:57,960
Speaker 3: people give us tell us about the misconceptions that they

719
00:36:58,000 --> 00:37:00,319
Speaker 3: have in a way that we can then diagnose, oh,

720
00:37:00,440 --> 00:37:02,080
Speaker 3: you know, this is why this person is thinking this

721
00:37:02,120 --> 00:37:04,160
Speaker 3: particular thing. And so I think there's lots of room

722
00:37:04,200 --> 00:37:06,520
Speaker 3: for coming up with better ways of evaluating our AI

723
00:37:06,560 --> 00:37:09,600
Speaker 3: systems that look more like cognitive science experiments. We're really

724
00:37:09,640 --> 00:37:12,719
Speaker 3: targeting understanding what's going on rather than just trying to

725
00:37:12,719 --> 00:37:15,480
Speaker 3: get some brute sort of you know, performance score.

726
00:37:16,320 --> 00:37:19,920
Speaker 1: Okay, good, And you have talked about curiosity as a

727
00:37:19,960 --> 00:37:25,800
Speaker 1: computational problem. So how do you think about what curiosity

728
00:37:25,960 --> 00:37:30,320
Speaker 1: is and how we might measure real curiosity in a machine?

729
00:37:30,680 --> 00:37:34,400
Speaker 3: What problem is curiosity trying to solve? Yeah, this is

730
00:37:34,440 --> 00:37:36,000
Speaker 3: this is a good question. You can you can ask

731
00:37:36,040 --> 00:37:38,480
Speaker 3: this kind of question that we call rational analysis. Right,

732
00:37:38,520 --> 00:37:42,640
Speaker 3: if you have a system that's solving a problem, you

733
00:37:42,680 --> 00:37:44,000
Speaker 3: know what's what's the problem?

734
00:37:44,000 --> 00:37:44,800
Speaker 4: What's the ideal solution?

735
00:37:44,960 --> 00:37:48,279
Speaker 3: Okay, So for curiosity, we've argued and this is work

736
00:37:48,320 --> 00:37:49,399
Speaker 3: with wretched debate.

737
00:37:49,400 --> 00:37:52,200
Speaker 4: Who is that UCLA That.

738
00:37:53,800 --> 00:37:56,960
Speaker 3: One way I think about curiosity is that you're trying

739
00:37:56,960 --> 00:38:02,760
Speaker 3: to find things that are good in increasing your long

740
00:38:03,120 --> 00:38:06,880
Speaker 3: run probability of being able to solve problems in the future. Right,

741
00:38:07,000 --> 00:38:10,120
Speaker 3: So you know, it's sort of like you want data

742
00:38:10,200 --> 00:38:13,840
Speaker 3: which for which the derivative of your total knowledge is

743
00:38:13,960 --> 00:38:18,960
Speaker 3: high relative to that particular data point. And so that

744
00:38:19,040 --> 00:38:21,800
Speaker 3: explanation captures some of the things that happen in human cognition,

745
00:38:22,000 --> 00:38:25,239
Speaker 3: where you know, in some circumstances, we're interested in the

746
00:38:25,280 --> 00:38:29,239
Speaker 3: newest thing, something we've never seen before, Right, But in

747
00:38:29,280 --> 00:38:31,560
Speaker 3: a lot of circumstances, those things aren't the things that

748
00:38:31,600 --> 00:38:34,520
Speaker 3: grab our attention. It's more things that maybe we've seen

749
00:38:34,520 --> 00:38:38,640
Speaker 3: a few times, and you know, we just sort of

750
00:38:38,680 --> 00:38:41,319
Speaker 3: noticed that they're starting to occur. If something happens once,

751
00:38:41,360 --> 00:38:43,000
Speaker 3: you're just say okay, that was weird, and you sort

752
00:38:43,000 --> 00:38:45,160
Speaker 3: of dismiss it. But when something happens a few times

753
00:38:45,200 --> 00:38:47,279
Speaker 3: and it's unfamiliar to you, you say, okay, maybe I

