1
00:00:05,080 --> 00:00:10,440
Speaker 1: Humans are really smart? But how did intelligence evolve? If

2
00:00:10,440 --> 00:00:13,760
Speaker 1: we're trying to look back at the history of intelligent brains,

3
00:00:14,360 --> 00:00:16,520
Speaker 1: do we have to look all the way back to

4
00:00:16,640 --> 00:00:20,880
Speaker 1: our common ancestors with the apes, or all mammals or

5
00:00:21,000 --> 00:00:25,640
Speaker 1: all reptiles, or can the origins of intelligence be traced

6
00:00:25,680 --> 00:00:30,000
Speaker 1: back even further? And now that our species is good

7
00:00:30,040 --> 00:00:33,320
Speaker 1: and smart, what does the knowledge of our past mean

8
00:00:33,479 --> 00:00:41,640
Speaker 1: for us as we work to build intelligence artificially? Welcome

9
00:00:41,640 --> 00:00:45,640
Speaker 1: to Inner Cosmos with me David Eagleman. I'm a neuroscientist

10
00:00:45,680 --> 00:00:48,919
Speaker 1: and an author at Stanford and in these episodes, we

11
00:00:49,000 --> 00:00:53,440
Speaker 1: sail deeply into our three pound universe to understand why

12
00:00:53,560 --> 00:01:05,160
Speaker 1: and how our lives look the way they do. Today's

13
00:01:05,160 --> 00:01:10,200
Speaker 1: episode is about intelligence and the history of intelligence. How

14
00:01:10,240 --> 00:01:13,360
Speaker 1: did human intelligence arrive on the scene?

15
00:01:13,880 --> 00:01:14,080
Speaker 2: Now?

16
00:01:14,120 --> 00:01:16,800
Speaker 1: This is an important question because we seem to be

17
00:01:17,000 --> 00:01:21,120
Speaker 1: operating at a different level than our neighbors in the

18
00:01:21,200 --> 00:01:24,520
Speaker 1: animal kingdom. We are the only ones, as far as

19
00:01:24,520 --> 00:01:29,160
Speaker 1: we can tell, who compose symphonies and launch mars rover

20
00:01:29,280 --> 00:01:34,520
Speaker 1: missions and discover DNA and build courthouses and have congresses

21
00:01:34,560 --> 00:01:39,480
Speaker 1: and construct windmills and write novels and build screws and

22
00:01:39,520 --> 00:01:42,880
Speaker 1: screwdrivers to hold things together, and so on and so on,

23
00:01:43,200 --> 00:01:46,679
Speaker 1: none of which any other animal does. And this is

24
00:01:46,720 --> 00:01:50,000
Speaker 1: how we've taken over the whole planet. But how the

25
00:01:50,040 --> 00:01:56,559
Speaker 1: heck did this happen? Why are humans such a runaway species? Well,

26
00:01:56,640 --> 00:02:00,120
Speaker 1: traditionally the explanation has been something like this is a

27
00:02:00,200 --> 00:02:04,080
Speaker 1: special gift from your deity, whichever deity your family believed

28
00:02:04,120 --> 00:02:09,200
Speaker 1: in at whatever moment in history. But centuries of people

29
00:02:09,240 --> 00:02:12,480
Speaker 1: looking at this carefully, sometimes in a microscope, sometimes in

30
00:02:12,520 --> 00:02:18,000
Speaker 1: the brain scanner, sometimes at autopsy, careful examination has made

31
00:02:18,080 --> 00:02:21,600
Speaker 1: something very clear. When you look at the brains of

32
00:02:21,720 --> 00:02:26,800
Speaker 1: other animals, those brains are very similar to our own. Now,

33
00:02:26,880 --> 00:02:30,240
Speaker 1: this shouldn't be surprising. It's the same when you look

34
00:02:30,280 --> 00:02:34,320
Speaker 1: at other animals hearts or lungs or kidneys. It's the

35
00:02:34,440 --> 00:02:39,600
Speaker 1: same good idea, and it's conserved throughout evolution, and so

36
00:02:39,800 --> 00:02:44,960
Speaker 1: it goes with brains, with neurons and cerebellum and thalamus

37
00:02:44,960 --> 00:02:48,360
Speaker 1: and hippocampus and cortex and blah blah blah. It looks

38
00:02:48,520 --> 00:02:52,320
Speaker 1: pretty similar everywhere. And this leads to a point which

39
00:02:52,360 --> 00:02:56,160
Speaker 1: should be fairly obvious when you look across the evolution

40
00:02:56,480 --> 00:03:00,480
Speaker 1: of the vast Kingdom of animals. You don't find that

41
00:03:00,520 --> 00:03:05,120
Speaker 1: there was no intelligence and suddenly humans popped up like

42
00:03:05,240 --> 00:03:09,440
Speaker 1: hairless geniuses. That's not what happened. Instead, what you find

43
00:03:10,080 --> 00:03:14,000
Speaker 1: is there are versions of intelligence all around us. As

44
00:03:14,040 --> 00:03:17,880
Speaker 1: one example, I always admire the squirrels hopping in my

45
00:03:18,080 --> 00:03:22,040
Speaker 1: tree in the garden. They perform these sophisticated acrobatics and

46
00:03:22,080 --> 00:03:25,280
Speaker 1: do the kind of stuff that human gymnasts would never

47
00:03:25,320 --> 00:03:29,680
Speaker 1: even attempt. And crows show intelligence that's closer to our own.

48
00:03:29,720 --> 00:03:33,959
Speaker 1: They can solve really sophisticated puzzles, and dolphins have some

49
00:03:34,000 --> 00:03:37,440
Speaker 1: sort of societies and language, though again not quite as

50
00:03:37,480 --> 00:03:41,440
Speaker 1: sophisticated as ours. And in episode thirty four, I explored

51
00:03:41,480 --> 00:03:44,760
Speaker 1: what it would be like to have different levels of intelligence,

52
00:03:45,160 --> 00:03:47,480
Speaker 1: So please check out that episode if you're interested in that.

53
00:03:47,960 --> 00:03:52,120
Speaker 1: So back to this question. When we ask how intelligence

54
00:03:52,200 --> 00:03:55,480
Speaker 1: got here, it ends up being a question about an

55
00:03:55,520 --> 00:03:59,720
Speaker 1: evolutionary journey, like when we ask how did Homo sapiens

56
00:03:59,760 --> 00:04:02,440
Speaker 1: start walking on our rear legs? Or how did we

57
00:04:02,520 --> 00:04:06,040
Speaker 1: become hairless? Or why do we get pimples and other

58
00:04:06,120 --> 00:04:09,440
Speaker 1: primates don't, or even deeper things like how did any

59
00:04:09,480 --> 00:04:12,640
Speaker 1: of us we and other land dwelling animals, how did

60
00:04:12,680 --> 00:04:16,159
Speaker 1: we get kidneys or lungs? We can ask the same

61
00:04:16,279 --> 00:04:19,560
Speaker 1: sort of questions about the brain. The brain has a

62
00:04:19,800 --> 00:04:25,159
Speaker 1: very rich evolutionary history, a long and sometimes branching pathway

63
00:04:25,240 --> 00:04:28,719
Speaker 1: that has led from early brains swimming around looking for

64
00:04:28,839 --> 00:04:33,520
Speaker 1: food to brains now that build skyscrapers and launch rocket

65
00:04:33,520 --> 00:04:37,520
Speaker 1: ships and try to figure themselves out. This is the

66
00:04:37,600 --> 00:04:39,560
Speaker 1: kind of stuff that none of our neighbors in the

67
00:04:39,600 --> 00:04:41,919
Speaker 1: animal kingdom do, as far as we can tell. And

68
00:04:41,960 --> 00:04:45,680
Speaker 1: there's clearly something special about the human brain that allows

69
00:04:45,720 --> 00:04:49,560
Speaker 1: that to happen. In other words, we find smarts all

70
00:04:49,600 --> 00:04:52,800
Speaker 1: across the animal kingdom, but there is something very special

71
00:04:52,920 --> 00:04:58,320
Speaker 1: about human intelligence. There's an evolutionary biologist named Theodosius Dubzanski,

72
00:04:58,560 --> 00:05:03,440
Speaker 1: and he once said all species are unique, but humans

73
00:05:03,680 --> 00:05:06,920
Speaker 1: are the uniquest. So I've just told you two things.

74
00:05:07,040 --> 00:05:10,640
Speaker 1: On the one hand, we have very similar brains to

75
00:05:10,720 --> 00:05:14,039
Speaker 1: all our animal cousins, and on the other hand, we

76
00:05:14,200 --> 00:05:19,000
Speaker 1: have a runaway intelligence. So what has happened here? One

77
00:05:19,040 --> 00:05:23,080
Speaker 1: person who has devoted himself to this question is Max Bennett,

78
00:05:23,080 --> 00:05:27,080
Speaker 1: who wrote a wonderful book called A Brief History of Intelligence,

79
00:05:27,480 --> 00:05:30,800
Speaker 1: And in this book, Max distills an enormous amount of

80
00:05:30,880 --> 00:05:34,960
Speaker 1: data about the history of animal species to reveal a

81
00:05:35,160 --> 00:05:40,839
Speaker 1: clear path that stretches from very ancient ancestors to us.

82
00:05:41,520 --> 00:05:45,839
Speaker 1: He attributes the story of human intelligence not just to

83
00:05:45,920 --> 00:05:50,719
Speaker 1: a single breakthrough, but to five breakthroughs. I really loved

84
00:05:50,720 --> 00:05:58,840
Speaker 1: his books, so I called him to join us today. So, Max,

85
00:05:58,880 --> 00:06:02,240
Speaker 1: when we're talking about the origins of intelligence, you might

86
00:06:02,320 --> 00:06:04,400
Speaker 1: think that what we need to do is look all

87
00:06:04,400 --> 00:06:07,279
Speaker 1: the way back to our common ancestors with the apes,

88
00:06:07,440 --> 00:06:10,640
Speaker 1: or maybe farther back to mammals, or maybe even as

89
00:06:10,640 --> 00:06:13,320
Speaker 1: far back as reptiles. But you suggest in your book

90
00:06:13,560 --> 00:06:16,560
Speaker 1: that we have to look back much farther than that.

91
00:06:16,720 --> 00:06:19,479
Speaker 1: Even so tell us where you think the sparks of

92
00:06:19,520 --> 00:06:20,600
Speaker 1: intelligence began.

93
00:06:21,320 --> 00:06:24,320
Speaker 2: So what's so interesting in trying to understand how the

94
00:06:24,360 --> 00:06:28,360
Speaker 2: human brain works is not only how much we've learned,

95
00:06:28,400 --> 00:06:31,880
Speaker 2: but also how much we've still failed to learn because

96
00:06:31,880 --> 00:06:34,240
Speaker 2: of how complicated the human brain is. I mean, the

97
00:06:34,279 --> 00:06:36,600
Speaker 2: human brain has eighty six billion neurons and one hundred

98
00:06:36,640 --> 00:06:40,360
Speaker 2: trillion connections, and so one strategy for trying to understand

99
00:06:40,400 --> 00:06:43,760
Speaker 2: the brain is to look at the series of steps

100
00:06:43,800 --> 00:06:46,200
Speaker 2: by which it came to be. Even if we only

101
00:06:46,240 --> 00:06:50,280
Speaker 2: go as far back as the first vertebrates, with whom

102
00:06:50,320 --> 00:06:53,039
Speaker 2: our common ancestors are around five hundred million years ago.

103
00:06:53,560 --> 00:06:56,880
Speaker 2: Our ancestors had brains somewhat akin to a modern fish,

104
00:06:57,279 --> 00:07:00,680
Speaker 2: and even in a fish brain there are a lot

105
00:07:00,760 --> 00:07:04,400
Speaker 2: of complicated structures and a lot of neurons. So I

106
00:07:04,400 --> 00:07:06,280
Speaker 2: think it behooves us to go back all the way

107
00:07:06,320 --> 00:07:09,200
Speaker 2: to the very first brains, which have brains akin to

108
00:07:09,840 --> 00:07:13,560
Speaker 2: a modern nematode and a modern Some species of modern nematodes,

109
00:07:13,600 --> 00:07:16,120
Speaker 2: like C. Elegans, only have three hundred two neurons, and

110
00:07:16,160 --> 00:07:18,119
Speaker 2: we can learn a lot about what the very first

111
00:07:18,120 --> 00:07:22,040
Speaker 2: brain did by understanding what a nematode brain does.

112
00:07:22,120 --> 00:07:25,400
Speaker 1: So tell us what a nematode is for listeners who don't.

113
00:07:25,280 --> 00:07:27,960
Speaker 2: Know, there's many different species of nematodes, but the most

114
00:07:27,960 --> 00:07:30,760
Speaker 2: well studied is something called Cea elegans, and it is

115
00:07:30,800 --> 00:07:34,040
Speaker 2: a small wormlike creature. You could fit a few on

116
00:07:34,080 --> 00:07:37,880
Speaker 2: your fingertip. And they have no eyes, they have no ears,

117
00:07:37,920 --> 00:07:40,640
Speaker 2: they can't render an image of the external world. They

118
00:07:40,680 --> 00:07:43,600
Speaker 2: only have three hundred two neurons in its entire nervous system,

119
00:07:44,040 --> 00:07:46,360
Speaker 2: and yet it can do some really impressive stuff and

120
00:07:46,360 --> 00:07:48,840
Speaker 2: that teach us a lot about the foundations of the

121
00:07:48,920 --> 00:07:49,680
Speaker 2: very first brains.

122
00:07:49,960 --> 00:07:52,680
Speaker 1: Okay, so give us a sense of what C. Elegans

123
00:07:52,720 --> 00:07:53,000
Speaker 1: can do.

124
00:07:53,480 --> 00:07:56,000
Speaker 2: One thing that's really interesting about C. Elegans is how

125
00:07:56,080 --> 00:07:59,400
Speaker 2: well it navigates the world and the absence of a

126
00:07:59,520 --> 00:08:03,640
Speaker 2: complex sensory apparatus. So one might think that in order

127
00:08:03,680 --> 00:08:07,679
Speaker 2: to find food or avoid predators, one needs to build

128
00:08:07,680 --> 00:08:11,280
Speaker 2: a map of space, or have eyes that enable them

129
00:08:11,320 --> 00:08:15,120
Speaker 2: to see into the distance, or have complex ears that

130
00:08:15,160 --> 00:08:18,640
Speaker 2: allow them to detect things through sound. But the elegance

131
00:08:18,640 --> 00:08:21,040
Speaker 2: has none of this. And yet if you put sea

132
00:08:21,040 --> 00:08:24,720
Speaker 2: elegans in a peatrie dish, it finds food rapidly. And

133
00:08:24,760 --> 00:08:27,720
Speaker 2: if you put them in the wild, they eminently find

134
00:08:27,800 --> 00:08:31,400
Speaker 2: optimal temperatures, and they eminently find ways to avoid predators.

135
00:08:31,800 --> 00:08:34,720
Speaker 2: And so the ways that their brain does this seems

136
00:08:34,720 --> 00:08:37,120
Speaker 2: to be quite similar to the way that a rumba works.

137
00:08:38,280 --> 00:08:42,800
Speaker 2: So a rumba, if folks aren't familiar, is the sort

138
00:08:42,800 --> 00:08:46,280
Speaker 2: of classic vacuum cleaning robot, and it also has no

139
00:08:46,360 --> 00:08:49,840
Speaker 2: eyes or ears, and yet somehow it cleans up everything

140
00:08:49,880 --> 00:08:53,240
Speaker 2: in your house. And so what a rumba does is

141
00:08:53,880 --> 00:08:55,600
Speaker 2: when it hits the wall, it sort of backs away

142
00:08:55,600 --> 00:08:58,079
Speaker 2: and turns randomly, and it keeps doing this randomly enough

143
00:08:58,160 --> 00:09:00,880
Speaker 2: until it reaches all the corners of your house. And

144
00:09:00,920 --> 00:09:04,640
Speaker 2: what nematoad does in some ways actually more advanced, where

145
00:09:04,679 --> 00:09:07,640
Speaker 2: it has sensory neurons around its head, and all these

146
00:09:07,679 --> 00:09:10,600
Speaker 2: sensory neurons do is they get excited when a good

147
00:09:10,640 --> 00:09:13,320
Speaker 2: thing like a smell, is increasing in concentration like a

148
00:09:13,320 --> 00:09:17,679
Speaker 2: food smell, and those drive forward movements, or another set

149
00:09:17,720 --> 00:09:22,000
Speaker 2: of neurons gets excited when something bad increases or something

150
00:09:22,000 --> 00:09:24,520
Speaker 2: good decreases, in other words, a decreasing concentration of a

151
00:09:24,559 --> 00:09:28,920
Speaker 2: food smell. And just by detecting these changes, a brain

152
00:09:29,040 --> 00:09:31,120
Speaker 2: can decide I'm going to keep going forward if good

153
00:09:31,160 --> 00:09:33,959
Speaker 2: things are increasing, or I'm going to turn randomly if

154
00:09:34,000 --> 00:09:37,800
Speaker 2: good things are decreasing, And this is classically called taxis navigation.

