1
00:00:03,710 --> 00:00:01,850
Michael Levin's work on regulating

2
00:00:05,329 --> 00:00:03,720
intractable pattern formation in living

3
00:00:07,610 --> 00:00:05,339
systems has made him one of the most

4
00:00:09,169 --> 00:00:07,620
compelling biologists of our time in

5
00:00:10,549 --> 00:00:09,179
Translation this means that his team is

6
00:00:12,530 --> 00:00:10,559
sussing out how to develop limbs

7
00:00:14,629 --> 00:00:12,540
regenerate limbs how to generate minds

8
00:00:16,730 --> 00:00:14,639
and even life extension by manipulating

9
00:00:18,410 --> 00:00:16,740
electric signals rather than genetics or

10
00:00:20,090 --> 00:00:18,420
epigenetics his work is something that I

11
00:00:21,109 --> 00:00:20,100
consider to be worthy of a Nobel Prize

12
00:00:23,150 --> 00:00:21,119
and I don't think I've said that about

13
00:00:24,950 --> 00:00:23,160
anyone on the podcast Michael Levin's

14
00:00:26,509 --> 00:00:24,960
previous podcast on toe is in the

15
00:00:28,070 --> 00:00:26,519
description that's a solo episode with

16
00:00:30,170 --> 00:00:28,080
him where we go into a two-hour Deep

17
00:00:31,910 --> 00:00:30,180
dive as well as there's a Theo locution

18
00:00:34,010 --> 00:00:31,920
so that is him and another guest just

19
00:00:35,510 --> 00:00:34,020
like today except between Carl friston

20
00:00:37,610 --> 00:00:35,520
and Chris fields on Consciousness

21
00:00:39,590 --> 00:00:37,620
yoshibak is widely considered to be the

22
00:00:41,569 --> 00:00:39,600
Pinnacle of an AI researcher dealing

23
00:00:44,150 --> 00:00:41,579
with emotion modeling and multi-agent

24
00:00:46,190 --> 00:00:44,160
systems a large focus of boxes to build

25
00:00:48,049 --> 00:00:46,200
a model of the Mind from strong AI

26
00:00:50,029 --> 00:00:48,059
speaking of Mind Bach is one of the most

27
00:00:51,950 --> 00:00:50,039
inventive Minds in the field of computer

28
00:00:53,750 --> 00:00:51,960
science and has appeared several times

29
00:00:55,610 --> 00:00:53,760
on toll prior again there's a solo

30
00:00:57,470 --> 00:00:55,620
episode there's also a theologution

31
00:00:59,090 --> 00:00:57,480
between yoshibak and Donald Hoffman on

32
00:01:01,970 --> 00:00:59,100
Consciousness and yo shabach and John

33
00:01:03,830 --> 00:01:01,980
vervaki also on in reality biology has

34
00:01:05,690 --> 00:01:03,840
much to teach us about artificial

35
00:01:07,310 --> 00:01:05,700
intelligence and vice versa this

36
00:01:08,570 --> 00:01:07,320
discussion between two brilliant

37
00:01:09,710 --> 00:01:08,580
researchers is something that I'm

38
00:01:12,170 --> 00:01:09,720
extremely

39
00:01:14,570 --> 00:01:12,180
lucky blessed fortunate to be a part of

40
00:01:15,830 --> 00:01:14,580
as well as us as collective as an

41
00:01:17,570 --> 00:01:15,840
audience that are fortunate enough to

42
00:01:19,730 --> 00:01:17,580
witness approximately 30 minutes into

43
00:01:22,310 --> 00:01:19,740
the conversation you'll see two sponsors

44
00:01:24,410 --> 00:01:22,320
they are Masterworks and Roman I implore

45
00:01:26,330 --> 00:01:24,420
you not to skip it as firstly watching

46
00:01:27,890 --> 00:01:26,340
it supports toe directly and then

47
00:01:29,990 --> 00:01:27,900
secondly they're fascinating companies

48
00:01:31,850 --> 00:01:30,000
in and of themselves additionally you'll

49
00:01:33,890 --> 00:01:31,860
hear from one more Trade coffee around

50
00:01:35,630 --> 00:01:33,900
the one and a half hour mark thank you

51
00:01:38,929 --> 00:01:35,640
and enjoy this theolocution between

52
00:01:41,090 --> 00:01:38,939
yoshabak and Michael Levin welcome both

53
00:01:42,649 --> 00:01:41,100
Professor Michael Evan and yoshabak it's

54
00:01:44,210 --> 00:01:42,659
an honor to have you on the toll podcast

55
00:01:46,069 --> 00:01:44,220
again both of you and then together

56
00:01:47,030 --> 00:01:46,079
right now thank you it's great to be

57
00:01:51,830 --> 00:01:47,040
here

58
00:01:53,330 --> 00:01:51,840
and look forward to this conversation I

59
00:01:55,910 --> 00:01:53,340
look forward to it as well so we'll

60
00:01:57,590 --> 00:01:55,920
start off with the question of what is

61
00:01:59,749 --> 00:01:57,600
it that you Michael find most

62
00:02:01,910 --> 00:01:59,759
interesting about yosho's work and then

63
00:02:03,230 --> 00:02:01,920
Yoshi will go for you toward Michael

64
00:02:05,870 --> 00:02:03,240
yeah

65
00:02:07,249 --> 00:02:05,880
um I really enjoy the uh the breadth so

66
00:02:08,930 --> 00:02:07,259
so I've been looking I think I've

67
00:02:10,609 --> 00:02:08,940
probably read almost everything on your

68
00:02:12,830 --> 00:02:10,619
on your website you know the short kind

69
00:02:15,830 --> 00:02:12,840
of blog pieces and everything and uh

70
00:02:17,990 --> 00:02:15,840
yeah I'm I'm uh I'm I'm a big fan of of

71
00:02:19,970 --> 00:02:18,000
the breadth of of tackling a lot of the

72
00:02:22,790 --> 00:02:19,980
different issues that you do with

73
00:02:25,369 --> 00:02:22,800
respect to uh you know computation and

74
00:02:27,050 --> 00:02:25,379
and cognition and Ai and uh you know and

75
00:02:28,729 --> 00:02:27,060
ethics and everything I I really like

76
00:02:32,270 --> 00:02:28,739
that aspect of it so

77
00:02:36,470 --> 00:02:32,280
then you're sure yeah my apologies my

78
00:02:39,229 --> 00:02:36,480
blog is not up to date I haven't uh

79
00:02:40,670 --> 00:02:39,239
done any updates for a few years now on

80
00:02:44,449 --> 00:02:40,680
it I think

81
00:02:46,610 --> 00:02:44,459
um so of course I'm still in the Pro

82
00:02:50,449 --> 00:02:46,620
process of progressing and having new

83
00:02:53,630 --> 00:02:50,459
ideas and the ideas that I had in recent

84
00:02:55,430 --> 00:02:53,640
years I have a great overlap with a lot

85
00:02:59,930 --> 00:02:55,440
of the things that you are working on

86
00:03:02,030 --> 00:02:59,940
and when I listen to your Lex podcast

87
00:03:04,850 --> 00:03:02,040
last night there were many thoughts that

88
00:03:05,630 --> 00:03:04,860
you had that I had stumbled on

89
00:03:07,670 --> 00:03:05,640
um

90
00:03:09,170 --> 00:03:07,680
uh that I've never heard from anybody

91
00:03:11,030 --> 00:03:09,180
else and so I found this very

92
00:03:12,770 --> 00:03:11,040
fascinating and thought

93
00:03:14,449 --> 00:03:12,780
um maybe let's look at some of these

94
00:03:16,009 --> 00:03:14,459
thoughts first and then go from there

95
00:03:19,670 --> 00:03:16,019
and expand

96
00:03:22,610 --> 00:03:19,680
um Beyond those ideas but uh I for

97
00:03:23,869 --> 00:03:22,620
instance found after thinking about how

98
00:03:25,790 --> 00:03:23,879
cells work

99
00:03:27,530 --> 00:03:25,800
kind of obvious but missed by most

100
00:03:29,930 --> 00:03:27,540
people in Neuroscience or in science in

101
00:03:33,290 --> 00:03:29,940
general is that every cell has the

102
00:03:35,509 --> 00:03:33,300
ability to send multiple message types

103
00:03:37,190 --> 00:03:35,519
and receive multiple message types and

104
00:03:38,750 --> 00:03:37,200
do this conditionally and learn an

105
00:03:41,210 --> 00:03:38,760
average conditions to do that and to

106
00:03:43,430 --> 00:03:41,220
modulate this also every cell is an

107
00:03:45,170 --> 00:03:43,440
individual reinforcement learning agent

108
00:03:47,210 --> 00:03:45,180
single-sided animal that tries to

109
00:03:48,890 --> 00:03:47,220
survive by cooperating with its

110
00:03:50,210 --> 00:03:48,900
environment gets most of its rewards

111
00:03:53,750 --> 00:03:50,220
from its environment

112
00:03:55,789 --> 00:03:53,760
and as a result this means that every

113
00:03:58,490 --> 00:03:55,799
cell can in principle function like a

114
00:04:00,289 --> 00:03:58,500
neuron it can fulfill the same learning

115
00:04:02,690 --> 00:04:00,299
and information processing tasks as a

116
00:04:04,550 --> 00:04:02,700
new one the only difference that exists

117
00:04:07,550 --> 00:04:04,560
with respect to neurons or the main

118
00:04:09,589 --> 00:04:07,560
difference is that they cannot do this

119
00:04:12,589 --> 00:04:09,599
over very long distances because they

120
00:04:14,750 --> 00:04:12,599
are mostly connected only to the cells

121
00:04:16,729 --> 00:04:14,760
that are directly adjacent of course new

122
00:04:18,710 --> 00:04:16,739
ones also only communicate to adjacent

123
00:04:20,569 --> 00:04:18,720
cells but the adjacency of neurons is

124
00:04:22,610 --> 00:04:20,579
such that they have Excellence parts of

125
00:04:24,950 --> 00:04:22,620
the cell that reach very far through the

126
00:04:28,189 --> 00:04:24,960
organism so in some sense in your one is

127
00:04:31,310 --> 00:04:28,199
a telegraph cell that uses very specific

128
00:04:33,530 --> 00:04:31,320
messages that are encoded in a way like

129
00:04:35,330 --> 00:04:33,540
Morsi signals in extremely short high

130
00:04:37,129 --> 00:04:35,340
energy bursts that allow to send

131
00:04:40,490 --> 00:04:37,139
messages over the very long distances

132
00:04:42,830 --> 00:04:40,500
very quickly to move the muscles of an

133
00:04:43,790 --> 00:04:42,840
animal at the limit of what physics

134
00:04:45,770 --> 00:04:43,800
allows

135
00:04:48,350 --> 00:04:45,780
so it can compete with other animals and

136
00:04:50,030 --> 00:04:48,360
search for food and in order to make

137
00:04:52,010 --> 00:04:50,040
that happen it also needs to have a

138
00:04:54,110 --> 00:04:52,020
model of the world that gets updated at

139
00:04:56,330 --> 00:04:54,120
this higher rate so there is going to be

140
00:04:58,909 --> 00:04:56,340
an information processing system that is

141
00:05:00,230 --> 00:04:58,919
duplicating basically this cellular

142
00:05:03,230 --> 00:05:00,240
brain that is made from all the other

143
00:05:04,610 --> 00:05:03,240
cells in the body of the organism and at

144
00:05:06,469 --> 00:05:04,620
some point these two systems get

145
00:05:09,110 --> 00:05:06,479
decoupled they have their own codes

146
00:05:11,810 --> 00:05:09,120
their own language so to speak and but

147
00:05:14,150 --> 00:05:11,820
it still makes sense I guess to see the

148
00:05:16,670 --> 00:05:14,160
brain as a telegraphic extension of the

149
00:05:19,430 --> 00:05:16,680
community of cells in the body and for

150
00:05:21,830 --> 00:05:19,440
me this Insight that I stumbled on just

151
00:05:23,570 --> 00:05:21,840
because means and motive that Evolution

152
00:05:25,189 --> 00:05:23,580
would equip cells with doing that

153
00:05:27,770 --> 00:05:25,199
information processing if the organism

154
00:05:29,870 --> 00:05:27,780
is long enough lives long enough and if

155
00:05:31,790 --> 00:05:29,880
the cells share common genetic Destiny

156
00:05:34,189 --> 00:05:31,800
so they can get attuned to each other in

157
00:05:36,710 --> 00:05:34,199
an organism that basically every

158
00:05:39,890 --> 00:05:36,720
organism has the potential to become

159
00:05:43,550 --> 00:05:39,900
intelligent and if it gets old enough to

160
00:05:46,430 --> 00:05:43,560
process enough data to get to a very

161
00:05:49,010 --> 00:05:46,440
high degree of understanding of its

162
00:05:51,170 --> 00:05:49,020
environment in principle so of course a

163
00:05:53,810 --> 00:05:51,180
normal houseplant is not going to get

164
00:05:55,490 --> 00:05:53,820
very old with compared to us because the

165
00:05:57,890 --> 00:05:55,500
information processing is so much slower

166
00:06:00,409 --> 00:05:57,900
so they're not going to be very smart

167
00:06:02,210 --> 00:06:00,419
but at the level of ecosystems it's

168
00:06:04,310 --> 00:06:02,220
conceivable that there is quite

169
00:06:06,409 --> 00:06:04,320
considerable intelligence and then I

170
00:06:09,409 --> 00:06:06,419
stumbled on this notion that our

171
00:06:11,270 --> 00:06:09,419
ancestors thought that one day in

172
00:06:15,290 --> 00:06:11,280
Fairyland equals seven years in human

173
00:06:18,170 --> 00:06:15,300
land which is taught in the Artemis and

174
00:06:20,090 --> 00:06:18,180
also I just at some point I revised my

175
00:06:22,610 --> 00:06:20,100
notion of what the spirit is for

176
00:06:24,409 --> 00:06:22,620
instance is a spirit is it's an old bird

177
00:06:26,330 --> 00:06:24,419
for the operating system for an

178
00:06:29,029 --> 00:06:26,340
autonomous robot

179
00:06:31,010 --> 00:06:29,039
and the event this word was invented the

180
00:06:33,110 --> 00:06:31,020
um only autonomous robots that were

181
00:06:35,809 --> 00:06:33,120
known were people and plants and animals

182
00:06:38,029 --> 00:06:35,819
and nation states and ecosystems

183
00:06:41,510 --> 00:06:38,039
right whatever no robots built by people

184
00:06:44,629 --> 00:06:41,520
yet but uh there was this pattern of

185
00:06:46,550 --> 00:06:44,639
control in it that people could observe

186
00:06:48,469 --> 00:06:46,560
that was not directly tied to the

187
00:06:51,409 --> 00:06:48,479
hardware that was realized by the

188
00:06:53,150 --> 00:06:51,419
hardware but disembodied in a way and

189
00:06:55,850 --> 00:06:53,160
this notion of several is something that

190
00:06:57,830 --> 00:06:55,860
we lost after the enlightenment when we

191
00:07:00,529 --> 00:06:57,840
tried to deal with the

192
00:07:03,050 --> 00:07:00,539
um wrong Christian metaphysics and

193
00:07:04,430 --> 00:07:03,060
Superstition that came with it and threw

194
00:07:06,050 --> 00:07:04,440
out a lot of babies with the bath water

195
00:07:08,510 --> 00:07:06,060
and suddenly we basically lost a lot of

196
00:07:11,090 --> 00:07:08,520
Concepts especially this concept of

197
00:07:13,730 --> 00:07:11,100
software that existed before in a way

198
00:07:16,309 --> 00:07:13,740
this a software being a control pattern

199
00:07:17,570 --> 00:07:16,319
or a pattern of course a structure that

200
00:07:20,330 --> 00:07:17,580
exists at a certain level of course

201
00:07:22,850 --> 00:07:20,340
graining as some type of very very

202
00:07:26,230 --> 00:07:22,860
specific physical law that he exists by

203
00:07:29,990 --> 00:07:26,240
looking at reality from a certain angle

204
00:07:33,950 --> 00:07:30,000
and what I liked about your work is that

205
00:07:36,350 --> 00:07:33,960
you systematically have focused on this

206
00:07:38,270 --> 00:07:36,360
direction of what their can do that a

207
00:07:40,309 --> 00:07:38,280
cell is an agent and that levels of

208
00:07:43,730 --> 00:07:40,319
agency emerge in the interaction between

209
00:07:46,430 --> 00:07:43,740
cells and and you use a very clear

210
00:07:48,830 --> 00:07:46,440
language and uh clear Concepts and you

211
00:07:50,450 --> 00:07:48,840
obviously are driven by questions that

212
00:07:52,790 --> 00:07:50,460
you want to answer which is unusual in

213
00:07:55,730 --> 00:07:52,800
science I found that most of our

214
00:07:57,830 --> 00:07:55,740
contemporaries in science get broken it

215
00:08:01,189 --> 00:07:57,840
if it doesn't happen earlier during the

216
00:08:04,070 --> 00:08:01,199
PHD into people who apply methods in

217
00:08:05,570 --> 00:08:04,080
teams instead of people who join

218
00:08:06,830 --> 00:08:05,580
Academia because they think it's the

219
00:08:08,990 --> 00:08:06,840
most valuable thing they can do with

220
00:08:10,430 --> 00:08:09,000
their lives to pursue questions that

221
00:08:12,589 --> 00:08:10,440
they are interested and want to make

222
00:08:14,089 --> 00:08:12,599
progress on

223
00:08:16,790 --> 00:08:14,099
all right Michael there's plenty to

224
00:08:17,450 --> 00:08:16,800
respond to yeah yeah lots of lots of

225
00:08:19,550 --> 00:08:17,460
ideas

226
00:08:21,409 --> 00:08:19,560
um yeah I think I think it's very uh

227
00:08:23,150 --> 00:08:21,419
your point is very interesting about you

228
00:08:24,589 --> 00:08:23,160
know what what it really what really

229
00:08:26,110 --> 00:08:24,599
fundamentally is the difference between

230
00:08:28,969 --> 00:08:26,120
neurons and other cells of course

231
00:08:30,589 --> 00:08:28,979
evolutionarily they're reusing Machinery

232
00:08:32,209 --> 00:08:30,599
that has been around for a very long

233
00:08:34,010 --> 00:08:32,219
time since the time of bacteria

234
00:08:36,469 --> 00:08:34,020
basically right so our multi our

235
00:08:38,990 --> 00:08:36,479
unicellular ancestors had a lot of the

236
00:08:40,790 --> 00:08:39,000
same machinery and and even I mean of

237
00:08:43,850 --> 00:08:40,800
course axons are very can be very long

238
00:08:45,350 --> 00:08:43,860
but but uh there are sort of

239
00:08:47,150 --> 00:08:45,360
intermediate structures right there are

240
00:08:49,790 --> 00:08:47,160
tunneling nanotubes and things that

241
00:08:51,650 --> 00:08:49,800
allow cells to connect to maybe five or

242
00:08:53,690 --> 00:08:51,660
ten diameter cell diameters away right

243
00:08:55,610 --> 00:08:53,700
so so not terribly long but but also not

244
00:08:57,710 --> 00:08:55,620
immediate neighbors necessarily so that

245
00:08:59,930 --> 00:08:57,720
kind of architecture has been around for

246
00:09:00,889 --> 00:08:59,940
a while and people like roll so well

247
00:09:01,790 --> 00:09:00,899
look at

248
00:09:04,130 --> 00:09:01,800
um

249
00:09:06,530 --> 00:09:04,140
but a very brain-like electrical

250
00:09:09,769 --> 00:09:06,540
signaling in bacterial Colony so this is

251
00:09:12,769 --> 00:09:09,779
you know I think I think Evolution began

252
00:09:15,290 --> 00:09:12,779
to reuse this toolkit specifically of

253
00:09:16,550 --> 00:09:15,300
using this kind of communication to

254
00:09:18,290 --> 00:09:16,560
scale up

255
00:09:22,070 --> 00:09:18,300
um with the computational and other

256
00:09:24,050 --> 00:09:22,080
kinds of um other kinds of uh tricks uh

257
00:09:25,850 --> 00:09:24,060
yeah a really long time ago you know and

258
00:09:28,670 --> 00:09:25,860
I like to I like to imagine that if

259
00:09:30,230 --> 00:09:28,680
somebody had come to the people who were

260
00:09:32,389 --> 00:09:30,240
inventing connectionism and the first

261
00:09:33,949 --> 00:09:32,399
you know sort of perceptrons and neural

262
00:09:35,990 --> 00:09:33,959
networks and so on if somebody had come

263
00:09:37,910 --> 00:09:36,000
to them and said oh by the way sorry you

264
00:09:39,590 --> 00:09:37,920
know we're with the biologists we we got

265
00:09:41,509 --> 00:09:39,600
it wrong it's not the thinking isn't in

266
00:09:42,889 --> 00:09:41,519
the brain it's in the liver and so then

267
00:09:44,389 --> 00:09:42,899
the question is what would they do right

268
00:09:46,730 --> 00:09:44,399
what would they have changed anything

269
00:09:48,470 --> 00:09:46,740
about what they're doing and we said ah

270
00:09:50,210 --> 00:09:48,480
now we have to rethink our model or with

271
00:09:51,590 --> 00:09:50,220
is that fine who cares uh this is

272
00:09:53,750 --> 00:09:51,600
exactly the same the same model

273
00:09:55,130 --> 00:09:53,760
everything works just as well so I often

274
00:09:58,009 --> 00:09:55,140
I often think about that that that

275
00:10:01,190 --> 00:09:58,019
question what exactly do we mean by by

276
00:10:03,650 --> 00:10:01,200
neurons and isn't it interesting that we

277
00:10:07,250 --> 00:10:03,660
are able to to steal most of the tools

278
00:10:09,590 --> 00:10:07,260
the concepts the Frameworks the the math

279
00:10:12,110 --> 00:10:09,600
from neuroscience and apply it to

280
00:10:13,670 --> 00:10:12,120
problems in other spaces so not movement

281
00:10:15,170 --> 00:10:13,680
in three-dimensional space with muscles

282
00:10:16,850 --> 00:10:15,180
but for example movement through

283
00:10:18,829 --> 00:10:16,860
amorphous space right anatomical

284
00:10:20,329 --> 00:10:18,839
amorphous space the techniques can't

285
00:10:22,009 --> 00:10:20,339
tell the difference we use all the same

286
00:10:25,430 --> 00:10:22,019
stuff optogenetics neurotransmitter

287
00:10:28,850 --> 00:10:25,440
signaling we model active inference and

288
00:10:30,650 --> 00:10:28,860
uh we see perceptual by stability you

289
00:10:32,630 --> 00:10:30,660
name it you know we take Concepts from

290
00:10:35,389 --> 00:10:32,640
from neuroscience and we apply it out

291
00:10:37,850 --> 00:10:35,399
elsewhere in the body and uh generally

292
00:10:40,190 --> 00:10:37,860
speaking uh everything works exactly the

293
00:10:41,810 --> 00:10:40,200
same and that shows this I think what

294
00:10:45,949 --> 00:10:41,820
you were saying that there's this really

295
00:10:48,170 --> 00:10:45,959
interesting uh kind of um symmetry

296
00:10:49,310 --> 00:10:48,180
between these these uh that that a lot

297
00:10:50,930 --> 00:10:49,320
of the distinctions that we've been

298
00:10:52,550 --> 00:10:50,940
making are you know in terms of having

299
00:10:54,110 --> 00:10:52,560
different departments and different PHD

300
00:10:55,430 --> 00:10:54,120
programs and other things that say you

301
00:10:57,110 --> 00:10:55,440
know this is neuroscience this is

302
00:11:00,110 --> 00:10:57,120
developmental biology a lot of these

303
00:11:03,829 --> 00:11:00,120
things are just not um uh not as not as

304
00:11:06,350 --> 00:11:03,839
firm distinctions as we used to think

305
00:11:08,329 --> 00:11:06,360
yeah I suspect that people who insist on

306
00:11:12,170 --> 00:11:08,339
strong disciplinary boundaries do this

307
00:11:13,850 --> 00:11:12,180
out of of a protective impulse and what

308
00:11:16,850 --> 00:11:13,860
I noticed by I was studying many

309
00:11:19,190 --> 00:11:16,860
disciplines when I was young that

310
00:11:22,250 --> 00:11:19,200
um the different methodologies are so

311
00:11:23,870 --> 00:11:22,260
incompatible across fields that when I

312
00:11:26,150 --> 00:11:23,880
was studying philosophy

313
00:11:28,190 --> 00:11:26,160
um by or psychology I felt that computer

314
00:11:30,290 --> 00:11:28,200
scientists would be laughing about the

315
00:11:33,490 --> 00:11:30,300
methods that each of these fields are

316
00:11:36,650 --> 00:11:33,500
using to justify what they're doing and

317
00:11:39,650 --> 00:11:36,660
this I think is indicative of a defect

318
00:11:42,590 --> 00:11:39,660
because if you take science into the

319
00:11:44,690 --> 00:11:42,600
current regime of regulating it entirely

320
00:11:46,670 --> 00:11:44,700
by peer review there is no external

321
00:11:48,650 --> 00:11:46,680
Authority even the grand authorities are

322
00:11:51,170 --> 00:11:48,660
mostly fields or people who have been

323
00:11:52,970 --> 00:11:51,180
trained in The Sciences in existing

324
00:11:54,710 --> 00:11:52,980
paradigms and then are finding the

325
00:11:55,430 --> 00:11:54,720
continuation of those paradigms from the

326
00:11:57,250 --> 00:11:55,440
outside

327
00:12:00,110 --> 00:11:57,260
this um

328
00:12:02,210 --> 00:12:00,120
meta paradigmatic thinking it does not

329
00:12:04,550 --> 00:12:02,220
really exist that much in a

330
00:12:05,990 --> 00:12:04,560
peer-reviewed paradigm and ultimately

331
00:12:08,090 --> 00:12:06,000
when you do peer review for a couple

332
00:12:09,769 --> 00:12:08,100
Generations it also means that if your

333
00:12:11,329 --> 00:12:09,779
peers deteriorate there is nothing who

334
00:12:14,690 --> 00:12:11,339
pulls your science back

335
00:12:16,610 --> 00:12:14,700
and uh what I missed specifically in a

336
00:12:18,170 --> 00:12:16,620
lot of the way which Neuroscience is

337
00:12:19,430 --> 00:12:18,180
done is what you call the engineering

338
00:12:22,130 --> 00:12:19,440
stance

339
00:12:23,509 --> 00:12:22,140
and this engineering sense is very

340
00:12:25,130 --> 00:12:23,519
powerful and you get it automatically

341
00:12:26,930 --> 00:12:25,140
when you're a computer scientist because

342
00:12:28,850 --> 00:12:26,940
you don't really care about language is

343
00:12:31,430 --> 00:12:28,860
it written in what you care in is what

344
00:12:34,069 --> 00:12:31,440
causal pattern is realized and how can

345
00:12:36,410 --> 00:12:34,079
this be realized and how could I do it

346
00:12:38,389 --> 00:12:36,420
how would I do it how can Evolution do

347
00:12:40,850 --> 00:12:38,399
it what it means are it is disposal and

348
00:12:42,590 --> 00:12:40,860
this determines the search space for the

349
00:12:44,750 --> 00:12:42,600
things that I'm looking for but this

350
00:12:47,889 --> 00:12:44,760
requires that I think in causal systems

351
00:12:50,690 --> 00:12:47,899
and this thinking and causal systems is

352
00:12:52,430 --> 00:12:50,700
not impossible to it's possible not to

353
00:12:54,829 --> 00:12:52,440
do for a computer scientist

354
00:12:56,990 --> 00:12:54,839
but it is unusual outside of computer

355
00:12:59,210 --> 00:12:57,000
science and once you realize that it's

356
00:13:01,069 --> 00:12:59,220
it's very weird and suddenly you have

357
00:13:04,009 --> 00:13:01,079
Notions that try to replace causal

358
00:13:06,470 --> 00:13:04,019
structure with say evidence and then you

359
00:13:09,290 --> 00:13:06,480
notice that uh for instance evidence

360
00:13:11,030 --> 00:13:09,300
medicine is is not about uh

361
00:13:12,769 --> 00:13:11,040
probabilities of how something is

362
00:13:14,509 --> 00:13:12,779
realized and must work like you see

363
00:13:16,670 --> 00:13:14,519
people on the cruise ship getting

364
00:13:19,129 --> 00:13:16,680
infected over distances and you think oh

365
00:13:20,629 --> 00:13:19,139
this must be upward but no there is no

366
00:13:22,790 --> 00:13:20,639
peer control study so there is no

367
00:13:24,410 --> 00:13:22,800
evidence that it's airport and when you

368
00:13:26,090 --> 00:13:24,420
look at resistive planes from the

369
00:13:29,389 --> 00:13:26,100
outside like in this case the medical

370
00:13:32,269 --> 00:13:29,399
profession or the medical messaging and

371
00:13:35,150 --> 00:13:32,279
decision making or um I get terrified

372
00:13:37,190 --> 00:13:35,160
because it directly affects us and in

373
00:13:39,170 --> 00:13:37,200
terms of Neuroscience um of course

374
00:13:41,750 --> 00:13:39,180
there's more theoretical for the most

375
00:13:43,850 --> 00:13:41,760
part but there must be a reason why it's

376
00:13:46,910 --> 00:13:43,860
for the most part a theoretical because

377
00:13:49,670 --> 00:13:46,920
why there is no causal model that

378
00:13:51,949 --> 00:13:49,680
clinicians can use to explain uh what is

379
00:13:54,470 --> 00:13:51,959
happening in certain syndromes that

380
00:13:58,550 --> 00:13:54,480
people are exhibiting and I noticed this

381