754
00:38:47,320 --> 00:38:49,520
Speaker 3: need to figure that out. Right, And something that happens

755
00:38:49,520 --> 00:38:51,120
Speaker 3: all the time, you're not that curious about. That's just

756
00:38:51,120 --> 00:38:52,520
Speaker 3: the thing that happens all the time. And you can

757
00:38:52,560 --> 00:38:55,120
Speaker 3: explain that by thinking about this sort of derivative, right,

758
00:38:55,239 --> 00:38:58,920
Speaker 3: where if something just happens once, you shouldn't be interested

759
00:38:58,960 --> 00:39:00,480
Speaker 3: in it. Because it just happened once, it's probably never

760
00:39:00,480 --> 00:39:03,160
Speaker 3: going to happen again. If something happens a few times,

761
00:39:03,719 --> 00:39:06,640
Speaker 3: that's a clue that it's probably going to happen again

762
00:39:06,640 --> 00:39:09,160
Speaker 3: in the future, and you've not seen it enough to

763
00:39:09,200 --> 00:39:11,719
Speaker 3: actually know what's going on, And so paying attention to

764
00:39:11,760 --> 00:39:13,520
Speaker 3: that is good in terms of that derivative of your

765
00:39:13,520 --> 00:39:16,759
Speaker 3: future knowledge. And if something happens a lot, then it's

766
00:39:16,760 --> 00:39:18,680
Speaker 3: probably happened enough that you know something about what's going

767
00:39:18,680 --> 00:39:21,160
Speaker 3: on and it's not that interesting, right, And so so

768
00:39:21,280 --> 00:39:23,120
Speaker 3: that sort of sweet spot ends up being around the

769
00:39:23,120 --> 00:39:25,359
Speaker 3: things that are sort of like just happening to your

770
00:39:25,480 --> 00:39:27,719
Speaker 3: enough times that you're starting to realize, oh, this might be.

771
00:39:27,640 --> 00:39:28,880
Speaker 4: A thing that I need to pay attention to.

772
00:39:42,800 --> 00:39:45,880
Speaker 1: If you hadn't been on one capability that's going to

773
00:39:46,120 --> 00:39:50,319
Speaker 1: unlock a broader intelligence, unlock a jump to that.

774
00:39:50,360 --> 00:39:51,240
Speaker 2: What's your candidate?

775
00:39:52,160 --> 00:39:54,319
Speaker 3: I actually think the biggest obstacle at the moment is

776
00:39:54,400 --> 00:39:59,120
Speaker 3: more about generalizability of intelligence rather than any specific capacity, right.

777
00:39:59,239 --> 00:40:03,920
Speaker 3: And so people in the AI world talk about jagged intelligence, right,

778
00:40:03,920 --> 00:40:06,720
Speaker 3: the sort of phenomenon where you have an AI system

779
00:40:06,760 --> 00:40:08,799
Speaker 3: that can do something that's really smart and impress you,

780
00:40:09,280 --> 00:40:11,480
Speaker 3: and then five minutes later does something that's really dumb

781
00:40:11,520 --> 00:40:13,839
Speaker 3: on a problem that's like right next to it, and like,

782
00:40:13,880 --> 00:40:15,479
Speaker 3: if it's able to solve that first problem, it seems

783
00:40:15,480 --> 00:40:17,160
Speaker 3: obvious that it should be able to solve the second problem.

784
00:40:17,160 --> 00:40:19,120
Speaker 4: And you're just like, what happened? You know, why did

785
00:40:19,120 --> 00:40:19,799
Speaker 4: it go wrong there?