155
00:09:38,480 --> 00:09:41,000
Speaker 2: In simpler terms, you call this just steering. And in

156
00:09:41,040 --> 00:09:44,160
Speaker 2: the absence of any site, nematoads can find the origin

157
00:09:44,200 --> 00:09:48,040
Speaker 2: of food smells because food creates this gradience in water,

158
00:09:48,240 --> 00:09:51,280
Speaker 2: where the concentration of the smell is higher towards the source.

159
00:09:51,760 --> 00:09:55,200
Speaker 2: So the very first brain, its core function, was just

160
00:09:55,240 --> 00:09:57,520
Speaker 2: to categorize things in the world and too good and bad,

161
00:09:57,920 --> 00:09:59,760
Speaker 2: such that it would turn towards good things in a

162
00:09:59,800 --> 00:10:00,720
Speaker 2: way from bad things.

163
00:10:01,160 --> 00:10:03,760
Speaker 1: Now bacteria do that too, yes they do.

164
00:10:03,920 --> 00:10:09,800
Speaker 2: Clinokinesis absolutely what's almost mesmerizing about evolution is how this

165
00:10:10,080 --> 00:10:13,320
Speaker 2: exact same algorithm seems to have been recapitulated in a

166
00:10:13,320 --> 00:10:17,679
Speaker 2: completely different substrate. So single celled organisms do this exact

167
00:10:17,720 --> 00:10:20,920
Speaker 2: same type of taxis navigation, but it's implemented in sort

168
00:10:20,920 --> 00:10:24,319
Speaker 2: of the protein machinery of a single cell. And animatode

169
00:10:24,360 --> 00:10:27,400
Speaker 2: does the exact same algorithm, but not implemented within a

170
00:10:27,440 --> 00:10:29,360
Speaker 2: single cell, but through a web of neurons.

171
00:10:29,800 --> 00:10:32,240
Speaker 1: And so what you've proposed in your book, which is

172
00:10:32,280 --> 00:10:38,679
Speaker 1: an amazing book, is five breakthroughs that happened in evolutionary

173
00:10:38,760 --> 00:10:43,040
Speaker 1: time scales that led to intelligence the way that we

174
00:10:43,840 --> 00:10:47,840
Speaker 1: have and care about intelligence. So tell us about breakthrough

175
00:10:47,920 --> 00:10:48,400
Speaker 1: number one.

176
00:10:49,040 --> 00:10:52,679
Speaker 2: So breakthrough number one was this idea of steering. So

177
00:10:52,920 --> 00:10:57,000
Speaker 2: the animals before the first animals with brains, which are

178
00:10:57,000 --> 00:11:01,800
Speaker 2: classically called bileatrians because they have bilineateral symmetry, meaning they're

179
00:11:01,840 --> 00:11:04,959
Speaker 2: symmetric across the central plane. It is interesting to real

180
00:11:05,040 --> 00:11:07,320
Speaker 2: people don't realize this until they think about it, but

181
00:11:07,440 --> 00:11:09,800
Speaker 2: all animals that we think of as animals are symmetric

182
00:11:09,920 --> 00:11:11,600
Speaker 2: across the central line through their body.

183
00:11:11,840 --> 00:11:13,200
Speaker 1: So you mean they have a left side on the

184
00:11:13,280 --> 00:11:15,880
Speaker 1: right side, and they are a mirror image.

185
00:11:16,240 --> 00:11:19,439
Speaker 2: Yeah, and so, but not all animals have that. So

186
00:11:19,600 --> 00:11:21,920
Speaker 2: the very very first animals, we think, we don't have

187
00:11:21,960 --> 00:11:24,719
Speaker 2: perfect evidence for this, but we think we're probably more

188
00:11:24,760 --> 00:11:28,280
Speaker 2: akin to a coral polyp or a jellyfish, which has

189
00:11:28,400 --> 00:11:33,160
Speaker 2: radial symmetry, so they're symmetric across a central axis. And

190
00:11:33,240 --> 00:11:37,080
Speaker 2: so the transition from radial symmetry to bilateral symmetry seems

191
00:11:37,120 --> 00:11:39,960
Speaker 2: to be in part driven by the need to navigate. So,

192
00:11:40,440 --> 00:11:43,280
Speaker 2: although jellyfish are an interesting exception because some of them

193
00:11:43,360 --> 00:11:48,080
Speaker 2: independently seem to have evolved relatively complex navigational systems, most

194
00:11:48,400 --> 00:11:52,559
Speaker 2: evolutionary neuroscientists think the very first animals were more sensile,

195
00:11:52,800 --> 00:11:54,960
Speaker 2: like a coral polyp, where they sit in place. They

196
00:11:55,000 --> 00:11:57,040
Speaker 2: have tentacles and they just try to detect food that

197
00:11:57,080 --> 00:12:00,280
Speaker 2: pass by the tentacles. But the very first animal with

198
00:12:00,320 --> 00:12:05,040
Speaker 2: brains are bilateral ancestors. They use this brain to categorize

199
00:12:05,080 --> 00:12:07,040
Speaker 2: the world and to good and bad. To implement this

200
00:12:07,160 --> 00:12:10,280
Speaker 2: taxis navigation to find food and avoid predators.

201
00:12:10,679 --> 00:12:14,640
Speaker 1: So the existence of a brain correlates with having this

202
00:12:14,760 --> 00:12:17,120
Speaker 1: left right side. Is that correct?

203
00:12:17,480 --> 00:12:20,880
Speaker 2: There are all animals with brains descend have bilateral symmetry,

204
00:12:20,960 --> 00:12:24,000
Speaker 2: or descend from the bilaterally symmetric ancestor in which the

205
00:12:24,000 --> 00:12:27,000
Speaker 2: first brains evolved. And so we also see a suite

206
00:12:27,000 --> 00:12:30,199
Speaker 2: of other interesting things emerged with this first breakthrough of steering.

207
00:12:31,120 --> 00:12:34,440
Speaker 2: One is classically called affect, which is sort of the

208
00:12:34,440 --> 00:12:40,160
Speaker 2: first template of emotional states. And so a nematode actually

209
00:12:40,240 --> 00:12:44,400
Speaker 2: has dopamine neurons, and what these dopamine neurons do is

210
00:12:44,400 --> 00:12:48,160
Speaker 2: they detect the presence of bacteria outside of the nematode.

211
00:12:48,520 --> 00:12:51,240
Speaker 2: And what it does is it changed their behavioral repertoire

212
00:12:51,320 --> 00:12:54,720
Speaker 2: to search in their local area. And we see why

213
00:12:54,760 --> 00:12:56,840
Speaker 2: this exists in the rumba. So a rumba has something

214
00:12:56,880 --> 00:13:00,200
Speaker 2: called dirt detect and what dirt detect does is if

215
00:13:00,200 --> 00:13:03,840
Speaker 2: it bumps into dirt, it starts turning randomly in that area.

216
00:13:04,679 --> 00:13:06,839
Speaker 2: And the reason it does that is because the world

217
00:13:06,880 --> 00:13:09,600
Speaker 2: is clumpy, So if you detect dirt, it's likely that

218
00:13:09,640 --> 00:13:12,880
Speaker 2: there's other dirt nearby, even though you're not. Maybe detecting

219
00:13:12,960 --> 00:13:16,080
Speaker 2: dirt in the moment. So what anematod does is the

220
00:13:16,080 --> 00:13:18,800
Speaker 2: exact same thing. If it runs into food, even though

221
00:13:18,840 --> 00:13:21,480
Speaker 2: it might not detect food a second later, it's probably

222
00:13:21,520 --> 00:13:24,160
Speaker 2: the case there's other food nearby, and so this rush

223
00:13:24,160 --> 00:13:29,199
Speaker 2: of dopamine drives this local search in these very early brains. Similarly,

224
00:13:29,640 --> 00:13:33,200
Speaker 2: there are serotonin neurons, but they're in the throat, and

225
00:13:33,280 --> 00:13:36,640
Speaker 2: so what serotonin signals is the consumption of food, and

226
00:13:36,720 --> 00:13:40,199
Speaker 2: serotonin in these very early nematodes drives sort of satiation.

227
00:13:41,120 --> 00:13:44,800
Speaker 2: And of course those chemicals do much more complicated things

228
00:13:44,800 --> 00:13:48,000
Speaker 2: than human brains. That basic template of dopamine being the

229
00:13:48,040 --> 00:13:53,240
Speaker 2: seeking exploitation nearby reward signal and serotonin being the sort

230
00:13:53,280 --> 00:13:58,680
Speaker 2: of satiation consumption satisfaction signal. We do see hints of

231
00:13:58,760 --> 00:14:04,480
Speaker 2: that basic even in human brains. So we see categorizing

232
00:14:04,480 --> 00:14:07,000
Speaker 2: the world into good to bad, we see bilateral symmetry,

233
00:14:07,040 --> 00:14:09,800
Speaker 2: we see these very basic behavioral states. And then the

234
00:14:09,880 --> 00:14:11,920
Speaker 2: last thing we also see emerge in this breakthrough of

235
00:14:12,000 --> 00:14:15,520
Speaker 2: steering is the foundation of associative learning, and this is

236
00:14:15,520 --> 00:14:18,280
Speaker 2: the first form of real learning that we see emerge

237
00:14:18,440 --> 00:14:24,520
Speaker 2: in animal evolution. And anematode can associate a stimulus with

238
00:14:24,840 --> 00:14:27,240
Speaker 2: a good or bad thing. So if you put a

239
00:14:27,240 --> 00:14:30,720
Speaker 2: nematode in a peach redish and put salt on one side,

240
00:14:31,000 --> 00:14:35,120
Speaker 2: nematoads typically steer towards salts because salt tends to correlate

241
00:14:35,160 --> 00:14:37,720
Speaker 2: with food. But if you leave them in a peach redition,

242
00:14:37,920 --> 00:14:39,760
Speaker 2: starve them for a long period of time in the

243
00:14:39,760 --> 00:14:42,800
Speaker 2: presence of salt water, they change their opinion and they

244
00:14:42,800 --> 00:14:45,160
Speaker 2: will start steering away from salt in the future. And

245
00:14:45,200 --> 00:14:48,360
Speaker 2: it makes sense why associative learning would emerge if the

246
00:14:48,440 --> 00:14:51,080
Speaker 2: very first brain of steering, because you want to tweak

247
00:14:51,120 --> 00:14:54,000
Speaker 2: the goodness and badness of things, because deciding what to

248
00:14:54,040 --> 00:14:55,840
Speaker 2: turn towards and away from is a life or death

249
00:14:55,840 --> 00:14:58,720
Speaker 2: decision for anema TOD. So this first breakthrough of steering,

250
00:14:58,760 --> 00:15:01,640
Speaker 2: we see the suite of of new abilities from associative learning.

251
00:15:01,640 --> 00:15:05,440
Speaker 2: Bilateral symmetry categorizing things in a good and bad emerged

252
00:15:05,480 --> 00:15:08,240
Speaker 2: with the very first brain. So that was breakthrough number one.

253
00:15:08,080 --> 00:15:10,400
Speaker 1: Okay, terrific. And what was break through number two?

254
00:15:11,360 --> 00:15:14,400
Speaker 2: So if we fast forward about fifty million years or so,

255
00:15:14,840 --> 00:15:18,680
Speaker 2: we enter what's famously known as the Cambrian Period and

256
00:15:18,720 --> 00:15:23,200
Speaker 2: the Cambrian Explosion, is this huge diversification of life, which

257
00:15:23,280 --> 00:15:27,320
Speaker 2: actually is all of the children of this first bilateral animal.

258
00:15:27,520 --> 00:15:30,240
Speaker 2: So if you were to swim around the Cambrian Ocean,

259
00:15:30,560 --> 00:15:34,520
Speaker 2: you would see many ancestors of this bilateral wormlike creature

260
00:15:34,800 --> 00:15:37,960
Speaker 2: who had proliferated into what would look like crustaceans and

261
00:15:38,080 --> 00:15:42,240
Speaker 2: arthropods of today. There were huge insect like creatures in

262
00:15:42,280 --> 00:15:45,400
Speaker 2: the ocean, and then there were also our ancestors, which

263
00:15:45,440 --> 00:15:49,120
Speaker 2: were much smaller, modest creatures, but they were most akin

264
00:15:49,200 --> 00:15:51,240
Speaker 2: to a fish of today, and they were called the

265
00:15:51,280 --> 00:15:54,840
Speaker 2: first vertebrates. And the reason they're called vertebrates is because

266
00:15:54,840 --> 00:15:58,640
Speaker 2: in fossils, the most salient feature is the vertebral column,

267
00:15:58,760 --> 00:16:01,840
Speaker 2: so they had a spine. And in these first vertebrates,

268
00:16:02,360 --> 00:16:04,920
Speaker 2: we can get insight into what their brains did by

269
00:16:04,960 --> 00:16:09,160
Speaker 2: looking into the brains of fish today, because there's many

270
00:16:09,240 --> 00:16:12,920
Speaker 2: species of fish that evolutionary neuroscientists think have brains that

271
00:16:12,960 --> 00:16:15,480
Speaker 2: were quite similar to the very first vertebrates. And what

272
00:16:15,560 --> 00:16:17,840
Speaker 2: I found most surprising when I first started looking into

273
00:16:17,920 --> 00:16:20,880
Speaker 2: this is how similar fish brains are to human brains.

274
00:16:21,000 --> 00:16:22,880
Speaker 2: So I would have expected a fish brain to have

275
00:16:22,920 --> 00:16:25,600
Speaker 2: almost none of the features of a human brain, but

276
00:16:26,480 --> 00:16:29,680
Speaker 2: counter to that, intuition. Fish brain have all, with the

277
00:16:29,720 --> 00:16:32,040
Speaker 2: exception of a few things, have all of the major

278
00:16:32,080 --> 00:16:36,200
Speaker 2: brain structures that a human brain does. And also, counter

279
00:16:36,360 --> 00:16:39,400
Speaker 2: to what my expectations would have been, there's sort of

280
00:16:39,400 --> 00:16:42,440
Speaker 2: a stereotype that fish are really dumb, but the more

281
00:16:42,440 --> 00:16:45,520
Speaker 2: you look into the comparative psychology work done on fish,

282
00:16:45,680 --> 00:16:48,320
Speaker 2: fish are way smarter than we think. And for example,

283
00:16:48,720 --> 00:16:50,920
Speaker 2: fish can learn how to navigate out of a maze

284
00:16:50,960 --> 00:16:53,000
Speaker 2: and remember exactly how to do it a year later.

285
00:16:53,320 --> 00:16:55,680
Speaker 2: You can go to YouTube and find really funny cute

286
00:16:55,760 --> 00:16:58,560
Speaker 2: videos of people training fish to jump through hoops for treats,

287
00:16:59,040 --> 00:17:01,320
Speaker 2: and you can train a to push levers for food

288
00:17:01,360 --> 00:17:03,600
Speaker 2: and all of these sort of fun things. And so

289
00:17:04,040 --> 00:17:06,760
Speaker 2: when we look at these brain structures that emerged, there's

290
00:17:06,800 --> 00:17:09,320
Speaker 2: a lot of really good evidence that the key thing

291
00:17:09,400 --> 00:17:13,280
Speaker 2: that happened was these early vertebrate brains enabled the ability

292
00:17:13,320 --> 00:17:16,720
Speaker 2: to learn through reinforcement and AI. This is called reinforcement

293
00:17:16,840 --> 00:17:20,960
Speaker 2: learning and behavioral psychology is typically called trial and error learning.