00:14:01,069 --> 00:13:58,560
when I go to a doctor and even at a

382
00:14:02,870 --> 00:14:01,079
reputable institution like Stanford that

383
00:14:04,990 --> 00:14:02,880
most of the neuroscientists at some

384
00:14:07,009 --> 00:14:05,000
level there are of most of the

385
00:14:09,889 --> 00:14:07,019
neurologists that I'm talking towards us

386
00:14:13,790 --> 00:14:09,899
at some level dual lists that they uh

387
00:14:15,829 --> 00:14:13,800
don't have causal model of of the way in

388
00:14:18,530 --> 00:14:15,839
which the brain is realizing things and

389
00:14:21,050 --> 00:14:18,540
a lot of studies which discover that

390
00:14:22,910 --> 00:14:21,060
very simple mechanisms like the ability

391
00:14:24,949 --> 00:14:22,920
of human beings to use grammatical

392
00:14:26,750 --> 00:14:24,959
structure are actually reflected in the

393
00:14:28,550 --> 00:14:26,760
brain this is so amazing would have

394
00:14:31,870 --> 00:14:28,560
thought

395
00:14:35,269 --> 00:14:33,650
developments that existed in computer

396
00:14:36,769 --> 00:14:35,279
science have letters on a completely

397
00:14:40,430 --> 00:14:36,779
different track there is uh the

398
00:14:42,350 --> 00:14:40,440
perceptron is vaguely inspired by what

399
00:14:44,449 --> 00:14:42,360
the brain might be doing but I think

400
00:14:46,850 --> 00:14:44,459
it's it's really a toy model or a

401
00:14:48,650 --> 00:14:46,860
caricature of what cells are doing and

402
00:14:51,530 --> 00:14:48,660
not in the sense that it's inferior it's

403
00:14:54,110 --> 00:14:51,540
amazing what you can put forth with the

404
00:14:56,210 --> 00:14:54,120
modern perceptron variations right the

405
00:14:59,090 --> 00:14:56,220
current machine learning systems are

406
00:15:00,350 --> 00:14:59,100
mind-blowing and what they can do but

407
00:15:02,210 --> 00:15:00,360
they don't do it like biological

408
00:15:05,509 --> 00:15:02,220
organisms at all it's it's very

409
00:15:09,410 --> 00:15:05,519
different cells do not form change in

410
00:15:12,170 --> 00:15:09,420
which they wait uh sums of real numbers

411
00:15:14,210 --> 00:15:12,180
that there is something going on that is

412
00:15:15,769 --> 00:15:14,220
roughly similar to it but there's a

413
00:15:17,750 --> 00:15:15,779
self-organizing system that designs

414
00:15:19,370 --> 00:15:17,760
itself from the inside out not by a

415
00:15:21,410 --> 00:15:19,380
measure learning principle that applies

416
00:15:23,269 --> 00:15:21,420
to the outside and updates weights after

417
00:15:25,490 --> 00:15:23,279
reading and comparing them and Computing

418
00:15:27,949 --> 00:15:25,500
gradients to the system so this

419
00:15:30,650 --> 00:15:27,959
perspective of local surf organization

420
00:15:33,769 --> 00:15:30,660
by reinforcement agents that try to

421
00:15:35,090 --> 00:15:33,779
trade rewards with each other that is a

422
00:15:37,490 --> 00:15:35,100
perspective that I find totally

423
00:15:40,670 --> 00:15:37,500
fascinating and I wish this would have

424
00:15:43,009 --> 00:15:40,680
come and from Neuroscience into computer

425
00:15:44,449 --> 00:15:43,019
science but it hasn't there are some

426
00:15:46,310 --> 00:15:44,459
people which have thought about these

427
00:15:48,769 --> 00:15:46,320
ideas to some degree but there's been

428
00:15:50,990 --> 00:15:48,779
very little cross-pollination and I

429
00:15:53,090 --> 00:15:51,000
think all this talk of Neuroscience uh

430
00:15:55,389 --> 00:15:53,100
influencing computer science is mostly

431
00:15:58,910 --> 00:15:55,399
visual thinking

432
00:16:00,590 --> 00:15:58,920
yeah it's also uh I find uh this you

433
00:16:02,509 --> 00:16:00,600
know what you were saying about the

434
00:16:05,629 --> 00:16:02,519
different disciplines it's it's kind of

435
00:16:06,470 --> 00:16:05,639
amazing how uh well when I when I give a

436
00:16:07,910 --> 00:16:06,480
talk

437
00:16:10,189 --> 00:16:07,920
um I can always tell which department

438
00:16:12,470 --> 00:16:10,199
I'm in by by which part of the talk

439
00:16:13,910 --> 00:16:12,480
makes people uncomfortable and upset and

440
00:16:15,590 --> 00:16:13,920
it's always different depending on which

441
00:16:16,910 --> 00:16:15,600
department it is right so so there are

442
00:16:18,710 --> 00:16:16,920
things you can say in one Department

443
00:16:19,970 --> 00:16:18,720
that are completely obvious and you say

444
00:16:21,470 --> 00:16:19,980
this in the in another group of people

445
00:16:22,970 --> 00:16:21,480
and they they throw tomatoes they think

446
00:16:26,750 --> 00:16:22,980
this is just crazy for instance

447
00:16:28,730 --> 00:16:26,760
uh for for instance um I could I could

448
00:16:31,189 --> 00:16:28,740
uh say in you know in in a neuroscience

449
00:16:34,250 --> 00:16:31,199
Department I could say uh information

450
00:16:35,810 --> 00:16:34,260
can be processed is without changes in

451
00:16:36,949 --> 00:16:35,820
gene expression you don't need changes

452
00:16:38,769 --> 00:16:36,959
in gene expression to process

453
00:16:40,850 --> 00:16:38,779
information because the the the

454
00:16:42,530 --> 00:16:40,860
processing inside a neural network that

455
00:16:44,990 --> 00:16:42,540
runs on the physics of of action

456
00:16:46,370 --> 00:16:45,000
potentials right so I mean so right so

457
00:16:47,689 --> 00:16:46,380
so you can do all kinds of interesting

458
00:16:48,949 --> 00:16:47,699
information processing and you don't

459
00:16:50,269 --> 00:16:48,959
need genetic change sort of

460
00:16:51,829 --> 00:16:50,279
transcriptional or genetic change for

461
00:16:53,269 --> 00:16:51,839
that if I say the same thing in a

462
00:16:55,249 --> 00:16:53,279
molecular genetics department that say

463
00:16:57,050 --> 00:16:55,259
hey these cells could be processing tons

464
00:16:58,670 --> 00:16:57,060
of information long before the the

465
00:17:00,590 --> 00:16:58,680
transcriptome ever finds out about it

466
00:17:02,329 --> 00:17:00,600
this is this is considered you know just

467
00:17:04,970 --> 00:17:02,339
just completely wild because because

468
00:17:06,590 --> 00:17:04,980
it's thought that that that most of the

469
00:17:08,510 --> 00:17:06,600
hard work or in fact all of the hard

470
00:17:10,309 --> 00:17:08,520
work is done in in gene regulatory

471
00:17:12,770 --> 00:17:10,319
circuits and things like that right

472
00:17:15,770 --> 00:17:12,780
there are other other examples if I say

473
00:17:17,870 --> 00:17:15,780
um uh here's a uh here's a collection of

474
00:17:20,329 --> 00:17:17,880
cells that communicate electrically to

475
00:17:24,289 --> 00:17:20,339
to remember a particular spatial pattern

476
00:17:25,490 --> 00:17:24,299
again molecular cell biology that's what

477
00:17:27,409 --> 00:17:25,500
do you mean how can how can a collection

478
00:17:29,270 --> 00:17:27,419
of cells remember a spatial pattern but

479
00:17:30,530 --> 00:17:29,280
again in Neuroscience or in in an

480
00:17:32,690 --> 00:17:30,540
engineering department yeah of course

481
00:17:34,190 --> 00:17:32,700
the the of course they have electrical

482
00:17:36,110 --> 00:17:34,200
circuits that remember patterns and can

483
00:17:38,090 --> 00:17:36,120
do pattern completion and things like

484
00:17:40,190 --> 00:17:38,100
that so

485
00:17:42,110 --> 00:17:40,200
um you know views of views of causality

486
00:17:43,730 --> 00:17:42,120
views of uh just just lots of things

487
00:17:46,430 --> 00:17:43,740
like that that are that are very obvious

488
00:17:49,430 --> 00:17:46,440
uh to one group of people is completely

489
00:17:51,770 --> 00:17:49,440
taboo elsewhere so that that that

490
00:17:53,990 --> 00:17:51,780
distinction and and and and yeah and as

491
00:17:56,510 --> 00:17:54,000
as you as Joshua just said

492
00:17:58,549 --> 00:17:56,520
um it impacts it impacts everything it

493
00:18:01,970 --> 00:17:58,559
impacts education it impacts grants you

494
00:18:03,770 --> 00:18:01,980
know Grant reviews because when these

495
00:18:06,110 --> 00:18:03,780
kind of interdisciplinary grants come up

496
00:18:07,549 --> 00:18:06,120
the study sections have a really hard

497
00:18:09,409 --> 00:18:07,559
time finding people that can actually

498
00:18:11,630 --> 00:18:09,419
review them because what often happens

499
00:18:14,450 --> 00:18:11,640
is you'll find you'll you'll get a some

500
00:18:16,370 --> 00:18:14,460
kind of computational biology Grant and

501
00:18:17,990 --> 00:18:16,380
you put a proposal and you'll have some

502
00:18:19,850 --> 00:18:18,000
people on the on the panel who are

503
00:18:21,950 --> 00:18:19,860
biologists and some people who are the

504
00:18:23,510 --> 00:18:21,960
computational folks and it's very hard

505
00:18:25,070 --> 00:18:23,520
to get people that actually can

506
00:18:27,049 --> 00:18:25,080
appreciate both sides of it and

507
00:18:28,850 --> 00:18:27,059
understand what's happening together

508
00:18:30,409 --> 00:18:28,860
right so they will sort of each critique

509
00:18:32,570 --> 00:18:30,419
a certain part of it and the other part

510
00:18:35,390 --> 00:18:32,580
they say I don't know what this is and

511
00:18:37,310 --> 00:18:35,400
and as a result grants like that don't

512
00:18:39,230 --> 00:18:37,320
tend to not have a champion you know one

513
00:18:40,549 --> 00:18:39,240
person who can say no I get I get the

514
00:18:42,049 --> 00:18:40,559
whole thing and I I think it's really

515
00:18:44,990 --> 00:18:42,059
good or not

516
00:18:47,510 --> 00:18:45,000
so yeah it's uh even to the point where

517
00:18:49,970 --> 00:18:47,520
I I'm often asked you know when when

518
00:18:52,190 --> 00:18:49,980
people want to you list me somewhere

519
00:18:53,750 --> 00:18:52,200
they'll say so so what what what are you

520
00:18:55,850 --> 00:18:53,760
you know what kind of what kind of uh

521
00:18:56,990 --> 00:18:55,860
what's your field and I I never know how

522
00:18:58,250 --> 00:18:57,000
to answer that question you know this

523
00:18:59,870 --> 00:18:58,260
day it's been 30 years I still don't

524
00:19:01,909 --> 00:18:59,880
know how to answer that question I I

525
00:19:03,529 --> 00:19:01,919
just can't boil it down to one you know

526
00:19:06,110 --> 00:19:03,539
it just wouldn't make any sense to to

527
00:19:07,909 --> 00:19:06,120
say any of the traditional um you know

528
00:19:09,830 --> 00:19:07,919
the traditional fields

529
00:19:11,630 --> 00:19:09,840
so what do you say yosha when someone

530
00:19:13,970 --> 00:19:11,640
asks you what field you're in

531
00:19:18,049 --> 00:19:13,980
and it depends on who's asking

532
00:19:20,630 --> 00:19:18,059
so for instance I found it quite useful

533
00:19:23,150 --> 00:19:20,640
to sometimes say and

534
00:19:23,990 --> 00:19:23,160
sorry I'm not a philosopher

535
00:19:26,690 --> 00:19:24,000
um

536
00:19:28,970 --> 00:19:26,700
but uh but this or

537
00:19:31,610 --> 00:19:28,980
um I'm not that interested in machine

538
00:19:33,830 --> 00:19:31,620
learning and I'm I did publish papers in

539
00:19:36,409 --> 00:19:33,840
philosophy and in machine learning but

540
00:19:39,289 --> 00:19:36,419
it's not my specialty in the sense that

541
00:19:41,210 --> 00:19:39,299
I need to identify with it and

542
00:19:43,610 --> 00:19:41,220
um in some sense I guess that these

543
00:19:46,610 --> 00:19:43,620
categories are important when you try to

544
00:19:48,409 --> 00:19:46,620
write a grant proposal or when you try

545
00:19:50,450 --> 00:19:48,419
to find a job in a particular

546
00:19:53,450 --> 00:19:50,460
institution and they need to fill a

547
00:19:55,190 --> 00:19:53,460
position but for me it's more what

548
00:19:56,570 --> 00:19:55,200
questions am I interested and what is

549
00:19:58,010 --> 00:19:56,580
the thing that I want to make progress

550
00:19:59,330 --> 00:19:58,020
on or what is the thing that I want to

551
00:20:03,710 --> 00:19:59,340
build right now

552
00:20:07,250 --> 00:20:03,720
and uh I guess that in terms of the

553
00:20:10,130 --> 00:20:07,260
intersection I'm a cognitive scientist

554
00:20:12,110 --> 00:20:10,140
so I was asking Michael prior to you

555
00:20:14,210 --> 00:20:12,120
joining yosha why is that Michael that

556
00:20:15,409 --> 00:20:14,220
you were doing podcasts and if I

557
00:20:16,970 --> 00:20:15,419
understand correctly part of the reason

558
00:20:18,289 --> 00:20:16,980
was because you think out loud and you

559
00:20:20,390 --> 00:20:18,299
like to hear the other person's thoughts

560
00:20:21,950 --> 00:20:20,400
and take notes and it Spurs your own and

561
00:20:23,510 --> 00:20:21,960
firstly like Michael you can correct me

562
00:20:25,730 --> 00:20:23,520
if that's incorrect and then secondly

563
00:20:27,710 --> 00:20:25,740
yoksha I'm curious for an answer for

564
00:20:30,049 --> 00:20:27,720
this the same question what is it that

565
00:20:32,510 --> 00:20:30,059
you get out of doing podcasts other than

566
00:20:33,710 --> 00:20:32,520
say some marketing for if you were

567
00:20:35,510 --> 00:20:33,720
promoting something which I don't

568
00:20:39,230 --> 00:20:35,520
imagine you are currently

569
00:20:41,990 --> 00:20:39,240
uh no I'm not bucketing anything I uh

570
00:20:46,130 --> 00:20:42,000
what I like about podcasts is the

571
00:20:48,289 --> 00:20:46,140
ability to publish something in a format

572
00:20:50,450 --> 00:20:48,299
that is engaging to an interesting to

573
00:20:54,409 --> 00:20:50,460
people who actually care about it

574
00:20:56,270 --> 00:20:54,419
I like this informal way of uh holding

575
00:20:58,310 --> 00:20:56,280
on to some ideas and also like

576
00:21:01,930 --> 00:20:58,320
conversations is a medium to develop

577
00:21:04,190 --> 00:21:01,940
thought is this space in which we can

578
00:21:06,710 --> 00:21:04,200
reflect on each other look into each

579
00:21:09,650 --> 00:21:06,720
other's minds interact with the ideas of

580
00:21:11,990 --> 00:21:09,660
others in real time uh the production

581
00:21:13,730 --> 00:21:12,000
format of a podcast creates a certain

582
00:21:16,190 --> 00:21:13,740
focus of the conversation that can be

583
00:21:19,370 --> 00:21:16,200
useful and it's a pleasant kind of

584
00:21:23,690 --> 00:21:19,380
tangent that focuses you to stay on task

585
00:21:25,970 --> 00:21:23,700
and I also found that uh it's genuinely

586
00:21:28,549 --> 00:21:25,980
useful to some people like uh with the

587
00:21:31,370 --> 00:21:28,559
feedback that I get is that uh people

588
00:21:34,010 --> 00:21:31,380
tell me I had this really important

589
00:21:36,169 --> 00:21:34,020
question and I found this allowed me to

590
00:21:38,810 --> 00:21:36,179
make a progress on it and I feel much

591
00:21:41,270 --> 00:21:38,820
better now about these questions I uh

592
00:21:43,250 --> 00:21:41,280
just clarified something for me that has

593
00:21:45,770 --> 00:21:43,260
plagued me for years and

594
00:21:48,289 --> 00:21:45,780
um put me on track to solving it or this

595
00:21:50,570 --> 00:21:48,299
is inspired the following work so it's a

596
00:21:53,330 --> 00:21:50,580
form of uh publishing ideas and getting

597
00:21:56,450 --> 00:21:53,340
them into circulation in our Global

598
00:21:59,390 --> 00:21:56,460
Health Minds that is very informal in a

599
00:22:02,750 --> 00:21:59,400
way but it's not useless and also it's

600
00:22:05,090 --> 00:22:02,760
uh believes me in in this instance at

601
00:22:07,250 --> 00:22:05,100
least of the work of cutting editing and

602
00:22:09,409 --> 00:22:07,260
so on uh but anyway so I'm very grateful

603
00:22:12,169 --> 00:22:09,419
that you provide the service of curating

604
00:22:14,990 --> 00:22:12,179
our conversation and putting a little

605
00:22:17,270 --> 00:22:15,000
form that is useful to other people

606
00:22:19,250 --> 00:22:17,280
yeah yeah there's something well the two

607
00:22:21,409 --> 00:22:19,260
two things I was thinking of one is that

608
00:22:22,970 --> 00:22:21,419
uh you know I mean I have conversations

609
00:22:24,230 --> 00:22:22,980
with people all day long about these

610
00:22:26,330 --> 00:22:24,240
issues right so people in my lab

611
00:22:27,830 --> 00:22:26,340
collaborators whatever and most of

612
00:22:29,630 --> 00:22:27,840
course the vast majority of those

613
00:22:31,250 --> 00:22:29,640
conversations are not recorded and they

614
00:22:33,230 --> 00:22:31,260
just sort of disappear into the ether

615
00:22:34,250 --> 00:22:33,240
and then I I take something away from it

616
00:22:36,649 --> 00:22:34,260
and the other person takes something

617
00:22:38,149 --> 00:22:36,659
away from it but I've often thought that

618
00:22:40,669 --> 00:22:38,159
wouldn't it be wouldn't isn't it a shame

619
00:22:42,649 --> 00:22:40,679
that all of this is uh just just kind of

620
00:22:44,450 --> 00:22:42,659
disappears and it would be amazing to to

621
00:22:45,950 --> 00:22:44,460
have have a record of it and of course

622
00:22:48,649 --> 00:22:45,960
not every conversation is you know as

623
00:22:51,110 --> 00:22:48,659
gold but but but a lot of them are are

624
00:22:53,510 --> 00:22:51,120
useful and interesting and uh there are

625
00:22:55,370 --> 00:22:53,520
plenty of people that um could could be

626
00:22:57,590 --> 00:22:55,380
interested in could could benefit from

627
00:22:59,510 --> 00:22:57,600
it so so I really like this aspect that

628
00:23:00,830 --> 00:22:59,520
uh we can have conversations and then

629
00:23:03,230 --> 00:23:00,840
they are you know sort of canned and

630
00:23:04,730 --> 00:23:03,240
they're out there for uh for for people

631
00:23:06,830 --> 00:23:04,740
who are who are interested the other

632
00:23:08,570 --> 00:23:06,840
kind of um aspect of it which I don't

633
00:23:11,390 --> 00:23:08,580
really understand but but it's kind of

634
00:23:13,610 --> 00:23:11,400
neat is that if when somebody when

635
00:23:16,669 --> 00:23:13,620
somebody asks me to uh pre-record a talk

636
00:23:19,010 --> 00:23:16,679
it takes crazy amount of time because

637
00:23:20,330 --> 00:23:19,020
because I keep stopping and realizing I

638
00:23:21,950 --> 00:23:20,340
could have said that better let me start

639
00:23:23,990 --> 00:23:21,960
from the beginning and and it's just

640
00:23:26,450 --> 00:23:24,000
it's it's an incredible ordeal whereas

641
00:23:29,210 --> 00:23:26,460
something like this that's real time uh

642
00:23:30,289 --> 00:23:29,220
I'm sure has as many you know mistakes

643
00:23:31,730 --> 00:23:30,299
and things that I would have rather

644
00:23:33,289 --> 00:23:31,740
fixed later but but you can't do that

645
00:23:34,549 --> 00:23:33,299
right so you just sort of go with it and

646
00:23:36,470 --> 00:23:34,559
that's it and then it's done and you can

647
00:23:38,930 --> 00:23:36,480
move on so so I like I like that

648
00:23:40,490 --> 00:23:38,940
real-time aspect of it because it just

649
00:23:42,470 --> 00:23:40,500
it just helps you to get the ideas out

650
00:23:45,169 --> 00:23:42,480
without getting hung up and and trying

651
00:23:46,789 --> 00:23:45,179
to redo things 50 times

652
00:23:50,930 --> 00:23:46,799
yeah it's a format that allows

653
00:23:54,470 --> 00:23:50,940
tentativity if we have published we have

654
00:23:56,450 --> 00:23:54,480
culture in sciences that requires us to

655
00:23:58,610 --> 00:23:56,460
publish the things that we can hope to

656
00:24:00,649 --> 00:23:58,620
prove and make the best proof that we

657
00:24:02,630 --> 00:24:00,659
can but when we have anything

658
00:24:05,690 --> 00:24:02,640
complicated especially when we take our

659
00:24:07,130 --> 00:24:05,700
engineering stance we often cannot prove

660
00:24:10,070 --> 00:24:07,140
how things work

661
00:24:11,750 --> 00:24:10,080
instead our answers are in the realm of

662
00:24:14,750 --> 00:24:11,760
the possible and we need to discuss the

663
00:24:16,970 --> 00:24:14,760
possibilities and there is a value in

664
00:24:19,130 --> 00:24:16,980
understanding these possibilities to

665
00:24:21,409 --> 00:24:19,140
direct our future experiments and the

666
00:24:23,810 --> 00:24:21,419
Practical work that we do to see what's

667
00:24:25,850 --> 00:24:23,820
actually the case and we don't really

668
00:24:29,270 --> 00:24:25,860
have a publication format for that right

669
00:24:30,770 --> 00:24:29,280
we don't get neuroscientists to publish

670
00:24:32,270 --> 00:24:30,780
their ideas on how the mind works

671
00:24:34,310 --> 00:24:32,280
because nobody has a theory that they

672
00:24:35,930 --> 00:24:34,320
can prove and as a result there is

673
00:24:37,850 --> 00:24:35,940
basically a vacuum where theories should

674
00:24:40,370 --> 00:24:37,860
be and the theory building happens

675
00:24:41,930 --> 00:24:40,380
informally in conversations that

676
00:24:45,169 --> 00:24:41,940
basically requires personal contact

677
00:24:46,850 --> 00:24:45,179
which is a big issue once conference has

678
00:24:49,370 --> 00:24:46,860
been virtual because that contact

679
00:24:50,690 --> 00:24:49,380
diminished and you get a lot of

680
00:24:53,450 --> 00:24:50,700
important ideas by reading the

681
00:24:55,370 --> 00:24:53,460
Publications and so on but this uh what

682
00:24:57,289 --> 00:24:55,380
could be or connecting the dots or

683
00:24:59,029 --> 00:24:57,299
possibilities or ideas that might be

684
00:25:01,490 --> 00:24:59,039
proven wrong later that we just exchange

685
00:25:03,049 --> 00:25:01,500
as in the status of ideas that is

686
00:25:04,850 --> 00:25:03,059
something that has a good place and a

687
00:25:11,029 --> 00:25:04,860
podcast

688
00:25:13,190 --> 00:25:11,039
something new so for instance I was

689
00:25:14,870 --> 00:25:13,200
thinking about this and I well podcasts

690
00:25:16,850 --> 00:25:14,880
go back a while and Brogan invented this

691
00:25:19,970 --> 00:25:16,860
long-form format or popularized it

692
00:25:21,649 --> 00:25:19,980
however on television there are

693
00:25:23,630 --> 00:25:21,659
interviews so there's Oprah and those

694
00:25:26,570 --> 00:25:23,640
are long one hour they're 60 minutes and

695
00:25:28,370 --> 00:25:26,580
then back in the 90s there was a three

696
00:25:30,169 --> 00:25:28,380
and a half hour it's essentially a

697
00:25:32,029 --> 00:25:30,179
podcast it's like Charlie Rose three and

698
00:25:34,669 --> 00:25:32,039
a half hour conversation it's like a

699
00:25:38,149 --> 00:25:34,679
Theo location with Freeman Dyson Daniel

700
00:25:41,330 --> 00:25:38,159
Dennett Stephen J Gould like the Rupert

701
00:25:43,370 --> 00:25:41,340
sheldrake all of those on the same one

702
00:25:46,370 --> 00:25:43,380
format it's essentially a podcast

703
00:25:47,810 --> 00:25:46,380
talking about metaphysics like man oh

704
00:25:49,250 --> 00:25:47,820
man I can't believe that got published

705
00:25:51,230 --> 00:25:49,260
and then also I think about it well did

706
00:25:52,669 --> 00:25:51,240
Plato have the first podcast

707
00:25:54,049 --> 00:25:52,679
because he's just publishing these

708
00:25:55,549 --> 00:25:54,059
dialogues and you read them but it's not

709
00:25:57,110 --> 00:25:55,559
as if maybe he would have published it

710
00:25:58,669 --> 00:25:57,120
in video I think Plato was the first

711
00:26:00,950 --> 00:25:58,679
podcaster

712
00:26:02,930 --> 00:26:00,960
so is there something new about this

713
00:26:05,090 --> 00:26:02,940
format of podcasting that wasn't there

714
00:26:07,130 --> 00:26:05,100
before or what's new about it I think

715
00:26:09,769 --> 00:26:07,140
it's like it's like vlogging uh blogging

716
00:26:12,950 --> 00:26:09,779
is also not new right being able to

717
00:26:15,890 --> 00:26:12,960
write text that you publish uh and uh

718
00:26:18,529 --> 00:26:15,900
people can follow about your writing and

719
00:26:20,570 --> 00:26:18,539
so on did exist in some sense before but

720
00:26:22,730 --> 00:26:20,580
uh the internet made it possible to

721
00:26:25,250 --> 00:26:22,740
publish this for everyone do you don't

722
00:26:26,870 --> 00:26:25,260
need a publisher anymore and you don't

723
00:26:28,490 --> 00:26:26,880
know your TV studio anymore you don't

724
00:26:30,529 --> 00:26:28,500
need a broadcast station that is

725
00:26:32,630 --> 00:26:30,539
recording your talk show and sends it to

726
00:26:35,450 --> 00:26:32,640
an audience there is no competition with

727
00:26:37,549 --> 00:26:35,460
other talk Source because uh there is no

728
00:26:40,190 --> 00:26:37,559
limitations on how many people can

729
00:26:42,710 --> 00:26:40,200
broadcast at the same time and this

730
00:26:45,649 --> 00:26:42,720
allows an enormous diversity and of

731
00:26:48,529 --> 00:26:45,659
thoughts and small Productions that are

732
00:26:50,570 --> 00:26:48,539
done at a very low cost lowering the

733
00:26:52,730 --> 00:26:50,580
threshold for putting something out

734
00:26:54,830 --> 00:26:52,740
there and seeing what happens so in this

735
00:26:57,110 --> 00:26:54,840
sense it's the ecosystem that emerged is

736
00:26:59,210 --> 00:26:57,120
new because a variable change that

737
00:27:00,529 --> 00:26:59,220
changed the cost of producing a talk

738
00:27:02,149 --> 00:27:00,539
show

739
00:27:03,950 --> 00:27:02,159
all right

740
00:27:06,470 --> 00:27:03,960
Michael you agree

741
00:27:09,590 --> 00:27:06,480
um yeah yeah I mean yes that and and all

742
00:27:12,169 --> 00:27:09,600
of that and also just the fact that you

743
00:27:13,669 --> 00:27:12,179
know uh as you as you just said uh these

744
00:27:15,769 --> 00:27:13,679
kind of like long form things were

745
00:27:16,850 --> 00:27:15,779
fairly rare so most of the time if

746
00:27:18,649 --> 00:27:16,860
you're going to be in one of the

747
00:27:20,630 --> 00:27:18,659
traditional media they tell you okay

748
00:27:22,310 --> 00:27:20,640
you've got the you've got you know three

749
00:27:23,570 --> 00:27:22,320
minutes uh you know we're gonna cut all

750
00:27:25,549 --> 00:27:23,580
this stuff and we're gonna boil it down

751
00:27:27,169 --> 00:27:25,559
to three minutes and this this is often

752
00:27:28,669 --> 00:27:27,179
incredibly frustrating and and I

753
00:27:31,010 --> 00:27:28,679
understand I mean we're drowned in

754
00:27:32,570 --> 00:27:31,020
information and so so there's obviously

755
00:27:34,610 --> 00:27:32,580
a place for very short statement on

756
00:27:36,950 --> 00:27:34,620
things but the kind of stuff that we're

757
00:27:39,529 --> 00:27:36,960
talking about uh cannot be boiled down