786
00:40:20,200 --> 00:40:24,680
Speaker 3: And so that lack of generalizability is also a consequence

787
00:40:24,680 --> 00:40:27,680
Speaker 3: of these kinds of inductive biases, right, So these human

788
00:40:27,719 --> 00:40:30,719
Speaker 3: inductive biases that steer us towards a solution and let

789
00:40:30,840 --> 00:40:34,640
Speaker 3: us learn from limited amounts of data, they constrain the

790
00:40:34,680 --> 00:40:36,160
Speaker 3: kinds of solutions that we find are The kinds of

791
00:40:36,200 --> 00:40:37,719
Speaker 3: solutions that we find are the ones that are sort

792
00:40:37,719 --> 00:40:40,640
Speaker 3: of like generalizable at least to us, right. They are

793
00:40:40,640 --> 00:40:42,239
Speaker 3: things that kind of make sense where if someone's able

794
00:40:42,280 --> 00:40:43,239
Speaker 3: to do one thing, they'll be able to do the

795
00:40:43,320 --> 00:40:45,799
Speaker 3: other thing. And because the AI systems are approaching these

796
00:40:45,840 --> 00:40:48,440
Speaker 3: problems just in a completely different way from a different

797
00:40:48,440 --> 00:40:50,800
Speaker 3: starting point and then getting tons of data that's allowing

798
00:40:50,800 --> 00:40:52,919
Speaker 3: them to sort of approximate what the human solutions are.

799
00:40:53,120 --> 00:40:55,600
Speaker 3: But they're coming at it from another angle. That's the

800
00:40:55,600 --> 00:40:58,160
Speaker 3: thing that makes them jagged. It's not that they don't

801
00:40:58,160 --> 00:41:00,400
Speaker 3: have sort of these same compatible inducted bis is that

802
00:41:00,400 --> 00:41:03,560
Speaker 3: we have that are informed by having evolved in certain

803
00:41:03,640 --> 00:41:06,279
Speaker 3: environments and having had experience of the world, and you know,

804
00:41:06,440 --> 00:41:08,000
Speaker 3: all of these other things that are part of what

805
00:41:08,040 --> 00:41:11,239
Speaker 3: it means to you know, sort of learn anything as

806
00:41:11,239 --> 00:41:15,080
Speaker 3: a human being. And so because they are coming with

807
00:41:15,160 --> 00:41:17,640
Speaker 3: this different set of inductive biases, they're very influenced by

808
00:41:17,680 --> 00:41:20,560
Speaker 3: their training data, they end up doing things that are

809
00:41:20,560 --> 00:41:24,800
Speaker 3: sort of inscrutable to us because they are, you know, yeah,

810
00:41:24,840 --> 00:41:26,720
Speaker 3: like coming at these problems in a way that doesn't

811
00:41:27,400 --> 00:41:29,759
Speaker 3: make sense to us. You know, from the starting point

812
00:41:29,800 --> 00:41:30,680
Speaker 3: that humans come from.

813
00:41:30,960 --> 00:41:33,960
Speaker 1: After writing this book, what do you think we understand

814
00:41:34,080 --> 00:41:36,960
Speaker 1: now about minds that we didn't understand let's say, a

815
00:41:37,040 --> 00:41:37,879
Speaker 1: decade or two ago.

816
00:41:39,239 --> 00:41:43,480
Speaker 3: So it's funny because when I have taught this material

817
00:41:43,680 --> 00:41:47,680
Speaker 3: for you know, twenty twenty years at this point, I

818
00:41:47,800 --> 00:41:51,200
Speaker 3: normally start my cognitive science classes saying, you know, welcome

819
00:41:51,239 --> 00:41:53,279
Speaker 3: to cognitive science. This is going to be different from

820
00:41:53,320 --> 00:41:55,480
Speaker 3: your other science classes. Normally, when you take a science class,

821
00:41:55,480 --> 00:41:57,160
Speaker 3: someone is going to stand up and say, Okay, here's

822
00:41:57,200 --> 00:41:58,520
Speaker 3: all the things that we figured out. Here are the

823
00:41:58,520 --> 00:42:02,319
Speaker 3: answers to the questions. And in cognito science, it's more

824
00:42:02,320 --> 00:42:04,640
Speaker 3: that we figured out how to get better at asking

825
00:42:04,680 --> 00:42:05,240
Speaker 3: the questions.

826
00:42:05,239 --> 00:42:06,359
Speaker 4: We haven't answered them. We don't.