294
00:17:21,280 --> 00:17:24,360
Speaker 2: So they could learn to perform arbitrary sequences of actions

295
00:17:24,440 --> 00:17:27,600
Speaker 2: on the basis of whether or not it led to

296
00:17:27,640 --> 00:17:30,440
Speaker 2: a reward at the end. So when we go into

297
00:17:30,440 --> 00:17:33,439
Speaker 2: the fish brain. There are two key structures that are

298
00:17:33,520 --> 00:17:36,879
Speaker 2: useful to know about because they will keep coming up

299
00:17:36,880 --> 00:17:39,400
Speaker 2: through our story and the evolution of the human brain.

300
00:17:39,680 --> 00:17:42,320
Speaker 2: One is something called the basal ganglia, and the basil

301
00:17:42,320 --> 00:17:45,600
Speaker 2: ganglia of a fish has almost exactly the same structure

302
00:17:46,240 --> 00:17:49,520
Speaker 2: and network as the basil ganglia of a human, and

303
00:17:49,960 --> 00:17:53,920
Speaker 2: computational neuroscientists have gone to painstaking efforts to show that

304
00:17:53,960 --> 00:17:58,000
Speaker 2: the basil gangly is implementing a reinforcement learning algorithm almost

305
00:17:58,000 --> 00:18:01,240
Speaker 2: identical to the reinforcement learning algorithm we use in AI

306
00:18:01,359 --> 00:18:05,040
Speaker 2: system today. And the way that it works in principle

307
00:18:05,280 --> 00:18:08,920
Speaker 2: is it trains itself based on the exciting the excitement

308
00:18:08,920 --> 00:18:12,600
Speaker 2: of dopamine, and it learns to repeat behaviors that drive

309
00:18:12,640 --> 00:18:16,439
Speaker 2: dopamine release and inhibit behaviors that drive dopamine decreasing. And

310
00:18:16,480 --> 00:18:18,680
Speaker 2: what's so fascinating is if you look at how this

311
00:18:19,000 --> 00:18:21,920
Speaker 2: system came to be, you can see how reinforcement learning

312
00:18:22,000 --> 00:18:25,560
Speaker 2: is only possible if brains first had the foundation of steering.

313
00:18:26,040 --> 00:18:29,359
Speaker 2: Because the foundation of steering gives us the categorization of

314
00:18:29,359 --> 00:18:31,600
Speaker 2: things in the world and to good and bad, and

315
00:18:31,640 --> 00:18:34,720
Speaker 2: that is repurposed to create this reward signal that the

316
00:18:34,720 --> 00:18:38,159
Speaker 2: basal ganglia then can use to create arbitrary sequences of

317
00:18:38,200 --> 00:18:41,400
Speaker 2: behavior on the basis of what leads to reward or none.

318
00:18:41,960 --> 00:18:44,120
Speaker 2: And this is how a fish can learn really complex

319
00:18:44,160 --> 00:18:47,600
Speaker 2: sequences of actions on the basis of what leads to reward.

320
00:18:47,640 --> 00:18:50,320
Speaker 2: In the end, the second key structure in a fish

321
00:18:50,359 --> 00:18:53,480
Speaker 2: brain is something called the cortex, and we do have

322
00:18:53,520 --> 00:18:55,919
Speaker 2: a version of a cortex. There's a portion of our

323
00:18:55,960 --> 00:18:58,439
Speaker 2: cortex that we'll talk about that's way more advanced. But

324
00:18:58,480 --> 00:19:01,800
Speaker 2: a phish cortex can still do something incredible that the

325
00:19:01,800 --> 00:19:05,399
Speaker 2: first nematodes could not, which is it recognizes things in

326
00:19:05,440 --> 00:19:08,679
Speaker 2: the world on the basis of patterns. So in the

327
00:19:08,720 --> 00:19:11,919
Speaker 2: first nema, in the first bilateral brain, it could not

328
00:19:12,000 --> 00:19:13,679
Speaker 2: detect things in the world on the basis of a

329
00:19:13,680 --> 00:19:16,080
Speaker 2: pattern of activation. So when you look at a horse,

330
00:19:16,440 --> 00:19:19,000
Speaker 2: you recognize a horse not because of any single neuron

331
00:19:19,040 --> 00:19:21,760
Speaker 2: in your brain, but because your brain is somehow decoding

332
00:19:21,800 --> 00:19:24,480
Speaker 2: the pattern of activation on your retina, the neurons in

333
00:19:24,520 --> 00:19:27,760
Speaker 2: your retina. And so the first brains could not do

334
00:19:27,840 --> 00:19:30,639
Speaker 2: anything like this. They only detected things when a single

335
00:19:30,680 --> 00:19:34,040
Speaker 2: neuron got excited in the presence of some stimulus. But fish,

336
00:19:34,400 --> 00:19:36,760
Speaker 2: fish can even recognize human faces. There have been some

337
00:19:36,800 --> 00:19:39,040
Speaker 2: amazing studies that show a fish can recognize a human

338
00:19:39,040 --> 00:19:41,520
Speaker 2: face and learn which face leads to a reward in

339
00:19:41,560 --> 00:19:44,640
Speaker 2: which face does not. Even when that face is rotated

340
00:19:44,680 --> 00:19:48,320
Speaker 2: in space, they still recognize it. So the cortex somehow,

341
00:19:48,320 --> 00:19:51,600
Speaker 2: this is still an outstanding mystery in the field of neuroscience.

342
00:19:51,920 --> 00:19:56,320
Speaker 2: Somehow the cortex recognizes patterns and fish eminently well. And

343
00:19:56,359 --> 00:19:59,160
Speaker 2: in some ways the cortex of a fish recognizes patterns

344
00:19:59,200 --> 00:20:01,359
Speaker 2: better than even our our best vision systems in AI,

345
00:20:01,480 --> 00:20:03,680
Speaker 2: because we can. They've done studies that show that a

346
00:20:03,760 --> 00:20:07,200
Speaker 2: fish can recognize objects in one shots even though it's

347
00:20:07,240 --> 00:20:10,159
Speaker 2: been rotated in space, and AI systems typically don't do that.

348
00:20:10,160 --> 00:20:11,600
Speaker 2: You need a lot of data to get into that.

349
00:20:12,000 --> 00:20:15,560
Speaker 2: So at the first fish brain, we see reinforcement learning emerge,

350
00:20:15,840 --> 00:20:19,000
Speaker 2: which can recognize patterns in the world and can learn

351
00:20:19,040 --> 00:20:21,359
Speaker 2: to take actions in the presence of those patterns based

352
00:20:21,400 --> 00:20:24,200
Speaker 2: on rewards. We see reinforcement learning as breakthrough number two.

353
00:20:24,800 --> 00:20:26,879
Speaker 1: Excellent, Okay, how about number three?

354
00:20:27,160 --> 00:20:30,199
Speaker 2: Then we're going to fast forward through a long period

355
00:20:30,240 --> 00:20:34,160
Speaker 2: of evolutionary time, all the way until about one hundred

356
00:20:34,160 --> 00:20:36,240
Speaker 2: and fifty million years ago. Between hundred and hundred million

357
00:20:36,280 --> 00:20:40,480
Speaker 2: years ago. This is the era of dinosaurs. Our ancestors

358
00:20:40,480 --> 00:20:44,600
Speaker 2: were very, very humble, tiny squirrel like creatures that lived underground,

359
00:20:45,080 --> 00:20:48,560
Speaker 2: and we only came out at nights to hunt for insects.

360
00:20:49,240 --> 00:20:51,800
Speaker 2: But these were the first mammals. We know a lot

361
00:20:51,800 --> 00:20:54,120
Speaker 2: about mammal brains, way more than we actually know about

362
00:20:54,160 --> 00:20:57,959
Speaker 2: fish brains, because the main stay of neuroscience research typically

363
00:20:58,000 --> 00:21:02,760
Speaker 2: happens in rats and mice when we go into these brains. Interestingly,

364
00:21:03,000 --> 00:21:06,399
Speaker 2: the fundamental difference between a mammal brain and a fish

365
00:21:06,440 --> 00:21:09,840
Speaker 2: brain is the presence of one key new structure, which

366
00:21:09,920 --> 00:21:13,440
Speaker 2: is a part of the cortex elaborates into what's famously

367
00:21:13,480 --> 00:21:17,560
Speaker 2: called the neocortex NEO for new, and under a microscope

368
00:21:17,560 --> 00:21:20,800
Speaker 2: there's some really interesting things. So we have remnants of

369
00:21:20,840 --> 00:21:24,800
Speaker 2: the old cortex of fish they are called the olfactory cortex,

370
00:21:24,920 --> 00:21:28,440
Speaker 2: and humans and mammals they're called the hippocampus, and they're

371
00:21:28,440 --> 00:21:32,000
Speaker 2: called the cortical amygdala. These are all ancestral remnants of

372
00:21:32,040 --> 00:21:35,240
Speaker 2: the very first cortex. But the neocortex is entirely new.

373
00:21:35,400 --> 00:21:38,560
Speaker 2: This is something that only occurred within mammals, and it

374
00:21:38,600 --> 00:21:42,120
Speaker 2: looks way more complicated under a microscope and so there's

375
00:21:42,119 --> 00:21:47,080
Speaker 2: this grand question what did this neocortex do? And classically,

376
00:21:47,400 --> 00:21:50,240
Speaker 2: when we study the neocortex, we look at a lot

377
00:21:50,280 --> 00:21:52,200
Speaker 2: of humans, and when you look at a human brain,

378
00:21:52,880 --> 00:21:54,960
Speaker 2: the whole thing seems to be neocortex. So when you

379
00:21:55,000 --> 00:21:57,160
Speaker 2: look at human brain, all of this all is full,

380
00:21:57,240 --> 00:22:01,720
Speaker 2: that's all neocortex bunched together. It's this sort of has

381
00:22:01,800 --> 00:22:06,879
Speaker 2: this sort of surface area, and the neocortex seems to

382
00:22:06,920 --> 00:22:09,960
Speaker 2: do everything, which is this funny perplexing thing in neuroscience.

383
00:22:10,040 --> 00:22:12,840
Speaker 2: Because there's one region that seems to do vision. If

384
00:22:12,840 --> 00:22:15,960
Speaker 2: it gets damaged, people can't see. There's another area that

385
00:22:16,000 --> 00:22:18,520
Speaker 2: seems to do audition. If it gets damaged, people struggle

386
00:22:18,520 --> 00:22:20,919
Speaker 2: to hear things. There's a region that seems to do attention.

387
00:22:20,960 --> 00:22:23,000
Speaker 2: If it get damaged, you can't perceive things on one

388
00:22:23,000 --> 00:22:26,400
Speaker 2: side or visual of view. There's an area for movements.

389
00:22:26,480 --> 00:22:28,560
Speaker 2: If it gets damage, you get paralyzed, so on and

390
00:22:28,600 --> 00:22:30,840
Speaker 2: so forth. So it's this grand sort of mystery of

391
00:22:30,840 --> 00:22:33,480
Speaker 2: what the neocortex does, but most of it seems to

392
00:22:33,520 --> 00:22:35,679
Speaker 2: have been based on this idea of perception. A lot

393
00:22:35,720 --> 00:22:38,040
Speaker 2: of the neocortex seems to enable us to perceive things

394
00:22:38,040 --> 00:22:40,240
Speaker 2: in the world, But what's odd is if we think

395
00:22:40,240 --> 00:22:44,639
Speaker 2: about this from an evolutionary perspective, there's no clear perceptual

396
00:22:44,840 --> 00:22:49,040
Speaker 2: improvements or very salient at least perceptual improvements, and a

397
00:22:49,080 --> 00:22:51,680
Speaker 2: mammal relative to a fish, so a fish can recognize

398
00:22:51,680 --> 00:22:54,760
Speaker 2: faces as well as a rat can. It recognizes them

399
00:22:54,800 --> 00:22:57,600
Speaker 2: when rotated in space, So it's not so clear from

400
00:22:57,600 --> 00:23:01,680
Speaker 2: an evolutionary perspective that the neocortech evolve for better perception.

401
00:23:02,119 --> 00:23:05,639
Speaker 2: If we really examine the fundamental differences in the abilities

402
00:23:05,720 --> 00:23:09,520
Speaker 2: of simple mammals with fish. There are, however, four things

403
00:23:09,520 --> 00:23:11,639
Speaker 2: that are seen, and I think these are great clues

404
00:23:11,680 --> 00:23:14,240
Speaker 2: as to what the first the neo cortex did. One

405
00:23:14,280 --> 00:23:16,880
Speaker 2: thing that mammals can do very well is they can

406
00:23:16,920 --> 00:23:20,800
Speaker 2: imagine the future. So there's some really wonderful studies done

407
00:23:20,840 --> 00:23:23,200
Speaker 2: by David Reddish that show you can put a mouse

408
00:23:23,200 --> 00:23:25,080
Speaker 2: in amaze and you can watch a mouse imagining its

409
00:23:25,119 --> 00:23:30,359
Speaker 2: possible futures. Another thing you can do is mammals, even rats,

410
00:23:30,720 --> 00:23:33,400
Speaker 2: are eminently good at having regret. So if you put

411
00:23:33,400 --> 00:23:36,400
Speaker 2: them in a situation where they have to make irreversible choices,

412
00:23:36,640 --> 00:23:38,920
Speaker 2: they will often regret their decision and you can watch

413
00:23:38,960 --> 00:23:41,919
Speaker 2: them in their brain imagining themselves taking prior past choices.

414
00:23:42,520 --> 00:23:45,880
Speaker 2: Mammals also have something akin to episodic memory. You can

415
00:23:45,880 --> 00:23:49,359
Speaker 2: put rats in experiments where they have to imagine some

416
00:23:49,400 --> 00:23:51,399
Speaker 2: recent past event in order to solve a puzzle in

417
00:23:51,440 --> 00:23:52,879
Speaker 2: front of them, and you can watch them do that.

418
00:23:53,560 --> 00:23:55,760
Speaker 2: And then the fourth is they have really great fine

419
00:23:55,760 --> 00:24:00,000
Speaker 2: motor skills. So in the reptile literature, there's some good

420
00:24:00,080 --> 00:24:04,000
Speaker 2: evidence that most lizards, with the exception of birds, which

421
00:24:04,040 --> 00:24:07,320
Speaker 2: is a non mammalian vertebrate that has amazing find motor skills.

422
00:24:07,640 --> 00:24:12,639
Speaker 2: But reptiles don't even sort of anticipate our movements to

423
00:24:12,760 --> 00:24:15,600
Speaker 2: get over obstacles. They're very sloppy in their movements. And

424
00:24:15,680 --> 00:24:18,639
Speaker 2: yet a squirrel, watch a squirrel run across sort of

425
00:24:18,680 --> 00:24:21,440
Speaker 2: tree branches, has find motor skills that blow away any

426
00:24:21,480 --> 00:24:26,240
Speaker 2: modern robotic system. So these four things actually can be

427
00:24:26,320 --> 00:24:30,200
Speaker 2: seen as different applications of what I would call simulating

428
00:24:30,840 --> 00:24:34,160
Speaker 2: an AI. This is called planning. Typically, so mammal brains

429
00:24:34,200 --> 00:24:37,800
Speaker 2: are good at simulating possible states of the worlds and

430
00:24:37,840 --> 00:24:40,359
Speaker 2: then making choices on the basis of that simulation. They

431
00:24:40,400 --> 00:24:43,640
Speaker 2: can simulate the future, that's imagination. They can simulate past

432
00:24:43,640 --> 00:24:46,679
Speaker 2: events that's episodic memory, they can simulate and plan their

433
00:24:47,520 --> 00:24:50,840
Speaker 2: hand motions, which is effectively enabling them to find motor skills.

434
00:24:51,440 --> 00:24:54,560
Speaker 2: And so this mental simulation we even see in humans.

435
00:24:54,600 --> 00:24:56,560
Speaker 2: I mean, we are eminently capable of doing this. Close

436
00:24:56,560 --> 00:24:58,920
Speaker 2: your eyes. You can imagine things in your mind's eye.

437
00:24:59,160 --> 00:25:01,439
Speaker 2: This lights up your neocortex the same way as if

438
00:25:01,480 --> 00:25:04,600
Speaker 2: you perceived those same objects. And so simulation was this

439
00:25:04,680 --> 00:25:07,840
Speaker 2: incredible skill given to these early mammals because it enabled

440
00:25:07,840 --> 00:25:10,600
Speaker 2: them to plan their movements ahead of time and sort

441
00:25:10,600 --> 00:25:15,200
Speaker 2: of outsmart the dinosaurs. In AI today, this is classically

442
00:25:15,200 --> 00:25:19,320
Speaker 2: called model based reinforcement learning. And so in AAI there's

443
00:25:19,320 --> 00:25:22,359
Speaker 2: this big division between model free, which means learning to

444
00:25:22,400 --> 00:25:25,399
Speaker 2: take actions without any planning at all. You just see

445
00:25:25,440 --> 00:25:27,400
Speaker 2: sort of the current state and then you make a choice.