758
00:27:41,570 --> 00:27:39,539
to to you know TV sound bites or

759
00:27:43,370 --> 00:27:41,580
anything it's just not and and so so the

760
00:27:45,350 --> 00:27:43,380
ability to have these long-form things

761
00:27:47,510 --> 00:27:45,360
so that anybody who wants to really dig

762
00:27:49,310 --> 00:27:47,520
in can hear what the actual thought is

763
00:27:52,430 --> 00:27:49,320
as opposed to something that's that's

764
00:27:54,950 --> 00:27:52,440
been just uh you know boiled into into a

765
00:27:57,049 --> 00:27:54,960
very um very very short statement I

766
00:27:58,430 --> 00:27:57,059
think is is invaluable just being able

767
00:28:01,430 --> 00:27:58,440
to have it out there for people to find

768
00:28:03,289 --> 00:28:01,440
now a brief note from two sponsors this

769
00:28:05,149 --> 00:28:03,299
is a considerably easy sponsor because

770
00:28:06,409 --> 00:28:05,159
some of you watching this are losing

771
00:28:07,850 --> 00:28:06,419
your hair and it's not because you have

772
00:28:09,169 --> 00:28:07,860
a you have trouble understanding the

773
00:28:10,789 --> 00:28:09,179
content so you're pulling it out it

774
00:28:12,830 --> 00:28:10,799
happens to approximately half of all men

775
00:28:15,590 --> 00:28:12,840
it's uncomfortable but there are methods

776
00:28:17,390 --> 00:28:15,600
to help and even to stop the process of

777
00:28:19,850 --> 00:28:17,400
balding Roman offers clinically proven

778
00:28:21,230 --> 00:28:19,860
medication to help treat hair loss all

779
00:28:22,730 --> 00:28:21,240
of course from the comfort and privacy

780
00:28:24,409 --> 00:28:22,740
of your own home Roman offers

781
00:28:26,149 --> 00:28:24,419
prescription medication and

782
00:28:28,190 --> 00:28:26,159
over-the-counter treatments they also

783
00:28:29,810 --> 00:28:28,200
offer specially formulated shampoos and

784
00:28:31,610 --> 00:28:29,820
conditioners with ingredients that

785
00:28:34,430 --> 00:28:31,620
fortify and moisturize the hair to look

786
00:28:35,930 --> 00:28:34,440
fuller research shows that 80 of men who

787
00:28:37,909 --> 00:28:35,940
use prescription hair loss medication

788
00:28:40,549 --> 00:28:37,919
had no further hair loss after two years

789
00:28:42,890 --> 00:28:40,559
Roman is licensed and the whole process

790
00:28:44,149 --> 00:28:42,900
is straightforward and discreet plans

791
00:28:46,250 --> 00:28:44,159
start at twenty dollars a month

792
00:28:49,029 --> 00:28:46,260
currently Roman has a special offer for

793
00:28:53,149 --> 00:28:49,039
toe listeners that is you use the link

794
00:28:56,230 --> 00:28:53,159
ro.co Kurt c-u-r-t to get 20 off your

795
00:28:58,669 --> 00:28:56,240
first order again that's ro.co

796
00:29:00,950 --> 00:28:58,679
c-u-r-t the link is in the description

797
00:29:02,870 --> 00:29:00,960
and you get 20 off as the toe project

798
00:29:04,610 --> 00:29:02,880
grows we get plenty of sponsor is coming

799
00:29:06,649 --> 00:29:04,620
and I thought you know this one is a

800
00:29:08,750 --> 00:29:06,659
fascinating company our new sponsor is

801
00:29:10,610 --> 00:29:08,760
Masterworks Masterworks is the only

802
00:29:12,769 --> 00:29:10,620
platform that allows you to invest in

803
00:29:15,529 --> 00:29:12,779
multi-million dollar works of art by

804
00:29:17,149 --> 00:29:15,539
Picasso bansky and more Masterworks has

805
00:29:18,950 --> 00:29:17,159
given you access to invest in Fine Art

806
00:29:20,930 --> 00:29:18,960
which is usually only accessible to

807
00:29:23,149 --> 00:29:20,940
multi-millionaires or billionaires the

808
00:29:25,010 --> 00:29:23,159
art that you see hanging in museums can

809
00:29:26,330 --> 00:29:25,020
now be partially owned by you the

810
00:29:28,130 --> 00:29:26,340
inventive Parts is that you don't need

811
00:29:29,630 --> 00:29:28,140
to know the details of art or investing

812
00:29:31,549 --> 00:29:29,640
Masterworks makes the whole process

813
00:29:33,769 --> 00:29:31,559
straightforward with a clean interface

814
00:29:35,210 --> 00:29:33,779
and exceptional customer service they

815
00:29:36,889 --> 00:29:35,220
are innovating as more traditional

816
00:29:39,409 --> 00:29:36,899
Investments suffer last month we

817
00:29:41,269 --> 00:29:39,419
verified a cell which had a 21.5 percent

818
00:29:42,350 --> 00:29:41,279
return so for instance if you were to

819
00:29:43,850 --> 00:29:42,360
put ten thousand dollars in you would

820
00:29:45,230 --> 00:29:43,860
now have twelve thousand Welcome to our

821
00:29:46,730 --> 00:29:45,240
new sponsor Masterworks and the link to

822
00:29:48,649 --> 00:29:46,740
them is in the description just so you

823
00:29:50,810 --> 00:29:48,659
know there's in fact a wait list to join

824
00:29:52,669 --> 00:29:50,820
their platform right now however total

825
00:29:54,529 --> 00:29:52,679
listeners can skip the wait list by

826
00:29:56,750 --> 00:29:54,539
visiting the link in the description as

827
00:29:58,669 --> 00:29:56,760
well as by using the promo code theories

828
00:30:00,110 --> 00:29:58,679
of everything again the link is in the

829
00:30:01,909 --> 00:30:00,120
description you can see important

830
00:30:05,330 --> 00:30:01,919
regulation A disclosures at Masterworks

831
00:30:07,549 --> 00:30:05,340
dot com slash CD what some stance of

832
00:30:09,529 --> 00:30:07,559
yours some belief that has changed most

833
00:30:11,389 --> 00:30:09,539
drastically in the past few years let's

834
00:30:13,250 --> 00:30:11,399
say three and it could be anywhere from

835
00:30:14,930 --> 00:30:13,260
something abstruse and academic to more

836
00:30:17,090 --> 00:30:14,940
colloquial like I didn't realize the

837
00:30:18,230 --> 00:30:17,100
value of children or I overvalued

838
00:30:20,750 --> 00:30:18,240
children now I'm stuck with them like

839
00:30:22,789 --> 00:30:20,760
geez that was a mistake

840
00:30:25,490 --> 00:30:22,799
yeah so something where I changed my

841
00:30:28,909 --> 00:30:25,500
mind was RNA based memory transfer

842
00:30:31,250 --> 00:30:28,919
and uh I think it's a super interesting

843
00:30:33,830 --> 00:30:31,260
idea in this context because it's uh

844
00:30:36,710 --> 00:30:33,840
close to uh stuff that Michael has been

845
00:30:38,450 --> 00:30:36,720
working on and is interested in uh there

846
00:30:41,510 --> 00:30:38,460
have been some experiments in the Soviet

847
00:30:46,070 --> 00:30:41,520
Union I think in the 70s where

848
00:30:47,930 --> 00:30:46,080
scientists look planaria uh trained them

849
00:30:49,430 --> 00:30:47,940
to learn something I think they were

850
00:30:51,649 --> 00:30:49,440
learned how to be afraid of electric

851
00:30:53,810 --> 00:30:51,659
shocks and things like that and then

852
00:30:55,970 --> 00:30:53,820
they put their uh the brains into a

853
00:30:57,710 --> 00:30:55,980
blender extracted the iron a injected

854
00:30:59,510 --> 00:30:57,720
other planaria visit and these other

855
00:31:02,870 --> 00:30:59,520
planaria had learned it

856
00:31:05,029 --> 00:31:02,880
and I I learned about this as a kid when

857
00:31:06,529 --> 00:31:05,039
I in the 1980s read Soviet science

858
00:31:09,889 --> 00:31:06,539
fiction into Twitter I grew up in

859
00:31:14,570 --> 00:31:09,899
Eastern Germany and the evil scientist

860
00:31:17,149 --> 00:31:14,580
uh harvested the brains of geniuses and

861
00:31:20,029 --> 00:31:17,159
injected himself is RNA extracted from

862
00:31:22,549 --> 00:31:20,039
these brains and thereby acquired the

863
00:31:25,430 --> 00:31:22,559
skills and even though I'm pretty sure

864
00:31:27,669 --> 00:31:25,440
it probably doesn't work if you if you

865
00:31:29,810 --> 00:31:27,679
do it at this level

866
00:31:33,110 --> 00:31:29,820
this was inspired by this original

867
00:31:37,210 --> 00:31:33,120
research and I later heard nothing about

868
00:31:41,029 --> 00:31:37,220
this anymore and so I dismissed it as uh

869
00:31:42,529 --> 00:31:41,039
similar things as I read in Sputnik and

870
00:31:44,630 --> 00:31:42,539
other Russian Publications which create

871
00:31:47,990 --> 00:31:44,640
their own methodological Universe about

872
00:31:50,330 --> 00:31:48,000
ball lightning that is authentic and

873
00:31:52,250 --> 00:31:50,340
possibly sentient and so on and

874
00:31:54,230 --> 00:31:52,260
dismissed this all as basically Another

875
00:31:56,570 --> 00:31:54,240
Universe of another Reader's Digest

876
00:31:58,970 --> 00:31:56,580
culture that is producing its own ideas

877
00:32:00,169 --> 00:31:58,980
that then later on get dissolved once

878
00:32:02,210 --> 00:32:00,179
science advanced answers because

879
00:32:03,649 --> 00:32:02,220
everybody knows it's it's synapses it's

880
00:32:05,690 --> 00:32:03,659
connections between neurons that matter

881
00:32:07,789 --> 00:32:05,700
the RNA is not that important for the

882
00:32:09,710 --> 00:32:07,799
information processing might change some

883
00:32:11,810 --> 00:32:09,720
state but you cannot learn something by

884
00:32:13,909 --> 00:32:11,820
extracting RNA and re-injecting it into

885
00:32:16,250 --> 00:32:13,919
the next organism because how would that

886
00:32:18,470 --> 00:32:16,260
work if it's done in the synapses

887
00:32:20,690 --> 00:32:18,480
and uh then recently there were some

888
00:32:22,490 --> 00:32:20,700
papers which replicated uh the original

889
00:32:24,190 --> 00:32:22,500
research and has been replicated from

890
00:32:28,730 --> 00:32:24,200
time to time

891
00:32:31,730 --> 00:32:28,740
in different types of organisms but to

892
00:32:34,850 --> 00:32:31,740
my knowledge not in of course macaques

893
00:32:36,830 --> 00:32:34,860
or not even mice but so it's not clear

894
00:32:39,409 --> 00:32:36,840
if their brains work according to the

895
00:32:41,350 --> 00:32:39,419
same principles as planaria but planaria

896
00:32:43,669 --> 00:32:41,360
are not

897
00:32:45,289 --> 00:32:43,679
extremely simple organisms only a

898
00:32:47,289 --> 00:32:45,299
handful of neurons there are something

899
00:32:49,789 --> 00:32:47,299
intermediate with some their main

900
00:32:51,230 --> 00:32:49,799
architecture is different from ours and

901
00:32:53,149 --> 00:32:51,240
their functioning principles of their

902
00:32:55,130 --> 00:32:53,159
neurons might be slightly different but

903
00:32:58,190 --> 00:32:55,140
it's worse following this idea and going

904
00:33:00,169 --> 00:32:58,200
down that rabbit hole and then I looked

905
00:33:02,090 --> 00:33:00,179
for my computer science engineering

906
00:33:04,250 --> 00:33:02,100
perspective and I realized that there

907
00:33:07,789 --> 00:33:04,260
are always things about the synaptic

908
00:33:10,250 --> 00:33:07,799
story that uh I find confusing because

909
00:33:12,649 --> 00:33:10,260
they're very difficult to implement

910
00:33:14,750 --> 00:33:12,659
so for instance weight sharing as a

911
00:33:16,070 --> 00:33:14,760
computer scientist I require weight

912
00:33:19,009 --> 00:33:16,080
sharing I don't know how to get around

913
00:33:21,049 --> 00:33:19,019
this if I want to entrain myself as

914
00:33:23,090 --> 00:33:21,059
computational Primitives in a local area

915
00:33:25,850 --> 00:33:23,100
of of my brain for instance the ability

916
00:33:28,909 --> 00:33:25,860
to rotate something which rotation is

917
00:33:31,549 --> 00:33:28,919
some operator that I apply on a pattern

918
00:33:32,990 --> 00:33:31,559
that allows this pattern to be

919
00:33:35,870 --> 00:33:33,000
represented in a slightly different way

920
00:33:38,090 --> 00:33:35,880
to have this object rotated a few

921
00:33:39,710 --> 00:33:38,100
degrees but an object doesn't consist of

922
00:33:42,830 --> 00:33:39,720
a single point it consists of many

923
00:33:44,630 --> 00:33:42,840
features that all need to get the same

924
00:33:47,509 --> 00:33:44,640
rotation applied to them using the same

925
00:33:49,970 --> 00:33:47,519
mathematical Primitives so how do you

926
00:33:53,269 --> 00:33:49,980
implement the same operator across an

927
00:33:55,909 --> 00:33:53,279
entire brain area do you make many many

928
00:33:57,769 --> 00:33:55,919
copies of the same pattern and computer

929
00:33:59,389 --> 00:33:57,779
scientists solve that with so-called

930
00:34:02,149 --> 00:33:59,399
convolutional neural networks which

931
00:34:04,810 --> 00:34:02,159
basically use the same weights again and

932
00:34:06,769 --> 00:34:04,820
again in different areas using only

933
00:34:08,570 --> 00:34:06,779
training them once and making them

934
00:34:11,149 --> 00:34:08,580
available everywhere and that would be

935
00:34:14,030 --> 00:34:11,159
very difficult to implement and synapses

936
00:34:16,430 --> 00:34:14,040
maybe there are ways but it's not

937
00:34:18,829 --> 00:34:16,440
straightforward another thing is if we

938
00:34:21,290 --> 00:34:18,839
see how training Works in babies they

939
00:34:22,909 --> 00:34:21,300
learn something and then they get rid of

940
00:34:24,290 --> 00:34:22,919
the Surplus synopsis initially they have

941
00:34:27,649 --> 00:34:24,300
much more connectivity than they need

942
00:34:29,930 --> 00:34:27,659
and uh when they get uh after they've

943
00:34:32,270 --> 00:34:29,940
trained they optimize the way in which

944
00:34:34,010 --> 00:34:32,280
the wiring works by discarding the

945
00:34:36,409 --> 00:34:34,020
things they don't need to compute what

946
00:34:39,470 --> 00:34:36,419
they want to compute so it's like

947
00:34:41,329 --> 00:34:39,480
calling the synapses it does not uh

948
00:34:43,369 --> 00:34:41,339
freeze or Edge the learning into the

949
00:34:45,730 --> 00:34:43,379
brain but it optimizes the energy usage

950
00:34:47,810 --> 00:34:45,740
of the brain another issue is that

951
00:34:49,490 --> 00:34:47,820
patterns of activation are not

952
00:34:51,589 --> 00:34:49,500
completely stable in the brain they in

953
00:34:53,389 --> 00:34:51,599
the cortex if you look you find that

954
00:34:55,190 --> 00:34:53,399
they might be moving the next day or

955
00:34:57,109 --> 00:34:55,200
even rotate a little bit which is also

956
00:34:58,790 --> 00:34:57,119
difficult to do with synapses you cannot

957
00:35:00,230 --> 00:34:58,800
read out the weights and copy them

958
00:35:02,810 --> 00:35:00,240
somewhere else in an easy

959
00:35:05,390 --> 00:35:02,820
straightforward fashion and another

960
00:35:07,370 --> 00:35:05,400
issue is defragmentation if you learn

961
00:35:08,089 --> 00:35:07,380
for instance your body map into a brain

962
00:35:10,310 --> 00:35:08,099
area

963
00:35:11,930 --> 00:35:10,320
and then somebody changes your body map

964
00:35:13,970 --> 00:35:11,940
because you have an accident and lose a

965
00:35:15,589 --> 00:35:13,980
finger or somebody gives you an

966
00:35:17,810 --> 00:35:15,599
artificial limp and you start to

967
00:35:19,490 --> 00:35:17,820
integrate this into your body map how do

968
00:35:21,589 --> 00:35:19,500
you shift all the representations around

969
00:35:23,810 --> 00:35:21,599
how do you make space for something else

970
00:35:26,150 --> 00:35:23,820
and move it or also initially when you

971
00:35:27,470 --> 00:35:26,160
set up your Maps via happier learning

972
00:35:29,870 --> 00:35:27,480
how do you make sure that the

973
00:35:31,550 --> 00:35:29,880
neighborhoods are always correct and you

974
00:35:33,589 --> 00:35:31,560
don't need to realign anything and I

975
00:35:35,810 --> 00:35:33,599
guess you need some kind of realignment

976
00:35:37,550 --> 00:35:35,820
and uh all these things seem to be

977
00:35:38,750 --> 00:35:37,560
possible when you switch to a different

978
00:35:41,750 --> 00:35:38,760
paradigm

979
00:35:44,030 --> 00:35:41,760
and uh this so if you take this RNA

980
00:35:47,210 --> 00:35:44,040
basically seriously go down this wrapper

981
00:35:49,609 --> 00:35:47,220
tool what you get is

982
00:35:52,190 --> 00:35:49,619
um the neurons are not learning a local

983
00:35:53,870 --> 00:35:52,200
function over its neighbors but they are

984
00:35:55,910 --> 00:35:53,880
learning how to respond to the shape of

985
00:35:57,650 --> 00:35:55,920
an incoming activation front

986
00:35:59,270 --> 00:35:57,660
right the spatial temporal pattern in

987
00:36:01,609 --> 00:35:59,280
their neighborhood and they are densely

988
00:36:04,849 --> 00:36:01,619
enough connected so the enabled is just

989
00:36:07,250 --> 00:36:04,859
the space around them and in this space

990
00:36:10,069 --> 00:36:07,260
they basically interpret this according

991
00:36:12,290 --> 00:36:10,079
to a certain topology to say this uh is

992
00:36:14,810 --> 00:36:12,300
maybe a convolution that gives me two

993
00:36:16,670 --> 00:36:14,820
and a half D or it gives me 2D or 1D or

994
00:36:20,750 --> 00:36:16,680
whatever the type of function is that

995
00:36:23,750 --> 00:36:20,760
they want to compute and they learn how

996
00:36:25,730 --> 00:36:23,760
to fire in this in response to those

997
00:36:27,170 --> 00:36:25,740
patterns and thereby modulate the

998
00:36:28,670 --> 00:36:27,180
patterns when they're passed on so the

999
00:36:31,670 --> 00:36:28,680
neurons act as something like a

1000
00:36:34,150 --> 00:36:31,680
self-modulating ether so which waveforms

1001
00:36:37,609 --> 00:36:34,160
propagate that perform the computations

1002
00:36:39,829 --> 00:36:37,619
and they store the responses to the

1003
00:36:43,250 --> 00:36:39,839
distributions of incoming signals

1004
00:36:45,829 --> 00:36:43,260
possibly in RNA so you have little mix

1005
00:36:48,230 --> 00:36:45,839
tapes little tape fragments that they

1006
00:36:50,390 --> 00:36:48,240
store in the Soma and that you can make

1007
00:36:52,310 --> 00:36:50,400
more of very cheaply and easily if they

1008
00:36:53,930 --> 00:36:52,320
are successful mixtapes and their useful

1009
00:36:55,490 --> 00:36:53,940
computational Primitives that they

1010
00:36:57,410 --> 00:36:55,500
discovered and they can distribute this

1011
00:37:00,230 --> 00:36:57,420
to other neurons through the entire

1012
00:37:02,030 --> 00:37:00,240
cortex so new ones of the same type will

1013
00:37:04,250 --> 00:37:02,040
gain the knowledge to apply the same

1014
00:37:06,650 --> 00:37:04,260
computational Primitives

1015
00:37:08,510 --> 00:37:06,660
and that is something I don't know if

1016
00:37:10,190 --> 00:37:08,520
the brain is doing that and of the human

1017
00:37:11,750 --> 00:37:10,200
brain is using these principles or if

1018
00:37:13,670 --> 00:37:11,760
it's using them a lot and how important

1019
00:37:15,829 --> 00:37:13,680
this is and how many other mechanisms

1020
00:37:17,630 --> 00:37:15,839
exist but it's a mechanism that we

1021
00:37:19,730 --> 00:37:17,640
haven't so my knowledge tried very much

1022
00:37:22,310 --> 00:37:19,740
in Ai and computer science

1023
00:37:25,130 --> 00:37:22,320
and it would work there is something

1024
00:37:27,170 --> 00:37:25,140
that is a very close analog that is a

1025
00:37:29,390 --> 00:37:27,180
neural cellular automaton

1026
00:37:31,430 --> 00:37:29,400
here so basically instead of learning

1027
00:37:33,290 --> 00:37:31,440
weight uh weight shifts or weight

1028
00:37:34,849 --> 00:37:33,300
changes between adjacent new ones what

1029
00:37:37,670 --> 00:37:34,859
you learn is

1030
00:37:40,069 --> 00:37:37,680
um Global functions that tell neurons on

1031
00:37:41,450 --> 00:37:40,079
how to respond to patterns in the

1032
00:37:44,030 --> 00:37:41,460
neighborhood

1033
00:37:46,550 --> 00:37:44,040
and these functions are the same for

1034
00:37:48,890 --> 00:37:46,560
every point in your Matrix

1035
00:37:51,650 --> 00:37:48,900
and you can learn arbitrary functions in

1036
00:37:53,329 --> 00:37:51,660
this way and what's nice about is that

1037
00:37:56,630 --> 00:37:53,339
you only need to learn computational

1038
00:37:59,510 --> 00:37:56,640
Primitives once our kubernetworks need

1039
00:38:01,130 --> 00:37:59,520
to learn the same linear algebra over

1040
00:38:03,650 --> 00:38:01,140
and over again in many different corners

1041
00:38:05,150 --> 00:38:03,660
of the neural network because you need a

1042
00:38:07,370 --> 00:38:05,160
vector algebra for many kinds of

1043
00:38:08,930 --> 00:38:07,380
operations that we perform like for

1044
00:38:11,450 --> 00:38:08,940
instance operations in space

1045
00:38:14,390 --> 00:38:11,460
where we shift things around or rotate

1046
00:38:15,650 --> 00:38:14,400
them and if they could change exchange

1047
00:38:17,870 --> 00:38:15,660
these useful

1048
00:38:19,430 --> 00:38:17,880
um operations with each other and just

1049
00:38:21,710 --> 00:38:19,440
apply an operator whenever the

1050
00:38:23,390 --> 00:38:21,720
environment uh dictates that this would

1051
00:38:25,310 --> 00:38:23,400
be a good idea to try to apply this

1052
00:38:27,530 --> 00:38:25,320
operator right now in this context that

1053
00:38:28,849 --> 00:38:27,540
could speed up learning it could make uh

1054
00:38:31,550 --> 00:38:28,859
training much more sample efficient

1055
00:38:33,109 --> 00:38:31,560
right so something super interesting to

1056
00:38:35,450 --> 00:38:33,119
try and this is one of the rabbit holes

1057
00:38:38,450 --> 00:38:35,460
I recently fell down whenever I changed

1058
00:38:41,810 --> 00:38:38,460
my thinking based on some experiment

1059
00:38:43,609 --> 00:38:41,820
from Neuroscience that doesn't have very

1060
00:38:46,250 --> 00:38:43,619
big impact for for the mainstream of

1061
00:38:49,329 --> 00:38:46,260
Neuroscience but that I found reflected

1062
00:38:51,470 --> 00:38:49,339
in Michael's work with planaria

1063
00:38:54,530 --> 00:38:51,480
yeah that's that's super interesting

1064
00:38:57,829 --> 00:38:54,540
stuff um I I can sprinkle a few a few

1065
00:39:00,589 --> 00:38:57,839
details onto this um uh the original so

1066
00:39:02,569 --> 00:39:00,599
so the original finding in planaria was

1067
00:39:04,609 --> 00:39:02,579
um was a guy named James McConnell in

1068
00:39:06,710 --> 00:39:04,619
the in in Michigan actually in the US

1069
00:39:08,450 --> 00:39:06,720
and and then that was in the 60s the

1070
00:39:10,670 --> 00:39:08,460
early 60s and then there were some

1071
00:39:12,650 --> 00:39:10,680
really interesting Russian work um that

1072
00:39:14,810 --> 00:39:12,660
that picked it up after that uh we

1073
00:39:17,150 --> 00:39:14,820
reproduced some of it recently in using

1074
00:39:19,069 --> 00:39:17,160
modern uh modern quantitative uh

1075
00:39:20,810 --> 00:39:19,079
Automation and and things like this but

1076
00:39:22,190 --> 00:39:20,820
one of the one of the really cool

1077
00:39:24,109 --> 00:39:22,200
aspects of this and there's a whole

1078
00:39:26,089 --> 00:39:24,119
Community by the way with um you know

1079
00:39:27,829 --> 00:39:26,099
people like uh Randy gallistel and Sam

1080
00:39:30,290 --> 00:39:27,839
gershman and um you know of course

1081
00:39:32,750 --> 00:39:30,300
glansman David glansman and people who

1082
00:39:34,970 --> 00:39:32,760
are that that story of memory in the in

1083
00:39:36,530 --> 00:39:34,980
the precise details of the um of the

1084
00:39:38,089 --> 00:39:36,540
synapses is that that story is really

1085
00:39:39,770 --> 00:39:38,099
starting to crack actually for a number

1086
00:39:41,329 --> 00:39:39,780
of reasons but one of the one of the

1087
00:39:43,550 --> 00:39:41,339
cool things that that was done in the

1088
00:39:46,430 --> 00:39:43,560
Russian work and it was also done later

1089
00:39:48,470 --> 00:39:46,440
on by um uh Doug Blackiston who's who's

1090
00:39:50,589 --> 00:39:48,480
in my lab now as a staff scientist and

1091
00:39:52,730 --> 00:39:50,599
other people is this this you can

1092
00:39:54,470 --> 00:39:52,740
certain certain animals that go through

1093
00:39:56,390 --> 00:39:54,480
larval stages right so you can take so

1094
00:39:58,790 --> 00:39:56,400
what the Russians were using um beetle

1095
00:40:01,190 --> 00:39:58,800
beetle larvae and uh and Doug and other

1096
00:40:03,109 --> 00:40:01,200
people used uh used used moths and

1097
00:40:05,690 --> 00:40:03,119
butterflies so what happens is you train

1098
00:40:07,550 --> 00:40:05,700
you train the larva right so so here

1099
00:40:09,109 --> 00:40:07,560
you've got a butterfly a caterpillar so

1100
00:40:10,849 --> 00:40:09,119
so this caterpillar lives in a

1101
00:40:12,290 --> 00:40:10,859
two-dimensional world it's a soft-bodied

1102
00:40:13,849 --> 00:40:12,300
robot it lives in a two-dimensional

1103
00:40:15,530 --> 00:40:13,859
world that eats leaves and so on right

1104
00:40:17,150 --> 00:40:15,540
and so you train this thing for a

1105
00:40:19,730 --> 00:40:17,160
particular task well during

1106
00:40:21,950 --> 00:40:19,740
metamorphosis it needs to become a moth

1107
00:40:23,210 --> 00:40:21,960
or butterfly which it lives in a

1108
00:40:24,770 --> 00:40:23,220
three-dimensional world plus it's a

1109
00:40:26,210 --> 00:40:24,780
hard-bodied creature so so the

1110
00:40:27,770 --> 00:40:26,220
controller is completely different right

1111
00:40:29,630 --> 00:40:27,780
for running this for running a

1112
00:40:31,550 --> 00:40:29,640
caterpillar versus a butterfly so so

1113
00:40:34,010 --> 00:40:31,560
during that process what happens is the

1114
00:40:35,450 --> 00:40:34,020
brain is basically dissolved so most of

1115
00:40:37,790 --> 00:40:35,460
the connections are broken most of the

1116
00:40:39,290 --> 00:40:37,800
cells are gone they die uh you you put

1117
00:40:41,450 --> 00:40:39,300
together a brand new brain itself

1118
00:40:43,130 --> 00:40:41,460
assembles and you can ask all sorts of

1119
00:40:44,930 --> 00:40:43,140
interesting philosophical questions of

1120
00:40:46,430 --> 00:40:44,940
what it's like to be a creature whose

1121
00:40:49,490 --> 00:40:46,440
brain is undergoing this massive change

1122
00:40:51,650 --> 00:40:49,500
uh but the information remains and and

1123
00:40:53,210 --> 00:40:51,660
so one can ask okay this is you know

1124
00:40:55,550 --> 00:40:53,220
certainly for computer science it's it's

1125
00:40:57,710 --> 00:40:55,560
amazing to have a cop to have to have a

1126
00:41:00,310 --> 00:40:57,720
memory medium that can survive this

1127
00:41:03,470 --> 00:41:00,320
radical re Remodeling and reconstruction

1128
00:41:04,730 --> 00:41:03,480
and there's there's the RNA story but

1129
00:41:06,710 --> 00:41:04,740
but also

1130
00:41:08,569 --> 00:41:06,720
um uh you had mentioned you know does

1131