827
00:42:06,360 --> 00:42:08,200
Speaker 3: It's not like you have a consensus across the whole

828
00:42:08,200 --> 00:42:10,520
Speaker 3: field about what those answers look like. And so I

829
00:42:10,520 --> 00:42:13,040
Speaker 3: think that's important that we're still very much And this

830
00:42:13,120 --> 00:42:14,759
Speaker 3: is what got me interested in cognres science in the

831
00:42:14,760 --> 00:42:17,120
Speaker 3: first place. You know, still a field that has deep

832
00:42:17,160 --> 00:42:20,640
Speaker 3: mysteries and lots of opportunities to learn and discover interesting things.

833
00:42:21,200 --> 00:42:25,560
Speaker 3: But I think over the last ten years, like so,

834
00:42:25,640 --> 00:42:28,560
Speaker 3: as I was working on this book, I wrote the

835
00:42:28,640 --> 00:42:32,919
Speaker 3: first chapter, and I had that disclaimer in the first

836
00:42:32,960 --> 00:42:37,520
Speaker 3: chapter and said, okay, look, you know I'm not bromising

837
00:42:37,560 --> 00:42:40,080
Speaker 3: you answers. Well, well, we're going to see if we

838
00:42:40,120 --> 00:42:42,839
Speaker 3: can get a good handle on the questions. But by

839
00:42:42,880 --> 00:42:44,440
Speaker 3: the time I got to the end of the book,

840
00:42:44,520 --> 00:42:46,680
Speaker 3: right after that sort of process of working on it

841
00:42:46,680 --> 00:42:49,160
Speaker 3: for years, I felt like things that actually, you know,

842
00:42:49,360 --> 00:42:51,480
Speaker 3: me going through the process of writing it and exploring

843
00:42:51,480 --> 00:42:53,120
Speaker 3: all these things and thinking about how they fit together,

844
00:42:53,440 --> 00:42:56,040
Speaker 3: but also just where the field was, you know, having

845
00:42:56,040 --> 00:42:58,399
Speaker 3: moved forward, I actually started to feel like, actually, these

846
00:42:58,400 --> 00:43:00,480
Speaker 3: things do fit together in a way where you can

847
00:43:00,480 --> 00:43:03,040
Speaker 3: see the glimpses of what answers are going to look

848
00:43:03,080 --> 00:43:05,040
Speaker 3: like in a way that I think really wasn't there

849
00:43:05,040 --> 00:43:08,200
Speaker 3: ten years ago. And it's that story of Okay, we

850
00:43:08,239 --> 00:43:11,200
Speaker 3: sort of know what the goals are, right, we know

851
00:43:11,239 --> 00:43:13,880
Speaker 3: what the right mathematical systems are for describing what intelligent

852
00:43:13,880 --> 00:43:16,520
Speaker 3: system should be doing. You have these ingredients of symbolic

853
00:43:16,560 --> 00:43:19,920
Speaker 3: systems and probablistic inference, and we've discovered that in fact,

854
00:43:19,960 --> 00:43:22,760
Speaker 3: you can get a remarkable way just using these artificial

855
00:43:22,760 --> 00:43:25,680
Speaker 3: neural networks to learn to approximate those things, and so

856
00:43:26,560 --> 00:43:29,839
Speaker 3: that demonstration I think has shown first of all, that

857
00:43:30,440 --> 00:43:36,120
Speaker 3: language is a extremely good substrate for intelligence right in

858
00:43:36,160 --> 00:43:38,319
Speaker 3: a way that I think people had not anticipated before

859
00:43:38,400 --> 00:43:41,520
Speaker 3: large language models, and that you can make big neural

860
00:43:41,520 --> 00:43:44,799
Speaker 3: networks that can learn to approximate really complex probability distributions.