446
00:25:27,640 --> 00:25:30,680
Speaker 2: Our self driving cars, the AI algorithm that keeps you

447
00:25:30,720 --> 00:25:32,600
Speaker 2: in the lane is a model free system just sees

448
00:25:32,640 --> 00:25:34,520
Speaker 2: a picture of the road and decides how to put

449
00:25:34,560 --> 00:25:38,960
Speaker 2: the seering wheel. Model based systems are ones that imagine

450
00:25:39,000 --> 00:25:43,000
Speaker 2: possible futures before making a choice. So Alpha Go, that

451
00:25:43,080 --> 00:25:45,720
Speaker 2: one classically be the best go player in the world,

452
00:25:45,800 --> 00:25:48,600
Speaker 2: was a model based reinforcement learning system. It actually within

453
00:25:48,640 --> 00:25:51,800
Speaker 2: a matter of seconds simulated thousands of possible games before

454
00:25:51,800 --> 00:25:55,080
Speaker 2: making a choice, and so there's this really nice synergy

455
00:25:55,520 --> 00:25:58,280
Speaker 2: with AI. Where in early vertebrates, with breakthrough two, we

456
00:25:58,359 --> 00:26:01,359
Speaker 2: see model free reinforcement learning. There's no evidence of fish

457
00:26:01,400 --> 00:26:04,240
Speaker 2: being able to imagine the future, but with early mammals

458
00:26:04,240 --> 00:26:07,000
Speaker 2: we see model based reinforcement learning, which is them being

459
00:26:07,000 --> 00:26:10,200
Speaker 2: able to imagine futures before acting. And what is also

460
00:26:10,240 --> 00:26:13,760
Speaker 2: really interesting is how you can't have simulation without first

461
00:26:13,840 --> 00:26:16,639
Speaker 2: having trial and error learning, because the way that simulation

462
00:26:16,880 --> 00:26:20,960
Speaker 2: cascades into action is you're training yourself in your mind's eye.

463
00:26:21,240 --> 00:26:24,040
Speaker 2: When a rat closes its eyes and imagines itself taking

464
00:26:24,359 --> 00:26:28,360
Speaker 2: multiple paths, a little dopamine gets released when it imagines

465
00:26:28,400 --> 00:26:31,199
Speaker 2: taking the path that actually leads to food. And so

466
00:26:31,320 --> 00:26:35,040
Speaker 2: then the way that the simulation leads to action is

467
00:26:35,040 --> 00:26:37,280
Speaker 2: because you already have this trial and error system in

468
00:26:37,280 --> 00:26:41,159
Speaker 2: place that you're training vicariously with your mind. This is

469
00:26:41,200 --> 00:26:44,040
Speaker 2: also why they've shown this with athletes too. This is

470
00:26:44,080 --> 00:26:48,239
Speaker 2: why mental rehearsal dramatically improves performance. Surgeons also, they've done

471
00:26:48,280 --> 00:26:51,360
Speaker 2: studies that show mental rehearsal improves performance. Okay, so that's

472
00:26:51,400 --> 00:26:52,439
Speaker 2: break through number three.

473
00:26:52,600 --> 00:26:54,800
Speaker 1: Yeah, this is something I talked about on this podcast

474
00:26:54,840 --> 00:26:57,320
Speaker 1: A lot is the way that we unhook from the

475
00:26:57,400 --> 00:27:00,199
Speaker 1: here and now and we go to the therein and then,

476
00:27:00,240 --> 00:27:02,760
Speaker 1: whether that's in the future or the past. As the

477
00:27:02,800 --> 00:27:06,760
Speaker 1: philosopher Carl Popper said, this is what allows our hypotheses

478
00:27:06,800 --> 00:27:09,960
Speaker 1: to die in our stead. And we're going to come

479
00:27:10,000 --> 00:27:29,159
Speaker 1: back to internal models a little bit tell us about

480
00:27:29,240 --> 00:27:30,120
Speaker 1: the next breakthrough.

481
00:27:30,600 --> 00:27:34,040
Speaker 2: Okay, So moving forward from early mammals, a huge asteroid

482
00:27:34,119 --> 00:27:38,080
Speaker 2: hits Earth, which tragically kills off all the dinosaurs and

483
00:27:38,119 --> 00:27:41,159
Speaker 2: opens up the world for what is sometimes called the

484
00:27:41,160 --> 00:27:44,040
Speaker 2: Age of mammals because our ancestors took over from that

485
00:27:44,080 --> 00:27:46,720
Speaker 2: point forward. It is an interesting quirk that if that

486
00:27:46,800 --> 00:27:49,840
Speaker 2: asteroid never heard Earth, there would almost certainly be no humans,

487
00:27:49,880 --> 00:27:51,919
Speaker 2: and it would likely be that we would still be

488
00:27:52,040 --> 00:27:55,639
Speaker 2: tiny little squirrels hiding in the dirt. So that is

489
00:27:55,720 --> 00:28:00,840
Speaker 2: just an interesting accident of the universe. But as mammals

490
00:28:00,840 --> 00:28:05,080
Speaker 2: started proliferating throughout Earth, our ancestors were the ones that

491
00:28:05,200 --> 00:28:08,320
Speaker 2: stayed in the trees and they became the first primates.

492
00:28:09,160 --> 00:28:12,399
Speaker 2: And primates are known for having really really big brains,

493
00:28:12,720 --> 00:28:16,600
Speaker 2: you know. The modern primates include monkeys, non human apes,

494
00:28:16,680 --> 00:28:20,320
Speaker 2: and of course humans, of whom are apes? And these

495
00:28:20,359 --> 00:28:24,400
Speaker 2: primates have really big brains for a perplexing reason. So

496
00:28:24,600 --> 00:28:28,560
Speaker 2: it's been open question in primatology for a lot, or

497
00:28:28,640 --> 00:28:30,080
Speaker 2: was an open question for a long time, why do

498
00:28:30,119 --> 00:28:32,360
Speaker 2: primates have such big brains. They don't seem to have

499
00:28:32,760 --> 00:28:36,960
Speaker 2: such a complex lifestyle that requires them this massive neocortex

500
00:28:37,040 --> 00:28:40,840
Speaker 2: that evolved. But several decades ago some theories emerged that

501
00:28:40,840 --> 00:28:43,920
Speaker 2: have been proven out, which it seems to be something

502
00:28:43,920 --> 00:28:47,360
Speaker 2: about the social lives of primates that drive their really

503
00:28:47,360 --> 00:28:50,760
Speaker 2: big brains. And so Robin Dunbar is one of the

504
00:28:50,800 --> 00:28:53,320
Speaker 2: early people that came up with this idea, And what

505
00:28:53,400 --> 00:28:55,880
Speaker 2: he did is he looked at the size of the

506
00:28:55,880 --> 00:28:59,240
Speaker 2: social group of primates and compared it to the relative

507
00:28:59,280 --> 00:29:01,440
Speaker 2: size of their new cortex relatives to the rest of

508
00:29:01,440 --> 00:29:05,240
Speaker 2: the brain. And you see this almost beautiful curve where

509
00:29:05,280 --> 00:29:07,520
Speaker 2: the bigger the social group, the bigger the relative size

510
00:29:07,560 --> 00:29:11,120
Speaker 2: their neocortex. This relationship does not hold for other mammals.

511
00:29:11,160 --> 00:29:14,080
Speaker 2: So this is not some universal principle, but something about

512
00:29:14,120 --> 00:29:18,160
Speaker 2: primate societies are such that they require really big neo courtices.

513
00:29:18,680 --> 00:29:21,360
Speaker 2: And so the more we examine the primate society, we

514
00:29:21,400 --> 00:29:25,520
Speaker 2: see some interesting features primate societies are very political, so

515
00:29:25,800 --> 00:29:31,760
Speaker 2: unlike a troop of gazelles and a troop of gazelle's,

516
00:29:32,000 --> 00:29:35,280
Speaker 2: whoever is the top ranking gazelle is typically the one

517
00:29:35,280 --> 00:29:38,640
Speaker 2: that's the strongest. So there's very explicit hierarchies in many

518
00:29:38,680 --> 00:29:41,360
Speaker 2: mammal groupings, but they're based on who's the toughest and

519
00:29:41,400 --> 00:29:44,200
Speaker 2: the strongest. But if you look at primate societies, it's

520
00:29:44,200 --> 00:29:47,360
Speaker 2: typically not the strongest. It's the most socially savvy one.

521
00:29:47,680 --> 00:29:49,920
Speaker 2: It's the one that cozies up to the most allies,

522
00:29:50,120 --> 00:29:53,080
Speaker 2: it's the one that builds the most friendships, that build

523
00:29:53,160 --> 00:29:56,040
Speaker 2: sort of this political regime that enables them to be

524
00:29:56,160 --> 00:30:00,320
Speaker 2: the top ranking chimpanzee, their top ranking bnobo. So there's

525
00:30:00,320 --> 00:30:03,360
Speaker 2: been some also amazing studies of the ways in which

526
00:30:03,480 --> 00:30:08,280
Speaker 2: these apes and monkeys reason about other people's mind states

527
00:30:08,280 --> 00:30:11,240
Speaker 2: when making choices on how to befriend them or how

528
00:30:11,240 --> 00:30:14,800
Speaker 2: to deceive them. So you can see non human apes

529
00:30:14,840 --> 00:30:18,280
Speaker 2: do things like they will hide transgressions from other people

530
00:30:18,320 --> 00:30:20,800
Speaker 2: to try and prevent themselves from getting in trouble. There's

531
00:30:20,800 --> 00:30:24,760
Speaker 2: this famous study that I love by Emil Menzel. I

532
00:30:24,760 --> 00:30:27,720
Speaker 2: think it was in the seventies where he put two

533
00:30:27,760 --> 00:30:30,640
Speaker 2: chimpanzees in the sort of one acre forest, and he

534
00:30:30,800 --> 00:30:34,680
Speaker 2: showed the location of treats to one of the chimpanzees

535
00:30:34,720 --> 00:30:39,080
Speaker 2: named Belle, and she initially would share the treat with

536
00:30:39,160 --> 00:30:42,840
Speaker 2: another chimpanzee named Rock, but then Rock started stealing the

537
00:30:42,880 --> 00:30:45,920
Speaker 2: treat from her. So what she started doing is, when

538
00:30:45,960 --> 00:30:48,200
Speaker 2: she knew the location of the treat, she would wait

539
00:30:48,280 --> 00:30:50,000
Speaker 2: for a rock to look away, and then she would

540
00:30:50,080 --> 00:30:53,320
Speaker 2: run over and grab it. So then Rock, in response

541
00:30:53,360 --> 00:30:56,680
Speaker 2: to this, decided to pretend to look away so that

542
00:30:56,760 --> 00:30:58,880
Speaker 2: when she started running, then he would turn around and run.

543
00:30:59,440 --> 00:31:01,640
Speaker 2: Then respect to this, what she would do is she

544
00:31:01,640 --> 00:31:04,320
Speaker 2: would pretend to run in the wrong direction, lead him

545
00:31:04,320 --> 00:31:06,400
Speaker 2: to the wrong place, and then run back. And so

546
00:31:06,520 --> 00:31:10,720
Speaker 2: this cycle of deception and counter deeception is very very unique,

547
00:31:10,760 --> 00:31:13,520
Speaker 2: with impossible exceptions of a few very very smart non

548
00:31:13,560 --> 00:31:17,400
Speaker 2: primate mammals like dolphins, seems to be unique to primates,

549
00:31:17,480 --> 00:31:20,040
Speaker 2: and so this gives us a clue as to what

550
00:31:20,160 --> 00:31:23,240
Speaker 2: might be new in the brains of primates. When we

551
00:31:23,280 --> 00:31:26,880
Speaker 2: go into the primate brain, we see these suite of

552
00:31:26,960 --> 00:31:29,920
Speaker 2: new neocortical regions sort of The most sizable one is

553
00:31:29,960 --> 00:31:31,920
Speaker 2: something in the front of the brain called the granular

554
00:31:32,000 --> 00:31:37,120
Speaker 2: prefrontal cortex, and when we do sort of neuroscience to

555
00:31:37,120 --> 00:31:40,240
Speaker 2: try and understand what does the structure do, it lights

556
00:31:40,320 --> 00:31:42,960
Speaker 2: up a ton when we reason about our own mind,

557
00:31:43,320 --> 00:31:45,360
Speaker 2: so how we would feel in certain states, or we

558
00:31:45,440 --> 00:31:49,200
Speaker 2: reason about other people's minds. So in tests of what's

559
00:31:49,240 --> 00:31:52,480
Speaker 2: called theory of mind, when I need to guess what

560
00:31:52,480 --> 00:31:54,920
Speaker 2: someone else is thinking about, or what their intention is,

561
00:31:55,040 --> 00:31:56,640
Speaker 2: or what knowledge they might have, this part of the

562
00:31:56,680 --> 00:31:59,720
Speaker 2: brain lights up a ton. And they've done some cool

563
00:31:59,720 --> 00:32:02,360
Speaker 2: study on macaque monkeys that show that in order for

564
00:32:02,400 --> 00:32:04,960
Speaker 2: a monkey to make a correct assessment of what someone

565
00:32:04,960 --> 00:32:07,120
Speaker 2: else knows or doesn't know, they need this part of

566
00:32:07,160 --> 00:32:09,720
Speaker 2: their brain active. If you temporarily inhibit it, they lose

567
00:32:09,720 --> 00:32:12,880
Speaker 2: their ability to reason about other people's minds. So you

568
00:32:12,920 --> 00:32:15,360
Speaker 2: get theory of mind. And so the idea is break

569
00:32:15,400 --> 00:32:19,480
Speaker 2: through four is mentalizing, which is also called metacognition, thinking

570
00:32:19,520 --> 00:32:23,240
Speaker 2: about thinking, reasoning about your own mind and other people's minds.

571
00:32:23,480 --> 00:32:26,040
Speaker 2: But there's two unique things about primates that are not

572
00:32:26,160 --> 00:32:30,400
Speaker 2: classically thought about as being related to mentalizing that I

573
00:32:30,440 --> 00:32:33,440
Speaker 2: would argue are are only possible in prime it's because

574
00:32:33,480 --> 00:32:37,080
Speaker 2: of mentalizing. One is imitation learning we know that primates

575
00:32:37,120 --> 00:32:39,840
Speaker 2: are exceptionally good imitation learners. So if you take a

576
00:32:39,920 --> 00:32:42,440
Speaker 2: chimpanzee out of their group and teach them how to

577
00:32:42,520 --> 00:32:45,960
Speaker 2: open a puzzle box or do some clever motor skill,

578
00:32:46,320 --> 00:32:48,840
Speaker 2: and then you release them back into their troop, within

579
00:32:48,920 --> 00:32:50,880
Speaker 2: thirty to sixty days, the whole troop will know the

580
00:32:50,880 --> 00:32:54,800
Speaker 2: same exact skill. So chimpanzees are very good at learning

581
00:32:54,840 --> 00:32:58,240
Speaker 2: skills through observation. This is part of why apes are

582
00:32:58,280 --> 00:33:01,000
Speaker 2: such good tool users, because once one member learns how

583
00:33:01,000 --> 00:33:03,640
Speaker 2: to use a tool, they all adopt the skill, and

584
00:33:03,680 --> 00:33:07,320
Speaker 2: then they cascade it through generations. In AI, we have

585
00:33:07,880 --> 00:33:11,440
Speaker 2: tried to teach systems through imitation. We've discovered something really interesting.