00:41:10,910 --> 00:41:08,579
this does this work from for mammals so

1132
00:41:12,530 --> 00:41:10,920
there was a guy in the 70s and 80s there

1133
00:41:14,510 --> 00:41:12,540
was a there was a guy named George Ungar

1134
00:41:18,530 --> 00:41:14,520
who did tons of he's got tons of papers

1135
00:41:20,990 --> 00:41:18,540
he uh reproduced it in rats so so his

1136
00:41:23,210 --> 00:41:21,000
was fear of the dark and he actually

1137
00:41:24,170 --> 00:41:23,220
um by by establishing this essay and

1138
00:41:25,970 --> 00:41:24,180
then

1139
00:41:28,010 --> 00:41:25,980
um uh you know fractionating their

1140
00:41:30,349 --> 00:41:28,020
brains and and extracting this this

1141
00:41:32,750 --> 00:41:30,359
activity now he thought it was a peptide

1142
00:41:34,609 --> 00:41:32,760
not not RNA so he he ended up with a

1143
00:41:36,410 --> 00:41:34,619
with a thing called uh scottophobin

1144
00:41:39,230 --> 00:41:36,420
which turns out to be I think an eight

1145
00:41:40,490 --> 00:41:39,240
mer peptide or something and the claim

1146
00:41:42,770 --> 00:41:40,500
was that you can transfer this

1147
00:41:44,750 --> 00:41:42,780
scotophobia you can synthesize it and

1148
00:41:46,490 --> 00:41:44,760
then transfer it from brain to brain and

1149
00:41:47,870 --> 00:41:46,500
and that's and that's you know that's

1150
00:41:49,790 --> 00:41:47,880
that's that's what he thought it was and

1151
00:41:52,609 --> 00:41:49,800
then of course I think David glansman

1152
00:41:53,930 --> 00:41:52,619
favors RNA again but yeah I agree with

1153
00:41:56,630 --> 00:41:53,940
you I think I think that's a that's a

1154
00:41:59,150 --> 00:41:56,640
that's a super important story of how it

1155
00:42:01,910 --> 00:41:59,160
is that this kind of information can

1156
00:42:03,530 --> 00:42:01,920
survive uh just just massive uh

1157
00:42:05,870 --> 00:42:03,540
remodeling of the of the cognitive

1158
00:42:07,490 --> 00:42:05,880
substrate in planaria what we what we

1159
00:42:09,349 --> 00:42:07,500
did and and planaria you know they they

1160
00:42:11,390 --> 00:42:09,359
have a true centralized brain they have

1161
00:42:12,710 --> 00:42:11,400
all the same neurotransmitters that we

1162
00:42:14,210 --> 00:42:12,720
have they're not as simple a simple

1163
00:42:15,950 --> 00:42:14,220
organism

1164
00:42:17,930 --> 00:42:15,960
um what we did was McConnell's first

1165
00:42:19,490 --> 00:42:17,940
experiments which is to train them on

1166
00:42:22,849 --> 00:42:19,500
something and we train them to recognize

1167
00:42:24,349 --> 00:42:22,859
a laser etched um kind of bumpy pattern

1168
00:42:25,609 --> 00:42:24,359
on the bottom of the dish to recognize

1169
00:42:26,690 --> 00:42:25,619
that that's where their food was going

1170
00:42:28,970 --> 00:42:26,700
to be found so they made this

1171
00:42:30,950 --> 00:42:28,980
association between this pattern and the

1172
00:42:33,230 --> 00:42:30,960
and getting food and then we cut their

1173
00:42:34,910 --> 00:42:33,240
heads off and and we took the Tails and

1174
00:42:36,650 --> 00:42:34,920
the Tails sit there for 10 days doing

1175
00:42:38,569 --> 00:42:36,660
nothing and then eventually they grow a

1176
00:42:40,730 --> 00:42:38,579
new brain and what happens is that

1177
00:42:41,870 --> 00:42:40,740
information is then imprinted onto the

1178
00:42:44,030 --> 00:42:41,880
new brain and then and then you can

1179
00:42:46,030 --> 00:42:44,040
recover behavioral you know evidence

1180
00:42:48,829 --> 00:42:46,040
that they remember the information

1181
00:42:51,170 --> 00:42:48,839
so so that's pretty cool too because it

1182
00:42:53,569 --> 00:42:51,180
suggests that well we don't know if the

1183
00:42:55,609 --> 00:42:53,579
information is everywhere or if it's in

1184
00:42:57,290 --> 00:42:55,619
other places in the peripheral nervous

1185
00:42:58,670 --> 00:42:57,300
system or in this you know in the Nerf

1186
00:43:01,190 --> 00:42:58,680
core that we don't know where it is yet

1187
00:43:03,109 --> 00:43:01,200
but it's clear that it can move around

1188
00:43:04,910 --> 00:43:03,119
that the information can move around in

1189
00:43:06,710 --> 00:43:04,920
the body because it can be in the

1190
00:43:08,210 --> 00:43:06,720
posterior half and then imprinted onto

1191
00:43:09,650 --> 00:43:08,220
the brain which actually drives all the

1192
00:43:11,750 --> 00:43:09,660
all the behaviors

1193
00:43:13,790 --> 00:43:11,760
so thinking about that I I totally agree

1194
00:43:16,069 --> 00:43:13,800
that this is this is a really important

1195
00:43:17,750 --> 00:43:16,079
rabbit hole for asking but but it has it

1196
00:43:20,390 --> 00:43:17,760
there's an interesting puzzle here which

1197
00:43:24,190 --> 00:43:20,400
which is which is this you know it's one

1198
00:43:27,710 --> 00:43:24,200
thing to remember things that are

1199
00:43:29,210 --> 00:43:27,720
evolutionarily uh adaptive like fear of

1200
00:43:31,190 --> 00:43:29,220
the dark and things like this but

1201
00:43:33,829 --> 00:43:31,200
imagine and this hasn't really been done

1202
00:43:35,210 --> 00:43:33,839
well but imagine for a moment if we

1203
00:43:37,910 --> 00:43:35,220
could train them to something that is

1204
00:43:40,190 --> 00:43:37,920
completely novel let's say let's say we

1205
00:43:41,990 --> 00:43:40,200
train them um uh three yellow life

1206
00:43:43,430 --> 00:43:42,000
flashes means so take a step to your

1207
00:43:44,750 --> 00:43:43,440
left otherwise you get shocked something

1208
00:43:46,130 --> 00:43:44,760
like that and let's say they learn to do

1209
00:43:47,990 --> 00:43:46,140
it we haven't done this yet but let's

1210
00:43:50,690 --> 00:43:48,000
say let's say this could be could work

1211
00:43:52,609 --> 00:43:50,700
one of the big uh puzzles is going to be

1212
00:43:54,530 --> 00:43:52,619
when you extract whatever it is that you

1213
00:43:56,450 --> 00:43:54,540
extract let's say it's RNA or protein

1214
00:43:59,809 --> 00:43:56,460
whatever it is you stick it into the

1215
00:44:01,370 --> 00:43:59,819
brain of a recipient host and in order

1216
00:44:02,569 --> 00:44:01,380
for that memory to transfer one of the

1217
00:44:04,550 --> 00:44:02,579
things that the host has to be able to

1218
00:44:06,829 --> 00:44:04,560
do is has to be able to decode it in

1219
00:44:08,390 --> 00:44:06,839
order to decode it it's one thing if we

1220
00:44:10,130 --> 00:44:08,400
share the same code book and by

1221
00:44:11,569 --> 00:44:10,140
Evolution we could have this same code

1222
00:44:13,309 --> 00:44:11,579
book for things that come up all the

1223
00:44:16,490 --> 00:44:13,319
time like fear of the dark fear you know

1224
00:44:18,470 --> 00:44:16,500
things like that but how do you how how

1225
00:44:20,270 --> 00:44:18,480
would the recipient look at a a a a

1226
00:44:22,609 --> 00:44:20,280
weird sort of you know some some kind of

1227
00:44:24,230 --> 00:44:22,619
crazy hairpin RNA structure and look at

1228
00:44:27,109 --> 00:44:24,240
and and analyze it and be like oh yes

1229
00:44:28,970 --> 00:44:27,119
that's three light flashes and then uh

1230
00:44:31,550 --> 00:44:28,980
step to the left I see so you would need

1231
00:44:33,470 --> 00:44:31,560
to be able to interpret somehow this the

1232
00:44:35,210 --> 00:44:33,480
structure and convert it back to the

1233
00:44:37,970 --> 00:44:35,220
behavior and for and for behaviors that

1234
00:44:39,530 --> 00:44:37,980
are truly arbitrary that might be I I

1235
00:44:41,990 --> 00:44:39,540
don't know actually how that would work

1236
00:44:45,650 --> 00:44:42,000
and so so I think the frontier of this

1237
00:44:47,990 --> 00:44:45,660
field is going to be to have a really uh

1238
00:44:50,329 --> 00:44:48,000
convincing demonstration of of a

1239
00:44:52,670 --> 00:44:50,339
transfer of a memory that doesn't have a

1240
00:44:54,710 --> 00:44:52,680
plausible pre-existing shared

1241
00:44:56,030 --> 00:44:54,720
evolutionary decoding because because

1242
00:44:57,650 --> 00:44:56,040
otherwise you have a you have a real

1243
00:44:59,750 --> 00:44:57,660
puzzle as to how the how the the

1244
00:45:01,490 --> 00:44:59,760
decoding is going to work so this this

1245
00:45:03,410 --> 00:45:01,500
idea and and then and then even without

1246
00:45:06,230 --> 00:45:03,420
the transfer you can also think of it a

1247
00:45:08,210 --> 00:45:06,240
different way every memory is like a

1248
00:45:09,530 --> 00:45:08,220
message is like basically a transplanted

1249
00:45:11,390 --> 00:45:09,540
message from your past self to your

1250
00:45:12,710 --> 00:45:11,400
future self meaning that you still have

1251
00:45:14,510 --> 00:45:12,720
to decode your memories whatever your

1252
00:45:16,010 --> 00:45:14,520
memories are in an important sense you

1253
00:45:18,410 --> 00:45:16,020
have to you know those engrams you have

1254
00:45:20,870 --> 00:45:18,420
to decode them somehow so

1255
00:45:22,550 --> 00:45:20,880
um that that whole issue of of encoding

1256
00:45:24,230 --> 00:45:22,560
and decoding whatever the substrate of

1257
00:45:26,890 --> 00:45:24,240
memory is is you know maybe one of the

1258
00:45:29,329 --> 00:45:26,900
most important questions there are

1259
00:45:31,190 --> 00:45:29,339
one of the ways I we can think about

1260
00:45:34,630 --> 00:45:31,200
these engrams I think that there are

1261
00:45:37,430 --> 00:45:34,640
priors that condition what kinds of

1262
00:45:40,910 --> 00:45:37,440
features are being spawned in which

1263
00:45:42,589 --> 00:45:40,920
context for instance when we see a new

1264
00:45:46,309 --> 00:45:42,599
scene the way that perception seems to

1265
00:45:47,750 --> 00:45:46,319
be working is that we uh spawn lots lots

1266
00:45:49,670 --> 00:45:47,760
of feature controllers that then

1267
00:45:51,770 --> 00:45:49,680
organize into objects that are

1268
00:45:54,470 --> 00:45:51,780
controlled at the level of the scene

1269
00:45:56,870 --> 00:45:54,480
and uh this is basically a game engine

1270
00:45:59,089 --> 00:45:56,880
that is forming in our brain that is

1271
00:46:02,630 --> 00:45:59,099
creating a population of interacting

1272
00:46:04,430 --> 00:46:02,640
objects that are tuned to track our

1273
00:46:05,690 --> 00:46:04,440
perceptual data at the lowest level so

1274
00:46:09,109 --> 00:46:05,700
all the patterns that we get from our

1275
00:46:10,130 --> 00:46:09,119
retina and so on uh are samples noisy

1276
00:46:12,050 --> 00:46:10,140
samples

1277
00:46:13,609 --> 00:46:12,060
um that are difficult to interpret but

1278
00:46:15,829 --> 00:46:13,619
we are matching them into these

1279
00:46:18,290 --> 00:46:15,839
hierarchies of features that are

1280
00:46:20,750 --> 00:46:18,300
translated into objects that assign

1281
00:46:23,210 --> 00:46:20,760
every feed a feature to exactly one

1282
00:46:25,190 --> 00:46:23,220
object and every pixel so to speak to

1283
00:46:28,010 --> 00:46:25,200
exactly one except in the case of

1284
00:46:29,690 --> 00:46:28,020
transparency and use this to interpret

1285
00:46:31,370 --> 00:46:29,700
the scene that is happening in front of

1286
00:46:33,589 --> 00:46:31,380
us and when we are in the dark

1287
00:46:35,569 --> 00:46:33,599
what happens is that we spawn lots of

1288
00:46:38,030 --> 00:46:35,579
object controllers without being able to

1289
00:46:40,910 --> 00:46:38,040
disprove them because there is no data

1290
00:46:43,370 --> 00:46:40,920
that forces us to reject them and if you

1291
00:46:45,230 --> 00:46:43,380
have a vivid imagination especially as a

1292
00:46:47,329 --> 00:46:45,240
child you will feel this Darkness

1293
00:46:49,309 --> 00:46:47,339
automatically with lots of objects many

1294
00:46:52,670 --> 00:46:49,319
of which will be scary

1295
00:46:54,349 --> 00:46:52,680
and so I think that lots of the fear of

1296
00:46:56,990 --> 00:46:54,359
the dark doesn't need a lot of encoding

1297
00:46:59,569 --> 00:46:57,000
in our brain it is just an artifact of

1298
00:47:01,069 --> 00:46:59,579
the fact that there are scary things in

1299
00:47:03,230 --> 00:47:01,079
the world which I learned to represent

1300
00:47:04,849 --> 00:47:03,240
at an early age and that we cannot

1301
00:47:06,170 --> 00:47:04,859
disprove them that they just will just

1302
00:47:08,630 --> 00:47:06,180
spawn

1303
00:47:10,250 --> 00:47:08,640
I remember this vividly as a child that

1304
00:47:14,089 --> 00:47:10,260
whenever I had to go into the dark

1305
00:47:17,510 --> 00:47:14,099
basement to uh get some food in our

1306
00:47:19,370 --> 00:47:17,520
house uh in the countryside that this

1307
00:47:20,870 --> 00:47:19,380
Darkness automatically filled with all

1308
00:47:22,550 --> 00:47:20,880
sorts of shapes and things and

1309
00:47:25,970 --> 00:47:22,560
possibilities

1310
00:47:27,829 --> 00:47:25,980
and it took me later to learn that uh

1311
00:47:30,829 --> 00:47:27,839
you need to be much more afraid of the

1312
00:47:32,870 --> 00:47:30,839
ghosts that can hide in the light

1313
00:47:34,670 --> 00:47:32,880
so what would be the implications of if

1314
00:47:35,690 --> 00:47:34,680
you were able to transfer memory for

1315
00:47:37,790 --> 00:47:35,700
something that's

1316
00:47:40,550 --> 00:47:37,800
not trivial so nothing that's like an

1317
00:47:44,030 --> 00:47:40,560
archetype of fear of the dark between

1318
00:47:46,069 --> 00:47:44,040
a mammal like rats

1319
00:47:47,630 --> 00:47:46,079
and when I say transfer memory I mean in

1320
00:47:49,069 --> 00:47:47,640
this way that you blend up the brain or

1321
00:47:51,050 --> 00:47:49,079
you and also can you explain what's

1322
00:47:52,550 --> 00:47:51,060
meant by I think I understand what it

1323
00:47:53,690 --> 00:47:52,560
means to blend the brain of a planaria

1324
00:47:55,910 --> 00:47:53,700
but I don't think that's the same

1325
00:47:58,309 --> 00:47:55,920
process that's going on in rats maybe it

1326
00:48:00,290 --> 00:47:58,319
is they they well ongar did exactly the

1327
00:48:01,970 --> 00:48:00,300
same thing he would train rats for

1328
00:48:04,550 --> 00:48:01,980
particular tasks he would extract the

1329
00:48:06,829 --> 00:48:04,560
brain literally liquefy to extract the

1330
00:48:09,050 --> 00:48:06,839
chemical contents uh he would then

1331
00:48:10,609 --> 00:48:09,060
either inject the whole extract or a

1332
00:48:12,829 --> 00:48:10,619
filtered extract where you would divide

1333
00:48:14,990 --> 00:48:12,839
it up you'd fractionated so here's the

1334
00:48:17,329 --> 00:48:15,000
rnas here's the proteins here you know

1335
00:48:19,790 --> 00:48:17,339
other things uh and and then he would

1336
00:48:22,670 --> 00:48:19,800
inject that liquid directly into the

1337
00:48:25,130 --> 00:48:22,680
brains of recipient rats so you know

1338
00:48:27,349 --> 00:48:25,140
when you do that you lose you lose

1339
00:48:28,970 --> 00:48:27,359
spatial structure on the input because

1340
00:48:30,349 --> 00:48:28,980
you just Blended your brain whatever

1341
00:48:34,190 --> 00:48:30,359
spatial structure otherwise you just

1342
00:48:36,349 --> 00:48:34,200
destroyed it also on the recipient you

1343
00:48:37,910 --> 00:48:36,359
just you just inject it it's not you

1344
00:48:39,349 --> 00:48:37,920
know you're not finding that particular

1345
00:48:41,329 --> 00:48:39,359
place where you're going to stick that

1346
00:48:42,470 --> 00:48:41,339
you just inject this this thing right in

1347
00:48:44,390 --> 00:48:42,480
the middle of the brain who knows where

1348
00:48:45,770 --> 00:48:44,400
it goes you know where the fluid goes is

1349
00:48:48,890 --> 00:48:45,780
it sort of there's no spatial

1350
00:48:52,190 --> 00:48:48,900
specificity there whatsoever so if that

1351
00:48:54,290 --> 00:48:52,200
works uh what you're counting on is the

1352
00:48:57,710 --> 00:48:54,300
ability of the brain to take up

1353
00:48:59,630 --> 00:48:57,720
information via a completely novel route

1354
00:49:01,250 --> 00:48:59,640
so it's not information that's for

1355
00:49:03,470 --> 00:49:01,260
example visual right visual information

1356
00:49:05,150 --> 00:49:03,480
the accounts it comes in exactly the

1357
00:49:06,650 --> 00:49:05,160
same place all the time right there are

1358
00:49:08,210 --> 00:49:06,660
there are there are optic nerves that

1359
00:49:09,290 --> 00:49:08,220
connect to the same place in the in the

1360
00:49:12,410 --> 00:49:09,300
brain and that's where that information

1361
00:49:15,710 --> 00:49:12,420
arrives if you bathe the brain in some

1362
00:49:17,809 --> 00:49:15,720
sort of uh informational extract you're

1363
00:49:19,730 --> 00:49:17,819
basically asking the cells to take it up

1364
00:49:21,650 --> 00:49:19,740
almost as as a primitive animal would

1365
00:49:23,210 --> 00:49:21,660
with taste or touch you right that's

1366
00:49:24,710 --> 00:49:23,220
kind of distributed all over the body

1367
00:49:26,089 --> 00:49:24,720
and you can sort of pick it up anywhere

1368
00:49:28,190 --> 00:49:26,099
and then you have to process this

1369
00:49:29,690 --> 00:49:28,200
information so so you've got those

1370
00:49:31,370 --> 00:49:29,700
issues right off the bat right that

1371
00:49:33,710 --> 00:49:31,380
you've destroyed the incoming spatial

1372
00:49:35,690 --> 00:49:33,720
structure you you can't really count on

1373
00:49:38,630 --> 00:49:35,700
where it's going to land in the brain

1374
00:49:40,030 --> 00:49:38,640
and and then the third thing is as you

1375
00:49:41,990 --> 00:49:40,040
just mentioned is the idea that

1376
00:49:45,710 --> 00:49:42,000
especially if we start with information

1377
00:49:50,510 --> 00:49:45,720
that isn't any that is is that is so um

1378
00:49:51,490 --> 00:49:50,520
uh kind of uh specific and

1379
00:49:54,950 --> 00:49:51,500
um

1380
00:49:56,510 --> 00:49:54,960
you know kind of uh invented you know

1381
00:49:57,829 --> 00:49:56,520
that free light flashes means move to

1382
00:49:59,510 --> 00:49:57,839
your left I mean there's never there's

1383
00:50:01,490 --> 00:49:59,520
never been an evolutionary reason to

1384
00:50:02,930 --> 00:50:01,500
have that encoded like as you just said

1385
00:50:04,609 --> 00:50:02,940
having a fear of the dark is absolutely

1386
00:50:06,470 --> 00:50:04,619
a natural kind of thing that shows that

1387
00:50:08,030 --> 00:50:06,480
that you can expect but and then there

1388
00:50:10,430 --> 00:50:08,040
are many other things like that but but

1389
00:50:12,349 --> 00:50:10,440
something something as as contrived as

1390
00:50:13,910 --> 00:50:12,359
as you know three light flashes and then

1391
00:50:16,130 --> 00:50:13,920
you move to your left there's no reason

1392
00:50:18,349 --> 00:50:16,140
to think that we have a built-in way to

1393
00:50:20,930 --> 00:50:18,359
recognize that so when you as a as a

1394
00:50:23,150 --> 00:50:20,940
recipient brain are handed this weird uh

1395
00:50:25,430 --> 00:50:23,160
molecule with a particular structure or

1396
00:50:27,950 --> 00:50:25,440
or a set of molecules uh being able to

1397
00:50:29,809 --> 00:50:27,960
analyze that having the cells in your

1398
00:50:31,790 --> 00:50:29,819
brain or or other parts of the body

1399
00:50:34,010 --> 00:50:31,800
actually that could Analyze That and

1400
00:50:35,809 --> 00:50:34,020
recover that original information would

1401
00:50:37,910 --> 00:50:35,819
be extremely puzzling I actually don't

1402
00:50:40,130 --> 00:50:37,920
know how that would work and I'm a big

1403
00:50:41,750 --> 00:50:40,140
fan of of unlikely sounding experiments

1404
00:50:44,510 --> 00:50:41,760
that have implications if they would

1405
00:50:47,030 --> 00:50:44,520
work so this is something that uh that I

1406
00:50:48,349 --> 00:50:47,040
think should absolutely be done and um

1407
00:50:51,109 --> 00:50:48,359
and you know at some point we'll we'll

1408
00:50:52,849 --> 00:50:51,119
do it but we haven't done it yet but so

1409
00:50:55,010 --> 00:50:52,859
how far did the research in my school

1410
00:50:58,910 --> 00:50:55,020
what is the complexity of things that

1411
00:51:01,370 --> 00:50:58,920
could be transmitted via this road the I

1412
00:51:06,170 --> 00:51:01,380
as as I I don't remember everything that

1413
00:51:07,069 --> 00:51:06,180
he did uh the vast majority of he he did

1414
00:51:08,690 --> 00:51:07,079
not

1415
00:51:11,150 --> 00:51:08,700
um go

1416
00:51:12,589 --> 00:51:11,160
far to test all the complexities what he

1417
00:51:14,569 --> 00:51:12,599
tried to do was because as you can

1418
00:51:16,790 --> 00:51:14,579
imagine he faced incredible opposition

1419
00:51:18,770 --> 00:51:16,800
right so so everybody you know sort of

1420
00:51:21,770 --> 00:51:18,780
wanted to critique this thing so he

1421
00:51:23,329 --> 00:51:21,780
spent all of his time on he picked one

1422
00:51:25,670 --> 00:51:23,339
simple assay which was this fear of the

1423
00:51:28,730 --> 00:51:25,680
dark thing and then he just he just

1424
00:51:30,589 --> 00:51:28,740
bashed it for for 20 years uh to just

1425
00:51:32,690 --> 00:51:30,599
finally try to kind of crack that into

1426
00:51:34,970 --> 00:51:32,700
the into the Paradigm he did not

1427
00:51:37,309 --> 00:51:34,980
um as far as I know do lots of different

1428
00:51:39,109 --> 00:51:37,319
assays to try and make it more complex I

1429
00:51:42,109 --> 00:51:39,119
think uh I think it you know it's it's

1430
00:51:43,430 --> 00:51:42,119
very ripe for for investigation uh this

1431
00:51:44,870 --> 00:51:43,440
is the kind of did anyone else build

1432
00:51:46,609 --> 00:51:44,880
upon his work

1433
00:51:48,890 --> 00:51:46,619
um not that I know I mean I mean David

1434
00:51:50,569 --> 00:51:48,900
glansman is the best modern person who

1435
00:51:53,510 --> 00:51:50,579
works on this right so he does aplizia

1436
00:51:55,970 --> 00:51:53,520
and he you know and he does RNA so so he

1437
00:51:57,910 --> 00:51:55,980
favors um he favors RNA

1438
00:52:01,309 --> 00:51:57,920
um there's a little bit of work from um

1439
00:52:04,250 --> 00:52:01,319
in Israel with C elegans uh he's kind of

1440
00:52:06,109 --> 00:52:04,260
looking into that it uh there's related

1441
00:52:08,630 --> 00:52:06,119
work that has to do with cryogenics

1442
00:52:11,750 --> 00:52:08,640
which is this this you know this idea

1443
00:52:13,490 --> 00:52:11,760
idea that that if if it is if if

1444
00:52:16,970 --> 00:52:13,500
memories are a particular kind of

1445
00:52:18,470 --> 00:52:16,980
dynamic electrical State then some sort

1446
00:52:20,870 --> 00:52:18,480
of cryogenic freezing is probably going

1447
00:52:24,410 --> 00:52:20,880
to disrupt that whereas if it's a stable

1448
00:52:25,970 --> 00:52:24,420
molecule then it should survive so again

1449
00:52:27,890 --> 00:52:25,980
I think there are people interested in

1450
00:52:29,329 --> 00:52:27,900
that aspect of it but I'm not sure

1451
00:52:32,390 --> 00:52:29,339
um I'm not sure they've done anything

1452
00:52:35,329 --> 00:52:32,400
with it there's also a goraf uh

1453
00:52:38,569 --> 00:52:35,339
venkataraman I think he's oh yes yes

1454
00:52:40,730 --> 00:52:38,579
yeah yeah yeah yeah he uh taught me that

1455
00:52:41,990 --> 00:52:40,740
uh he has been working on this for

1456
00:52:45,349 --> 00:52:42,000
several years but he said It's

1457
00:52:47,990 --> 00:52:45,359
sociologically Tricky and that's to me

1458
00:52:51,530 --> 00:52:48,000
fascinating that we should care about

1459
00:52:53,510 --> 00:52:51,540
that what does he mean by that what do

1460
00:52:56,270 --> 00:52:53,520
you care about what stupid people think

1461
00:52:58,549 --> 00:52:56,280
uh if this uh possibility exists that

1462
00:53:01,370 --> 00:52:58,559
this works the upside is so big that

1463
00:53:03,829 --> 00:53:01,380
it's criminal to not research this I I

1464
00:53:05,450 --> 00:53:03,839
think it's a disaster that you can read

1465
00:53:07,490 --> 00:53:05,460
uh introductory textbooks on your

1466
00:53:09,710 --> 00:53:07,500
residence and never ever hear about any

1467
00:53:11,630 --> 00:53:09,720
of these experiments yeah yeah like uh

1468
00:53:13,910 --> 00:53:11,640
everybody who gets the introductory

1469
00:53:16,069 --> 00:53:13,920
stuff on Neuroscience only knows about

1470
00:53:19,910 --> 00:53:16,079
uh information stored in the connect

1471
00:53:22,910 --> 00:53:19,920
home and this leads to for instance the

1472
00:53:24,829 --> 00:53:22,920
blue brain project if that if RNA based

1473
00:53:27,589 --> 00:53:24,839
memory transfer is a thing then this

1474
00:53:29,750 --> 00:53:27,599
entire project is doomed and because you

1475
00:53:31,849 --> 00:53:29,760
cannot get the story out of uh just

1476
00:53:33,710 --> 00:53:31,859
recording the contact home or the most

1477
00:53:36,049 --> 00:53:33,720
of the research right now is focused on

1478
00:53:38,630 --> 00:53:36,059
reconstructing The Chronic tone as as it

1479
00:53:40,670 --> 00:53:38,640
was circuitry and hoping that we can get

1480
00:53:43,670 --> 00:53:40,680
the functionality of information

1481
00:53:45,950 --> 00:53:43,680
processing and to use the specificity of

1482
00:53:48,170 --> 00:53:45,960
the particular brain what it has learned

1483
00:53:50,210 --> 00:53:48,180
from the connections between neurons but

1484
00:53:53,150 --> 00:53:50,220
what if it turns out this doesn't matter

1485
00:53:55,250 --> 00:53:53,160
what is if you just need a connection

1486
00:53:56,930 --> 00:53:55,260
that dense enough and so basically

1487
00:53:59,329 --> 00:53:56,940
stochastic lattice that is somewhat

1488
00:54:00,470 --> 00:53:59,339
randomly wired and what matters is what

1489
00:54:02,030 --> 00:54:00,480
the new ones are doing with the

1490
00:54:04,130 --> 00:54:02,040
information that they're getting through

1491
00:54:05,930 --> 00:54:04,140
this ether to this lattice right this

1492
00:54:07,970 --> 00:54:05,940
changes the entire way in which we need

1493
00:54:09,530 --> 00:54:07,980
to look at things and if this

1494
00:54:11,630 --> 00:54:09,540
possibility exists and if this

1495
00:54:13,309 --> 00:54:11,640
possibility it's just one percent but

1496
00:54:16,010 --> 00:54:13,319
there is some experimentice points in

1497
00:54:17,809 --> 00:54:16,020
this direction it is not

1498
00:54:20,390 --> 00:54:17,819
um it's ridiculous to not pursue this

1499
00:54:22,130 --> 00:54:20,400
with high pressure and focus on it and

1500
00:54:25,190 --> 00:54:22,140
support research that goes in this

