1
00:00:01,006 --> 00:00:06,156
We have historical records that allow us
to know how the ancient Greeks dressed,

2
00:00:06,180 --> 00:00:07,434
how they lived,

3
00:00:07,458 --> 00:00:08,980
how they fought...

4
00:00:09,004 --> 00:00:10,528
but how did they think?

5
00:00:11,432 --> 00:00:15,872
One natural idea is that the deepest
aspects of human thought --

6
00:00:15,896 --> 00:00:17,768
our ability to imagine,

7
00:00:17,792 --> 00:00:19,189
to be conscious,

8
00:00:19,213 --> 00:00:20,444
to dream --

9
00:00:20,468 --> 00:00:22,087
have always been the same.

10
00:00:22,872 --> 00:00:24,371
Another possibility

11
00:00:24,395 --> 00:00:28,118
is that the social transformations
that have shaped our culture

12
00:00:28,142 --> 00:00:31,927
may have also changed
the structural columns of human thought.

13
00:00:32,911 --> 00:00:35,435
We may all have different
opinions about this.

14
00:00:35,459 --> 00:00:38,176
Actually, it's a long-standing
philosophical debate.

15
00:00:38,644 --> 00:00:41,371
But is this question
even amenable to science?

16
00:00:42,834 --> 00:00:45,340
Here I'd like to propose

17
00:00:45,364 --> 00:00:50,136
that in the same way we can reconstruct
how the ancient Greek cities looked like,

18
00:00:50,160 --> 00:00:52,548
just based on a few bricks,

19
00:00:52,572 --> 00:00:56,698
that the writings of a culture
are the archaeological records --

20
00:00:56,722 --> 00:00:58,865
the fossils of human thought.

21
00:00:59,905 --> 00:01:01,079
And in fact,

22
00:01:01,103 --> 00:01:03,309
doing some form of psychological analysis

23
00:01:03,333 --> 00:01:06,877
of some of the most ancient
books of human culture,

24
00:01:06,901 --> 00:01:12,856
Julian Jaynes came up in the '70s
with a very wild and radical hypothesis:

25
00:01:12,880 --> 00:01:15,293
that only 3,000 years ago,

26
00:01:15,317 --> 00:01:20,205
humans were what today
we would call schizophrenics.

27
00:01:21,753 --> 00:01:23,261
And he made this claim

28
00:01:23,285 --> 00:01:26,586
based on the fact that the first
humans described in these books

29
00:01:26,610 --> 00:01:28,514
behaved consistently,

30
00:01:28,538 --> 00:01:31,554
in different traditions
and in different places of the world,

31
00:01:31,578 --> 00:01:35,110
as if they were hearing and obeying voices

32
00:01:35,134 --> 00:01:38,174
that they perceived
as coming from the Gods,

33
00:01:38,198 --> 00:01:39,396
or from the muses...

34
00:01:40,063 --> 00:01:42,832
what today we would call hallucinations.

35
00:01:43,888 --> 00:01:46,514
And only then, as time went on,

36
00:01:46,538 --> 00:01:50,189
they began to recognize
that they were the creators --

37
00:01:50,213 --> 00:01:52,728
the owners of these inner voices.

38
00:01:53,316 --> 00:01:56,031
And with this, they gained introspection:

39
00:01:56,055 --> 00:01:58,538
the ability to think
about their own thoughts.

40
00:01:59,785 --> 00:02:03,182
So Jaynes' theory is that consciousness --

41
00:02:03,206 --> 00:02:06,372
at least in the way we perceive it today,

42
00:02:06,396 --> 00:02:09,936
where we feel that we are the pilots
of our own existence --

43
00:02:09,960 --> 00:02:12,697
is a quite recent cultural development.

44
00:02:13,456 --> 00:02:15,242
And this theory is quite spectacular,

45
00:02:15,266 --> 00:02:16,699
but it has an obvious problem

46
00:02:16,723 --> 00:02:20,715
which is that it's built on just a few
and very specific examples.

47
00:02:21,085 --> 00:02:22,848
So the question is whether the theory

48
00:02:22,872 --> 00:02:27,623
that introspection built up in human
history only about 3,000 years ago

49
00:02:27,647 --> 00:02:30,631
can be examined in a quantitative
and objective manner.

50
00:02:31,543 --> 00:02:35,106
And the problem of how
to go about this is quite obvious.