586
00:33:12,520 --> 00:33:15,880
Speaker 2: We've learned that direct imitation of other people's actions does

587
00:33:15,920 --> 00:33:18,600
Speaker 2: not work. So we've tried this in self driving cars,

588
00:33:18,680 --> 00:33:21,480
Speaker 2: where we try to teach an AI system to drive

589
00:33:21,520 --> 00:33:24,520
Speaker 2: a car by watching a human drive a car. And

590
00:33:24,600 --> 00:33:28,560
Speaker 2: the reason it fails is because when you watch an expert,

591
00:33:28,600 --> 00:33:31,520
Speaker 2: you never see the expert recover from mistakes. So the

592
00:33:31,560 --> 00:33:34,680
Speaker 2: second this AI system started veering off the road, nothing

593
00:33:34,720 --> 00:33:37,160
Speaker 2: in its training set taught it how to recover from

594
00:33:37,240 --> 00:33:39,000
Speaker 2: veering off the road, because it only watched from an

595
00:33:39,000 --> 00:33:41,440
Speaker 2: expert of who never veered off the road. The way

596
00:33:41,480 --> 00:33:44,160
Speaker 2: we get this to work in AI systems, which was

597
00:33:45,240 --> 00:33:49,280
Speaker 2: most famously invented by Andrew Aang, it's called inverse reinforcement learning.

598
00:33:49,720 --> 00:33:51,920
Speaker 2: And so what you do is you first try to

599
00:33:52,000 --> 00:33:55,680
Speaker 2: infer what the person you're imitating is trying to do.

600
00:33:55,720 --> 00:33:58,920
Speaker 2: You infer their reward function. So if you watch someone drive,

601
00:33:59,000 --> 00:34:01,880
Speaker 2: you say, oh, they're trying to stay in the center

602
00:34:01,920 --> 00:34:04,520
Speaker 2: of the road, and then I train myself in my

603
00:34:04,600 --> 00:34:06,800
Speaker 2: mind's eye to do the same thing that they're trying

604
00:34:06,800 --> 00:34:11,040
Speaker 2: to do, and that works. So Entering in the early

605
00:34:11,040 --> 00:34:14,520
Speaker 2: two thousands trained a helicopter to do all these crazy

606
00:34:14,560 --> 00:34:18,760
Speaker 2: aerobatic tricks through watching other trained experts do those tricks,

607
00:34:18,800 --> 00:34:21,600
Speaker 2: but not by directly copying them, by first inferring what

608
00:34:21,600 --> 00:34:24,759
Speaker 2: they're trying to do, and so it eliminates all the

609
00:34:24,800 --> 00:34:29,279
Speaker 2: extraneous behaviors. This is part of why imitation learning requires mentalizing,

610
00:34:29,760 --> 00:34:32,439
Speaker 2: because in order for me to really understand what you're

611
00:34:32,480 --> 00:34:36,000
Speaker 2: trying to do with certain tool usage behaviors, I need

612
00:34:36,040 --> 00:34:38,480
Speaker 2: to reason about your mind and infer what your intent is.

613
00:34:38,600 --> 00:34:41,680
Speaker 2: And that's part of why I would argue that primates

614
00:34:41,680 --> 00:34:44,560
Speaker 2: are so good at imitation learning, they repurpose. It's mentalizing

615
00:34:44,560 --> 00:34:49,440
Speaker 2: for that. The last one is something called anticipating future needs.

616
00:34:50,040 --> 00:34:53,200
Speaker 2: So when we go grocery shopping for the week, we're

617
00:34:53,239 --> 00:34:57,160
Speaker 2: actually doing something really remarkable. We are taking an action

618
00:34:57,239 --> 00:35:00,239
Speaker 2: today to satiate a need that we do not currently have.

619
00:35:00,600 --> 00:35:02,480
Speaker 2: I might not be hungry, and yet I'm going to

620
00:35:02,480 --> 00:35:04,440
Speaker 2: take an hour out of my day to fill up

621
00:35:04,520 --> 00:35:08,320
Speaker 2: my refrigerator. And it's not so clear how many animals

622
00:35:08,360 --> 00:35:12,160
Speaker 2: are capable of doing that. So, for example, in mice,

623
00:35:12,600 --> 00:35:15,520
Speaker 2: you see hoarding behavior before the winter, but we now

624
00:35:15,560 --> 00:35:18,280
Speaker 2: know that that is genetically hard coded. They're not mentally

625
00:35:18,320 --> 00:35:21,279
Speaker 2: imagining the winter and realizing they'll be hungry. A rat

626
00:35:21,320 --> 00:35:24,040
Speaker 2: that is, or a mouse who has never experienced hunger

627
00:35:24,040 --> 00:35:26,520
Speaker 2: in the winter, never even experienced a winter at all.

628
00:35:26,960 --> 00:35:29,400
Speaker 2: If you turn down the temperature, we'll start hoarding. But

629
00:35:29,960 --> 00:35:33,480
Speaker 2: primates seem to be capable of doing this, So they've

630
00:35:33,520 --> 00:35:36,000
Speaker 2: done some fun studies on squirrel monkeys that show that

631
00:35:36,080 --> 00:35:39,920
Speaker 2: they will actually choose having less treats today to reduce

632
00:35:40,000 --> 00:35:42,840
Speaker 2: their future thirsts even when they're not thirsty today, whereas

633
00:35:43,040 --> 00:35:45,879
Speaker 2: a rat is incapable of doing that, And so this guy.

634
00:35:45,960 --> 00:35:48,879
Speaker 2: Tom and Sudendorff came up with this theory that maybe

635
00:35:48,960 --> 00:35:53,560
Speaker 2: anticipating our own future needs uses the same machinery in

636
00:35:53,560 --> 00:35:56,800
Speaker 2: our brains as reasoning about other minds, because if you

637
00:35:56,840 --> 00:35:59,280
Speaker 2: think about it, it's really the same thing. For me to ask,

638
00:35:59,520 --> 00:36:02,680
Speaker 2: what will David feel like if he didn't drink for

639
00:36:02,920 --> 00:36:05,880
Speaker 2: a week is really the same question as what I

640
00:36:05,920 --> 00:36:08,640
Speaker 2: feel like if I didn't drink for a week, And

641
00:36:08,719 --> 00:36:12,480
Speaker 2: so this might also explain why apes and other primates

642
00:36:12,520 --> 00:36:14,960
Speaker 2: are so good at anticipating their own future needs and

643
00:36:15,000 --> 00:36:18,640
Speaker 2: making these really long term plans. So breakthrough FORO is mentalizing.

644
00:36:19,120 --> 00:36:21,920
Speaker 2: It is the building a sort of model of your

645
00:36:21,920 --> 00:36:24,439
Speaker 2: own inner mind, and it enables you to reason about

646
00:36:24,480 --> 00:36:27,840
Speaker 2: other minds. It enables you to learn through imitation, and

647
00:36:27,880 --> 00:36:30,440
Speaker 2: it allows you to anticipate your own future needs.

648
00:36:31,000 --> 00:36:34,279
Speaker 1: Great tell us about the final breakthrough that led to

649
00:36:34,400 --> 00:36:36,720
Speaker 1: the kind of intelligence that we enjoy.

650
00:36:37,560 --> 00:36:41,959
Speaker 2: So there's been throughout the ages so many thinkers, philosophers,

651
00:36:41,960 --> 00:36:45,360
Speaker 2: and scientists have tried to draw a hard line between

652
00:36:45,440 --> 00:36:48,200
Speaker 2: humans and other animals and articulate what is the thing

653
00:36:48,239 --> 00:36:51,839
Speaker 2: that makes humans unique? And after writing this book, one

654
00:36:51,880 --> 00:36:55,560
Speaker 2: of the most like clear things to me is how

655
00:36:55,840 --> 00:36:58,759
Speaker 2: little difference there really is between us and other animals.

656
00:36:59,280 --> 00:37:02,239
Speaker 2: So people used to think only humans could imagine things.

657
00:37:02,280 --> 00:37:04,880
Speaker 2: I think the evidence is very strong that other mammals

658
00:37:05,239 --> 00:37:09,160
Speaker 2: and probably birds regularly have imagination. Some people thought only

659
00:37:09,239 --> 00:37:12,960
Speaker 2: humans think about thinking. I think there's pretty good evidence

660
00:37:13,000 --> 00:37:15,719
Speaker 2: that other primates do the same, and so there's been

661
00:37:15,719 --> 00:37:18,919
Speaker 2: this long laundry list of stuff. I think the main

662
00:37:19,239 --> 00:37:23,840
Speaker 2: feature of human intelligence that there is this good evidence

663
00:37:23,920 --> 00:37:27,279
Speaker 2: is uniquely human, or at least uniquely evolved in the

664
00:37:27,360 --> 00:37:31,000
Speaker 2: human lineage and was not present in other primates is language.

665
00:37:32,000 --> 00:37:33,880
Speaker 2: And so it's language is not the same thing as communication.

666
00:37:34,120 --> 00:37:37,840
Speaker 2: Even single celled organisms engage in communication, but language is

667
00:37:37,960 --> 00:37:42,080
Speaker 2: unique on two counts. Human language has what's called declarative labels.

668
00:37:42,640 --> 00:37:45,719
Speaker 2: It allows us to assign an arbitrary symbol to a

669
00:37:45,800 --> 00:37:49,120
Speaker 2: thing or an action in the world. So when you

670
00:37:49,120 --> 00:37:52,000
Speaker 2: tell a dog to sit, now what it's learning is

671
00:37:52,080 --> 00:37:54,440
Speaker 2: when I hear the symbol sit, if I take this

672
00:37:54,520 --> 00:37:57,319
Speaker 2: action sit, I get a reward. That's something linguists call

673
00:37:57,640 --> 00:38:01,560
Speaker 2: imperative labels. A declarative label is if I say sit,

674
00:38:02,080 --> 00:38:06,000
Speaker 2: we're all imagining the action of sitting. And it's not

675
00:38:06,200 --> 00:38:09,279
Speaker 2: clear that other animals are capable of these types of

676
00:38:09,320 --> 00:38:14,239
Speaker 2: declarative labels. There's been painstaking attempts to train non human primates,

677
00:38:14,320 --> 00:38:18,279
Speaker 2: specifically apes, to use language. Typically it's sign language because

678
00:38:18,320 --> 00:38:20,640
Speaker 2: they don't actually have the sort of vocal apparatus for

679
00:38:20,760 --> 00:38:24,000
Speaker 2: verbal language, And it's still controversial the extent to which

680
00:38:24,239 --> 00:38:27,760
Speaker 2: what they were able to do could be called language.

681
00:38:28,080 --> 00:38:30,280
Speaker 2: But even if you would classify it as a primitive

682
00:38:30,320 --> 00:38:33,640
Speaker 2: form of language, it's very clear that non human apes

683
00:38:33,880 --> 00:38:37,480
Speaker 2: are not nearly as good at learning languages as human children.

684
00:38:38,080 --> 00:38:42,160
Speaker 2: The second thing that's unique about human language is grammar.

685
00:38:42,680 --> 00:38:46,759
Speaker 2: So we can switch the ordering of these symbols to

686
00:38:46,880 --> 00:38:52,000
Speaker 2: change their meaning in seemingly arbitrary ways. So Max jumped

687
00:38:52,040 --> 00:38:56,200
Speaker 2: over Charlie means something different than Charlie jumped over Max,

688
00:38:56,239 --> 00:38:59,560
Speaker 2: and by ordering the symbols, the meaning totally shifts. And

689
00:38:59,760 --> 00:39:03,319
Speaker 2: so one might think, okay, language is this unique thing,

690
00:39:03,920 --> 00:39:06,080
Speaker 2: that there'd be some unique structures in the human brain

691
00:39:06,160 --> 00:39:09,880
Speaker 2: that enabled language, and to my surprise, also looking to

692
00:39:09,920 --> 00:39:12,800
Speaker 2: the neuroscience, that's not at all the case. So there

693
00:39:12,880 --> 00:39:17,040
Speaker 2: are two regions of the neocortex and humans that are

694
00:39:17,280 --> 00:39:21,480
Speaker 2: very implicated in language, famously called Wernicke's area and Broker's area.

695
00:39:22,160 --> 00:39:26,800
Speaker 2: But interestingly, those same exact neocortical regions exist in other primates,

696
00:39:27,080 --> 00:39:30,360
Speaker 2: they're just not used in communication. So for some reason,

697
00:39:30,480 --> 00:39:34,000
Speaker 2: it wasn't that some new structure emerged in the human brain.

698
00:39:34,360 --> 00:39:38,880
Speaker 2: It's that we repurpose an existing structure to use in language.

699
00:39:39,320 --> 00:39:42,680
Speaker 2: And what seems to have happened is a new learning

700
00:39:42,680 --> 00:39:48,000
Speaker 2: curriculum evolved in humans that enabled us to learn language.

701
00:39:48,040 --> 00:39:51,600
Speaker 2: And so if we compare chimpanzee children to human children,

702
00:39:52,000 --> 00:39:55,440
Speaker 2: there's two very unique traits of human children. One is

703
00:39:55,480 --> 00:39:58,319
Speaker 2: they engage in something called joint attention at a very

704
00:39:58,400 --> 00:40:01,960
Speaker 2: very young preverbal age, which means children get a unique

705
00:40:02,239 --> 00:40:05,680
Speaker 2: burst of excitement when they can confirm by looking at

706
00:40:05,719 --> 00:40:08,239
Speaker 2: your eyes that we are that they and you are

707
00:40:08,239 --> 00:40:11,080
Speaker 2: attending to the same object. So they've done lots of

708
00:40:11,160 --> 00:40:14,200
Speaker 2: painstaking studies to show that the child is not excited

709
00:40:14,200 --> 00:40:16,359
Speaker 2: because they think they're going to get the object. They're

710
00:40:16,400 --> 00:40:19,919
Speaker 2: not excited because the parent is excited. They are specifically

711
00:40:19,960 --> 00:40:22,680
Speaker 2: happy and satisfied when they confirm that they are looking

712
00:40:22,680 --> 00:40:24,960
Speaker 2: at the same object that the parent is looking at.

713
00:40:25,280 --> 00:40:27,680
Speaker 2: And what does this enable us to do? This enables

714
00:40:27,760 --> 00:40:29,960
Speaker 2: us to render a simulation of the same object in

715
00:40:30,000 --> 00:40:32,160
Speaker 2: our head, so we can assign a symbol to it.

716
00:40:32,520 --> 00:40:34,440
Speaker 2: If we all look at a cat and I confirm

717
00:40:34,480 --> 00:40:36,400
Speaker 2: you're looking at a cat, and then the parent says

718
00:40:36,440 --> 00:40:39,440
Speaker 2: the symbol cat, whether it's verbal or a sign or

719
00:40:39,480 --> 00:40:43,520
Speaker 2: a written word, it creates this sort of basic foundation

720
00:40:44,040 --> 00:40:47,440
Speaker 2: for labels to be constructed. And the other thing that's

721
00:40:47,520 --> 00:40:50,960
Speaker 2: unique in human children is proto conversation. So they've shown

722
00:40:50,960 --> 00:40:53,759
Speaker 2: that very young human infants will match the duration of

723
00:40:53,840 --> 00:40:57,960
Speaker 2: babbling before words with their parents. So if the parent

724
00:40:58,000 --> 00:41:00,360
Speaker 2: babbles for four seconds, the child tends to bet for

725
00:41:00,400 --> 00:41:02,600
Speaker 2: four seconds and then pause and wait for the parent

726
00:41:02,680 --> 00:41:06,400
Speaker 2: to do that. These two things are not naturally occurring

727
00:41:06,520 --> 00:41:09,080
Speaker 2: in non human primates, so it's very hard to get

728
00:41:09,160 --> 00:41:11,719
Speaker 2: a chimpanzee to attend to the same object and for

729
00:41:11,800 --> 00:41:14,560
Speaker 2: them to confirm that we're all attending to the same thing. Okay,

730
00:41:14,600 --> 00:41:17,280
Speaker 2: so we get language, But why does language make humans

731
00:41:17,280 --> 00:41:21,000
Speaker 2: so special? So this has been well discussed in linguistics

732
00:41:21,000 --> 00:41:23,800
Speaker 2: in Uvall's books Sapiens, I think he speaks to a

733
00:41:23,840 --> 00:41:28,200
Speaker 2: lot of this. What makes language so incredible? This enables

734
00:41:28,280 --> 00:41:32,160
Speaker 2: us to share our inner simulations, and so it transforms

735
00:41:32,160 --> 00:41:34,920
Speaker 2: the human brain from just sort of the epicenter of

736
00:41:34,960 --> 00:41:38,319
Speaker 2: intelligence to being the medium through which ideas can flow

737
00:41:38,360 --> 00:41:42,319
Speaker 2: through time. So because I can share what's going on

738
00:41:42,400 --> 00:41:45,759
Speaker 2: in my mind, culture canform or a more advanced form

739
00:41:45,800 --> 00:41:48,920
Speaker 2: of culture because I can learn certain skills and then

740
00:41:49,040 --> 00:41:52,040
Speaker 2: describe the skill to you, or the five of us

741
00:41:52,040 --> 00:41:54,960
Speaker 2: can go on a hunt together, and I can imagine

742
00:41:54,960 --> 00:41:57,520
Speaker 2: a plan and then share the plan in my mind

743
00:41:57,520 --> 00:41:59,640
Speaker 2: with you through symbols, and then we all have the

744
00:41:59,640 --> 00:42:01,839
Speaker 2: same plan in in our minds, and then we can

745
00:42:01,840 --> 00:42:04,400
Speaker 2: coordinate and do the same thing together. Without the ability

746
00:42:04,440 --> 00:42:07,440
Speaker 2: to share inner simulations, you don't get this type of flexibility.