861
00:43:45,360 --> 00:43:48,080
Speaker 3: And so it gives us some of these ingredients for

862
00:43:48,360 --> 00:43:52,080
Speaker 3: seeing how what originally worth three very different views of

863
00:43:52,120 --> 00:43:54,719
Speaker 3: the mind might start to fit together to make something

864
00:43:54,760 --> 00:43:56,239
Speaker 3: that's a little bit more of a unified hole.

865
00:43:57,440 --> 00:44:00,839
Speaker 1: Excellent, And when you wrote the book what struck You

866
00:44:00,960 --> 00:44:04,920
Speaker 1: is the most beautiful idea in the whole quest, in

867
00:44:04,960 --> 00:44:07,480
Speaker 1: the whole history of this gosh.

868
00:44:07,560 --> 00:44:10,960
Speaker 3: Okay, I mean, I'm a big probability for theory fan,

869
00:44:11,120 --> 00:44:16,560
Speaker 3: so going to you're gonna get me endorsing Bays rule,

870
00:44:16,560 --> 00:44:19,040
Speaker 3: which I really do think is like it's it's when

871
00:44:19,080 --> 00:44:21,000
Speaker 3: you learn it, take it a probability class. It's just

872
00:44:21,040 --> 00:44:23,359
Speaker 3: like it's just a dumb principle of probability theory. But

873
00:44:23,400 --> 00:44:26,279
Speaker 3: when you make this move of saying probability theory isn't

874
00:44:26,360 --> 00:44:30,360
Speaker 3: just about dice and cards, it's about you know, beliefs,

875
00:44:30,640 --> 00:44:34,440
Speaker 3: it suddenly becomes a very deep and insightful sort of principle,

876
00:44:34,480 --> 00:44:37,640
Speaker 3: And in the book, I also show probability theory kind

877
00:44:37,640 --> 00:44:40,480
Speaker 3: of subsumes logic, like everything that's a valid logical inference

878
00:44:40,520 --> 00:44:43,040
Speaker 3: is also a valid inference in probability theory. Probability theory

879
00:44:43,080 --> 00:44:45,879
Speaker 3: just kind of extends the surmountics of logic to these

880
00:44:45,920 --> 00:44:49,200
Speaker 3: cases of uncertainty. So to me, I think that's a

881
00:44:49,320 --> 00:44:50,840
Speaker 3: that's a that's a big one. I kind of like

882
00:44:50,880 --> 00:44:51,520
Speaker 3: that's where I live.

883
00:44:51,640 --> 00:44:55,080
Speaker 1: Yeah, excellent, And and somebody, if we do have a

884
00:44:55,480 --> 00:44:58,520
Speaker 1: mature physics of thought, let's say fifty years from now,

885
00:44:58,560 --> 00:45:01,400
Speaker 1: what is that change from us in terms of education,

886
00:45:02,040 --> 00:45:04,040
Speaker 1: in terms of the way we build machines.

887
00:45:05,000 --> 00:45:07,640
Speaker 3: So I think this is this is exactly where we

888
00:45:07,680 --> 00:45:11,400
Speaker 3: can go, right, which is, once you figure out the

889
00:45:11,440 --> 00:45:13,880
Speaker 3: scientific principles of a domain, you can start to think

890
00:45:13,920 --> 00:45:16,680
Speaker 3: about how to do engineering right. So like you know,

891
00:45:17,040 --> 00:45:19,680
Speaker 3: when you're an engineer and you go to engineering school,

892
00:45:20,000 --> 00:45:22,480
Speaker 3: you take physics, right, and then you learn in your

893
00:45:22,480 --> 00:45:24,799
Speaker 3: physics class what these principles are, and then you take

894
00:45:24,840 --> 00:45:27,280
Speaker 3: your applied engineering classes, which are like taking those physical

895
00:45:27,320 --> 00:45:29,040
Speaker 3: principles and telling you how to build a bridge right

896
00:45:29,080 --> 00:45:32,239
Speaker 3: and explaining that you know not in terms of heuristics

897
00:45:32,760 --> 00:45:34,520
Speaker 3: for what makes a good bridge, but in terms of