51
00:02:35,130 --> 00:02:38,590
It's not like Plato woke up one day
and then he wrote,

52
00:02:38,614 --> 00:02:40,273
"Hello, I'm Plato,

53
00:02:40,297 --> 00:02:43,186
and as of today, I have
a fully introspective consciousness."

54
00:02:43,210 --> 00:02:45,503
(Laughter)

55
00:02:45,527 --> 00:02:48,860
And this still is actually
what is the essence of the problem.

56
00:02:49,467 --> 00:02:53,522
We need to find the emergence
of a concept that's never said.

57
00:02:54,434 --> 00:02:58,744
The word introspection
does not appear a single time

58
00:02:58,768 --> 00:03:00,687
in the books we want to analyze.

59
00:03:01,728 --> 00:03:05,815
So our way to solve this
is to build the space of words.

60
00:03:06,571 --> 00:03:09,858
This is a huge space
that contains all words

61
00:03:09,882 --> 00:03:12,684
in such a way that the distance
between any two of them

62
00:03:12,708 --> 00:03:15,591
is indicative of how
closely related they are.

63
00:03:16,460 --> 00:03:17,611
So for instance,

64
00:03:17,635 --> 00:03:20,532
you want the words "dog" and "cat"
to be very close together,

65
00:03:20,556 --> 00:03:24,387
but the words "grapefruit" and "logarithm"
to be very far away.

66
00:03:24,809 --> 00:03:28,705
And this has to be true
for any two words within the space.

67
00:03:29,626 --> 00:03:32,967
And there are different ways
that we can construct the space of words.

68
00:03:32,991 --> 00:03:34,634
One is just asking the experts,

69
00:03:34,658 --> 00:03:36,554
a bit like we do with dictionaries.

70
00:03:36,896 --> 00:03:38,324
Another possibility

71
00:03:38,348 --> 00:03:42,063
is following the simple assumption
that when two words are related,

72
00:03:42,087 --> 00:03:44,436
they tend to appear in the same sentences,

73
00:03:44,460 --> 00:03:45,913
in the same paragraphs,

74
00:03:45,937 --> 00:03:47,707
in the same documents,

75
00:03:47,731 --> 00:03:50,913
more often than would be expected
just by pure chance.

76
00:03:52,231 --> 00:03:54,281
And this simple hypothesis,

77
00:03:54,305 --> 00:03:55,611
this simple method,

78
00:03:55,635 --> 00:03:57,242
with some computational tricks

79
00:03:57,266 --> 00:03:58,655
that have to do with the fact

80
00:03:58,679 --> 00:04:01,743
that this is a very complex
and high-dimensional space,

81
00:04:01,767 --> 00:04:03,432
turns out to be quite effective.

82
00:04:04,155 --> 00:04:06,957
And just to give you a flavor
of how well this works,

83
00:04:06,981 --> 00:04:10,893
this is the result we get when
we analyze this for some familiar words.

84
00:04:11,607 --> 00:04:12,792
And you can see first

85
00:04:12,816 --> 00:04:16,094
that words automatically organize
into semantic neighborhoods.

86
00:04:16,118 --> 00:04:18,335
So you get the fruits, the body parts,

87
00:04:18,359 --> 00:04:20,784
the computer parts,
the scientific terms and so on.

88
00:04:21,119 --> 00:04:25,341
The algorithm also identifies
that we organize concepts in a hierarchy.

89
00:04:25,852 --> 00:04:27,003
So for instance,

90
00:04:27,027 --> 00:04:30,624
you can see that the scientific terms
break down into two subcategories

91
00:04:30,648 --> 00:04:32,748
of the astronomic and the physics terms.

92
00:04:33,338 --> 00:04:35,584
And then there are very fine things.

93
00:04:35,608 --> 00:04:37,513
For instance, the word astronomy,

94
00:04:37,537 --> 00:04:39,352
which seems a bit bizarre where it is,

95
00:04:39,376 --> 00:04:41,424
is actually exactly where it should be,

96
00:04:41,448 --> 00:04:43,043
between what it is --

97
00:04:43,067 --> 00:04:44,337
an actual science --

98
00:04:44,361 --> 00:04:45,897
and between what it describes --

99
00:04:45,921 --> 00:04:47,413
the astronomical terms.

100
00:04:48,182 --> 00:04:50,073
And we could go on and on with this.