747
00:42:07,719 --> 00:42:10,520
Speaker 2: So that's one of the fundamental things that enables language

748
00:42:10,560 --> 00:42:14,080
Speaker 2: to make humans so powerful, because as generations go on,

749
00:42:14,280 --> 00:42:17,000
Speaker 2: the ideas sort of ratchet up and get more and

750
00:42:17,000 --> 00:42:21,920
Speaker 2: more complex over time, versus in chimpanzee societies. Because they

751
00:42:21,920 --> 00:42:25,879
Speaker 2: can't reliably share ideas, they can only observe through learn

752
00:42:25,960 --> 00:42:28,719
Speaker 2: from each other through observation. There's a limit to how

753
00:42:28,760 --> 00:42:31,239
Speaker 2: complex these ideas can get over generations. And so that's

754
00:42:31,239 --> 00:42:33,640
Speaker 2: one of the leading theories, not my theory. Lots of

755
00:42:33,719 --> 00:42:37,760
Speaker 2: linguists and primatologists talk about this as to why humans

756
00:42:37,920 --> 00:42:40,200
Speaker 2: sort of took over the world, which is ideas got

757
00:42:40,200 --> 00:42:42,440
Speaker 2: to get more complex over time until they reach this

758
00:42:42,520 --> 00:42:45,839
Speaker 2: sort of critical point. And so break through five was

759
00:42:46,160 --> 00:42:49,560
Speaker 2: speaking or language. And the last point I'll make on

760
00:42:49,600 --> 00:42:52,960
Speaker 2: this is how you one can see how even speaking

761
00:42:52,960 --> 00:42:57,600
Speaker 2: in language is dependent on the prior breakthroughs. So as

762
00:42:57,640 --> 00:43:00,640
Speaker 2: we now know in AI systems, when the leading problems

763
00:43:00,640 --> 00:43:04,319
Speaker 2: with an AI system bound by just language is how

764
00:43:04,360 --> 00:43:07,719
Speaker 2: hard it is to actually describe our desires in the

765
00:43:07,760 --> 00:43:12,279
Speaker 2: form of language. So Nick Bostrom has this really great

766
00:43:12,320 --> 00:43:16,480
Speaker 2: allegory where suppose there is an AI that manages a

767
00:43:16,480 --> 00:43:20,360
Speaker 2: paper clip factory, a super intelligent AI, and the instruction

768
00:43:20,840 --> 00:43:24,040
Speaker 2: US humans give that AI is maximize paper clip production.

769
00:43:24,239 --> 00:43:26,959
Speaker 2: That's the we give that a natural language, maximize paper

770
00:43:26,960 --> 00:43:30,480
Speaker 2: clip production. In his allegory, what he imagines if the

771
00:43:30,520 --> 00:43:34,520
Speaker 2: superintelligent AI were actually to just optimize for the explicit

772
00:43:34,560 --> 00:43:38,160
Speaker 2: request it was given, it would start to take over

773
00:43:38,320 --> 00:43:42,600
Speaker 2: Earth and convert everything it could observe into paper clips.

774
00:43:42,680 --> 00:43:44,600
Speaker 2: And when it was done with Earth, it would expand

775
00:43:44,600 --> 00:43:46,279
Speaker 2: to Mars and it would start to try and take

776
00:43:46,280 --> 00:43:48,760
Speaker 2: over the universe to convert all of it into paper clips.

777
00:43:49,280 --> 00:43:52,640
Speaker 2: And as silly as that example is, as almost nonsensical

778
00:43:52,800 --> 00:43:57,000
Speaker 2: as it seems, it reveals why mentalizing is required for

779
00:43:57,120 --> 00:44:00,520
Speaker 2: language to work. Because when you tell a human maximize

780
00:44:00,560 --> 00:44:03,400
Speaker 2: production of paper clips, what a human is doing is

781
00:44:03,400 --> 00:44:06,040
Speaker 2: it's they're inferring what you actually mean by what you say.

782
00:44:06,640 --> 00:44:09,880
Speaker 2: I'm simulating your mind and I'm trying to infer your preferences,

783
00:44:09,920 --> 00:44:12,879
Speaker 2: and I'm doing this really complex inference task to take

784
00:44:12,880 --> 00:44:15,160
Speaker 2: the symbols that you gave me and convert it into

785
00:44:15,480 --> 00:44:18,040
Speaker 2: a really complex reward function that I'm going to try

786
00:44:18,040 --> 00:44:21,040
Speaker 2: and optimize for. But if all system does is take

787
00:44:21,080 --> 00:44:22,759
Speaker 2: our words for what we say them to be and

788
00:44:22,840 --> 00:44:25,359
Speaker 2: doesn't have a model of our minds, then you can

789
00:44:25,360 --> 00:44:29,040
Speaker 2: get these really wacky outcomes where they would try and

790
00:44:29,040 --> 00:44:32,600
Speaker 2: convert Earth into paper clips. And so the reason why

791
00:44:32,880 --> 00:44:36,160
Speaker 2: language requires mentalizing is when we're going back and forth

792
00:44:36,200 --> 00:44:38,560
Speaker 2: trading symbols all the time, we're trying to guess what

793
00:44:38,600 --> 00:44:40,880
Speaker 2: the other person means by what they say. We're trying

794
00:44:40,920 --> 00:44:43,920
Speaker 2: to tell them information to update their knowledge given what

795
00:44:43,960 --> 00:44:46,319
Speaker 2: we know they know and they don't know. It's so

796
00:44:46,480 --> 00:44:48,400
Speaker 2: natural for us we don't realize it. But this is

797
00:44:48,440 --> 00:44:50,520
Speaker 2: one of the key things that human brains are so

798
00:44:50,600 --> 00:44:54,239
Speaker 2: good at that. Aisystems, at least in the same way,

799
00:44:54,320 --> 00:44:54,960
Speaker 2: don't solve.

800
00:45:11,239 --> 00:45:13,400
Speaker 1: You know. One of the things that always has amazed

801
00:45:13,440 --> 00:45:16,600
Speaker 1: me is the existence of literature. The thing I hadn't

802
00:45:16,600 --> 00:45:20,400
Speaker 1: realized until I thought about it was how low bandwidth

803
00:45:20,400 --> 00:45:24,360
Speaker 1: literature is. The author tells you a few sentences about

804
00:45:24,360 --> 00:45:27,120
Speaker 1: this and that, the description and the emotions and all

805
00:45:27,200 --> 00:45:29,920
Speaker 1: the rest depends on the reader. The reader is bringing

806
00:45:30,040 --> 00:45:34,160
Speaker 1: everything to the table. The author can't put what he's

807
00:45:34,200 --> 00:45:37,880
Speaker 1: imagining directly into the mind of the reader because every

808
00:45:37,920 --> 00:45:42,239
Speaker 1: reader is going to imagine something differently predicated totally on

809
00:45:42,280 --> 00:45:45,920
Speaker 1: this issue that you know, it's all about mentalizing and

810
00:45:46,040 --> 00:45:49,400
Speaker 1: language is just a very few bits of information that

811
00:45:49,840 --> 00:45:54,319
Speaker 1: you know, get thrown over the transom to inspire something

812
00:45:54,320 --> 00:45:55,520
Speaker 1: in someone else's mind.

813
00:45:55,680 --> 00:45:57,400
Speaker 2: One hundred percent. I think one thing that just to

814
00:45:57,400 --> 00:45:59,640
Speaker 2: add to that I think is really cool is it

815
00:45:59,640 --> 00:46:04,120
Speaker 2: almost is a neuroscience or AI perspective on why many

816
00:46:04,239 --> 00:46:07,799
Speaker 2: artists talk about how art is an active process. In

817
00:46:08,000 --> 00:46:12,000
Speaker 2: the sort of consumer of art, when we read a book,

818
00:46:12,640 --> 00:46:16,399
Speaker 2: we are participating in that artistic creation because we are

819
00:46:16,400 --> 00:46:19,560
Speaker 2: filling in the gaps. And that's why people can interpret

820
00:46:19,640 --> 00:46:22,640
Speaker 2: art so differently, and in some ways that's why art

821
00:46:23,120 --> 00:46:26,920
Speaker 2: is so beautiful, because it's this like message, but it's

822
00:46:27,000 --> 00:46:31,680
Speaker 2: not fixed. We as consumers get to sort of explore

823
00:46:31,680 --> 00:46:33,640
Speaker 2: it in our own way. I think it's also in

824
00:46:33,680 --> 00:46:37,080
Speaker 2: some ways why reading feels harder than watching a movie

825
00:46:37,200 --> 00:46:39,120
Speaker 2: because you don't realize it, but your mind is doing

826
00:46:39,120 --> 00:46:42,360
Speaker 2: a lot of work when you read, because it's turning

827
00:46:42,360 --> 00:46:44,920
Speaker 2: what you read into a mental movie, and that translation

828
00:46:45,080 --> 00:46:48,239
Speaker 2: takes effort versus watching a movie requires less sort of

829
00:46:48,280 --> 00:46:49,080
Speaker 2: cognitive overlook.

830
00:46:49,560 --> 00:46:53,480
Speaker 1: Now returning to the primates and the humans. So one

831
00:46:53,520 --> 00:46:55,319
Speaker 1: of the things that people have pointed out is that

832
00:46:55,400 --> 00:46:59,880
Speaker 1: humans are the only species that teach. So a prime,

833
00:47:00,080 --> 00:47:03,960
Speaker 1: a young primate will watch his mother, you know, crushing

834
00:47:04,080 --> 00:47:07,239
Speaker 1: rocks and doing something, and the primate will imitate that.

835
00:47:07,719 --> 00:47:10,799
Speaker 1: But the mother never gives feedback. The mother never says, oh,

836
00:47:10,840 --> 00:47:13,680
Speaker 1: you're doing it wrong, do it this way, and grabs

837
00:47:13,680 --> 00:47:16,160
Speaker 1: his hands and does the right way. But humans do

838
00:47:16,200 --> 00:47:18,799
Speaker 1: that all the time. We actually teach, and that's something

839
00:47:18,920 --> 00:47:22,160
Speaker 1: unique to our species. What is the basis of that?

840
00:47:22,640 --> 00:47:25,160
Speaker 2: I would argue in my framework, I would argue the

841
00:47:25,160 --> 00:47:29,640
Speaker 2: basic machinery for teaching exists in mentalizing, but it teaching

842
00:47:29,719 --> 00:47:33,160
Speaker 2: might be such a complex version of mentalizing because it's

843
00:47:33,200 --> 00:47:35,399
Speaker 2: two steps. Not only do I need to render what's

844
00:47:35,400 --> 00:47:37,560
Speaker 2: in your mind, but then I need to be able

845
00:47:37,600 --> 00:47:40,040
Speaker 2: to think about what actions can I take to update

846
00:47:40,080 --> 00:47:42,759
Speaker 2: something in your mind. You know, that's a complex act.

847
00:47:43,080 --> 00:47:46,040
Speaker 2: So I think even if the machinery exists in mentalizing,

848
00:47:46,040 --> 00:47:47,960
Speaker 2: when you scale up the brain, I mean, the human

849
00:47:48,000 --> 00:47:50,160
Speaker 2: brain is about you know, three x bigger than a

850
00:47:50,200 --> 00:47:54,040
Speaker 2: chimpanzee brain, or then the cortex area, you start getting

851
00:47:54,080 --> 00:47:57,120
Speaker 2: some of the machinery that's there in a very lightweight,

852
00:47:57,239 --> 00:48:00,439
Speaker 2: primitive form. So I think in my frame, I would

853
00:48:00,480 --> 00:48:03,879
Speaker 2: argue that some very primitive version of teaching exists in mentalizing,

854
00:48:03,920 --> 00:48:07,400
Speaker 2: but it doesn't really get rendered more effective and so

855
00:48:07,480 --> 00:48:08,880
Speaker 2: it scales up in human brains.

856
00:48:09,160 --> 00:48:11,680
Speaker 1: Okay, So that puts us at today, and what we

857
00:48:11,760 --> 00:48:16,120
Speaker 1: have today is this incredible explosion of AI, which is

858
00:48:16,160 --> 00:48:21,920
Speaker 1: something that you know, my whole career in neuroscience, neuroscientists

859
00:48:21,960 --> 00:48:24,120
Speaker 1: generally looked at AI and said, well, it's you know,

860
00:48:24,160 --> 00:48:26,960
Speaker 1: it's not very good. It's not able to do X,

861
00:48:27,080 --> 00:48:29,640
Speaker 1: y Z. But we've all been surprised in the last

862
00:48:29,680 --> 00:48:32,200
Speaker 1: few years about what it is able to do. The

863
00:48:32,320 --> 00:48:36,840
Speaker 1: interesting thing is still the stuff that it's not able

864
00:48:36,880 --> 00:48:41,000
Speaker 1: to do and why. So let's talk about AI. Tell

865
00:48:41,040 --> 00:48:43,680
Speaker 1: me your take on where it is currently and what

866
00:48:43,760 --> 00:48:47,160
Speaker 1: all of your study about the history of intelligence tells us.

867
00:48:47,760 --> 00:48:51,920
Speaker 2: So one thing that's interesting is AI today, and this

868
00:48:52,120 --> 00:48:55,120
Speaker 2: moment seems to be almost taking the exact opposite path

869
00:48:55,320 --> 00:48:58,360
Speaker 2: as our brains. It's starting from language, at least the

870
00:48:58,840 --> 00:49:02,120
Speaker 2: sort of explosion general AI has at its foundation been

871
00:49:02,239 --> 00:49:05,200
Speaker 2: language models been these things called transformers that are trained

872
00:49:05,239 --> 00:49:08,200
Speaker 2: on huge amounts of language text. And what has been

873
00:49:08,239 --> 00:49:12,160
Speaker 2: surprising is the degree with which language seems to be

874
00:49:12,239 --> 00:49:15,560
Speaker 2: so informationally rich that from going from the top of

875
00:49:15,600 --> 00:49:18,600
Speaker 2: this pyramid of the five breakthroughs, you actually can start

876
00:49:18,640 --> 00:49:22,800
Speaker 2: going down. So if you ask a large language model

877
00:49:23,000 --> 00:49:26,200
Speaker 2: questions that require theory of mind, which just to remind

878
00:49:26,239 --> 00:49:29,280
Speaker 2: the listeners, is being able to reason about other people's

879
00:49:29,320 --> 00:49:33,240
Speaker 2: knowledge or intent, language models do very good at correctly

880
00:49:33,280 --> 00:49:36,680
Speaker 2: predicting what someone might do, given that they're missing certain information,

881
00:49:37,080 --> 00:49:39,600
Speaker 2: and so one might have thought that in the absence

882
00:49:40,080 --> 00:49:42,560
Speaker 2: of having a mind themselves, they would be quite bad

883
00:49:42,560 --> 00:49:44,920
Speaker 2: at that. But what seems to actually be the case

884
00:49:45,239 --> 00:49:48,759
Speaker 2: is by reading all of the texts that exists effectively

885
00:49:48,800 --> 00:49:52,160
Speaker 2: in the world, it has started to infer things about

886
00:49:52,640 --> 00:49:57,320
Speaker 2: other people's minds. Similarly, I would have thought that common

887
00:49:57,400 --> 00:50:01,880
Speaker 2: sense questions so questions about are three redimensional worlds. For example,

888
00:50:02,239 --> 00:50:04,680
Speaker 2: if you threw a baseball one hundred feet above my

889
00:50:04,719 --> 00:50:07,080
Speaker 2: head and I jumped up, could I catch it? It's

890
00:50:07,120 --> 00:50:09,680
Speaker 2: such a simple question for a child to answer. But

891
00:50:09,719 --> 00:50:11,960
Speaker 2: what you're doing in your mind is you're rendering a

892
00:50:12,000 --> 00:50:14,279
Speaker 2: three D simulation of the world, and you're looking at

893
00:50:14,280 --> 00:50:16,239
Speaker 2: the ball one hundred feet above my head, seeing me jump,

894
00:50:16,239 --> 00:50:18,880
Speaker 2: and realizing you'd know way you could solve that. I

895
00:50:18,880 --> 00:50:21,200
Speaker 2: would have thought these types of common sense questions would

896
00:50:21,200 --> 00:50:24,040
Speaker 2: fail in language models, and they did up until you

897
00:50:24,120 --> 00:50:27,200
Speaker 2: get the most recent update, GBT four. It answers these

898
00:50:27,239 --> 00:50:31,560
Speaker 2: common sense questions really well. However, all of that said,

899
00:50:31,680 --> 00:50:34,759
Speaker 2: the way IT solves these problems are completely different than

900
00:50:34,800 --> 00:50:37,680
Speaker 2: the way that human brains solve these problems, and those

901
00:50:37,760 --> 00:50:41,200
Speaker 2: differences do matter. Two key things that I think AI

902
00:50:41,320 --> 00:50:44,919
Speaker 2: is missing that mammal brains can do, even some fish

903
00:50:44,960 --> 00:50:47,080
Speaker 2: brands can do that I think AI can learn from

904
00:50:47,080 --> 00:50:50,760
Speaker 2: neuroscience is the following. The first is something called continual learning,

905
00:50:51,520 --> 00:50:54,680
Speaker 2: So we don't realize it. But all AI systems today

906
00:50:54,760 --> 00:50:58,640
Speaker 2: are largely trained all at once, so chat GBT doesn't

907
00:50:58,680 --> 00:51:02,040
Speaker 2: update its information as it reads new articles. The way

908
00:51:02,080 --> 00:51:04,840
Speaker 2: they update the system is, by and large, they retake

909
00:51:04,880 --> 00:51:07,480
Speaker 2: the entire data set and they rebuild the model from scratch.