898
00:45:34,600 --> 00:45:37,520
Speaker 3: those fundamental physical principles. So I think that's a thing

899
00:45:37,520 --> 00:45:40,640
Speaker 3: that's incredibly exciting here is that as we start to

900
00:45:40,719 --> 00:45:45,440
Speaker 3: converge on what these laws of thought look like, it

901
00:45:45,480 --> 00:45:48,359
Speaker 3: gives us the opportunity to do a much more sort

902
00:45:48,400 --> 00:45:54,719
Speaker 3: of science based form of engineering applied to human cognition,

903
00:45:55,000 --> 00:45:58,480
Speaker 3: thinking about how do we make an optimal you know,

904
00:45:58,680 --> 00:46:02,239
Speaker 3: sort of learning environment, how do we support human decision making.

905
00:46:02,280 --> 00:46:03,840
Speaker 3: That's something that I work on in my lab is like,

906
00:46:04,239 --> 00:46:08,480
Speaker 3: how do we put computation into human environments to overcome

907
00:46:08,520 --> 00:46:12,440
Speaker 3: whatever computational constraints we have as individual decision makers and

908
00:46:12,480 --> 00:46:17,160
Speaker 3: help us make better decisions. And how do we understand,

909
00:46:17,320 --> 00:46:21,359
Speaker 3: you know, the kinds of things that people are doing

910
00:46:21,360 --> 00:46:23,319
Speaker 3: in a way that allows us to then sort of

911
00:46:23,360 --> 00:46:27,160
Speaker 3: like make suggestions about, you know, how they might do

912
00:46:27,239 --> 00:46:30,160
Speaker 3: them better. Right, And so I think there's a there's

913
00:46:30,160 --> 00:46:32,680
Speaker 3: a lot of potential for you know, sort of human

914
00:46:32,760 --> 00:46:35,440
Speaker 3: upside as we start to be able to answer these

915
00:46:35,440 --> 00:46:36,400
Speaker 3: scientific questions.

916
00:46:41,200 --> 00:46:44,480
Speaker 1: That was my interview with Tom Griffith's. To quickly summarize

917
00:46:44,520 --> 00:46:49,120
Speaker 1: his framework, Tom sees three major scientific approaches that all

918
00:46:49,120 --> 00:46:53,000
Speaker 1: try to capture the mind. You've got rules and symbols,

919
00:46:53,440 --> 00:46:57,240
Speaker 1: you've got artificial neural networks, and you've got probability theory.

920
00:46:57,600 --> 00:47:01,239
Speaker 1: These very different approaches, and each which one has delivered

921
00:47:01,280 --> 00:47:05,759
Speaker 1: something a little different. Rules and symbols give us language

922
00:47:05,920 --> 00:47:10,040
Speaker 1: like machinery where pieces can be assembled and reassembled into

923
00:47:10,080 --> 00:47:15,759
Speaker 1: complex ideas. Artificial neural networks they give us graded concepts,

924
00:47:15,880 --> 00:47:20,160
Speaker 1: meaning ideas can be fuzzy, and probability theory gives us

925
00:47:20,160 --> 00:47:24,160
Speaker 1: a language for dealing with uncertainty. Now, what's interesting is

926
00:47:24,200 --> 00:47:28,000
Speaker 1: that human minds seem to traffic in all of these modes.

927
00:47:28,120 --> 00:47:32,719
Speaker 1: We use structured symbols, we also use graded concepts. We

928
00:47:32,760 --> 00:47:36,480
Speaker 1: also revise our belief as new evidence comes in. And

929
00:47:36,560 --> 00:47:38,400
Speaker 1: part of that is that we move through the world

930
00:47:38,400 --> 00:47:43,040
Speaker 1: with prior beliefs shaped by our history, our culture, our language,

931
00:47:43,040 --> 00:47:46,799
Speaker 1: our neighborhood, our moment in time. So none of these

932
00:47:46,840 --> 00:47:50,719
Speaker 1: models by themselves are the final answer. And what this