101
00:04:50,097 --> 00:04:52,157
Actually, if you stare
at this for a while,

102
00:04:52,181 --> 00:04:54,039
and you just build random trajectories,

103
00:04:54,063 --> 00:04:57,229
you will see that it actually feels
a bit like doing poetry.

104
00:04:58,018 --> 00:04:59,900
And this is because, in a way,

105
00:04:59,924 --> 00:05:02,864
walking in this space
is like walking in the mind.

106
00:05:04,027 --> 00:05:05,644
And the last thing

107
00:05:05,668 --> 00:05:09,708
is that this algorithm also identifies
what are our intuitions

108
00:05:09,732 --> 00:05:13,628
of which words should lead
in the neighborhood of introspection.

109
00:05:13,652 --> 00:05:14,875
So for instance,

110
00:05:14,899 --> 00:05:18,878
words such as "self," "guilt,"
"reason," "emotion"

111
00:05:18,902 --> 00:05:20,791
are very close to "introspection,"

112
00:05:20,815 --> 00:05:21,966
but other words,

113
00:05:21,990 --> 00:05:24,157
such as "red," "football,"
"candle," "banana,"

114
00:05:24,181 --> 00:05:25,633
are just very far away.

115
00:05:26,054 --> 00:05:28,816
And so once we've built the space,

116
00:05:28,840 --> 00:05:31,666
the question of the history
of introspection,

117
00:05:31,690 --> 00:05:34,023
or of the history of any concept

118
00:05:34,047 --> 00:05:38,826
which before could seem abstract
and somehow vague,

119
00:05:38,850 --> 00:05:40,454
becomes concrete --

120
00:05:40,478 --> 00:05:43,216
becomes amenable to quantitative science.

121
00:05:44,216 --> 00:05:46,978
All that we have to do is take the books,

122
00:05:47,002 --> 00:05:48,383
we digitize them,

123
00:05:48,407 --> 00:05:51,216
and we take this stream
of words as a trajectory

124
00:05:51,240 --> 00:05:53,209
and project them into the space,

125
00:05:53,233 --> 00:05:56,987
and then we ask whether this trajectory
spends significant time

126
00:05:57,011 --> 00:06:00,003
circling closely to the concept
of introspection.

127
00:06:00,760 --> 00:06:01,956
And with this,

128
00:06:01,980 --> 00:06:04,092
we could analyze
the history of introspection

129
00:06:04,116 --> 00:06:06,037
in the ancient Greek tradition,

130
00:06:06,061 --> 00:06:08,663
for which we have the best
available written record.

131
00:06:09,631 --> 00:06:11,886
So what we did is we took all the books --

132
00:06:11,910 --> 00:06:14,194
we just ordered them by time --

133
00:06:14,218 --> 00:06:15,970
for each book we take the words

134
00:06:15,994 --> 00:06:17,955
and we project them to the space,

135
00:06:17,979 --> 00:06:21,011
and then we ask for each word
how close it is to introspection,

136
00:06:21,035 --> 00:06:22,265
and we just average that.

137
00:06:22,590 --> 00:06:25,788
And as time goes on and on,

138
00:06:25,812 --> 00:06:29,064
these books get closer,
and closer and closer

139
00:06:29,088 --> 00:06:30,842
to the concept of introspection.

140
00:06:30,866 --> 00:06:34,667
And this is exactly what happens
in the ancient Greek tradition.

141
00:06:35,698 --> 00:06:38,825
So you can see that for the oldest books
in the Homeric tradition,

142
00:06:38,849 --> 00:06:42,261
there is a small increase with books
getting closer to introspection,

143
00:06:42,285 --> 00:06:44,491
but about four centuries before Christ,

144
00:06:44,515 --> 00:06:49,223
this starts ramping up very rapidly
to an almost five-fold increase

145
00:06:49,247 --> 00:06:51,747
of books getting closer,
and closer and closer

146
00:06:51,771 --> 00:06:53,453
to the concept of introspection.

147
00:06:54,159 --> 00:06:56,583
And one of the nice things about this

148
00:06:56,607 --> 00:06:57,805
is that now we can ask

149
00:06:57,829 --> 00:07:01,976
whether this is also true
in a different, independent tradition.

150
00:07:02,962 --> 00:07:06,138
So we just ran this same analysis
on the Judeo-Christian tradition,

151
00:07:06,162 --> 00:07:08,883
and we got virtually the same pattern.