1501
00:54:26,870 --> 00:54:25,200
direction basically what's useful is not

1502
00:54:28,849 --> 00:54:26,880
so much answering questions in science

1503
00:54:31,370 --> 00:54:28,859
it's discovering questions it's

1504
00:54:32,990 --> 00:54:31,380
discovering new uncertainty reducing the

1505
00:54:34,609 --> 00:54:33,000
uncertainty is much easier than

1506
00:54:36,230 --> 00:54:34,619
discovering new areas of where you

1507
00:54:39,349 --> 00:54:36,240
thought that you were certain

1508
00:54:41,450 --> 00:54:39,359
but that allow you to get new insights

1509
00:54:43,609 --> 00:54:41,460
and it seems to me that a lot of

1510
00:54:45,650 --> 00:54:43,619
Neuroscience is stuck that does not

1511
00:54:48,770 --> 00:54:45,660
produce uh results that seem to

1512
00:54:50,270 --> 00:54:48,780
accumulate in an obvious way towards a

1513
00:54:53,390 --> 00:54:50,280
theory on how the brain processes

1514
00:54:56,109 --> 00:54:53,400
information yeah so the Neuroscience I

1515
00:54:58,430 --> 00:54:56,119
just don't deliver input to the

1516
00:55:00,890 --> 00:54:58,440
researchers and add the Transformer is

1517
00:55:03,410 --> 00:55:00,900
not the result of uh reading a lot of

1518
00:55:05,450 --> 00:55:03,420
Neuroscience it's really mostly the

1519
00:55:09,170 --> 00:55:05,460
result of people thinking about

1520
00:55:11,690 --> 00:55:09,180
statistics of data processing and it

1521
00:55:14,750 --> 00:55:11,700
would be great if we would focus on

1522
00:55:16,730 --> 00:55:14,760
ideas that are promising and new and

1523
00:55:19,430 --> 00:55:16,740
that have the power to shake existing

1524
00:55:21,650 --> 00:55:19,440
paradigms yeah this is you know this is

1525
00:55:24,410 --> 00:55:21,660
this is so important and it's not just

1526
00:55:26,510 --> 00:55:24,420
Neuroscience in developmental biology we

1527
00:55:27,710 --> 00:55:26,520
have exactly the same thing and and I'll

1528
00:55:29,930 --> 00:55:27,720
just give you two two very simple

1529
00:55:31,790 --> 00:55:29,940
examples of it where and I tell the

1530
00:55:33,710 --> 00:55:31,800
students when I when I give talks to

1531
00:55:36,170 --> 00:55:33,720
students I say isn't it amazing that

1532
00:55:37,970 --> 00:55:36,180
that in in your whole

1533
00:55:39,290 --> 00:55:37,980
course of biology in your developmental

1534
00:55:42,010 --> 00:55:39,300
biology textbook there's not a mention

1535
00:55:44,809 --> 00:55:42,020
of any of this because it completely um

1536
00:55:46,130 --> 00:55:44,819
uh just undermines a lot of the basic

1537
00:55:49,549 --> 00:55:46,140
assumptions so here's a couple of

1538
00:55:51,650 --> 00:55:49,559
examples one one example is that as of

1539
00:55:53,809 --> 00:55:51,660
trophic memory and deer so there are

1540
00:55:55,430 --> 00:55:53,819
species of deer that every year they

1541
00:55:56,630 --> 00:55:55,440
regenerate that you know the whole so

1542
00:55:57,890 --> 00:55:56,640
they make this antler rack on their

1543
00:56:01,549 --> 00:55:57,900
heads the whole thing falls off and then

1544
00:56:04,250 --> 00:56:01,559
it regrows the next year so these two

1545
00:56:05,990 --> 00:56:04,260
guys bubenic which are the Father and

1546
00:56:08,270 --> 00:56:06,000
Son team that did these experiments for

1547
00:56:09,950 --> 00:56:08,280
40 years and actually have all these

1548
00:56:12,950 --> 00:56:09,960
antlers in my lab now because when when

1549
00:56:14,329 --> 00:56:12,960
the younger one retired I asked he sent

1550
00:56:16,910 --> 00:56:14,339
me all these things all these antlers

1551
00:56:19,190 --> 00:56:16,920
the idea is this um what you can do is

1552
00:56:21,710 --> 00:56:19,200
you take a knife and somewhere in this

1553
00:56:23,870 --> 00:56:21,720
Branch structure you you make a wound

1554
00:56:25,130 --> 00:56:23,880
and the bone will heal and you get a

1555
00:56:26,750 --> 00:56:25,140
little callous and that's you know

1556
00:56:30,049 --> 00:56:26,760
that's it for that year then the whole

1557
00:56:31,849 --> 00:56:30,059
thing drops off and then next year

1558
00:56:34,069 --> 00:56:31,859
it starts to grow and it will make an

1559
00:56:36,410 --> 00:56:34,079
ectopic tine an ectopic Branch at the

1560
00:56:38,150 --> 00:56:36,420
point where you injured it last year and

1561
00:56:40,430 --> 00:56:38,160
this goes on for five or six years and

1562
00:56:43,130 --> 00:56:40,440
then eventually it goes away and and and

1563
00:56:45,770 --> 00:56:43,140
you get a normal normal rack again and

1564
00:56:47,990 --> 00:56:45,780
so the the amazing thing about it is

1565
00:56:50,270 --> 00:56:48,000
that you know as you you could you can

1566
00:56:52,370 --> 00:56:50,280
you can the the the standard models for

1567
00:56:53,990 --> 00:56:52,380
for patterning for morphogenesis are

1568
00:56:56,510 --> 00:56:54,000
these kind of um Gene regulatory

1569
00:56:58,730 --> 00:56:56,520
networks and and you know genetic um

1570
00:57:00,829 --> 00:56:58,740
Kinds of Kinds of uh biochemical

1571
00:57:02,809 --> 00:57:00,839
gradients and so on you know if you try

1572
00:57:05,750 --> 00:57:02,819
to try to come up with a with a with a

1573
00:57:07,190 --> 00:57:05,760
model for this so for for encoding an

1574
00:57:09,410 --> 00:57:07,200
arbitrary Point within a branch

1575
00:57:11,569 --> 00:57:09,420
structure that your cells at the scalp

1576
00:57:13,849 --> 00:57:11,579
have to remember for months after the

1577
00:57:15,650 --> 00:57:13,859
whole thing is dropped off and then not

1578
00:57:17,150 --> 00:57:15,660
only remember it but then implement it

1579
00:57:19,130 --> 00:57:17,160
so that when the bone starts to grow

1580
00:57:20,809 --> 00:57:19,140
something says oh yes that's the you

1581
00:57:22,549 --> 00:57:20,819
know the start another another time

1582
00:57:24,710 --> 00:57:22,559
growing to your left exactly exactly

1583
00:57:26,390 --> 00:57:24,720
here right trying to trying to make a

1584
00:57:29,510 --> 00:57:26,400
model of this using the standard tools

1585
00:57:31,010 --> 00:57:29,520
of the field is uh just just incredibly

1586
00:57:32,510 --> 00:57:31,020
difficult and and this is that that's

1587
00:57:33,790 --> 00:57:32,520
and there are other examples of this but

1588
00:57:35,750 --> 00:57:33,800
this kind of

1589
00:57:37,190 --> 00:57:35,760
non-non-genetic memory that's just very

1590
00:57:39,109 --> 00:57:37,200
difficult to explain with standard

1591
00:57:41,030 --> 00:57:39,119
models the other thing which is any I

1592
00:57:43,309 --> 00:57:41,040
think an even bigger Scandal is the

1593
00:57:43,970 --> 00:57:43,319
whole situation with planaria

1594
00:57:46,010 --> 00:57:43,980
um

1595
00:57:47,690 --> 00:57:46,020
planaria some species of planaria the

1596
00:57:49,670 --> 00:57:47,700
way they reproduce is they tear

1597
00:57:51,170 --> 00:57:49,680
themselves in half each half regenerates

1598
00:57:52,849 --> 00:57:51,180
the missing piece and now you've got two

1599
00:57:54,290 --> 00:57:52,859
that's how they reproduce so if you're

1600
00:57:56,930 --> 00:57:54,300
going to do that what you end up

1601
00:57:58,670 --> 00:57:56,940
avoiding is weissman's barrier this idea

1602
00:58:00,049 --> 00:57:58,680
that when we get mutations in our body

1603
00:58:02,390 --> 00:58:00,059
our children don't inherit those

1604
00:58:04,370 --> 00:58:02,400
mutations right so this means that any

1605
00:58:06,530 --> 00:58:04,380
mutation that doesn't kill the stem cell

1606
00:58:08,690 --> 00:58:06,540
in the body gets Amplified as that cell

1607
00:58:11,089 --> 00:58:08,700
contributes to regrowing the worm so

1608
00:58:12,530 --> 00:58:11,099
every as a result of this for for 400

1609
00:58:14,690 --> 00:58:12,540
million years these planaria have

1610
00:58:16,849 --> 00:58:14,700
accumulative mutations their genomes are

1611
00:58:18,890 --> 00:58:16,859
an incredible mess their cells are

1612
00:58:20,329 --> 00:58:18,900
basically mixoploid that meaning meaning

1613
00:58:21,410 --> 00:58:20,339
they're like a tumor every cell has a

1614
00:58:23,210 --> 00:58:21,420
different number of chromosomes

1615
00:58:25,490 --> 00:58:23,220
potentially they just they just look

1616
00:58:27,589 --> 00:58:25,500
look you know it just looks horrible as

1617
00:58:30,349 --> 00:58:27,599
an end result you've got an animal that

1618
00:58:33,349 --> 00:58:30,359
is Immortal incredibly good at

1619
00:58:36,410 --> 00:58:33,359
regenerating with 100 Fidelity and very

1620
00:58:37,970 --> 00:58:36,420
resistant to cancer now this is all of

1621
00:58:40,490 --> 00:58:37,980
this is the exact opposite of the

1622
00:58:42,109 --> 00:58:40,500
message you get uh from from from a

1623
00:58:43,730 --> 00:58:42,119
typical course through through biology

1624
00:58:45,829 --> 00:58:43,740
which it says that what is the genome

1625
00:58:47,569 --> 00:58:45,839
for The genome is for setting your body

1626
00:58:50,030 --> 00:58:47,579
structure if you mess with the genome

1627
00:58:52,309 --> 00:58:50,040
that information goes away you get you

1628
00:58:54,109 --> 00:58:52,319
get aging you get cancer right why does

1629
00:58:57,109 --> 00:58:54,119
the animal with the worst genome have

1630
00:58:58,490 --> 00:58:57,119
the the the best anatomical Fidelity I

1631
00:59:00,049 --> 00:58:58,500
mean that's just that and and I think we

1632
00:59:01,370 --> 00:59:00,059
actually as of as of you know a few

1633
00:59:03,349 --> 00:59:01,380
months ago we actually I think have some

1634
00:59:05,390 --> 00:59:03,359
insight into this but but but it's been

1635
00:59:07,309 --> 00:59:05,400
bugging me for for years and this is the

1636
00:59:09,589 --> 00:59:07,319
kind of thing that nobody ever talks

1637
00:59:11,089 --> 00:59:09,599
about because it it goes against the

1638
00:59:13,309 --> 00:59:11,099
general Assumption of what of what

1639
00:59:14,930 --> 00:59:13,319
genomes actually do and and what they're

1640
00:59:17,390 --> 00:59:14,940
for and you know this this complete lack

1641
00:59:18,770 --> 00:59:17,400
of uh correlation between the genome in

1642
00:59:20,930 --> 00:59:18,780
fact an anti-correlation between the

1643
00:59:23,270 --> 00:59:20,940
genome quality and the incredible

1644
00:59:25,730 --> 00:59:23,280
ability of this of this animal to uh to

1645
00:59:27,230 --> 00:59:25,740
to to to to have a healthy Anatomy yeah

1646
00:59:28,609 --> 00:59:27,240
what is that Insight that you mentioned

1647
00:59:31,609 --> 00:59:28,619
you acquired a few months ago

1648
00:59:33,650 --> 00:59:31,619
preliminary so so okay uh in the in the

1649
00:59:35,510 --> 00:59:33,660
name of uh you know throwing out uh kind

1650
00:59:37,670 --> 00:59:35,520
of new unproven ideas right so this is

1651
00:59:39,770 --> 00:59:37,680
you know this this is this is just my my

1652
00:59:41,270 --> 00:59:39,780
conjecture we've we've done some we've

1653
00:59:44,630 --> 00:59:41,280
done some computational modeling of it

1654
00:59:46,730 --> 00:59:44,640
which which I initially this was a um a

1655
00:59:49,010 --> 00:59:46,740
very uh clever student that I work with

1656
00:59:52,010 --> 00:59:49,020
named lakshman who I did did some models

1657
00:59:53,510 --> 00:59:52,020
with me and uh uh I initially thought it

1658
00:59:55,250 --> 00:59:53,520
was a bug and then I realized that no

1659
00:59:58,430 --> 00:59:55,260
actually this is this is this is the

1660
01:00:00,230 --> 00:59:58,440
feature the idea is this imagine so so

1661
01:00:02,630 --> 01:00:00,240
we've been working for a long a long

1662
01:00:05,270 --> 01:00:02,640
time on a concept of

1663
01:00:07,010 --> 01:00:05,280
um competency among embryonic parts and

1664
01:00:10,430 --> 01:00:07,020
what this means is basically the idea

1665
01:00:13,789 --> 01:00:10,440
that uh there are there are um

1666
01:00:16,190 --> 01:00:13,799
homeostatic feedback loops among various

1667
01:00:19,069 --> 01:00:16,200
cells and tissues and organs that

1668
01:00:21,049 --> 01:00:19,079
attempt to reach specific outcomes in

1669
01:00:23,210 --> 01:00:21,059
anatomical amorphous space despite

1670
01:00:25,190 --> 01:00:23,220
various perturbations so the idea is

1671
01:00:27,890 --> 01:00:25,200
that if you have a tadpole and you do

1672
01:00:29,569 --> 01:00:27,900
something to it whether by a mutation or

1673
01:00:31,430 --> 01:00:29,579
by a drug or something you do something

1674
01:00:33,530 --> 01:00:31,440
to it where the eye is a little

1675
01:00:35,750 --> 01:00:33,540
off-kilter or the mouth is a little off

1676
01:00:37,010 --> 01:00:35,760
all of these organs pretty much know

1677
01:00:39,109 --> 01:00:37,020
where they're supposed to be they will

1678
01:00:41,270 --> 01:00:39,119
try to minimize distance from other um

1679
01:00:42,829 --> 01:00:41,280
landmarks and they will Remodel and

1680
01:00:44,930 --> 01:00:42,839
eventually you get a normal frog so that

1681
01:00:48,230 --> 01:00:44,940
so that they will they will sort of

1682
01:00:49,789 --> 01:00:48,240
um a a recover the correct uh Anatomy

1683
01:00:51,589 --> 01:00:49,799
despite starting off in the wrong

1684
01:00:53,450 --> 01:00:51,599
position or even things like changes in

1685
01:00:55,130 --> 01:00:53,460
the number of cells or the size of cells

1686
01:00:58,069 --> 01:00:55,140
they're really good at getting their job

1687
01:00:59,809 --> 01:00:58,079
done despite various changes right so

1688
01:01:02,329 --> 01:00:59,819
okay so they have these competencies to

1689
01:01:04,490 --> 01:01:02,339
optimize specific uh things like like

1690
01:01:05,270 --> 01:01:04,500
their position and structure and things

1691
01:01:07,549 --> 01:01:05,280
like that

1692
01:01:09,109 --> 01:01:07,559
so uh so so that's so that's competency

1693
01:01:09,829 --> 01:01:09,119
now now here's here's the interesting

1694
01:01:13,970 --> 01:01:09,839
thing

1695
01:01:16,490 --> 01:01:13,980
imagine that you have a species that uh

1696
01:01:18,589 --> 01:01:16,500
has some some degree of that competency

1697
01:01:21,170 --> 01:01:18,599
and so you've got an individual of that

1698
01:01:23,390 --> 01:01:21,180
species comes up for selection uh

1699
01:01:25,190 --> 01:01:23,400
Fitness is high looks pretty good but

1700
01:01:27,530 --> 01:01:25,200
here's the problem selection doesn't

1701
01:01:30,470 --> 01:01:27,540
know whether the fitness is high because

1702
01:01:31,970 --> 01:01:30,480
his genome was amazing or the fitness is

1703
01:01:34,069 --> 01:01:31,980
high because the genome was actually so

1704
01:01:35,450 --> 01:01:34,079
so but the competency sort of made up

1705
01:01:38,030 --> 01:01:35,460
for it and now everything kind of got

1706
01:01:40,849 --> 01:01:38,040
back to where it needs to go so what the

1707
01:01:43,670 --> 01:01:40,859
competency apparently does is Shield

1708
01:01:45,710 --> 01:01:43,680
information from Evolution about the

1709
01:01:47,630 --> 01:01:45,720
actual genome it makes it harder to pick

1710
01:01:49,730 --> 01:01:47,640
the best genomes because you're

1711
01:01:51,829 --> 01:01:49,740
individuals that perform well don't

1712
01:01:55,069 --> 01:01:51,839
necessarily have the best genomes what

1713
01:01:56,450 --> 01:01:55,079
they do have is competency so what

1714
01:01:59,210 --> 01:01:56,460
happens and what happens in our

1715
01:02:00,650 --> 01:01:59,220
simulations is that when if you start

1716
01:02:04,370 --> 01:02:00,660
off with even a little bit of that

1717
01:02:06,829 --> 01:02:04,380
competency Evolution loses some power in

1718
01:02:08,390 --> 01:02:06,839
selecting the best genomes but it but

1719
01:02:10,190 --> 01:02:08,400
but where all the work tends to happen

1720
01:02:12,349 --> 01:02:10,200
is increasing the competency

1721
01:02:14,089 --> 01:02:12,359
so then the competency goes up so the

1722
01:02:15,530 --> 01:02:14,099
cells are even better at and the tissues

1723
01:02:18,470 --> 01:02:15,540
are even better at getting the job done

1724
01:02:20,930 --> 01:02:18,480
despite the bad genome that makes it

1725
01:02:23,089 --> 01:02:20,940
even worse that that that that that that

1726
01:02:25,069 --> 01:02:23,099
makes it even harder for for evolution

1727
01:02:26,809 --> 01:02:25,079
to see the best genomes which relieves

1728
01:02:29,210 --> 01:02:26,819
some of the pressure on having a good

1729
01:02:31,250 --> 01:02:29,220
genome but it basically puts all the

1730
01:02:32,870 --> 01:02:31,260
pressure on being really competent so

1731
01:02:36,349 --> 01:02:32,880
you've got this so so basically what

1732
01:02:39,109 --> 01:02:36,359
happens is that uh the the the genetic

1733
01:02:41,270 --> 01:02:39,119
fitness basically levels out at a really

1734
01:02:43,250 --> 01:02:41,280
sub-optimal level and in fact the

1735
01:02:45,770 --> 01:02:43,260
pressure is off of it so so it's

1736
01:02:47,450 --> 01:02:45,780
tolerant all kinds of craziness but the

1737
01:02:49,789 --> 01:02:47,460
competency and the mechanisms of

1738
01:02:53,030 --> 01:02:49,799
Competency get get pushed up really high

1739
01:02:54,410 --> 01:02:53,040
so in many animals and but there are

1740
01:02:56,270 --> 01:02:54,420
other factors that sort of push against

1741
01:02:57,710 --> 01:02:56,280
this ratchet but it becomes it becomes a

1742
01:03:00,650 --> 01:02:57,720
positive feedback loop it becomes a

1743
01:03:03,650 --> 01:03:00,660
ratchet for Optimal Performance despite

1744
01:03:06,470 --> 01:03:03,660
a sub-optimal genome and so in some

1745
01:03:07,549 --> 01:03:06,480
animals this sort of evens out at a

1746
01:03:10,130 --> 01:03:07,559
particular point but I think what

1747
01:03:12,349 --> 01:03:10,140
happened in planaria is that this whole

1748
01:03:14,930 --> 01:03:12,359
process ran away to its ultimate

1749
01:03:17,870 --> 01:03:14,940
conclusion the ultimate conclusion is

1750
01:03:20,390 --> 01:03:17,880
the competency algorithm became so good

1751
01:03:22,430 --> 01:03:20,400
that basically whatever the genome is

1752
01:03:24,829 --> 01:03:22,440
it's really good at creating and

1753
01:03:27,049 --> 01:03:24,839
maintaining a proper worm because it is

1754
01:03:28,970 --> 01:03:27,059
already being evolved in the presence of

1755
01:03:31,849 --> 01:03:28,980
a genome whose quality we cannot control

1756
01:03:33,829 --> 01:03:31,859
so so in computer science speak it's

1757
01:03:35,870 --> 01:03:33,839
kind of like um and Steve Frank put me

1758
01:03:37,670 --> 01:03:35,880
onto this analogy it's kind of like what

1759
01:03:40,130 --> 01:03:37,680
happens in raid arrays when you have a

1760
01:03:41,990 --> 01:03:40,140
nice raid array where the software makes

1761
01:03:44,150 --> 01:03:42,000
sure that you don't lose any data the

1762
01:03:47,210 --> 01:03:44,160
pressure is off to have really uh really

1763
01:03:48,950 --> 01:03:47,220
high quality media and so now you can

1764
01:03:50,450 --> 01:03:48,960
tolerate you can tolerate media with

1765
01:03:52,309 --> 01:03:50,460
lots of mistakes because because the

1766
01:03:54,589 --> 01:03:52,319
software takes care of it in the in the

1767
01:03:56,750 --> 01:03:54,599
raid and the on the architecture takes

1768
01:03:59,210 --> 01:03:56,760
care of it so so basically what happens

1769
01:04:02,150 --> 01:03:59,220
is you've got this animal where that uh

1770
01:04:05,270 --> 01:04:02,160
that runaway uh feedback loop went went

1771
01:04:08,510 --> 01:04:05,280
so far that the algorithm is amazing and

1772
01:04:10,309 --> 01:04:08,520
it's been it's been uh evolved

1773
01:04:12,589 --> 01:04:10,319
specifically for the ability to do what

1774
01:04:15,349 --> 01:04:12,599
it needs to do even though the hardware

1775
01:04:16,970 --> 01:04:15,359
is kind of crap and and it's it's

1776
01:04:18,710 --> 01:04:16,980
incredibly tolerant so this has a number

1777
01:04:19,789 --> 01:04:18,720
of implications that that to my

1778
01:04:22,849 --> 01:04:19,799
knowledge have never been explained

1779
01:04:24,829 --> 01:04:22,859
before for example in every kind every

1780
01:04:26,329 --> 01:04:24,839
other kind of animal you can you can

1781
01:04:28,789 --> 01:04:26,339
call a stock Center and you can get

1782
01:04:31,549 --> 01:04:28,799
mutants so you can get mice with with

1783
01:04:34,309 --> 01:04:31,559
kinky kind of Kink tails you can get

1784
01:04:36,289 --> 01:04:34,319
flies with red eyes and you can get uh

1785
01:04:38,329 --> 01:04:36,299
chickens without toes and you can get

1786
01:04:40,010 --> 01:04:38,339
you know humans come with that it has

1787
01:04:41,030 --> 01:04:40,020
various you know albinos and things you

1788
01:04:43,809 --> 01:04:41,040
can you can there's there's always

1789
01:04:47,150 --> 01:04:43,819
mutants that you can get uh planaria

1790
01:04:49,309 --> 01:04:47,160
there are no there are no abnormal lines

1791
01:04:51,349 --> 01:04:49,319
of planaria anywhere except for the only

1792
01:04:52,910 --> 01:04:51,359
exception is our two-headed line and

1793
01:04:55,730 --> 01:04:52,920
that that one's not genetic that one's

1794
01:04:57,770 --> 01:04:55,740
that one's bioelectric so so isn't it

1795
01:04:59,630 --> 01:04:57,780
amazing that that that nobody has been

1796
01:05:01,549 --> 01:04:59,640
able despite despite a hundred and you

1797
01:05:03,589 --> 01:05:01,559
know I don't know 120 years of

1798
01:05:07,010 --> 01:05:03,599
experiments with planaria nobody has

1799
01:05:08,690 --> 01:05:07,020
isolated a uh a line of planaria that is

1800
01:05:10,430 --> 01:05:08,700
anything other than a perfect planarian

1801
01:05:12,650 --> 01:05:10,440
I think this is why I think it's because

1802
01:05:13,910 --> 01:05:12,660
they have been actually selected for

1803
01:05:15,470 --> 01:05:13,920
being able to do what they need to do

1804
01:05:18,950 --> 01:05:15,480
despite the fact that there the the

1805
01:05:20,450 --> 01:05:18,960
hardware is just very junky and so so

1806
01:05:22,549 --> 01:05:20,460
that's my that's my current that's my

1807
01:05:26,270 --> 01:05:22,559
current current take on it and and and

1808
01:05:28,670 --> 01:05:26,280
really uh it puts more kind of more

1809
01:05:30,770 --> 01:05:28,680
emphasis on on on the algorithm and the

1810
01:05:33,170 --> 01:05:30,780
decision making among that cellular

1811
01:05:34,250 --> 01:05:33,180
Collective of what what you know what

1812
01:05:35,809 --> 01:05:34,260
are we going to build and what's the

1813
01:05:38,329 --> 01:05:35,819
algorithm for for making sure that we're

1814
01:05:40,010 --> 01:05:38,339
all working to build the correct thing

1815
01:05:42,710 --> 01:05:40,020
so if you translate this idea into

1816
01:05:45,230 --> 01:05:42,720
computer science a way to look at it is

1817
01:05:47,329 --> 01:05:45,240
imagine that you find some computers

1818
01:05:50,930 --> 01:05:47,339
that have

1819
01:05:52,730 --> 01:05:50,940
um artists that are very very noisy and

1820
01:05:54,349 --> 01:05:52,740
where the heart is basically makes lots

1821
01:05:57,109 --> 01:05:54,359
and lots of mistakes in encoding things

1822
01:05:58,910 --> 01:05:57,119
and bits often flip and so on and you

1823
01:06:00,950 --> 01:05:58,920
will find that these computers still

1824
01:06:02,750 --> 01:06:00,960
work and they work on pretty much the

1825
01:06:06,109 --> 01:06:02,760
same way as the other computers that you

1826
01:06:08,210 --> 01:06:06,119
have and there is an orthodox sect of

1827
01:06:10,430 --> 01:06:08,220
computer scientists that thinks uh it is

1828
01:06:13,370 --> 01:06:10,440
necessary that uh the every bit on the

1829
01:06:15,289 --> 01:06:13,380
hard disk is completely reliable or

1830
01:06:17,990 --> 01:06:15,299
reliable to such a degree that you only

1831
01:06:20,809 --> 01:06:18,000
have a mistake once every 100 trillion

1832
01:06:22,069 --> 01:06:20,819
copies and you can have an error

1833
01:06:23,750 --> 01:06:22,079
correction quote running on the hard

1834
01:06:25,549 --> 01:06:23,760
disk at the low level that corrects this

1835
01:06:27,170 --> 01:06:25,559
and after some point it doesn't become

1836
01:06:29,029 --> 01:06:27,180
efficient anymore so you need to have

1837
01:06:31,910 --> 01:06:29,039
reliable hard disks to be able to have

1838
01:06:34,130 --> 01:06:31,920
computers that work like this but uh how

1839
01:06:36,230 --> 01:06:34,140
would these other computers work and it

1840
01:06:38,329 --> 01:06:36,240
basically means that you create a

1841
01:06:40,430 --> 01:06:38,339
virtual structure on top of the noisy

1842
01:06:43,430 --> 01:06:40,440
structure that is correcting for

1843
01:06:45,230 --> 01:06:43,440
whatever degree of um uncertainty you

1844
01:06:48,410 --> 01:06:45,240
have or the degree of Randomness that

1845
01:06:51,109 --> 01:06:48,420
gets injected into your substrate David

1846
01:06:55,549 --> 01:06:51,119
Dave Eckley has a very nice metaphor for

1847
01:06:58,370 --> 01:06:55,559
this you know maybe yeah yeah he's a I

1848
01:07:00,349 --> 01:06:58,380
think a beautiful artist who explores

1849
01:07:02,390 --> 01:07:00,359
complexity by tinkering with

1850
01:07:04,670 --> 01:07:02,400
computational models I really find this

1851
01:07:06,950 --> 01:07:04,680
work very inspiring and he has this idea

1852
01:07:09,529 --> 01:07:06,960
of best effort Computing so in his view

1853
01:07:11,569 --> 01:07:09,539
our own nervous system is a best effort

1854
01:07:13,730 --> 01:07:11,579
computer it's one that does not rely on

1855
01:07:17,329 --> 01:07:13,740
the other neurons around you working

1856
01:07:20,210 --> 01:07:17,339
perfectly uh but make an effort to be

1857
01:07:22,309 --> 01:07:20,220
better than random and then you stack

1858
01:07:25,309 --> 01:07:22,319
the improbabilities empirically by

1859
01:07:28,970 --> 01:07:25,319
having a system that evolves to measure

1860
01:07:31,190 --> 01:07:28,980
in effect the unreliability of its

1861
01:07:32,990 --> 01:07:31,200
components and then stack the

1862
01:07:35,089 --> 01:07:33,000
probabilities until you get the system

1863
01:07:36,829 --> 01:07:35,099
to be deterministic enough to do what

1864
01:07:40,069 --> 01:07:36,839
you're doing with what to do with it

1865
01:07:42,770 --> 01:07:40,079
right so you uh if you have a system

1866
01:07:45,349 --> 01:07:42,780
that is as in a planaria inherently very

1867
01:07:47,750 --> 01:07:45,359
noisy where the genome is an unreliable

1868
01:07:50,750 --> 01:07:47,760
witness of what should be done and the