933
00:47:50,840 --> 00:47:54,120
Speaker 1: means is that, like most scientific stories, this is one

934
00:47:54,160 --> 00:47:58,880
Speaker 1: about humility. Tom's book illustrates how every generation arrives with

935
00:47:59,280 --> 00:48:03,319
Speaker 1: some new formalism, some new piece of math, some new

936
00:48:03,719 --> 00:48:08,240
Speaker 1: model that's powerful enough to illuminate an area of mental

937
00:48:08,280 --> 00:48:10,880
Speaker 1: life and for a moment it feels like, hey, the

938
00:48:10,880 --> 00:48:15,600
Speaker 1: whole mystery is finally collapsing. But then the spotlight widens

939
00:48:15,680 --> 00:48:19,160
Speaker 1: and we see more terrain. So what I love about

940
00:48:19,200 --> 00:48:21,640
Speaker 1: this conversation is that it can leave us with a

941
00:48:21,840 --> 00:48:25,120
Speaker 1: sense of progress and a sense of wonder. At the

942
00:48:25,160 --> 00:48:29,239
Speaker 1: same time, we feel a convergence of different fields, and

943
00:48:29,320 --> 00:48:34,280
Speaker 1: we can also feel how large this subject remains. Cognition

944
00:48:34,440 --> 00:48:37,560
Speaker 1: is still a field in motion, So let's look at

945
00:48:37,560 --> 00:48:40,799
Speaker 1: the big picture. When the field of physics matured, we

946
00:48:40,880 --> 00:48:45,399
Speaker 1: could then build bridges and airplanes and power grids because

947
00:48:45,440 --> 00:48:50,520
Speaker 1: we had firm principles to build on. So once the

948
00:48:50,640 --> 00:48:55,640
Speaker 1: laws of thought come into clearer view, what becomes possible

949
00:48:55,840 --> 00:49:00,520
Speaker 1: for education and for decision making, and for rules that

950
00:49:00,600 --> 00:49:04,839
Speaker 1: help us reason more effectively. So here we are at

951
00:49:04,880 --> 00:49:08,000
Speaker 1: a very cool moment in history where the old dream

952
00:49:08,640 --> 00:49:13,040
Speaker 1: of formalizing thought has escaped to the library and shown

953
00:49:13,120 --> 00:49:17,439
Speaker 1: up in everyone's laptop. The big thinkers of centuries ago

954
00:49:17,920 --> 00:49:21,239
Speaker 1: could sort of squint and see the outline of the project,

955
00:49:21,840 --> 00:49:24,839
Speaker 1: and now we're living much more squarely right in the

956
00:49:24,880 --> 00:49:27,799
Speaker 1: middle of it. If there truly are laws of thought,

957
00:49:27,840 --> 00:49:30,960
Speaker 1: they're going to teach us about our machines, but more importantly,

958
00:49:30,960 --> 00:49:35,399
Speaker 1: they're going to teach us about ourselves, because although it's

959
00:49:35,400 --> 00:49:39,320
Speaker 1: sometimes tempting to view the mind as a ghostly exception

960
00:49:39,440 --> 00:49:43,640
Speaker 1: to the universe, the mind, presumably is part of the universe,

961
00:49:44,120 --> 00:49:48,800
Speaker 1: and it is lawful and wondrous and discoverable, and every

962
00:49:48,880 --> 00:49:56,840
Speaker 1: step towards understanding it enlarges the human story. Go to

963
00:49:56,880 --> 00:49:59,680
Speaker 1: eagleman dot com slash podcast for more information and to

964
00:49:59,719 --> 00:50:03,520
Speaker 1: find further reading. Join the weekly discussions on my substack,

965
00:50:03,800 --> 00:50:06,800
Speaker 1: and check out and subscribe to Inner Cosmos on YouTube

966
00:50:06,800 --> 00:50:10,160
Speaker 1: for videos of each episode and to leave comments until

967
00:50:10,200 --> 00:50:13,640
Speaker 1: next time. I'm David Eagleman, and this is Inner Cosmos.