910
00:51:08,239 --> 00:51:11,680
Speaker 2: And the reason they do that is because AI systems

911
00:51:11,719 --> 00:51:15,080
Speaker 2: today suffer from what's called the problem of catastrophic forgetting.

912
00:51:15,280 --> 00:51:17,920
Speaker 2: All that means is when you train an AI system

913
00:51:18,000 --> 00:51:20,759
Speaker 2: with new data, it tends to overwrite its memories of

914
00:51:20,800 --> 00:51:24,800
Speaker 2: the old data. And somehow, mammal brands and even fish

915
00:51:24,840 --> 00:51:27,920
Speaker 2: brains don't forget things when they learn new information, at

916
00:51:28,000 --> 00:51:31,040
Speaker 2: least not to the extent that aisystems do. So for example,

917
00:51:31,360 --> 00:51:33,880
Speaker 2: if you learn to ride a bicycle, you don't forget

918
00:51:33,920 --> 00:51:37,319
Speaker 2: how to drive, or vice versa. And yet somehow AI

919
00:51:37,360 --> 00:51:41,760
Speaker 2: systems still suffer from this. So commercial AI systems ignore

920
00:51:41,800 --> 00:51:43,480
Speaker 2: this problem because they say, we're just going to throw

921
00:51:43,520 --> 00:51:45,719
Speaker 2: more money at the problem and just keep retraining systems.

922
00:51:46,040 --> 00:51:48,640
Speaker 2: That's also the approach in robotics, by the way, But

923
00:51:48,719 --> 00:51:50,920
Speaker 2: eventually we're going to want systems that can learn as

924
00:51:50,920 --> 00:51:53,759
Speaker 2: they go, that can get to know us, that can

925
00:51:53,840 --> 00:51:56,160
Speaker 2: change their approach based on how they interact with us

926
00:51:57,160 --> 00:51:59,440
Speaker 2: that can be around our home, and we can show

927
00:51:59,480 --> 00:52:01,600
Speaker 2: them new skills and they figure out the new skills

928
00:52:01,640 --> 00:52:04,560
Speaker 2: as they go, and that's something that's unique to mammals

929
00:52:04,560 --> 00:52:07,160
Speaker 2: that we have not yet figured out NA. So that's

930
00:52:07,200 --> 00:52:11,920
Speaker 2: one of the big problems. The second problem is mammals

931
00:52:12,239 --> 00:52:15,239
Speaker 2: have this internal model of the world, so they have

932
00:52:15,280 --> 00:52:18,400
Speaker 2: this sort of rendered world in their head that adheres

933
00:52:18,440 --> 00:52:20,440
Speaker 2: to the laws of physics. That's how I can imagine

934
00:52:20,440 --> 00:52:23,960
Speaker 2: myself do things, and the consequences of my actions in

935
00:52:24,000 --> 00:52:27,200
Speaker 2: my mind are relatively accurate for what would happen in

936
00:52:27,239 --> 00:52:31,560
Speaker 2: the real world. And this enables me to build hypotheses

937
00:52:31,920 --> 00:52:35,520
Speaker 2: and intervene in the world to test those hypotheses. And

938
00:52:36,400 --> 00:52:39,920
Speaker 2: the reason this is so important is these AI systems today,

939
00:52:40,400 --> 00:52:44,640
Speaker 2: the truthfulness of information is only as good as the

940
00:52:44,719 --> 00:52:48,240
Speaker 2: data you give it. So if you give articles about

941
00:52:48,239 --> 00:52:51,280
Speaker 2: the Earth being flat to the training set of chat SHEGBT,

942
00:52:51,560 --> 00:52:54,080
Speaker 2: it will start thinking the Earth is flat. But the

943
00:52:54,120 --> 00:52:56,799
Speaker 2: AI systems we want to create one day are going

944
00:52:56,840 --> 00:52:59,440
Speaker 2: to be ones that interact with the world, build their

945
00:52:59,480 --> 00:53:02,560
Speaker 2: own hypo aothesies about the world, and reject information that's

946
00:53:02,600 --> 00:53:06,000
Speaker 2: inconsistent with them. Model the world and so that's going

947
00:53:06,080 --> 00:53:07,960
Speaker 2: to be the way that we can get systems that

948
00:53:08,000 --> 00:53:10,560
Speaker 2: can contribute to science. That's the way we're going to

949
00:53:10,560 --> 00:53:15,000
Speaker 2: get systems that get more truthful over time. And that's

950
00:53:15,040 --> 00:53:17,200
Speaker 2: the way we're going to get systems that don't require

951
00:53:18,080 --> 00:53:20,799
Speaker 2: you know, humans to go in and manually curate these

952
00:53:20,880 --> 00:53:25,040
Speaker 2: data sets. So although CHATGBT has learned on its own,

953
00:53:25,760 --> 00:53:28,160
Speaker 2: the manual effort went into creating the data set on

954
00:53:28,200 --> 00:53:30,000
Speaker 2: which it learned and making sure that data sets rich.

955
00:53:30,080 --> 00:53:33,239
Speaker 2: So continual learning and world models that allow you to

956
00:53:33,239 --> 00:53:36,360
Speaker 2: build hypotheses, in my view, are the two big missing

957
00:53:36,400 --> 00:53:39,760
Speaker 2: gaps that mammal brains have. But aisystems today.

958
00:53:39,560 --> 00:53:42,000
Speaker 1: General I agree. You know, last year I wrote a

959
00:53:42,040 --> 00:53:45,560
Speaker 1: paper about how we would know if AI is really

960
00:53:45,800 --> 00:53:50,360
Speaker 1: intelligent as opposed to a statistical parrot. And my suggestion

961
00:53:50,480 --> 00:53:53,880
Speaker 1: is that scientific discovery is really the gold standard for that,

962
00:53:54,000 --> 00:53:56,960
Speaker 1: because yeah, this is what humans do, and what we

963
00:53:57,080 --> 00:53:59,600
Speaker 1: do with scientific discovery is not just piece facts together.

964
00:53:59,640 --> 00:54:03,520
Speaker 1: That's and chat GEPT can do that, but it's the

965
00:54:03,600 --> 00:54:08,000
Speaker 1: simulation of possible futures. It's what if I were writing

966
00:54:08,120 --> 00:54:11,160
Speaker 1: atop a photon, what would the world look like? And

967
00:54:11,480 --> 00:54:13,680
Speaker 1: you valuate that you simulate it out, and you come

968
00:54:13,719 --> 00:54:16,359
Speaker 1: up with a special theory of relativity. That's the kind

969
00:54:16,360 --> 00:54:19,120
Speaker 1: of thing that humans do all the time, not just Einstein,

970
00:54:19,200 --> 00:54:23,719
Speaker 1: but we do that when we mentalize and simulate anything

971
00:54:24,239 --> 00:54:26,919
Speaker 1: and evaluate it and say, okay, that's not going to work.

972
00:54:26,960 --> 00:54:29,359
Speaker 1: But this other strategy over here, maybe that is going

973
00:54:29,400 --> 00:54:32,600
Speaker 1: to yield something when I compare the results to other

974
00:54:32,719 --> 00:54:35,000
Speaker 1: things I know in the world. So that's what our

975
00:54:35,040 --> 00:54:39,000
Speaker 1: systems don't do currently. So this is what's really special

976
00:54:39,040 --> 00:54:42,600
Speaker 1: about human brains is being able to mentalize and having

977
00:54:43,160 --> 00:54:44,960
Speaker 1: and having a model of the world so that we

978
00:54:45,040 --> 00:54:48,040
Speaker 1: can evaluate the outcome compare it to what we know

979
00:54:48,320 --> 00:54:51,480
Speaker 1: in the world. Now you mentioned that as AI is

980
00:54:51,560 --> 00:54:55,359
Speaker 1: getting better. Let's say chatchept four and whatever will come out.

981
00:54:55,400 --> 00:54:57,480
Speaker 1: You know, a few months from now, you're saying that

982
00:54:57,480 --> 00:55:00,160
Speaker 1: it's better and better at answering these sort of of

983
00:55:00,239 --> 00:55:05,600
Speaker 1: mentalizing questions. But do you suppose it is because of

984
00:55:06,120 --> 00:55:09,919
Speaker 1: a lot of feedback from humans and a lot of

985
00:55:09,960 --> 00:55:14,480
Speaker 1: these examples appearing on the corpus of data that it

986
00:55:14,560 --> 00:55:17,239
Speaker 1: reads that it's able to do this as opposed to

987
00:55:17,719 --> 00:55:20,080
Speaker 1: actually mentalizing and having understanding.

988
00:55:20,880 --> 00:55:24,239
Speaker 2: Certainly, I think one of the key challenges with evaluating

989
00:55:24,280 --> 00:55:26,320
Speaker 2: these AI systems is we don't know what the training

990
00:55:26,400 --> 00:55:29,759
Speaker 2: data is, so it can be hard to know if

991
00:55:29,800 --> 00:55:32,120
Speaker 2: the solution to a problem or word problem you give

992
00:55:32,160 --> 00:55:35,719
Speaker 2: it is because it's effectively looking up what was in

993
00:55:35,760 --> 00:55:39,600
Speaker 2: the training data or actually generalizing. I do think though,

994
00:55:39,640 --> 00:55:42,000
Speaker 2: there's been lots of great work where like there was

995
00:55:42,040 --> 00:55:46,120
Speaker 2: a study out of Microsoft recently where they reformat some

996
00:55:46,160 --> 00:55:48,680
Speaker 2: of these mentalizing questions in way that it's very hard

997
00:55:48,680 --> 00:55:51,279
Speaker 2: to believe that it would be in the training data,

998
00:55:51,640 --> 00:55:55,359
Speaker 2: and it still solves the problems well. To me, this

999
00:55:55,400 --> 00:55:57,839
Speaker 2: is a question of how it solved the problems though,

1000
00:55:58,280 --> 00:56:02,120
Speaker 2: because the way that chatchebt solves these problems as it

1001
00:56:02,120 --> 00:56:04,880
Speaker 2: makes an inference over a whole series let's call it,

1002
00:56:04,920 --> 00:56:08,000
Speaker 2: millions of word problems about theory of mind questions, and

1003
00:56:08,080 --> 00:56:12,040
Speaker 2: so it probably builds some form of model how agents

1004
00:56:12,120 --> 00:56:14,319
Speaker 2: or humans act in the presence of information or lack

1005
00:56:14,360 --> 00:56:18,000
Speaker 2: of information. Certainly if it reads enough symbols that suggest

1006
00:56:18,120 --> 00:56:20,439
Speaker 2: that maybe it has some of that information in there,

1007
00:56:20,800 --> 00:56:22,680
Speaker 2: but that doesn't mean it solves the problem in the

1008
00:56:22,680 --> 00:56:25,440
Speaker 2: same way humans do. You know, when we mentalize, we

1009
00:56:25,560 --> 00:56:28,360
Speaker 2: compare the way our minds work and how we feel

1010
00:56:28,360 --> 00:56:30,759
Speaker 2: about things to how we would infer someone else does

1011
00:56:30,800 --> 00:56:33,960
Speaker 2: we put ourselves in someone else's shoes, And so although

1012
00:56:34,000 --> 00:56:37,160
Speaker 2: the performance on word problems might look the same, there

1013
00:56:37,200 --> 00:56:39,880
Speaker 2: might be very big differences in how we solve these problems,

1014
00:56:39,960 --> 00:56:42,520
Speaker 2: which might have very real consequences when we send these

1015
00:56:42,560 --> 00:56:45,239
Speaker 2: things out into the real world. For example, if we

1016
00:56:45,360 --> 00:56:48,440
Speaker 2: made a robot powered by chatchebt help one of our

1017
00:56:48,480 --> 00:56:51,719
Speaker 2: grandparents around the home, and we want them to empathize

1018
00:56:51,760 --> 00:56:54,840
Speaker 2: and understand how they feel, I would not be confidence

1019
00:56:55,120 --> 00:56:57,600
Speaker 2: based on the performance of word problems of theory of

1020
00:56:57,640 --> 00:57:01,520
Speaker 2: mind that chatsheebt is going to care infer about how

1021
00:57:01,520 --> 00:57:04,359
Speaker 2: my grandparent feels in this situation, versus I would feel

1022
00:57:04,440 --> 00:57:06,840
Speaker 2: confident that a human would because I know how a

1023
00:57:06,920 --> 00:57:09,600
Speaker 2: human brain is solving these tasks. So I think algorithmic

1024
00:57:09,680 --> 00:57:13,760
Speaker 2: differences matter the more and more we offload these TASKSDAIE systems,

1025
00:57:13,840 --> 00:57:17,280
Speaker 2: because otherwise performance in one task might not generalize well

1026
00:57:17,320 --> 00:57:18,160
Speaker 2: to these other tests.

1027
00:57:18,640 --> 00:57:20,919
Speaker 1: So what's interesting is I've spent a lot of time

1028
00:57:21,000 --> 00:57:25,320
Speaker 1: on GPT four seeing if it has theory of mind,

1029
00:57:25,840 --> 00:57:29,520
Speaker 1: you know, running tests on this and just for the audience,

1030
00:57:29,840 --> 00:57:32,320
Speaker 1: theory of mind tests would be something like Sally walks

1031
00:57:32,360 --> 00:57:34,960
Speaker 1: into the room and puts the baseball on the bed.