1869
01:07:52,789 --> 01:07:50,760
body you just need to interpret it in a

1870
01:07:55,190 --> 01:07:52,799
way that Stacks the probabilities that

1871
01:07:58,069 --> 01:07:55,200
is evaluating things with much more

1872
01:08:00,349 --> 01:07:58,079
error tolerance and maybe this is always

1873
01:08:02,150 --> 01:08:00,359
the case maybe there is a Continuum

1874
01:08:03,950 --> 01:08:02,160
um maybe not it's also possible that

1875
01:08:06,650 --> 01:08:03,960
there is some kind of phase shift where

1876
01:08:08,690 --> 01:08:06,660
your switch from organisms with reliable

1877
01:08:10,069 --> 01:08:08,700
genomes to organisms with noisy genomes

1878
01:08:11,450 --> 01:08:10,079
and you would basically use a completely

1879
01:08:13,789 --> 01:08:11,460
different way to construct the organism

1880
01:08:16,130 --> 01:08:13,799
as a result but it's a it's a very

1881
01:08:18,829 --> 01:08:16,140
interesting hypothesis then to see if if

1882
01:08:20,450 --> 01:08:18,839
this is a radical thing or a gradual

1883
01:08:23,030 --> 01:08:20,460
thing that happens on all organisms to

1884
01:08:25,070 --> 01:08:23,040
some degree yeah what I also like about

1885
01:08:28,130 --> 01:08:25,080
this uh description that you give about

1886
01:08:31,070 --> 01:08:28,140
how the organism emerges it maps on in

1887
01:08:34,430 --> 01:08:31,080
some sense also in how perception Works

1888
01:08:35,749 --> 01:08:34,440
in in our own mind at the moment measure

1889
01:08:37,610 --> 01:08:35,759
learning is mostly focused on

1890
01:08:40,669 --> 01:08:37,620
recognizing images

1891
01:08:42,709 --> 01:08:40,679
so our individual frames and the you

1892
01:08:44,090 --> 01:08:42,719
feed in information frame by frame and

1893
01:08:47,209 --> 01:08:44,100
the information is relatively

1894
01:08:49,550 --> 01:08:47,219
disconnected from a system like Delhi 2

1895
01:08:51,110 --> 01:08:49,560
is trained by giving it several hundreds

1896
01:08:52,789 --> 01:08:51,120
of millions of images

1897
01:08:54,530 --> 01:08:52,799
and they are disconnected the they're

1898
01:08:56,870 --> 01:08:54,540
not adjacent images in the space of

1899
01:08:59,390 --> 01:08:56,880
images and a baby could not probably

1900
01:09:01,309 --> 01:08:59,400
learn from giving 600 million images in

1901
01:09:02,689 --> 01:09:01,319
a dark room and only looking at this

1902
01:09:05,510 --> 01:09:02,699
introduce the structure of the world

1903
01:09:07,370 --> 01:09:05,520
from this whereas Delhi can which gives

1904
01:09:09,769 --> 01:09:07,380
Testament to the power of our

1905
01:09:11,749 --> 01:09:09,779
statistical methods and Hardware that we

1906
01:09:14,030 --> 01:09:11,759
have that far surpasses I think the

1907
01:09:15,289 --> 01:09:14,040
combined power and valuability of brains

1908
01:09:17,150 --> 01:09:15,299
which probably would not be able to

1909
01:09:18,289 --> 01:09:17,160
integrate so much information over such

1910
01:09:20,269 --> 01:09:18,299
a big distance

1911
01:09:22,729 --> 01:09:20,279
for us the world is learnable because

1912
01:09:25,309 --> 01:09:22,739
it's adjacent frames are correlated

1913
01:09:28,010 --> 01:09:25,319
basically information gets preserved in

1914
01:09:29,269 --> 01:09:28,020
the world so time and we only need to

1915
01:09:31,130 --> 01:09:29,279
learn the way in which the information

1916
01:09:33,349 --> 01:09:31,140
gets transmogorified

1917
01:09:34,910 --> 01:09:33,359
and these transformifications of

1918
01:09:37,010 --> 01:09:34,920
information means that we have a dynamic

1919
01:09:38,689 --> 01:09:37,020
World in which the static image is an

1920
01:09:40,849 --> 01:09:38,699
exception the identity function is a

1921
01:09:43,849 --> 01:09:40,859
special case of how the universe changes

1922
01:09:46,189 --> 01:09:43,859
and we mostly learn change we just got

1923
01:09:48,289 --> 01:09:46,199
visited by my cat and my cat has

1924
01:09:49,910 --> 01:09:48,299
difficulty to recognize static objects

1925
01:09:51,829 --> 01:09:49,920
compared to moving objects where it's

1926
01:09:54,229 --> 01:09:51,839
much much easier to see a moving ball

1927
01:09:55,850 --> 01:09:54,239
than a ball that is laying still yeah

1928
01:09:57,530 --> 01:09:55,860
and it's because it's much easier to

1929
01:10:00,290 --> 01:09:57,540
segment it out the environment when it

1930
01:10:01,910 --> 01:10:00,300
moves right so the task of learning on a

1931
01:10:03,890 --> 01:10:01,920
moving environmental Dynamic environment

1932
01:10:06,709 --> 01:10:03,900
is much easier because it imposes

1933
01:10:09,290 --> 01:10:06,719
constraints on the world and so how do

1934
01:10:11,570 --> 01:10:09,300
we represent a moving world compared to

1935
01:10:14,450 --> 01:10:11,580
a static well the semantics of features

1936
01:10:16,250 --> 01:10:14,460
changes and the uh an object is

1937
01:10:18,709 --> 01:10:16,260
basically composed of features that are

1938
01:10:21,169 --> 01:10:18,719
there can be objects themselves and the

1939
01:10:23,930 --> 01:10:21,179
scene is the decomposition of all the

1940
01:10:25,850 --> 01:10:23,940
features that we see into a complete set

1941
01:10:27,410 --> 01:10:25,860
of objects that explain the entirety of

1942
01:10:29,030 --> 01:10:27,420
the scene and the interaction between

1943
01:10:32,450 --> 01:10:29,040
them and causality is the interaction

1944
01:10:33,950 --> 01:10:32,460
between objects right and in a static

1945
01:10:35,510 --> 01:10:33,960
image these objects don't do anything

1946
01:10:37,130 --> 01:10:35,520
they don't interact with each other they

1947
01:10:38,870 --> 01:10:37,140
just stand in some kind of relationship

1948
01:10:40,610 --> 01:10:38,880
that you need to infer which is super

1949
01:10:43,729 --> 01:10:40,620
difficult because you only have this

1950
01:10:46,250 --> 01:10:43,739
static snapshot and so the features are

1951
01:10:48,890 --> 01:10:46,260
classifiers that tell you how to whether

1952
01:10:52,490 --> 01:10:48,900
a feature is a hand or a foot or a pen

1953
01:10:54,950 --> 01:10:52,500
or a Sun or a flashlight or whatever and

1954
01:10:56,209 --> 01:10:54,960
how they relate to the larger scene in

1955
01:10:57,590 --> 01:10:56,219
which again you have a static

1956
01:10:59,450 --> 01:10:57,600
relationship in which you need to

1957
01:11:01,490 --> 01:10:59,460
classify the objects based on the

1958
01:11:02,870 --> 01:11:01,500
features that contribute to them and you

1959
01:11:04,490 --> 01:11:02,880
need to find some kind of description

1960
01:11:05,930 --> 01:11:04,500
where you interpret the features which

1961
01:11:07,610 --> 01:11:05,940
are usually ambiguous and could be many

1962
01:11:09,350 --> 01:11:07,620
different things depending on the

1963
01:11:11,510 --> 01:11:09,360
context which you interpret them into

1964
01:11:14,450 --> 01:11:11,520
one optimal Global configuration right

1965
01:11:16,790 --> 01:11:14,460
but if scene is moving this changes a

1966
01:11:18,530 --> 01:11:16,800
little bit what happens now is that the

1967
01:11:20,689 --> 01:11:18,540
features become operators they're no

1968
01:11:22,669 --> 01:11:20,699
longer classifiers that tell you how

1969
01:11:24,050 --> 01:11:22,679
your internal State needs to change how

1970
01:11:25,669 --> 01:11:24,060
your world needs to change or your

1971
01:11:27,709 --> 01:11:25,679
simulation of the universe in your mind

1972
01:11:30,709 --> 01:11:27,719
needs to change to track the sensory

1973
01:11:33,350 --> 01:11:30,719
patterns right so a feature now is an

1974
01:11:36,110 --> 01:11:33,360
change operator a transformation

1975
01:11:38,270 --> 01:11:36,120
and uh the feature is in some sense a

1976
01:11:39,830 --> 01:11:38,280
controller that tells you how the bits

1977
01:11:42,649 --> 01:11:39,840
are moving and in your local model of

1978
01:11:44,930 --> 01:11:42,659
the universe and they are organized in a

1979
01:11:46,729 --> 01:11:44,940
hierarchy of controllers

1980
01:11:48,470 --> 01:11:46,739
and these controllers need to be turned

1981
01:11:50,030 --> 01:11:48,480
on and off at the level of the scene and

1982
01:11:51,590 --> 01:11:50,040
they have a lot of flexibility once you

1983
01:11:52,729 --> 01:11:51,600
have them they can move around in the

1984
01:11:54,590 --> 01:11:52,739
scene they're basically no

1985
01:11:57,050 --> 01:11:54,600
self-organizing self-stabilizing

1986
01:11:59,450 --> 01:11:57,060
entities in the same way as the mouse is

1987
01:12:01,310 --> 01:11:59,460
moving around in your organism a feature

1988
01:12:03,110 --> 01:12:01,320
can move around in the organism and

1989
01:12:05,450 --> 01:12:03,120
shift itself around to communicate with

1990
01:12:08,090 --> 01:12:05,460
other features until they negotiate a

1991
01:12:10,130 --> 01:12:08,100
valid interpretation of reality

1992
01:12:12,950 --> 01:12:10,140
that's that's incredibly interesting

1993
01:12:15,950 --> 01:12:12,960
because I you know as soon as you

1994
01:12:18,830 --> 01:12:15,960
started saying that uh I was starting to

1995
01:12:20,750 --> 01:12:18,840
think that the the virtualization that

1996
01:12:22,850 --> 01:12:20,760
enables right so the earlier part of

1997
01:12:24,250 --> 01:12:22,860
which we're saying the virtualization of

1998
01:12:27,410 --> 01:12:24,260
um

1999
01:12:28,370 --> 01:12:27,420
uh the the information that allows you

2000
01:12:30,229 --> 01:12:28,380
to

2001
01:12:33,410 --> 01:12:30,239
um deal with with unreliable hardware

2002
01:12:35,209 --> 01:12:33,420
and everything the the bioelectric uh

2003
01:12:36,770 --> 01:12:35,219
circuits that we deal with are a great

2004
01:12:39,229 --> 01:12:36,780
candidate for that because actually we

2005
01:12:41,209 --> 01:12:39,239
see exactly that we see a bioelectric

2006
01:12:43,310 --> 01:12:41,219
pattern that is very resistant to

2007
01:12:44,450 --> 01:12:43,320
changes in the details and make sure

2008
01:12:46,910 --> 01:12:44,460
that everybody does the right thing

2009
01:12:49,250 --> 01:12:46,920
under a wide range of you know different

2010
01:12:50,870 --> 01:12:49,260
defects and so on but but but even more

2011
01:12:52,729 --> 01:12:50,880
than that the other thing that what you

2012
01:12:54,410 --> 01:12:52,739
were just um emphasizing this the fact

2013
01:12:56,630 --> 01:12:54,420
that we learned the Delta right and that

2014
01:12:58,669 --> 01:12:56,640
and that we're looking for change very

2015
01:13:00,290 --> 01:12:58,679
interesting if you if you pivot the

2016
01:13:02,770 --> 01:13:00,300
whole thing from the temporal domain to

2017
01:13:05,930 --> 01:13:02,780
the spatial domain so so in development

2018
01:13:08,270 --> 01:13:05,940
what we we when we look at these

2019
01:13:10,070 --> 01:13:08,280
bioelectric patterns now now these

2020
01:13:11,870 --> 01:13:10,080
patterns are across space not across

2021
01:13:13,310 --> 01:13:11,880
time so so on like a neuroscience where

2022
01:13:15,410 --> 01:13:13,320
everything is kind of in the temporal

2023
01:13:17,870 --> 01:13:15,420
domain for neurons these things these

2024
01:13:19,490 --> 01:13:17,880
are voltage static voltage patterns

2025
01:13:22,729 --> 01:13:19,500
across tissue right across the whole

2026
01:13:24,709 --> 01:13:22,739
thing so for the longest time you know

2027
01:13:26,810 --> 01:13:24,719
we asked this question how are these

2028
01:13:28,790 --> 01:13:26,820
read out what how do cells actually read

2029
01:13:30,770 --> 01:13:28,800
these because because one possibility

2030
01:13:33,649 --> 01:13:30,780
early this was a very early hypothesis

2031
01:13:35,450 --> 01:13:33,659
you know 20 years ago was that maybe the

2032
01:13:37,189 --> 01:13:35,460
local voltage tells every cell what to

2033
01:13:39,490 --> 01:13:37,199
be so it's like a paint by numbers kind

2034
01:13:42,590 --> 01:13:39,500
of thing and every and and each voltage

2035
01:13:44,270 --> 01:13:42,600
you know each voltage value corresponds

2036
01:13:46,729 --> 01:13:44,280
to some kind of outcome that turned out

2037
01:13:48,530 --> 01:13:46,739
to be false what we did find is that

2038
01:13:51,050 --> 01:13:48,540
there and we have computational models

2039
01:13:54,410 --> 01:13:51,060
of how this works now

2040
01:13:56,570 --> 01:13:54,420
um what is read out is the Delta the

2041
01:13:58,010 --> 01:13:56,580
difference between regions it doesn't

2042
01:13:59,930 --> 01:13:58,020
care nobody cares about what the

2043
01:14:02,330 --> 01:13:59,940
absolute voltage is what what what is

2044
01:14:04,250 --> 01:14:02,340
read out in terms of outcomes for for

2045
01:14:06,290 --> 01:14:04,260
Downstream cell Behavior gene expression

2046
01:14:07,729 --> 01:14:06,300
all that what is actually right out is

2047
01:14:10,070 --> 01:14:07,739
the voltage difference between two

2048
01:14:11,689 --> 01:14:10,080
adjacent domains so that is exactly

2049
01:14:13,729 --> 01:14:11,699
actually what it's doing just in a

2050
01:14:16,970 --> 01:14:13,739
spatial domain it it only keys off of

2051
01:14:19,850 --> 01:14:16,980
the Delta and what is in what is learned

2052
01:14:21,709 --> 01:14:19,860
from that is um exactly just as as you

2053
01:14:23,810 --> 01:14:21,719
as you were saying it modifies uh the

2054
01:14:25,370 --> 01:14:23,820
controller for for what's Downstream of

2055
01:14:27,050 --> 01:14:25,380
that and there may be multiple ones that

2056
01:14:29,030 --> 01:14:27,060
are sort of moving around in Cohen

2057
01:14:30,169 --> 01:14:29,040
having I mean it's a very it's a very

2058
01:14:31,910 --> 01:14:30,179
compelling

2059
01:14:34,130 --> 01:14:31,920
um picture actually and way to look at

2060
01:14:36,410 --> 01:14:34,140
some of the some of the simulations that

2061
01:14:38,750 --> 01:14:36,420
that we've been doing about how the

2062
01:14:40,970 --> 01:14:38,760
bioelectric data are interpreted by the

2063
01:14:42,770 --> 01:14:40,980
rest of the cells you know it's very

2064
01:14:44,990 --> 01:14:42,780
interesting could we take a couple of

2065
01:14:46,790 --> 01:14:45,000
couple minute break yeah sure I'll get a

2066
01:14:48,890 --> 01:14:46,800
new coffee all right

2067
01:14:49,970 --> 01:14:48,900
speaking of coffee a brief note from our

2068
01:14:52,250 --> 01:14:49,980
sponsor

2069
01:14:54,169 --> 01:14:52,260
coffee helps me work it helps me fast

2070
01:14:56,270 --> 01:14:54,179
from carbs it's become one of the best

2071
01:14:57,649 --> 01:14:56,280
parts of my day consistently that's why

2072
01:14:59,750 --> 01:14:57,659
I'm delighted that we're collaborating

2073
01:15:01,910 --> 01:14:59,760
with Trade coffee they partner with top

2074
01:15:04,250 --> 01:15:01,920
independent Roasters to Freshly roast

2075
01:15:06,229 --> 01:15:04,260
and send the finest coffee in the

2076
01:15:08,330 --> 01:15:06,239
country directly to your home on your

2077
01:15:10,430 --> 01:15:08,340
preferred schedule this matters to me as

2078
01:15:12,649 --> 01:15:10,440
I work from home their team of experts

2079
01:15:15,229 --> 01:15:12,659
do all the work testing hundreds of

2080
01:15:18,590 --> 01:15:15,239
disparate coffees to land on a final

2081
01:15:20,930 --> 01:15:18,600
curated collection of 450 exceptional

2082
01:15:22,910 --> 01:15:20,940
coffees I chose these three and the team

2083
01:15:24,470 --> 01:15:22,920
at Trade coffee worked to create a

2084
01:15:26,510 --> 01:15:24,480
special lineup for theories of

2085
01:15:28,430 --> 01:15:26,520
everything for the tow audience based on

2086
01:15:30,350 --> 01:15:28,440
some questions they asked me such as how

2087
01:15:32,510 --> 01:15:30,360
much caffeine do I enjoy and what's the

2088
01:15:34,729 --> 01:15:32,520
bitterness ratio Etc you can get that

2089
01:15:36,770 --> 01:15:34,739
line up or if that's not let's say your

2090
01:15:39,169 --> 01:15:36,780
cup of coffee then you can take your own

2091
01:15:41,450 --> 01:15:39,179
quiz on their website to find a set that

2092
01:15:43,070 --> 01:15:41,460
matches your specific profile if you'd

2093
01:15:44,810 --> 01:15:43,080
like to support small businesses and

2094
01:15:47,390 --> 01:15:44,820
Brew the best cup of coffee you've ever

2095
01:15:49,550 --> 01:15:47,400
made at home then it's time to try Trade

2096
01:15:52,010 --> 01:15:49,560
coffee right now trade is offering our

2097
01:15:54,910 --> 01:15:52,020
listeners thirty dollars off your first

2098
01:15:58,450 --> 01:15:54,920
order plus free shipping at

2099
01:16:01,430 --> 01:15:58,460
drinktrade.com everything that's

2100
01:16:04,310 --> 01:16:01,440
drinktrade.com everything for thirty

2101
01:16:06,110 --> 01:16:04,320
dollars off so Professor Levin use the

2102
01:16:07,430 --> 01:16:06,120
word competence earlier and I'd like you

2103
01:16:08,510 --> 01:16:07,440
to Define that

2104
01:16:09,350 --> 01:16:08,520
yeah

2105
01:16:12,530 --> 01:16:09,360
um

2106
01:16:15,770 --> 01:16:12,540
I in order to Define it uh I I want to

2107
01:16:19,130 --> 01:16:15,780
put out two two concepts to this one one

2108
01:16:20,930 --> 01:16:19,140
idea is that to me and this goes back to

2109
01:16:24,649 --> 01:16:20,940
what we were talking about before is the

2110
01:16:28,610 --> 01:16:24,659
engineering stance on things I

2111
01:16:30,830 --> 01:16:28,620
think that useful cognitive claims such

2112
01:16:32,810 --> 01:16:30,840
as something you know when you say this

2113
01:16:34,790 --> 01:16:32,820
system has whatever or it can whatever

2114
01:16:36,709 --> 01:16:34,800
right as far as various types of

2115
01:16:38,870 --> 01:16:36,719
cognitive capacities I think those kind

2116
01:16:41,030 --> 01:16:38,880
of claims are really engineering uh

2117
01:16:43,189 --> 01:16:41,040
claims that is when you tell me that

2118
01:16:45,590 --> 01:16:43,199
something is competent at a particular

2119
01:16:47,450 --> 01:16:45,600
level maybe right so so you can think

2120
01:16:48,370 --> 01:16:47,460
about like like wiener and rosenberg's

2121
01:16:51,229 --> 01:16:48,380
scale of

2122
01:16:53,390 --> 01:16:51,239
cognition that goes from simple um you

2123
01:16:55,850 --> 01:16:53,400
know simple passive materials and then

2124
01:16:57,470 --> 01:16:55,860
reflexes and then all the way up to kind

2125
01:16:59,090 --> 01:16:57,480
of second order metacognition and all

2126
01:17:01,729 --> 01:16:59,100
that when you tell me that something is

2127
01:17:04,070 --> 01:17:01,739
on that ladder and where it is what

2128
01:17:06,709 --> 01:17:04,080
you're really telling me is if I want to

2129
01:17:08,689 --> 01:17:06,719
predict Its Behavior or I want to use it

2130
01:17:10,610 --> 01:17:08,699
in an engineering context or I want to

2131
01:17:13,070 --> 01:17:10,620
interact with it or relate to it in some

2132
01:17:14,390 --> 01:17:13,080
way this is what I can expect so right

2133
01:17:16,970 --> 01:17:14,400
so that's what you really telling me so

2134
01:17:19,010 --> 01:17:16,980
all of these terms what they really are

2135
01:17:22,430 --> 01:17:19,020
are engineering protocols so if you tell

2136
01:17:25,070 --> 01:17:22,440
me that something has the capacity to do

2137
01:17:26,750 --> 01:17:25,080
um a associate of learning or or

2138
01:17:28,550 --> 01:17:26,760
whatever what you're telling me is that

2139
01:17:30,229 --> 01:17:28,560
hey you can do something more with this

2140
01:17:33,350 --> 01:17:30,239
than you could with a mechanical clock

2141
01:17:35,209 --> 01:17:33,360
you can provide certain types of uh

2142
01:17:36,470 --> 01:17:35,219
stimuli or experiences and you can

2143
01:17:38,270 --> 01:17:36,480
expect that to do this or that

2144
01:17:40,970 --> 01:17:38,280
afterwards or if you tell me that

2145
01:17:43,370 --> 01:17:40,980
something is a um a homeous you know a

2146
01:17:46,010 --> 01:17:43,380
homeostat that means that hey I can

2147
01:17:47,990 --> 01:17:46,020
count on it to keep some variable in at

2148
01:17:50,209 --> 01:17:48,000
a particular range without having to be

2149
01:17:51,770 --> 01:17:50,219
there myself to control it all the way

2150
01:17:53,149 --> 01:17:51,780
it has a certain autonomy now how much

2151
01:17:55,550 --> 01:17:53,159
right and if you tell me that something

2152
01:17:57,530 --> 01:17:55,560
is really intelligent and it can do XYZ

2153
01:17:59,090 --> 01:17:57,540
then I know that okay you're telling me

2154
01:18:01,250 --> 01:17:59,100
that it has even more autonomous

2155
01:18:02,870 --> 01:18:01,260
behavior in certain contexts so so all

2156
01:18:04,850 --> 01:18:02,880
of these terms to me what they really

2157
01:18:06,350 --> 01:18:04,860
are they're not and that has an

2158
01:18:08,870 --> 01:18:06,360
important implication the implication is

2159
01:18:10,850 --> 01:18:08,880
that they're Observer dependent that

2160
01:18:13,370 --> 01:18:10,860
that that that you you've picked some

2161
01:18:15,290 --> 01:18:13,380
kind of problem space you've picked uh

2162
01:18:16,729 --> 01:18:15,300
some kind of perspective and from that

2163
01:18:18,590 --> 01:18:16,739
problem problem space in that

2164
01:18:21,229 --> 01:18:18,600
perspective you're telling me that with

2165
01:18:23,870 --> 01:18:21,239
with sort of given certain goal States

2166
01:18:25,610 --> 01:18:23,880
this system has that much competency to

2167
01:18:27,290 --> 01:18:25,620
pursue those gold States and different

2168
01:18:28,910 --> 01:18:27,300
observers can have different views on

2169
01:18:31,130 --> 01:18:28,920
this for any given system so for example

2170
01:18:33,229 --> 01:18:31,140
somebody might look at a brain like

2171
01:18:34,610 --> 01:18:33,239
let's say a human brain and say well I'm

2172
01:18:36,229 --> 01:18:34,620
pretty sure the only thing this is this

2173
01:18:38,090 --> 01:18:36,239
is a paperweight so it's really pretty

2174
01:18:39,890 --> 01:18:38,100
much just competent in going down

2175
01:18:41,570 --> 01:18:39,900
gravitational gradient so all I can do

2176
01:18:43,130 --> 01:18:41,580
is hold down paper that's it and

2177
01:18:44,270 --> 01:18:43,140
somebody else will look at and say ah

2178
01:18:45,110 --> 01:18:44,280
you missed the whole point you missed

2179
01:18:46,729 --> 01:18:45,120
the whole point this thing is

2180
01:18:49,550 --> 01:18:46,739
competencies in behavioral space and

2181
01:18:52,669 --> 01:18:49,560
linguistic space right so these are all

2182
01:18:54,350 --> 01:18:52,679
um empirically testable uh engineering

2183
01:18:56,390 --> 01:18:54,360
claims about what you can expect the

2184
01:18:59,810 --> 01:18:56,400
system to do so when I say competency

2185
01:19:01,430 --> 01:18:59,820
what I mean is we specify a space a

2186
01:19:02,450 --> 01:19:01,440
problem space and at the time when we

2187
01:19:05,090 --> 01:19:02,460
were talking about this the problem

2188
01:19:07,310 --> 01:19:05,100
space that I was talking about was the

2189
01:19:09,169 --> 01:19:07,320
the anatomical morphe space that was the

2190
01:19:10,610 --> 01:19:09,179
space we were talking about so so the

2191
01:19:12,590 --> 01:19:10,620
space of possible anatomical

2192
01:19:14,630 --> 01:19:12,600
configurations and specifically

2193
01:19:17,149 --> 01:19:14,640
navigating that morph of space so you

2194
01:19:19,790 --> 01:19:17,159
start off as an egg or you start off as

2195
01:19:22,070 --> 01:19:19,800
a damaged limb or whatever and you

2196
01:19:23,630 --> 01:19:22,080
navigate that morphe space into the

2197
01:19:25,790 --> 01:19:23,640
correct structure so so when I say

2198
01:19:28,850 --> 01:19:25,800
competency I mean you have the ability

2199
01:19:31,370 --> 01:19:28,860
to deploy certain kinds of tricks to

2200
01:19:34,910 --> 01:19:31,380
navigate that amorphous space with some

2201
01:19:36,530 --> 01:19:34,920
level of uh performance that I can count

2202
01:19:38,510 --> 01:19:36,540
on and so the competency might be really

2203
01:19:40,310 --> 01:19:38,520
low or it might be really high then I

2204
01:19:42,169 --> 01:19:40,320
would have to make specific claims about

2205
01:19:43,790 --> 01:19:42,179
what I mean here's an example of a

2206
01:19:44,930 --> 01:19:43,800
compound and there are many you know if

2207
01:19:46,729 --> 01:19:44,940
you just think about the Behavioral

2208
01:19:48,830 --> 01:19:46,739
Science of navigation there are many

2209
01:19:50,630 --> 01:19:48,840
competencies you can think about does it

2210
01:19:51,950 --> 01:19:50,640
you know does it does it know ahead of

2211
01:19:54,350 --> 01:19:51,960
time where it's going does it have a

2212
01:19:56,810 --> 01:19:54,360
memory of where it's been or is it a

2213
01:19:59,870 --> 01:19:56,820
very simple you know sort of reflex arc

2214
01:20:01,970 --> 01:19:59,880
is all it has or um here's one one

2215
01:20:04,070 --> 01:20:01,980
example of a pretty cool competency that

2216
01:20:04,729 --> 01:20:04,080
um that that a lot of biological systems

2217
01:20:08,570 --> 01:20:04,739
have

2218
01:20:11,050 --> 01:20:08,580
if we take some cells that are in the

2219
01:20:14,570 --> 01:20:11,060
tail of a tadpole and we

2220
01:20:16,910 --> 01:20:14,580
give them a particular we we modify

2221
01:20:19,550 --> 01:20:16,920
their ion channels uh with such that

2222
01:20:22,370 --> 01:20:19,560
they now acquire

2223
01:20:24,290 --> 01:20:22,380
um a a the goal of navigating to an eye

2224
01:20:25,490 --> 01:20:24,300
fate in more in in this morphe space

2225
01:20:27,590 --> 01:20:25,500
meaning that they're going to make an

2226
01:20:28,850 --> 01:20:27,600
eye these things in fact will create an

2227
01:20:30,950 --> 01:20:28,860
eye and they'll make an eye in the tail

2228
01:20:32,750 --> 01:20:30,960
on the gut wherever you want but one of

2229
01:20:34,729 --> 01:20:32,760
the cool and so that's already pretty

2230
01:20:38,630 --> 01:20:34,739
pretty cool but but one of the amazing

2231
01:20:40,550 --> 01:20:38,640
aspects is if I only modify a few cells

2232
01:20:42,290 --> 01:20:40,560
not enough to make an actual eye just

2233
01:20:43,970 --> 01:20:42,300
just a handful of cells and we've done

2234
01:20:46,850 --> 01:20:43,980
this and you can see this work

2235
01:20:48,590 --> 01:20:46,860
one of the competencies they have is to

2236
01:20:51,229 --> 01:20:48,600
recruit local neighbors that were

2237
01:20:53,030 --> 01:20:51,239
themselves not in any way manipulated to

2238
01:20:54,410 --> 01:20:53,040
help them achieve that goal it's a

2239