152
00:07:09,548 --> 00:07:14,183
Again, you see a small increase
for the oldest books in the Old Testament,

153
00:07:14,207 --> 00:07:16,121
and then it increases much more rapidly

154
00:07:16,145 --> 00:07:17,984
in the new books of the New Testament,

155
00:07:18,008 --> 00:07:20,040
and then we get the peak of introspection

156
00:07:20,064 --> 00:07:22,191
in "The Confessions of Saint Augustine,"

157
00:07:22,215 --> 00:07:24,072
about four centuries after Christ.

158
00:07:24,897 --> 00:07:26,841
And this was very important,

159
00:07:26,865 --> 00:07:30,238
because Saint Augustine
had been recognized by scholars,

160
00:07:30,262 --> 00:07:32,434
philologists, historians,

161
00:07:32,458 --> 00:07:34,536
as one of the founders of introspection.

162
00:07:35,060 --> 00:07:38,357
Actually, some believe him to be
the father of modern psychology.

163
00:07:39,012 --> 00:07:40,851
So our algorithm,

164
00:07:40,875 --> 00:07:43,717
which has the virtue
of being quantitative,

165
00:07:43,741 --> 00:07:45,004
of being objective,

166
00:07:45,028 --> 00:07:47,044
and of course of being extremely fast --

167
00:07:47,068 --> 00:07:49,465
it just runs in a fraction of a second --

168
00:07:49,489 --> 00:07:52,992
can capture some of the most
important conclusions

169
00:07:53,016 --> 00:07:55,238
of this long tradition of investigation.

170
00:07:56,317 --> 00:07:59,968
And this is in a way
one of the beauties of science,

171
00:07:59,992 --> 00:08:03,468
which is that now this idea
can be translated

172
00:08:03,492 --> 00:08:06,063
and generalized to a whole lot
of different domains.

173
00:08:06,769 --> 00:08:11,536
So in the same way that we asked
about the past of human consciousness,

174
00:08:11,560 --> 00:08:14,966
maybe the most challenging question
we can pose to ourselves

175
00:08:14,990 --> 00:08:19,127
is whether this can tell us something
about the future of our own consciousness.

176
00:08:19,550 --> 00:08:21,020
To put it more precisely,

177
00:08:21,044 --> 00:08:23,460
whether the words we say today

178
00:08:23,484 --> 00:08:28,681
can tell us something
of where our minds will be in a few days,

179
00:08:28,705 --> 00:08:29,856
in a few months,

180
00:08:29,880 --> 00:08:31,062
or a few years from now.

181
00:08:31,597 --> 00:08:34,617
And in the same way many of us
are now wearing sensors

182
00:08:34,641 --> 00:08:36,427
that detect our heart rate,

183
00:08:36,451 --> 00:08:37,720
our respiration,

184
00:08:37,744 --> 00:08:39,411
our genes,

185
00:08:39,435 --> 00:08:43,086
on the hopes that this may
help us prevent diseases,

186
00:08:43,110 --> 00:08:46,631
we can ask whether monitoring
and analyzing the words we speak,

187
00:08:46,655 --> 00:08:49,338
we tweet, we email, we write,

188
00:08:49,362 --> 00:08:54,170
can tell us ahead of time whether
something may go wrong with our minds.

189
00:08:55,087 --> 00:08:56,621
And with Guillermo Cecchi,

190
00:08:56,645 --> 00:08:59,646
who has been my brother in this adventure,

191
00:08:59,670 --> 00:09:01,225
we took on this task.

192
00:09:02,228 --> 00:09:07,760
And we did so by analyzing
the recorded speech of 34 young people

193
00:09:07,784 --> 00:09:10,585
who were at a high risk
of developing schizophrenia.

194
00:09:11,434 --> 00:09:14,315
And so what we did is,
we measured speech at day one,

195
00:09:14,339 --> 00:09:17,581
and then we asked whether the properties
of the speech could predict,

196
00:09:17,605 --> 00:09:20,101
within a window of almost three years,

197
00:09:20,125 --> 00:09:22,160
the future development of psychosis.

198
00:09:23,427 --> 00:09:25,793
But despite our hopes,

199
00:09:25,817 --> 00:09:28,934
we got failure after failure.