1032
00:57:35,360 --> 00:57:38,760
Speaker 1: Then she leaves and comes into the room, sees the

1033
00:57:38,760 --> 00:57:41,360
Speaker 1: baseball on the bed, picks it up, puts in the closet,

1034
00:57:41,560 --> 00:57:44,640
Speaker 1: and leaves. When Sally walks back in the room, where

1035
00:57:44,640 --> 00:57:47,360
Speaker 1: does she look for the ball? And the answer, of

1036
00:57:47,360 --> 00:57:49,000
Speaker 1: course is that she looks on the bed. But this

1037
00:57:49,080 --> 00:57:51,680
Speaker 1: requires us to be inside her head. If you ask

1038
00:57:51,720 --> 00:57:54,120
Speaker 1: a question like that to any of the big language models,

1039
00:57:54,280 --> 00:57:57,160
Speaker 1: it will get it right. But why. In part, it's

1040
00:57:57,240 --> 00:58:01,439
Speaker 1: because that particular test, the seal antest, is all over

1041
00:58:01,520 --> 00:58:04,960
Speaker 1: the Internet the gajillion places, and there are many many

1042
00:58:05,600 --> 00:58:08,400
Speaker 1: questions that have been asked about theory of mind that

1043
00:58:08,560 --> 00:58:12,160
Speaker 1: already exist on the Internet. The part that I have

1044
00:58:12,240 --> 00:58:15,840
Speaker 1: found so fascinating is that GPT gets this stuff right

1045
00:58:16,040 --> 00:58:19,200
Speaker 1: about I don't know, sixty percent of the time. So

1046
00:58:19,440 --> 00:58:22,360
Speaker 1: in other words, several times in a row, I'll try

1047
00:58:22,440 --> 00:58:24,320
Speaker 1: to make up some question that I think is new,

1048
00:58:24,600 --> 00:58:27,000
Speaker 1: and it gets it right, and I'm stunned, and I think, wow,

1049
00:58:27,400 --> 00:58:29,600
Speaker 1: I think it really has a sense of what it

1050
00:58:29,640 --> 00:58:31,720
Speaker 1: is to be a person. But then it will get

1051
00:58:31,800 --> 00:58:35,240
Speaker 1: one wrong, and it's the kind of mistake that a

1052
00:58:35,280 --> 00:58:39,040
Speaker 1: person wouldn't make if a person understands theory of mind,

1053
00:58:39,080 --> 00:58:41,760
Speaker 1: they wouldn't get this other version wrong. And that's why

1054
00:58:41,840 --> 00:58:44,200
Speaker 1: I find myself a little bit confused here in the

1055
00:58:44,240 --> 00:58:47,240
Speaker 1: middle of twenty twenty four about whether to conclude that

1056
00:58:47,320 --> 00:58:50,840
Speaker 1: AI has theory of mind capabilities or not.

1057
00:58:51,560 --> 00:58:55,000
Speaker 2: I think this goes to the semantics of how we

1058
00:58:55,080 --> 00:58:57,520
Speaker 2: measure this thing we call theory of mine, and this

1059
00:58:57,560 --> 00:58:59,120
Speaker 2: is actually what we're asking these in some ways a

1060
00:58:59,160 --> 00:59:03,160
Speaker 2: profound question and an open question in AI, because the

1061
00:59:03,360 --> 00:59:07,840
Speaker 2: entire field of machine learning operates on performance benchmarks. The

1062
00:59:08,000 --> 00:59:10,280
Speaker 2: entire field is based on this idea of give me

1063
00:59:10,320 --> 00:59:12,360
Speaker 2: an evaluation test, and then I'm going to see how

1064
00:59:12,360 --> 00:59:14,880
Speaker 2: well I perform on this test. But that's problematic for

1065
00:59:14,960 --> 00:59:17,480
Speaker 2: things like theory of mind because if you ask any

1066
00:59:17,520 --> 00:59:20,400
Speaker 2: scientist a theory of mind, theory of mind is defined

1067
00:59:20,400 --> 00:59:23,320
Speaker 2: in the mechanism, not the performance, but theory of mind

1068
00:59:23,480 --> 00:59:26,400
Speaker 2: is is the algorithm by which we imagine ourselves on

1069
00:59:26,400 --> 00:59:29,120
Speaker 2: other people's shoes. They don't define theory of mind as

1070
00:59:29,160 --> 00:59:31,880
Speaker 2: the ability to solve this word problem, and so we

1071
00:59:31,960 --> 00:59:35,920
Speaker 2: see this sort of challenge where just because it solves

1072
00:59:35,960 --> 00:59:38,360
Speaker 2: the word problems doesn't mean that it's solving them in

1073
00:59:38,400 --> 00:59:41,320
Speaker 2: the way that someone else might classify as theory of mind.

1074
00:59:41,480 --> 00:59:43,040
Speaker 2: So I think in some ways this is in the

1075
00:59:43,080 --> 00:59:45,040
Speaker 2: semantics of what do we mean when we say does

1076
00:59:45,080 --> 00:59:47,680
Speaker 2: the sing have theory of mind? I think it clearly

1077
00:59:47,720 --> 00:59:51,080
Speaker 2: is very good at solving theory of mind like word problems.

1078
00:59:51,200 --> 00:59:54,000
Speaker 2: I'm quite confident that it's not doing what primates do

1079
00:59:54,040 --> 00:59:56,160
Speaker 2: when they engage in theory of mind. And I'm also

1080
00:59:56,320 --> 00:59:59,880
Speaker 2: not confident that the solutions to these word problems will

1081
01:00:00,080 --> 01:00:03,400
Speaker 2: generalize well to other types of tasks that are not

1082
01:00:03,560 --> 01:00:07,120
Speaker 2: word based that require theory of mind, such as a

1083
01:00:07,240 --> 01:00:09,960
Speaker 2: robot around the house that has to infer how someone

1084
01:00:10,040 --> 01:00:13,560
Speaker 2: might feel in certain situations to proactively help them, proactively

1085
01:00:13,600 --> 01:00:17,960
Speaker 2: comfort them. I'm not confident that the theory of mind

1086
01:00:17,960 --> 01:00:20,480
Speaker 2: word problem success will translate to these other types of

1087
01:00:20,640 --> 01:00:21,560
Speaker 2: theory of mind problems.

1088
01:00:22,080 --> 01:00:25,520
Speaker 1: So to get to that robot that is like a

1089
01:00:25,600 --> 01:00:29,080
Speaker 1: human and really understands these things, what do you see

1090
01:00:29,160 --> 01:00:32,880
Speaker 1: from your framework of these five breakthroughs of intelligence? What

1091
01:00:33,040 --> 01:00:35,880
Speaker 1: needs to happen besides this language piece.

1092
01:00:36,240 --> 01:00:39,479
Speaker 2: So the big missing pieces are breakthrough three and four.

1093
01:00:39,720 --> 01:00:42,200
Speaker 2: We need these systems to have some form of internal

1094
01:00:42,240 --> 01:00:46,240
Speaker 2: world model that they're continuously updating based on interacting with

1095
01:00:46,280 --> 01:00:50,120
Speaker 2: the actual world. And I do think this grounding in

1096
01:00:50,200 --> 01:00:52,760
Speaker 2: reality is important for many of the features that we

1097
01:00:52,800 --> 01:00:55,720
Speaker 2: want these AI systems to have, but that will not

1098
01:00:55,840 --> 01:01:00,360
Speaker 2: be enough. That will maybe solve some very utilitarian functionals

1099
01:01:00,400 --> 01:01:03,640
Speaker 2: around the home, but I think we will quickly realize

1100
01:01:03,680 --> 01:01:08,320
Speaker 2: that understanding how to interact with humans and the social

1101
01:01:08,360 --> 01:01:12,520
Speaker 2: lives of humans will emerge as this other really important

1102
01:01:12,560 --> 01:01:15,120
Speaker 2: missing piece, which will require some form of mentalizing. In

1103
01:01:15,160 --> 01:01:18,480
Speaker 2: other words, understanding what's going on in human heads a

1104
01:01:18,480 --> 01:01:20,760
Speaker 2: fascinating open question that I don't have the answer to,

1105
01:01:21,360 --> 01:01:24,240
Speaker 2: but something we'll need to think about. One way in

1106
01:01:24,280 --> 01:01:28,400
Speaker 2: which humans build common ground is that our minds algorithmically

1107
01:01:28,440 --> 01:01:31,760
Speaker 2: are quite similar. So when I put myself in someone

1108
01:01:31,760 --> 01:01:34,880
Speaker 2: else's shoes, certainly there's lots of mistakes we make when

1109
01:01:34,920 --> 01:01:37,800
Speaker 2: trying to guess how other people feel in situations, but

1110
01:01:37,920 --> 01:01:41,840
Speaker 2: there is this basic grounding that we are all very similar.

1111
01:01:42,000 --> 01:01:45,600
Speaker 2: Our brains works relatively similarly in the scope of all

1112
01:01:45,640 --> 01:01:49,040
Speaker 2: possible preferences of life form could have. Humans are remarkably

1113
01:01:49,080 --> 01:01:51,440
Speaker 2: more similar than they are different. And yet when we

1114
01:01:51,440 --> 01:01:53,880
Speaker 2: build this AI system, it's not at all clear that

1115
01:01:53,920 --> 01:01:55,560
Speaker 2: the way it would feel about the world is going

1116
01:01:55,600 --> 01:01:57,480
Speaker 2: to be the way we feel about the world. And

1117
01:01:57,560 --> 01:02:00,720
Speaker 2: so the basic trick that it seems primate brains use,

1118
01:02:00,920 --> 01:02:03,240
Speaker 2: which is I reason about your mind by building a

1119
01:02:03,280 --> 01:02:06,240
Speaker 2: model of my own mind and projecting myself into your situation,

1120
01:02:06,720 --> 01:02:11,360
Speaker 2: won't work for an aisystem because it won't be the

1121
01:02:11,400 --> 01:02:13,840
Speaker 2: same as us. It won't necessarily have the same preferences.

1122
01:02:14,440 --> 01:02:17,160
Speaker 2: And so I do think that begets an interesting sort

1123
01:02:17,160 --> 01:02:19,320
Speaker 2: of safety challenge for us, which is, how do we

1124
01:02:19,400 --> 01:02:23,200
Speaker 2: make sure that they actually understand human preferences, how we

1125
01:02:23,200 --> 01:02:25,920
Speaker 2: feel about things, how we would feel about things, while

1126
01:02:25,960 --> 01:02:28,520
Speaker 2: not being grounded and having those same feelings themselves.

1127
01:02:33,320 --> 01:02:36,560
Speaker 1: That was Max Bennett diving into the six hundred million

1128
01:02:36,680 --> 01:02:40,800
Speaker 1: year history of how the human brain got here. As

1129
01:02:40,840 --> 01:02:43,040
Speaker 1: you can see, Max looks at evolution the way that

1130
01:02:43,080 --> 01:02:46,960
Speaker 1: you might look at technological innovation in the business world.

1131
01:02:47,080 --> 01:02:50,640
Speaker 1: When a new technology comes onto the scene, like the

1132
01:02:50,680 --> 01:02:55,240
Speaker 1: personal computer, it enables all kinds of new products, and

1133
01:02:55,320 --> 01:02:59,600
Speaker 1: it's the same when a new brain capability hits the scene,

1134
01:03:00,120 --> 01:03:04,240
Speaker 1: that opens the door to new sorts of skills. For example,

1135
01:03:04,560 --> 01:03:08,080
Speaker 1: once a brain can run internal simulations, then it can

1136
01:03:08,120 --> 01:03:13,760
Speaker 1: do things like remember the past, and envision possible futures.

1137
01:03:14,280 --> 01:03:17,040
Speaker 1: So I just wanted to summarize Max's framework here so

1138
01:03:17,080 --> 01:03:19,880
Speaker 1: that you can remember it. The first breakthrough happened in

1139
01:03:19,960 --> 01:03:23,560
Speaker 1: animals that have left right symmetry, like a human or

1140
01:03:23,560 --> 01:03:27,240
Speaker 1: a bird or a lizard as opposed to a starfish

1141
01:03:27,320 --> 01:03:30,400
Speaker 1: or a jellyfish. The first step was that these left

1142
01:03:30,440 --> 01:03:34,320
Speaker 1: right animals learned how to steer themselves through their environment.

1143
01:03:35,200 --> 01:03:38,600
Speaker 1: Break Through number two happened in vertebrates, those animals that

1144
01:03:38,640 --> 01:03:41,840
Speaker 1: have a spinal column. They figured out how to learn

1145
01:03:42,080 --> 01:03:47,600
Speaker 1: from trial and error. Break Through three happened in mammals.

1146
01:03:47,640 --> 01:03:52,600
Speaker 1: They learned to simulate internally, that's thinking about the past

1147
01:03:52,680 --> 01:03:55,919
Speaker 1: and running versions of the future. Break Through number four

1148
01:03:56,080 --> 01:04:00,920
Speaker 1: happened in primates in particular, and that was meant, in

1149
01:04:00,960 --> 01:04:04,560
Speaker 1: other words, imagining what it is like to be inside

1150
01:04:04,560 --> 01:04:08,400
Speaker 1: someone else's head to infer the intent of the other,

1151
01:04:08,840 --> 01:04:13,520
Speaker 1: and for that matter, thinking about your own thinking. And finally,

1152
01:04:13,560 --> 01:04:17,760
Speaker 1: break through number five happened in humans, and that was speech,

1153
01:04:18,320 --> 01:04:21,520
Speaker 1: which allows us to pass information rapidly from one to

1154
01:04:21,560 --> 01:04:24,919
Speaker 1: another and for that matter, from generation to generation. From

1155
01:04:24,960 --> 01:04:29,680
Speaker 1: the Library of Alexandria to the Inner Cosmos podcast, all

1156
01:04:29,720 --> 01:04:32,760
Speaker 1: of this is made possible by figuring out how to

1157
01:04:32,840 --> 01:04:36,880
Speaker 1: communicate at this high bandwidth. As a result of this,

1158
01:04:37,440 --> 01:04:40,520
Speaker 1: humans don't have to start from scratch every generation the

1159
01:04:40,520 --> 01:04:44,160
Speaker 1: way a cat or a horse does, but instead humans

1160
01:04:44,160 --> 01:04:47,840
Speaker 1: are able to springboard off the top of everything that

1161
01:04:47,880 --> 01:04:54,040
Speaker 1: has been discovered by previous humans. Collectively, these breakthroughs, which

1162
01:04:54,040 --> 01:04:57,560
Speaker 1: happened over hundreds of millions of years, gave us the

1163
01:04:57,720 --> 01:05:00,600
Speaker 1: kind of brains that we have us to do the

1164
01:05:00,720 --> 01:05:04,160
Speaker 1: kind of things that we do. A lot of questions remain.

1165
01:05:04,640 --> 01:05:08,240
Speaker 1: One of them is whether there are different paths to intelligence,

1166
01:05:08,560 --> 01:05:11,520
Speaker 1: as we suspect when we look at the octopus brain,

1167
01:05:11,680 --> 01:05:14,920
Speaker 1: which is a mollusc brain that somehow evolved along a

1168
01:05:15,120 --> 01:05:18,720
Speaker 1: very different sort of pathway and yet ended up at

1169
01:05:18,760 --> 01:05:22,640
Speaker 1: a similar spot. And once we find other sorts of

1170
01:05:22,680 --> 01:05:26,480
Speaker 1: intelligences in the universe, we may look back and realize

1171
01:05:26,520 --> 01:05:30,720
Speaker 1: there are many ways to get to intelligence from single

1172
01:05:30,760 --> 01:05:35,080
Speaker 1: celled organisms floating around. For all we know, intelligence is

1173
01:05:35,200 --> 01:05:39,440
Speaker 1: a path that is nudged into being by the pressures

1174
01:05:39,480 --> 01:05:43,280
Speaker 1: of evolution because of the advantages that it grants, so

1175
01:05:43,320 --> 01:05:47,160
Speaker 1: that things generally move in that direction. And if that's

1176
01:05:47,200 --> 01:05:51,840
Speaker 1: the case, if the pressures of evolution guide animals inexorably

1177
01:05:51,920 --> 01:05:56,200
Speaker 1: toward intelligence so they can outcompete their neighbors. Then what

1178
01:05:56,280 --> 01:05:59,680
Speaker 1: a pleasure it would be to visit the Earth six

1179
01:05:59,720 --> 01:06:04,439
Speaker 1: hundred million years from now, when lots of other species

1180
01:06:04,840 --> 01:06:09,640
Speaker 1: have reached new elevations in that long road. They've reached

1181
01:06:10,040 --> 01:06:13,560
Speaker 1: those heights that give them the kind of view that

1182
01:06:13,600 --> 01:06:19,120
Speaker 1: has allowed us to invent and create and discover and

1183
01:06:19,440 --> 01:06:29,080
Speaker 1: intellectually explore. Go to Eagleman dot com slash podcast for

1184
01:06:29,120 --> 01:06:32,720
Speaker 1: more information and to find further reading. Send me an

1185
01:06:32,760 --> 01:06:36,720
Speaker 1: email at podcasts at eagleman dot com with questions or discussion,

1186
01:06:37,200 --> 01:06:40,440
Speaker 1: and check out and subscribe to Inner Cosmos on YouTube

1187
01:06:40,560 --> 01:06:44,480
Speaker 1: for videos of each episode and to leave comments. Until

1188
01:06:44,520 --> 01:06:48,800
Speaker 1: next time. I'm David Eagleman, and this is Inner Cosmos