01:20:55,790 --> 01:20:54,420
little bit like in an ant colony right

2240
01:20:57,290 --> 01:20:55,800
there's a there's this idea of

2241
01:20:58,490 --> 01:20:57,300
Recruitment and ants and termites is an

2242
01:21:00,530 --> 01:20:58,500
idea of recruitment where where

2243
01:21:01,610 --> 01:21:00,540
individuals can can recruit others and

2244
01:21:03,350 --> 01:21:01,620
you know talk about a flexible

2245
01:21:05,750 --> 01:21:03,360
collective intelligence this is it this

2246
01:21:08,090 --> 01:21:05,760
is they they you you've re-specified the

2247
01:21:09,649 --> 01:21:08,100
goal for that set of cells but one of

2248
01:21:11,930 --> 01:21:09,659
the things that they do without us

2249
01:21:13,910 --> 01:21:11,940
telling them how to do it or having to

2250
01:21:16,070 --> 01:21:13,920
micromanage it they already have the

2251
01:21:18,950 --> 01:21:16,080
competency to recruit as many cells as

2252
01:21:20,630 --> 01:21:18,960
they need to get the job done so that's

2253
01:21:22,490 --> 01:21:20,640
a very nice for an engineer that's a

2254
01:21:24,830 --> 01:21:22,500
very nice competency because it means

2255
01:21:26,750 --> 01:21:24,840
that I don't need to worry about taking

2256
01:21:29,270 --> 01:21:26,760
care of getting exactly the right number

2257
01:21:31,250 --> 01:21:29,280
of cells if I'm if I'm a little bit over

2258
01:21:33,110 --> 01:21:31,260
that's fine if I'm way under also fine

2259
01:21:35,510 --> 01:21:33,120
the system has that competency of

2260
01:21:37,189 --> 01:21:35,520
recruiting other cells to get the job

2261
01:21:39,410 --> 01:21:37,199
done so so that's what I so that's what

2262
01:21:40,970 --> 01:21:39,420
I meant so to me to make an um any kind

2263
01:21:43,610 --> 01:21:40,980
of a cognitive claim you have to specify

2264
01:21:45,350 --> 01:21:43,620
the problem space you have to specify

2265
01:21:47,330 --> 01:21:45,360
the goal goal towards which it's

2266
01:21:49,310 --> 01:21:47,340
expressing competencies and that may

2267
01:21:51,229 --> 01:21:49,320
right and and then and then you can make

2268
01:21:53,390 --> 01:21:51,239
a claim about well how competent is it

2269
01:21:54,709 --> 01:21:53,400
to get to that goal and somebody I wish

2270
01:21:56,510 --> 01:21:54,719
I could remember who it was but somebody

2271
01:21:58,250 --> 01:21:56,520
made this really nice analogy about kind

2272
01:22:00,950 --> 01:21:58,260
of the ends of that spectrum they said

2273
01:22:03,229 --> 01:22:00,960
two magnets try to get together and

2274
01:22:05,030 --> 01:22:03,239
Romeo and Juliet try to get together but

2275
01:22:06,410 --> 01:22:05,040
the degree of flexible problem solving

2276
01:22:09,229 --> 01:22:06,420
that you can expect out of those two

2277
01:22:11,090 --> 01:22:09,239
systems is incredibly different and

2278
01:22:12,950 --> 01:22:11,100
within that range there are all kinds of

2279
01:22:14,570 --> 01:22:12,960
in-between systems that may be better or

2280
01:22:16,430 --> 01:22:14,580
worse and may deploy different kinds of

2281
01:22:18,470 --> 01:22:16,440
strategies you know can they avoid local

2282
01:22:20,390 --> 01:22:18,480
Optima can they have a memory of the

2283
01:22:22,010 --> 01:22:20,400
where they've been can they look further

2284
01:22:23,570 --> 01:22:22,020
than their local environment a million

2285
01:22:25,130 --> 01:22:23,580
different things so that's that's what I

2286
01:22:28,189 --> 01:22:25,140
meant by competency it's it's it's it's

2287
01:22:30,950 --> 01:22:28,199
a claim about what and what an engineer

2288
01:22:32,689 --> 01:22:30,960
can expect the system to do given a

2289
01:22:34,850 --> 01:22:32,699
particular problem space in a particular

2290
01:22:35,930 --> 01:22:34,860
goal that you think it's uh trying to

2291
01:22:37,669 --> 01:22:35,940
reach

2292
01:22:42,050 --> 01:22:37,679
so the way in which you use the word

2293
01:22:43,790 --> 01:22:42,060
competency could be treated as the

2294
01:22:46,130 --> 01:22:43,800
capacity of a system for adaptive

2295
01:22:51,110 --> 01:22:48,070
yeah yeah yeah

2296
01:22:53,209 --> 01:22:51,120
one uh issue that I have with the notion

2297
01:22:55,610 --> 01:22:53,219
of goals and goal directedness is that

2298
01:22:58,250 --> 01:22:55,620
sometimes you only have a tendency in a

2299
01:23:00,830 --> 01:22:58,260
system to go in a certain direction and

2300
01:23:03,350 --> 01:23:00,840
so it's it's directed but the goal is

2301
01:23:05,030 --> 01:23:03,360
something that can be emergent sometimes

2302
01:23:06,470 --> 01:23:05,040
it's not sometimes there is an explicit

2303
01:23:08,270 --> 01:23:06,480
representation in the system of a

2304
01:23:10,310 --> 01:23:08,280
discrete event that is associated or a

2305
01:23:11,990 --> 01:23:10,320
class of events this fulfilling a

2306
01:23:13,430 --> 01:23:12,000
certain condition that the system has

2307
01:23:14,990 --> 01:23:13,440
committed itself to and if you don't

2308
01:23:17,390 --> 01:23:15,000
have that you don't have a proper goal

2309
01:23:20,630 --> 01:23:17,400
but in real systems it's difficult to

2310
01:23:22,910 --> 01:23:20,640
say I mean when do we pursue goals right

2311
01:23:24,830 --> 01:23:22,920
sometimes we just are vaguely hungry are

2312
01:23:26,270 --> 01:23:24,840
moving towards the kitchen because we

2313
01:23:28,550 --> 01:23:26,280
hope that something will

2314
01:23:31,370 --> 01:23:28,560
opportunistically emerge that will deal

2315
01:23:33,229 --> 01:23:31,380
with this vague tendency in our Behavior

2316
01:23:35,390 --> 01:23:33,239
we could also say we have the goal of

2317
01:23:37,490 --> 01:23:35,400
finding food but uh that is a

2318
01:23:40,729 --> 01:23:37,500
rationalization that is maybe stretching

2319
01:23:42,530 --> 01:23:40,739
sick sometimes yeah so uh sometimes a

2320
01:23:44,870 --> 01:23:42,540
better distinction for me is

2321
01:23:47,750 --> 01:23:44,880
going from a simple controller to an

2322
01:23:49,610 --> 01:23:47,760
agent and I try to because we are very

2323
01:23:51,050 --> 01:23:49,620
good at discovering agency in the world

2324
01:23:53,149 --> 01:23:51,060
what does it actually mean when we

2325
01:23:56,330 --> 01:23:53,159
discover agency and when we discover our

2326
01:23:56,930 --> 01:23:56,340
own agency and start to amplify it

2327
01:23:59,209 --> 01:23:56,940
um

2328
01:24:00,950 --> 01:23:59,219
by making models of who we are and how

2329
01:24:02,990 --> 01:24:00,960
we deal with the world and with others

2330
01:24:04,729 --> 01:24:03,000
and so on the minimal definition of

2331
01:24:06,290 --> 01:24:04,739
agent that I found it's a controller for

2332
01:24:08,630 --> 01:24:06,300
future States

2333
01:24:11,390 --> 01:24:08,640
the thermostat doesn't have a goal by

2334
01:24:14,149 --> 01:24:11,400
itself right it just has a Target value

2335
01:24:15,890 --> 01:24:14,159
and a sensor that tells the deviation

2336
01:24:18,169 --> 01:24:15,900
from the target value and when that

2337
01:24:21,530 --> 01:24:18,179
exceeds a certain threshold the heating

2338
01:24:22,850 --> 01:24:21,540
is turned on and if it goes below a

2339
01:24:24,950 --> 01:24:22,860
certain threshold the heating is turned

2340
01:24:26,510 --> 01:24:24,960
off again and this is it so the

2341
01:24:29,570 --> 01:24:26,520
thermostat is not an age and it only

2342
01:24:30,830 --> 01:24:29,580
reacts to the present frame it's only a

2343
01:24:33,350 --> 01:24:30,840
reactive system

2344
01:24:36,169 --> 01:24:33,360
whereas an agent is proactive which

2345
01:24:38,930 --> 01:24:36,179
means that it's trying to um not just

2346
01:24:41,810 --> 01:24:38,940
minimize the current deviation from the

2347
01:24:43,370 --> 01:24:41,820
target value but the integral over at

2348
01:24:45,770 --> 01:24:43,380
the time span previously the future

2349
01:24:49,189 --> 01:24:45,780
deviation so it builds an expectation

2350
01:24:51,530 --> 01:24:49,199
about how uh an action is going to

2351
01:24:54,110 --> 01:24:51,540
change this trajectory of the universe

2352
01:24:57,110 --> 01:24:54,120
and over that trajectory it tries to

2353
01:24:59,750 --> 01:24:57,120
figure out some measure of how big the

2354
01:25:02,030 --> 01:24:59,760
compound Target deviation is going to be

2355
01:25:03,890 --> 01:25:02,040
and so as a result you get a branching

2356
01:25:05,390 --> 01:25:03,900
universe and the branches in this

2357
01:25:08,090 --> 01:25:05,400
universe some of these branches depend

2358
01:25:09,950 --> 01:25:08,100
on actions that are available to you and

2359
01:25:13,669 --> 01:25:09,960
that translate into decisions that you

2360
01:25:15,050 --> 01:25:13,679
can make that uh move you into more or

2361
01:25:16,729 --> 01:25:15,060
less preferable wealth States and

2362
01:25:19,610 --> 01:25:16,739
suddenly you have a system with emergent

2363
01:25:22,070 --> 01:25:19,620
beliefs desires and intentions but to

2364
01:25:25,010 --> 01:25:22,080
make that happen to move from a

2365
01:25:27,290 --> 01:25:25,020
controller to agency and agent has been

2366
01:25:30,470 --> 01:25:27,300
a controller with an integrated setpoint

2367
01:25:33,110 --> 01:25:30,480
generator and the ability to control

2368
01:25:36,590 --> 01:25:33,120
future states that requires that you can

2369
01:25:38,570 --> 01:25:36,600
make models that are counterfactual so

2370
01:25:41,030 --> 01:25:38,580
because the future universe doesn't

2371
01:25:43,130 --> 01:25:41,040
exist right now you need to create a

2372
01:25:45,350 --> 01:25:43,140
counterfactual universe The Future model

2373
01:25:46,910 --> 01:25:45,360
of the future Universe maybe even a

2374
01:25:48,350 --> 01:25:46,920
model of the past universe that allows

2375
01:25:50,750 --> 01:25:48,360
you to reason about possible for future

2376
01:25:53,030 --> 01:25:50,760
universes and so on and to make these

2377
01:25:54,890 --> 01:25:53,040
counter factual causal models of the

2378
01:25:55,970 --> 01:25:54,900
universe you need to have a turing

2379
01:25:58,129 --> 01:25:55,980
machine

2380
01:26:00,890 --> 01:25:58,139
so without a computer without something

2381
01:26:02,810 --> 01:26:00,900
that is too incomplete that insulates

2382
01:26:05,270 --> 01:26:02,820
you from the causal structure of your

2383
01:26:07,250 --> 01:26:05,280
substrate that allows you to build

2384
01:26:09,590 --> 01:26:07,260
representations regardless of what the

2385
01:26:11,510 --> 01:26:09,600
universe says right now around you right

2386
01:26:14,510 --> 01:26:11,520
you need to have that machine and the

2387
01:26:17,149 --> 01:26:14,520
simplest uh system in nature that has

2388
01:26:19,669 --> 01:26:17,159
turing machine integrated is the cell

2389
01:26:22,490 --> 01:26:19,679
so uh it's very difficult to find a

2390
01:26:25,250 --> 01:26:22,500
system in nature that is an agent that

2391
01:26:27,229 --> 01:26:25,260
is not made from cells as a result

2392
01:26:29,510 --> 01:26:27,239
maybe there are systems in nature that

2393
01:26:32,990 --> 01:26:29,520
are able to uh compute things and make

2394
01:26:35,810 --> 01:26:33,000
models but uh I I'm not aware of any so

2395
01:26:37,810 --> 01:26:35,820
the simplest one that I know that can do

2396
01:26:41,750 --> 01:26:37,820
this reliably is the cells or

2397
01:26:44,390 --> 01:26:41,760
arrangement of sets and that can possess

2398
01:26:46,970 --> 01:26:44,400
agency which is interesting thing that

2399
01:26:49,669 --> 01:26:46,980
explains this coincidence that living

2400
01:26:51,470 --> 01:26:49,679
things are agents and vice versa that

2401
01:26:53,750 --> 01:26:51,480
the agent that we discover are mostly

2402
01:26:56,570 --> 01:26:53,760
living things or there are robots that

2403
01:26:59,689 --> 01:26:56,580
have computers built into them or uh

2404
01:27:01,910 --> 01:26:59,699
virtual robots that have that rely on

2405
01:27:04,070 --> 01:27:01,920
computation so the ability to make

2406
01:27:06,770 --> 01:27:04,080
models of the future is the prerequisite

2407
01:27:09,770 --> 01:27:06,780
for agency and to make arbitrary models

2408
01:27:13,970 --> 01:27:09,780
which means structures that embody uh

2409
01:27:16,550 --> 01:27:13,980
causal uh simulations of some sort that

2410
01:27:19,610 --> 01:27:16,560
requires computation

2411
01:27:21,470 --> 01:27:19,620
yeah yeah um you know I'm on board with

2412
01:27:24,530 --> 01:27:21,480
that with that ladder of that that

2413
01:27:26,629 --> 01:27:24,540
taxonomy uh of of uh

2414
01:27:28,669 --> 01:27:26,639
of of goals and so on one one

2415
01:27:30,530 --> 01:27:28,679
interesting thing about uh about goals

2416
01:27:32,209 --> 01:27:30,540
and as you say some are emergent and and

2417
01:27:34,010 --> 01:27:32,219
some are not there's a there's an

2418
01:27:36,709 --> 01:27:34,020
interesting uh planarian

2419
01:27:37,850 --> 01:27:36,719
um uh version of this which is which is

2420
01:27:41,149 --> 01:27:37,860
this we

2421
01:27:42,770 --> 01:27:41,159
we made this hypothesis about so so with

2422
01:27:44,810 --> 01:27:42,780
implant area you chop it up into pieces

2423
01:27:46,610 --> 01:27:44,820
and every piece regenerates exactly the

2424
01:27:48,950 --> 01:27:46,620
right rest of the work right so if it's

2425
01:27:51,950 --> 01:27:48,960
if if you chop it into pieces each piece

2426
01:27:53,570 --> 01:27:51,960
will have one head one tail so and then

2427
01:27:55,310 --> 01:27:53,580
and then of course what happens is it

2428
01:27:57,649 --> 01:27:55,320
stops when it when it may when it

2429
01:28:00,830 --> 01:27:57,659
reaches a correct planarian then then it

2430
01:28:01,970 --> 01:28:00,840
stops and so so we started to think that

2431
01:28:03,709 --> 01:28:01,980
there are two there are a couple of

2432
01:28:06,110 --> 01:28:03,719
possibilities one possibility is that

2433
01:28:08,030 --> 01:28:06,120
this is a purely emergent process and

2434
01:28:10,189 --> 01:28:08,040
that the goal of rebuilding ahead is is

2435
01:28:12,410 --> 01:28:10,199
an emergent thing that comes about as a

2436
01:28:14,870 --> 01:28:12,420
consequence of other things or could

2437
01:28:16,550 --> 01:28:14,880
there be a an actual explicit

2438
01:28:18,709 --> 01:28:16,560
representation of what a correct

2439
01:28:20,810 --> 01:28:18,719
planarian is that serves as a set point

2440
01:28:22,850 --> 01:28:20,820
as an encoded as an explicitly encoded

2441
01:28:24,970 --> 01:28:22,860
set point for these cells to follow and

2442
01:28:27,169 --> 01:28:24,980
and because it's a cellular a collective

2443
01:28:29,330 --> 01:28:27,179
we're communicating electrically we

2444
01:28:32,149 --> 01:28:29,340
thought well maybe maybe what it's doing

2445
01:28:34,790 --> 01:28:32,159
is basically storing a memory of what

2446
01:28:36,709 --> 01:28:34,800
did you like you would in in a neural

2447
01:28:38,390 --> 01:28:36,719
circuit you're storing a memory of what

2448
01:28:39,770 --> 01:28:38,400
it should be so we started looking for

2449
01:28:42,470 --> 01:28:39,780
this and this is what we found and this

2450
01:28:45,290 --> 01:28:42,480
is this is kind of I think one type of

2451
01:28:48,229 --> 01:28:45,300
one one important type of goal in a in a

2452
01:28:50,990 --> 01:28:48,239
goal seeking system is is a is a goal

2453
01:28:52,490 --> 01:28:51,000
that you can rewrite without changing

2454
01:28:54,110 --> 01:28:52,500
the hardware and the system will now

2455
01:28:56,330 --> 01:28:54,120
pursue that goal instead of something

2456
01:28:57,649 --> 01:28:56,340
else in a purely emergent system that

2457
01:28:59,750 --> 01:28:57,659
doesn't work right if you have a

2458
01:29:01,010 --> 01:28:59,760
cellular automaton or a fractal or

2459
01:29:02,810 --> 01:29:01,020
something that that does some kind of

2460
01:29:04,129 --> 01:29:02,820
complex thing if you want to change what

2461
01:29:05,689 --> 01:29:04,139
that complex thing is you have to figure

2462
01:29:07,550 --> 01:29:05,699
out what the local rule how to change

2463
01:29:09,770 --> 01:29:07,560
the local rules that's very hard in most

2464
01:29:12,350 --> 01:29:09,780
cases but what we found in planaria is

2465
01:29:15,290 --> 01:29:12,360
that we can literally using a voltage

2466
01:29:18,530 --> 01:29:15,300
reporter die we can look at the worm and

2467
01:29:19,970 --> 01:29:18,540
we can see now the pattern the the and

2468
01:29:21,590 --> 01:29:19,980
it's a distributed pattern but we can

2469
01:29:23,689 --> 01:29:21,600
see the pattern that tells this animal

2470
01:29:25,850 --> 01:29:23,699
how many heads it's supposed to have and

2471
01:29:29,810 --> 01:29:25,860
what you can do is you can go in and

2472
01:29:31,550 --> 01:29:29,820
using a a brief transient uh

2473
01:29:33,950 --> 01:29:31,560
manipulation of the ion Channels with

2474
01:29:35,450 --> 01:29:33,960
drugs with ION channel drugs that and we

2475
01:29:37,189 --> 01:29:35,460
have a computational model that tells

2476
01:29:39,770 --> 01:29:37,199
you what those drugs should be that

2477
01:29:41,629 --> 01:29:39,780
bring briefly changes the electrical

2478
01:29:43,550 --> 01:29:41,639
state of the circuit but the circuit is

2479
01:29:46,430 --> 01:29:43,560
amazing it's it it once you've changed

2480
01:29:48,169 --> 01:29:46,440
that state it holds so so by default it

2481
01:29:50,090 --> 01:29:48,179
in a standard planarian it always says

2482
01:29:51,410 --> 01:29:50,100
one head but but it's kind of like a

2483
01:29:53,629 --> 01:29:51,420
flip-flop in that when you you

2484
01:29:55,430 --> 01:29:53,639
temporarily shift it it holds and you

2485
01:29:57,050 --> 01:29:55,440
can push it to a to a state that says

2486
01:29:58,729 --> 01:29:57,060
two heads so now something very

2487
01:30:00,590 --> 01:29:58,739
interesting happens

2488
01:30:02,270 --> 01:30:00,600
um two interesting things one is that if

2489
01:30:04,669 --> 01:30:02,280
you if you take those worms and you cut

2490
01:30:06,290 --> 01:30:04,679
those into pieces you get two-headed

2491
01:30:07,850 --> 01:30:06,300
worms even though the genetics are the

2492
01:30:09,229 --> 01:30:07,860
hardware is all wild type there's

2493
01:30:10,669 --> 01:30:09,239
nothing wrong with Hardware all the

2494
01:30:12,649 --> 01:30:10,679
proteins are the same all the genetics

2495
01:30:14,750 --> 01:30:12,659
is the same but the electric circuit now

2496
01:30:17,570 --> 01:30:14,760
says make two heads instead of one and

2497
01:30:19,669 --> 01:30:17,580
so this is in in an interesting way it

2498
01:30:21,169 --> 01:30:19,679
is an explicit goal because you can

2499
01:30:22,610 --> 01:30:21,179
rewrite it because much like with your

2500
01:30:24,350 --> 01:30:22,620
thermostat there's an interface for

2501
01:30:25,550 --> 01:30:24,360
changing what the goal state is and then

2502
01:30:26,870 --> 01:30:25,560
you don't even need to know how the rest

2503
01:30:28,729 --> 01:30:26,880
of the thermostat works as long as you

2504
01:30:30,530 --> 01:30:28,739
know how to use your how to modify that

2505
01:30:32,750 --> 01:30:30,540
interface the system takes care of the

2506
01:30:34,070 --> 01:30:32,760
rest the other interesting thing is and

2507
01:30:37,250 --> 01:30:34,080
and I love what you said about the

2508
01:30:39,770 --> 01:30:37,260
counter factuals what you can do is you

2509
01:30:41,330 --> 01:30:39,780
can change that electrical pattern in an

2510
01:30:44,209 --> 01:30:41,340
intact worm

2511
01:30:45,410 --> 01:30:44,219
and not cut it for a long time and if

2512
01:30:47,030 --> 01:30:45,420
you do that when you look at that

2513
01:30:49,669 --> 01:30:47,040
pattern that is a counter factual

2514
01:30:52,610 --> 01:30:49,679
pattern because that two-headed pattern

2515
01:30:54,050 --> 01:30:52,620
is not a readout of the current state it

2516
01:30:56,510 --> 01:30:54,060
says two heads but the animal only has

2517
01:30:59,030 --> 01:30:56,520
one head it's a normal planarian so the

2518
01:31:01,430 --> 01:30:59,040
that pattern memory is not a readout of

2519
01:31:03,590 --> 01:31:01,440
what the animal is doing right now it is

2520
01:31:05,750 --> 01:31:03,600
a representation of what the animal will

2521
01:31:07,669 --> 01:31:05,760
do in the future if it happens to get

2522
01:31:09,530 --> 01:31:07,679
injured and you may never cut it or you

2523
01:31:11,570 --> 01:31:09,540
may cut it but if you do then the

2524
01:31:13,550 --> 01:31:11,580
pattern becomes uh then the cells

2525
01:31:15,229 --> 01:31:13,560
consult a pattern and build a two-headed

2526
01:31:17,390 --> 01:31:15,239
worm and then it becomes a you know the

2527
01:31:19,250 --> 01:31:17,400
current state but until then it's this

2528
01:31:21,229 --> 01:31:19,260
weird like primitive

2529
01:31:23,629 --> 01:31:21,239
um uh it's a primitive counterfactual

2530
01:31:26,930 --> 01:31:23,639
system because it's able to a body of a

2531
01:31:27,830 --> 01:31:26,940
planarian is able to store at least two

2532
01:31:30,169 --> 01:31:27,840
different

2533
01:31:32,330 --> 01:31:30,179
representations of what probably many

2534
01:31:33,770 --> 01:31:32,340
more but we found two so far what a

2535
01:31:35,330 --> 01:31:33,780
correct planarian should look like it

2536
01:31:36,890 --> 01:31:35,340
can have a memory of a one-headed

2537
01:31:39,770 --> 01:31:36,900
planarian or a memory of a two-headed

2538
01:31:41,750 --> 01:31:39,780
planarian and and and both of those can

2539
01:31:43,970 --> 01:31:41,760
live in exactly the same Hardware in

2540
01:31:46,250 --> 01:31:43,980
exactly the same the same body the other

2541
01:31:48,050 --> 01:31:46,260
kind of cool thing about this and um

2542
01:31:50,450 --> 01:31:48,060
I'll just mention this even though this

2543
01:31:52,669 --> 01:31:50,460
is this is you know disclaimer this is

2544
01:31:54,169 --> 01:31:52,679
not published yet so this is uh you know

2545
01:31:56,629 --> 01:31:54,179
take all this with a grain of salt the

2546
01:31:58,910 --> 01:31:56,639
the but but the the latest thing you can

2547
01:32:00,770 --> 01:31:58,920
do is you can actually treat it with

2548
01:32:03,350 --> 01:32:00,780
some of the same compounds that are used

2549
01:32:06,169 --> 01:32:03,360
in uh Neuroscience in in humans and in

2550
01:32:08,149 --> 01:32:06,179
rats as memory blockers so so things

2551
01:32:10,430 --> 01:32:08,159
that block recall or or memory

2552
01:32:11,990 --> 01:32:10,440
consolidation and when you do that you

2553
01:32:13,430 --> 01:32:12,000
can make the animal forget how many

2554
01:32:14,930 --> 01:32:13,440
heads it's supposed to have and then

2555
01:32:17,090 --> 01:32:14,940
they basically turn into a featureless

2556
01:32:18,709 --> 01:32:17,100
circle when when you can just wipe you

2557
01:32:20,270 --> 01:32:18,719
can just wipe the the pattern memory

2558
01:32:22,370 --> 01:32:20,280
completely with it using exactly the

2559
01:32:24,709 --> 01:32:22,380
same techniques you would use in a in a

2560
01:32:26,209 --> 01:32:24,719
rat or a human they just forget what to

2561
01:32:27,770 --> 01:32:26,219
do when they turn into they fail to

2562
01:32:31,550 --> 01:32:27,780
break Symmetry and they just become a

2563
01:32:33,470 --> 01:32:31,560
Circle so yeah I I think I think what

2564
01:32:35,870 --> 01:32:33,480
you were saying is is right on with with

2565
01:32:37,729 --> 01:32:35,880
this this ability to uh store

2566
01:32:39,890 --> 01:32:37,739
counterfactual states that are not true

2567
01:32:41,750 --> 01:32:39,900
now but may represent

2568
01:32:44,090 --> 01:32:41,760
aspects of the of the future I think

2569
01:32:46,510 --> 01:32:44,100
that's that's a very important capacity

2570
01:32:49,430 --> 01:32:46,520
another important notion is a constraint

2571
01:32:52,070 --> 01:32:49,440
uh constraint satisfaction a constraint

2572
01:32:54,110 --> 01:32:52,080
is a rule that tells you uh whether two

2573
01:32:55,490 --> 01:32:54,120
things are compatible or not and the

2574
01:32:57,410 --> 01:32:55,500
constraint is satisfied if they're

2575
01:32:59,030 --> 01:32:57,420
compatible so you basically have a

2576
01:33:01,330 --> 01:32:59,040
number of conditions that you establish

2577
01:33:04,310 --> 01:33:01,340
by measuring that somehow for instance

2578
01:33:07,250 --> 01:33:04,320
uh whether you have a head or multiple

2579
01:33:09,169 --> 01:33:07,260
heads and you try to find a solution

2580
01:33:11,390 --> 01:33:09,179
where you can end up with exactly one

2581
01:33:13,970 --> 01:33:11,400
head and if you end up with exactly one

2582
01:33:15,530 --> 01:33:13,980
head based on the starting State then

2583
01:33:18,770 --> 01:33:15,540
you have managed to find a way to

2584
01:33:21,530 --> 01:33:18,780
satisfy your constraints and so in a

2585
01:33:24,890 --> 01:33:21,540
sense uh what you call a competency is

2586
01:33:27,709 --> 01:33:24,900
the ability of a system to take a region

2587
01:33:30,290 --> 01:33:27,719
of this of states of the space of the

2588
01:33:32,090 --> 01:33:30,300
universe some basically some local

2589
01:33:36,470 --> 01:33:32,100
region of possible state that the Euros

2590
01:33:39,410 --> 01:33:36,480
can be in and move that region to a

2591
01:33:40,669 --> 01:33:39,420
smaller region that is acceptable so the

2592
01:33:42,110 --> 01:33:40,679
basically there is a region on the

2593
01:33:43,070 --> 01:33:42,120
universe State space where you have only

2594
01:33:44,689 --> 01:33:43,080
one head

2595
01:33:46,189 --> 01:33:44,699
and there's a larger region where you

2596
01:33:48,530 --> 01:33:46,199
don't have any head at all but the

2597
01:33:50,930 --> 01:33:48,540
starting state of your organism and then

2598
01:33:52,910 --> 01:33:50,940
you try to get from A to B so you get

2599
01:33:55,010 --> 01:33:52,920
from this larger region to the one in

2600
01:33:56,390 --> 01:33:55,020
which you want to be of course if you

2601
01:33:58,610 --> 01:33:56,400
have one head you want to stay in the

2602
01:34:00,709 --> 01:33:58,620
region in which you have one head which

2603
01:34:02,750 --> 01:34:00,719
of course is usually much easier but the

2604
01:34:06,290 --> 01:34:02,760
ability basically to condense the space

2605
01:34:09,290 --> 01:34:06,300
to get to Bridge Over many regions into

2606
01:34:11,090 --> 01:34:09,300
the target region is what comes down to

2607
01:34:13,490 --> 01:34:11,100
this this is what this competency is the

2608
01:34:15,470 --> 01:34:13,500