200
00:09:29,793 --> 00:09:33,675
There was just not enough
information in semantics

201
00:09:33,699 --> 00:09:36,492
to predict the future
organization of the mind.

202
00:09:36,516 --> 00:09:38,325
It was good enough

203
00:09:38,349 --> 00:09:42,524
to distinguish between a group
of schizophrenics and a control group,

204
00:09:42,548 --> 00:09:45,260
a bit like we had done
for the ancient texts,

205
00:09:45,284 --> 00:09:48,278
but not to predict the future
onto the psychosis.

206
00:09:49,164 --> 00:09:50,870
But then we realized

207
00:09:50,894 --> 00:09:54,982
that maybe the most important thing
was not so much what they were saying,

208
00:09:55,006 --> 00:09:56,679
but how they were saying it.

209
00:09:57,679 --> 00:09:58,899
More specifically,

210
00:09:58,923 --> 00:10:01,750
it was not in which semantic
neighborhoods the words were,

211
00:10:01,774 --> 00:10:04,374
but how far and fast they jumped

212
00:10:04,398 --> 00:10:06,699
from one semantic neighborhood
to the other one.

213
00:10:07,247 --> 00:10:08,978
And so we came up with this measure,

214
00:10:09,002 --> 00:10:11,391
which we termed semantic coherence,

215
00:10:11,415 --> 00:10:16,219
which essentially measures the persistence
of speech within one semantic topic,

216
00:10:16,243 --> 00:10:17,772
within one semantic category.

217
00:10:19,294 --> 00:10:23,341
And it turned out to be
that for this group of 34 people,

218
00:10:23,365 --> 00:10:27,024
the algorithm based on semantic
coherence could predict,

219
00:10:27,048 --> 00:10:29,548
with 100 percent accuracy,

220
00:10:29,572 --> 00:10:32,079
who developed psychosis and who will not.

221
00:10:32,976 --> 00:10:35,913
And this was something
that could not be achieved --

222
00:10:35,937 --> 00:10:37,445
not even close --

223
00:10:37,469 --> 00:10:40,595
with all the other
existing clinical measures.

224
00:10:42,525 --> 00:10:46,104
And I remember vividly,
while I was working on this,

225
00:10:46,128 --> 00:10:48,445
I was sitting at my computer

226
00:10:48,469 --> 00:10:51,104
and I saw a bunch of tweets by Polo --

227
00:10:51,128 --> 00:10:54,295
Polo had been my first student
back in Buenos Aires,

228
00:10:54,319 --> 00:10:56,389
and at the time,
he was living in New York.

229
00:10:56,413 --> 00:10:58,501
And there was something in this tweets --

230
00:10:58,525 --> 00:11:02,026
I could not tell exactly what
because nothing was said explicitly --

231
00:11:02,050 --> 00:11:04,071
but I got this strong hunch,

232
00:11:04,095 --> 00:11:07,050
this strong intuition
that something was going wrong.

233
00:11:08,347 --> 00:11:11,070
So I picked up the phone
and I called Polo,

234
00:11:11,094 --> 00:11:13,013
and in fact, he was not feeling well.

235
00:11:13,362 --> 00:11:15,299
And this simple fact,

236
00:11:15,323 --> 00:11:17,814
that reading in between the lines,

237
00:11:17,838 --> 00:11:22,100
I could sense --
through words -- his feelings,

238
00:11:22,124 --> 00:11:24,743
was a simple, but very
effective way to help.

239
00:11:25,987 --> 00:11:27,625
What I tell you today

240
00:11:27,649 --> 00:11:30,157
is that we're getting
close to understanding

241
00:11:30,181 --> 00:11:34,467
how we can convert this intuition
that we all have,

242
00:11:34,491 --> 00:11:35,856
that we all share,

243
00:11:35,880 --> 00:11:37,077
into an algorithm.

244
00:11:38,102 --> 00:11:39,563
And in doing so,

245
00:11:39,587 --> 00:11:44,237
we may be seeing in the future
a very different form of mental health,

246
00:11:44,261 --> 00:11:49,882
based on objective, quantitative
and automated analysis

247
00:11:49,906 --> 00:11:51,615
of the words we write,

248
00:11:51,639 --> 00:11:53,176
of the words we say.

249
00:11:53,200 --> 00:11:54,351
Gracias.

250
00:11:54,375 --> 00:11:58,239
(Applause)