system basically has an emergent wanting

2609
01:34:17,330 --> 01:34:15,480
to go in this region and it's trying to

2610
01:34:19,310 --> 01:34:17,340
move there and so there are constraints

2611
01:34:21,530 --> 01:34:19,320
on the level of the substrate

2612
01:34:23,330 --> 01:34:21,540
uh that are battling with the functional

2613
01:34:25,430 --> 01:34:23,340
constraints that the organism wants to

2614
01:34:28,189 --> 01:34:25,440
realize to fulfill its function and

2615
01:34:29,510 --> 01:34:28,199
sometimes uh you cannot satisfy this and

2616
01:34:30,830 --> 01:34:29,520
you end up with two heads because you

2617
01:34:33,890 --> 01:34:30,840
don't know which one you'll get rid of

2618
01:34:37,010 --> 01:34:33,900
or how to digest one of the heads and so

2619
01:34:40,310 --> 01:34:37,020
on and you end up with some Siamese twin

2620
01:34:42,169 --> 01:34:40,320
and uh so this this is an interesting

2621
01:34:44,629 --> 01:34:42,179
constraint that you have to solve for

2622
01:34:46,430 --> 01:34:44,639
when you are dealing with reality and

2623
01:34:49,550 --> 01:34:46,440
how you battle with the substrate until

2624
01:34:50,950 --> 01:34:49,560
you get to the functional solution that

2625
01:34:53,510 --> 01:34:50,960
you evolve for

2626
01:34:56,570 --> 01:34:53,520
yeah it's that's that's interesting I

2627
01:35:00,229 --> 01:34:56,580
mean we we've also found that there are

2628
01:35:02,390 --> 01:35:00,239
so so we look at exactly this uh the the

2629
01:35:03,830 --> 01:35:02,400
navigation this kind of navigation and

2630
01:35:05,629 --> 01:35:03,840
Morpher space how you get from here to

2631
01:35:07,790 --> 01:35:05,639
there and what paths are possible to get

2632
01:35:10,129 --> 01:35:07,800
from here to there and so on one of the

2633
01:35:12,350 --> 01:35:10,139
things that we found is that there are

2634
01:35:15,709 --> 01:35:12,360
regions of that space that belong to

2635
01:35:17,990 --> 01:35:15,719
other species and you can you can push a

2636
01:35:20,750 --> 01:35:18,000
a planarian with a standard wild type

2637
01:35:22,070 --> 01:35:20,760
genome into the goal state of a

2638
01:35:23,810 --> 01:35:22,080
completely different species so we can

2639
01:35:25,310 --> 01:35:23,820
get them to grow ahead so there's a

2640
01:35:27,050 --> 01:35:25,320
species that normally has a triangular

2641
01:35:28,970 --> 01:35:27,060
head you can make it grow a round head

2642
01:35:32,030 --> 01:35:28,980
like a like an like a different species

2643
01:35:35,450 --> 01:35:32,040
or or a flathead or whatever uh you can

2644
01:35:37,189 --> 01:35:35,460
so those are about 100 100 to 150

2645
01:35:40,250 --> 01:35:37,199
million years of evolutionary distance

2646
01:35:43,550 --> 01:35:40,260
and you can do it you know within within

2647
01:35:45,110 --> 01:35:43,560
a few days just by um by by perturbing

2648
01:35:47,689 --> 01:35:45,120
that electrical circuit so that it lands

2649
01:35:50,330 --> 01:35:47,699
in the wrong space and then outside of

2650
01:35:52,250 --> 01:35:50,340
that there are regions that don't belong

2651
01:35:54,169 --> 01:35:52,260
to planaria at all so planaria are

2652
01:35:56,270 --> 01:35:54,179
normally nice and flat we can make we've

2653
01:35:58,010 --> 01:35:56,280
made planarians that are they look like

2654
01:35:59,450 --> 01:35:58,020
a they are cylinder like like a ski cap

2655
01:36:02,090 --> 01:35:59,460
you know they become like a like a like

2656
01:36:04,070 --> 01:36:02,100
a hemisphere or really weird ones that

2657
01:36:05,510 --> 01:36:04,080
are spiky they're they're like a ball

2658
01:36:07,970 --> 01:36:05,520
with spikes on it they're all kinds of

2659
01:36:11,090 --> 01:36:07,980
other uh regions in that space that you

2660
01:36:13,189 --> 01:36:11,100
can that you can push them to uh and so

2661
01:36:15,229 --> 01:36:13,199
those are new those are not species that

2662
01:36:17,330 --> 01:36:15,239
they diverge from those are no one's

2663
01:36:18,649 --> 01:36:17,340
ever said to my knowledge that's yes

2664
01:36:19,910 --> 01:36:18,659
there's no such there are no such

2665
01:36:21,350 --> 01:36:19,920
species

2666
01:36:22,729 --> 01:36:21,360
um it's easier to you know and we've

2667
01:36:25,550 --> 01:36:22,739
done this in frog too you can you can

2668
01:36:27,410 --> 01:36:25,560
push tadpoles to make it to look like

2669
01:36:29,270 --> 01:36:27,420
those of other species

2670
01:36:30,709 --> 01:36:29,280
um or you can make you know I mean

2671
01:36:32,930 --> 01:36:30,719
that's a whole interesting thing for

2672
01:36:34,669 --> 01:36:32,940
evolution anyway right what one species

2673
01:36:36,770 --> 01:36:34,679
birth defect is a pretty is a perfectly

2674
01:36:39,169 --> 01:36:36,780
reasonable different species so we can

2675
01:36:42,590 --> 01:36:39,179
make we can make tadpoles with a rounded

2676
01:36:44,030 --> 01:36:42,600
tail which for a xenopus tadpole is a

2677
01:36:46,729 --> 01:36:44,040
terrible Tale But for zebrafish that's

2678
01:36:49,310 --> 01:36:46,739
exactly the right tale so you can sort

2679
01:36:51,610 --> 01:36:49,320
of Imagine imagine right Evolution uh

2680
01:36:53,629 --> 01:36:51,620
manipulating the different information

2681
01:36:55,010 --> 01:36:53,639
processing whether whether by the

2682
01:36:57,410 --> 01:36:55,020
bioelectrical circuits or other other

2683
01:37:00,229 --> 01:36:57,420
Machinery uh that help the system

2684
01:37:01,810 --> 01:37:00,239
explore that morphe space and you know

2685
01:37:05,870 --> 01:37:01,820
start to

2686
01:37:07,550 --> 01:37:05,880
from from whatever the you know that

2687
01:37:09,709 --> 01:37:07,560
speciation is moving away from your

2688
01:37:12,350 --> 01:37:09,719
standard attractor that you usually land

2689
01:37:15,410 --> 01:37:12,360
on how does this relate to intelligence

2690
01:37:17,390 --> 01:37:15,420
well intelligence is the ability to make

2691
01:37:19,129 --> 01:37:17,400
models and and usually in the service of

2692
01:37:22,310 --> 01:37:19,139
control at least that's the way I would

2693
01:37:24,350 --> 01:37:22,320
uh explain intelligence

2694
01:37:26,270 --> 01:37:24,360
there are other definitions but it's the

2695
01:37:27,350 --> 01:37:26,280
simplest one that I found it also

2696
01:37:29,209 --> 01:37:27,360
accounts for the fact that many

2697
01:37:31,550 --> 01:37:29,219
intelligent people are not very good at

2698
01:37:33,290 --> 01:37:31,560
getting things done uh that's it's

2699
01:37:36,110 --> 01:37:33,300
basically an intelligence and goal

2700
01:37:38,510 --> 01:37:36,120
rationality are support orthogonal

2701
01:37:41,750 --> 01:37:38,520
and excessive intelligence is often a

2702
01:37:44,090 --> 01:37:41,760
prosthesis for bad regulation have you

2703
01:37:47,270 --> 01:37:44,100
read the intelligence trap

2704
01:37:49,550 --> 01:37:47,280
no okay the author makes a similar case

2705
01:37:51,290 --> 01:37:49,560
and he's coming on shortly essentially

2706
01:37:53,209 --> 01:37:51,300
saying that there are certain traps that

2707
01:37:55,610 --> 01:37:53,219
people with high IQs have that are not

2708
01:37:58,070 --> 01:37:55,620
beneficial for them as biological beings

2709
01:37:59,810 --> 01:37:58,080
they're mainly cognitive biases so for

2710
01:38:01,729 --> 01:37:59,820
instance it's extremely interesting so

2711
01:38:03,530 --> 01:38:01,739
let's just give one of the biases to say

2712
01:38:05,450 --> 01:38:03,540
you're either liberal biased or you're

2713
01:38:07,250 --> 01:38:05,460
conservative biased and then you were to

2714
01:38:08,990 --> 01:38:07,260
give a test where there's some data that

2715
01:38:10,970 --> 01:38:09,000
says that on the surface it shows that

2716
01:38:13,790 --> 01:38:10,980
the data shows that gun control prevents

2717
01:38:15,649 --> 01:38:13,800
gun violence well the Liberals are more

2718
01:38:17,629 --> 01:38:15,659
likely to say yes this data does show

2719
01:38:19,070 --> 01:38:17,639
that but if you're conservative you're

2720
01:38:20,870 --> 01:38:19,080
more likely to find oh actually the

2721
01:38:23,870 --> 01:38:20,880
subtleties in the data show that gun

2722
01:38:25,129 --> 01:38:23,880
control increases gun violence and then

2723
01:38:26,510 --> 01:38:25,139
they thought okay well let's just switch

2724
01:38:28,370 --> 01:38:26,520
this to make it such that The

2725
01:38:31,010 --> 01:38:28,380
Superficial data suggests that gun

2726
01:38:32,390 --> 01:38:31,020
control increases violence you need to

2727
01:38:33,890 --> 01:38:32,400
look at the data carefully to show that

2728
01:38:35,209 --> 01:38:33,900
it actually prevents violence well the

2729
01:38:37,149 --> 01:38:35,219
conservatives in that case would be more

2730
01:38:39,590 --> 01:38:37,159
quickly to say oh look the gun control

2731
01:38:41,810 --> 01:38:39,600
increases violence and the Liberals

2732
01:38:43,010 --> 01:38:41,820
would find the the loophole well that's

2733
01:38:45,229 --> 01:38:43,020
one of the reasons why I don't mind

2734
01:38:47,629 --> 01:38:45,239
interviewing people who are biased

2735
01:38:49,970 --> 01:38:47,639
because to me they're more able to find

2736
01:38:52,970 --> 01:38:49,980
a justification for something that that

2737
01:38:55,010 --> 01:38:52,980
may be true but I or and others are so

2738
01:38:56,750 --> 01:38:55,020
well we all have our own biases we're so

2739
01:38:59,209 --> 01:38:56,760
inclined in some other direction that we

2740
01:39:00,350 --> 01:38:59,219
just we're blind to it but anyway the

2741
01:39:02,270 --> 01:39:00,360
point is to affirm what you're saying

2742
01:39:04,370 --> 01:39:02,280
yosha okay so I know Michael has a hard

2743
01:39:07,310 --> 01:39:04,380
cut off at 2PM so I want to ask the

2744
01:39:09,530 --> 01:39:07,320
question for AGI that is artificial

2745
01:39:11,810 --> 01:39:09,540
general intelligence it seems as though

2746
01:39:14,689 --> 01:39:11,820
we're far away or that our current

2747
01:39:16,669 --> 01:39:14,699
methods of machine learning and what we

2748
01:39:18,530 --> 01:39:16,679
learn in Neuroscience or or what we

2749
01:39:20,209 --> 01:39:18,540
learn in computer science is something

2750
01:39:22,129 --> 01:39:20,219
that we're missing some paradigm shift

2751
01:39:23,390 --> 01:39:22,139
or we're missing some new techniques is

2752
01:39:25,189 --> 01:39:23,400
there something from Michael's work

2753
01:39:26,750 --> 01:39:25,199
Yoshi I'm asking you this and then

2754
01:39:28,189 --> 01:39:26,760
Michael please respond is there

2755
01:39:30,169 --> 01:39:28,199
something from Michael's work that you

2756
01:39:33,709 --> 01:39:30,179
think can be applied to the development

2757
01:39:35,450 --> 01:39:33,719
of AGI if such a creature mind can exist

2758
01:39:37,070 --> 01:39:35,460
because there are some arguments against

2759
01:39:38,990 --> 01:39:37,080
it so first of all I don't know how far

2760
01:39:40,490 --> 01:39:39,000
we are for the HEI it could be that the

2761
01:39:41,750 --> 01:39:40,500
existing private names are sufficient to

2762
01:39:43,850 --> 01:39:41,760
brute force it

2763
01:39:45,470 --> 01:39:43,860
but we don't know that yet right so

2764
01:39:48,050 --> 01:39:45,480
we're going to find out in the next few

2765
01:39:49,550 --> 01:39:48,060
months uh but it could also be that we

2766
01:39:51,830 --> 01:39:49,560
need to rewrite the stack to build

2767
01:39:53,390 --> 01:39:51,840
systems that work in real time that are

2768
01:39:56,090 --> 01:39:53,400
entangled with the environment that can

2769
01:39:58,370 --> 01:39:56,100
build uh shared representations with the

2770
01:40:00,770 --> 01:39:58,380
environment and uh that we need to

2771
01:40:02,450 --> 01:40:00,780
rewrite the stack and there are actually

2772
01:40:04,490 --> 01:40:02,460
a number of questions that I'd like to

2773
01:40:07,610 --> 01:40:04,500
ask Michael

2774
01:40:09,709 --> 01:40:07,620
um what uh I noticed that Michael is

2775
01:40:11,750 --> 01:40:09,719
wisely reluctant to use certain words

2776
01:40:13,010 --> 01:40:11,760
like Consciousness a lot and it's

2777
01:40:14,750 --> 01:40:13,020
because a lot of people are very

2778
01:40:16,189 --> 01:40:14,760
opinionated about what these concepts

2779
01:40:18,470 --> 01:40:16,199
mean and you first have to deal with

2780
01:40:20,090 --> 01:40:18,480
these opinions before you come down to

2781
01:40:21,830 --> 01:40:20,100
saying oh here I have the following

2782
01:40:24,169 --> 01:40:21,840
proposal for implementing reflexive

2783
01:40:27,590 --> 01:40:24,179
attention as a tool to form coherence

2784
01:40:29,330 --> 01:40:27,600
and a representation and this leads to

2785
01:40:30,470 --> 01:40:29,340
the same phenomena as what you call

2786
01:40:33,290 --> 01:40:30,480
consciousness

2787
01:40:34,790 --> 01:40:33,300
right so uh that is is a detailed

2788
01:40:36,470 --> 01:40:34,800
discussion maybe you don't want to have

2789
01:40:37,970 --> 01:40:36,480
that discussion in every Foreman rather

2790
01:40:40,490 --> 01:40:37,980
than uh that and then having this

2791
01:40:42,470 --> 01:40:40,500
discussion you may be looking at uh how

2792
01:40:44,810 --> 01:40:42,480
to create coherence using a reflexive

2793
01:40:47,330 --> 01:40:44,820
attention process that makes a real-time

2794
01:40:48,890 --> 01:40:47,340
model of what it's attending to and the

2795
01:40:51,530 --> 01:40:48,900
fact that it's attending to it so it

2796
01:40:54,830 --> 01:40:51,540
remains coherent but for itself so this

2797
01:40:56,810 --> 01:40:54,840
is a concrete thing but uh I wonder how

2798
01:40:58,490 --> 01:40:56,820
to implement this in a self-organized

2799
01:41:00,950 --> 01:40:58,500
fashion if the substrate that you have

2800
01:41:04,250 --> 01:41:00,960
are individual agents and there is a

2801
01:41:07,430 --> 01:41:04,260
similarity here between societies and

2802
01:41:09,970 --> 01:41:07,440
brains and social networks

2803
01:41:12,229 --> 01:41:09,980
that is uh if you have

2804
01:41:13,370 --> 01:41:12,239
self-interested agents in a way that try

2805
01:41:15,290 --> 01:41:13,380
to survive

2806
01:41:17,330 --> 01:41:15,300
and that get their rewards from other

2807
01:41:19,189 --> 01:41:17,340
agents that are similar to them

2808
01:41:22,189 --> 01:41:19,199
structurally

2809
01:41:23,870 --> 01:41:22,199
and and they have the capacity to learn

2810
01:41:27,649 --> 01:41:23,880
to some degree

2811
01:41:30,229 --> 01:41:27,659
and that capacity is sufficient so they

2812
01:41:34,430 --> 01:41:30,239
can in the aggregate learn arbitrary

2813
01:41:37,129 --> 01:41:34,440
programs arbitrary computable functions

2814
01:41:38,930 --> 01:41:37,139
and it's efficient enough so they can

2815
01:41:42,770 --> 01:41:38,940
Converge on the functions that they need

2816
01:41:44,149 --> 01:41:42,780
to to as a group reap rewards that apply

2817
01:41:46,189 --> 01:41:44,159
to the whole group because they have a

2818
01:41:47,750 --> 01:41:46,199
shared Destiny like the poor little

2819
01:41:49,250 --> 01:41:47,760
cells that are locked in the same skull

2820
01:41:52,430 --> 01:41:49,260
and they are all going to die together

2821
01:41:54,169 --> 01:41:52,440
if they up so they have to get

2822
01:41:56,570 --> 01:41:54,179
along they have to form an organization

2823
01:41:58,790 --> 01:41:56,580
that is Distributing rewards among each

2824
01:42:01,729 --> 01:41:58,800
other and this gives us a search space

2825
01:42:04,250 --> 01:42:01,739
for possible systems that can exist and

2826
01:42:08,209 --> 01:42:04,260
the search space is mostly given I think

2827
01:42:09,830 --> 01:42:08,219
by the minimal agent that is able to

2828
01:42:11,090 --> 01:42:09,840
learn how to distribute rewards

2829
01:42:13,609 --> 01:42:11,100
efficiently

2830
01:42:15,649 --> 01:42:13,619
while doing something useful using these

2831
01:42:17,689 --> 01:42:15,659
rewards to change how you do something

2832
01:42:20,209 --> 01:42:17,699
useful so you have an emerging form of

2833
01:42:22,129 --> 01:42:20,219
governance in these systems there's not

2834
01:42:24,050 --> 01:42:22,139
some centralized control that is imposed

2835
01:42:25,550 --> 01:42:24,060
on the system from the outside as an

2836
01:42:28,310 --> 01:42:25,560
existing machine learning approaches and

2837
01:42:30,649 --> 01:42:28,320
AI approaches but this is this only is

2838
01:42:32,750 --> 01:42:30,659
an emergent pattern in the interactions

2839
01:42:35,270 --> 01:42:32,760
between the individual small units small

2840
01:42:37,070 --> 01:42:35,280
reinforcement learning agents and this

2841
01:42:39,169 --> 01:42:37,080
control architecture leads to

2842
01:42:41,030 --> 01:42:39,179
hierarchical government it's not fully

2843
01:42:42,770 --> 01:42:41,040
decentralized in any way there are

2844
01:42:44,390 --> 01:42:42,780
centralized structures that distribute

2845
01:42:46,729 --> 01:42:44,400
rewards for instance via the

2846
01:42:48,649 --> 01:42:46,739
dopaminergic system and very centralized

2847
01:42:50,629 --> 01:42:48,659
top-down Manner and that's because every

2848
01:42:52,490 --> 01:42:50,639
regulation has an optimal layer where it

2849
01:42:54,649 --> 01:42:52,500
needs to take place some stuff needs to

2850
01:42:56,450 --> 01:42:54,659
be decided very high up some stuff needs

2851
01:42:58,430 --> 01:42:56,460
to be optimally regulated very low down

2852
01:43:00,410 --> 01:42:58,440
depending on the incentives game

2853
01:43:02,750 --> 01:43:00,420
theoretically a government is an agent

2854
01:43:06,229 --> 01:43:02,760
that imposes an offset on your payoff

2855
01:43:08,270 --> 01:43:06,239
metrics to make your Nash equilibrium

2856
01:43:10,189 --> 01:43:08,280
compatible with the globally best

2857
01:43:12,530 --> 01:43:10,199
outcome

2858
01:43:14,330 --> 01:43:12,540
right so to to do this you need to have

2859
01:43:16,070 --> 01:43:14,340
agents that are sensitive to rewards

2860
01:43:18,950 --> 01:43:16,080
it's super interesting to think about

2861
01:43:20,870 --> 01:43:18,960
these reward infrastructures Elon Musk

2862
01:43:22,790 --> 01:43:20,880
has bought Twitter I think because he

2863
01:43:24,830 --> 01:43:22,800
has realized that Twitter is the network

2864
01:43:27,229 --> 01:43:24,840
of among all the social networks that is

2865
01:43:29,270 --> 01:43:27,239
closest to a global brain

2866
01:43:30,649 --> 01:43:29,280
it's totally mind-blowing to realize

2867
01:43:33,169 --> 01:43:30,659
that he basically trades a bunch of

2868
01:43:36,290 --> 01:43:33,179
frothy stock for the opportunity to

2869
01:43:38,270 --> 01:43:36,300
become Pope pope of a religion that has

2870
01:43:40,010 --> 01:43:38,280
more active participants than

2871
01:43:42,290 --> 01:43:40,020
Catholicism even right daily

2872
01:43:44,270 --> 01:43:42,300
practitioning people who enter this

2873
01:43:45,709 --> 01:43:44,280
church and think together and it's a

2874
01:43:46,970 --> 01:43:45,719
thing that is completely incoherent at

2875
01:43:49,129 --> 01:43:46,980
this point almost and completely

2876
01:43:50,450 --> 01:43:49,139
incorrect the bubbles of sentience but

2877
01:43:52,010 --> 01:43:50,460
for the most part this thing is just

2878
01:43:54,229 --> 01:43:52,020
screeching at itself

2879
01:43:56,149 --> 01:43:54,239
and now the SD question can we fix the

2880
01:43:57,410 --> 01:43:56,159
incentives of Twitter to turn it into a

2881
01:44:01,910 --> 01:43:57,420
global brain

2882
01:44:03,770 --> 01:44:01,920
believes that this is the case and

2883
01:44:05,510 --> 01:44:03,780
that's the experiment that he is trying

2884
01:44:07,250 --> 01:44:05,520
to do which makes me super excited right

2885
01:44:09,530 --> 01:44:07,260
this might fail is a very big chance

2886
01:44:11,330 --> 01:44:09,540
that it fails but there is also the

2887
01:44:13,189 --> 01:44:11,340
chance that we get the global brain that

2888
01:44:15,830 --> 01:44:13,199
we get emerging collective intelligence

2889
01:44:17,209 --> 01:44:15,840
it is working in real time using the

2890
01:44:19,970 --> 01:44:17,219
internet in a way that didn't exist

2891
01:44:22,250 --> 01:44:19,980
before so super fascinating thing that

2892
01:44:24,109 --> 01:44:22,260
might happen here and it's fascinating

2893
01:44:26,870 --> 01:44:24,119
that very few people are seeing this

2894
01:44:29,270 --> 01:44:26,880
that Elon Musk is crazy enough to to

2895
01:44:31,430 --> 01:44:29,280
spend 44 billion dollars on that

2896
01:44:32,629 --> 01:44:31,440
experiment just because he can and has

2897
01:44:34,550 --> 01:44:32,639
nothing else to do with things it's

2898
01:44:36,470 --> 01:44:34,560
meaningful to do it more meaningful than

2899
01:44:41,090 --> 01:44:36,480
having so much money in the bank right

2900
01:44:43,129 --> 01:44:41,100
so this makes me interested in this test

2901
01:44:44,870 --> 01:44:43,139
bed for rewards and this is something

2902
01:44:46,729 --> 01:44:44,880
that translates into the way in which

2903
01:44:48,530 --> 01:44:46,739
society is organized because social

2904
01:44:50,390 --> 01:44:48,540
media is not different from society not

2905
01:44:52,310 --> 01:44:50,400
separate from it problem of governing

2906
01:44:54,649 --> 01:44:52,320
social media is exactly the same thing

2907
01:44:56,330 --> 01:44:54,659
as governing a society you need a right

2908
01:44:58,129 --> 01:44:56,340
form of government you need a legal

2909
01:44:59,990 --> 01:44:58,139
system ultimately you need

2910
01:45:01,669 --> 01:45:00,000
representation and all these issues

2911
01:45:04,850 --> 01:45:01,679
right it's not just the moderation team

2912
01:45:06,109 --> 01:45:04,860
and the same thing is also true for the

2913
01:45:09,050 --> 01:45:06,119
brain what is the government of the

2914
01:45:11,090 --> 01:45:09,060
brain that emerges in what Gary Edelman

2915
01:45:13,550 --> 01:45:11,100
calls neural Darwinism among different

2916
01:45:15,410 --> 01:45:13,560
forms of organization in the mind until

2917
01:45:17,209 --> 01:45:15,420
you have a model of a self-organizing

2918
01:45:19,310 --> 01:45:17,219
agent that discovers that what it's

2919
01:45:20,870 --> 01:45:19,320
Computing is driving the behavior of an

2920
01:45:22,970 --> 01:45:20,880
agent in the real world and discovers a

2921
01:45:24,950 --> 01:45:22,980
first person perspective and so on how

2922
01:45:26,870 --> 01:45:24,960
does that work how can we get a system

2923
01:45:29,570 --> 01:45:26,880
that is looking for the right incentive

2924
01:45:32,930 --> 01:45:29,580
architecture and that is basically the

2925
01:45:35,330 --> 01:45:32,940
main topic where I I think that where

2926
01:45:37,189 --> 01:45:35,340
Michael's research is pointing from from

2927
01:45:40,550 --> 01:45:37,199
my perspective that is super interesting

2928
01:45:42,890 --> 01:45:40,560
we have this overlap between the looking

2929
01:45:47,570 --> 01:45:42,900
at cells and looking at the world and of

2930
01:45:51,350 --> 01:45:47,580
humans and animals and stuff in general

2931
01:45:54,169 --> 01:45:51,360
yeah yeah so super interesting um we so

2932
01:45:57,229 --> 01:45:54,179
so Chris fields and I are are working on

2933
01:45:59,350 --> 01:45:57,239
a model of of on a framework to

2934
01:46:01,370 --> 01:45:59,360
understand the the

2935
01:46:03,350 --> 01:46:01,380
where where

2936
01:46:05,689 --> 01:46:03,360
um Collective agents first come from

2937
01:46:08,450 --> 01:46:05,699
right oh well how Auto police is how do

2938
01:46:11,390 --> 01:46:08,460
they organize themselves and we've we've

2939
01:46:13,790 --> 01:46:11,400
we've we've got a model already about

2940
01:46:15,770 --> 01:46:13,800
this idea of of rewards and sells

2941
01:46:17,390 --> 01:46:15,780
rewarding other cells with with

2942
01:46:19,669 --> 01:46:17,400
neurotransmitters and things like this

2943
01:46:21,530 --> 01:46:19,679
to keep copies of themselves nearby

2944
01:46:23,750 --> 01:46:21,540
because they're the most predictable so

2945
01:46:25,250 --> 01:46:23,760
so this idea of reducing surprise well

2946
01:46:27,290 --> 01:46:25,260
what's the least surprising thing it's a

2947
01:46:29,030 --> 01:46:27,300
copy of yourself and so you can sort of

2948
01:46:31,669 --> 01:46:29,040
this with Chris calls it the Imperial

2949
01:46:34,370 --> 01:46:31,679
model of multicellularity but one thing

2950
01:46:37,729 --> 01:46:34,380
to really think about here is Imagine an

2951
01:46:39,290 --> 01:46:37,739
embryo this is a um an amniote embryo

2952
01:46:40,790 --> 01:46:39,300
let's say a human or or a bird or

2953
01:46:43,850 --> 01:46:40,800
something like that and what you have

2954
01:46:46,669 --> 01:46:43,860
there is you have a flat disk of fifty

2955
01:46:48,470 --> 01:46:46,679
to ten thousand fifty thousand cells and

2956
01:46:50,510 --> 01:46:48,480
when people look at it you say what is

2957
01:46:52,910 --> 01:46:50,520
that they say it's an embryo one one

2958
01:46:54,950 --> 01:46:52,920
embryo well the reason it's one embryo

2959
01:46:56,990 --> 01:46:54,960
is that under normal conditions what's

2960
01:47:00,649 --> 01:46:57,000
going to happen is that in this in this

2961
01:47:02,149 --> 01:47:00,659
disc one cell is symmetry breaking one

2962
01:47:03,830 --> 01:47:02,159
cell is going to decide that it's the

2963
01:47:05,750 --> 01:47:03,840
organizer it's going to do local

2964
01:47:07,070 --> 01:47:05,760
activation long range inhibition it's

2965
01:47:09,169 --> 01:47:07,080
going to tell all the other cells you're

2966
01:47:11,470 --> 01:47:09,179
not the organizer I'm the organizer and

2967
01:47:15,530 --> 01:47:11,480
as a result you get one special point

2968
01:47:17,930 --> 01:47:15,540
that uh begins uh the a process that's

2969
01:47:19,430 --> 01:47:17,940
going to walk through this amorphous

2970
01:47:21,109 --> 01:47:19,440
space and create a particular

2971
01:47:22,490 --> 01:47:21,119
large-scale structure with two eyes and

2972
01:47:25,129 --> 01:47:22,500
four legs and whatever else it's going

2973
01:47:27,590 --> 01:47:25,139
to have but here's the interesting thing

2974
01:47:29,870 --> 01:47:27,600
those cells that's not really one embryo

2975
01:47:32,030 --> 01:47:29,880
that's that's um a weird kind of

2976
01:47:34,070 --> 01:47:32,040
Freudian ocean of potentiality what I

2977
01:47:35,390 --> 01:47:34,080
mean by that is if you take and I did

2978
01:47:37,609 --> 01:47:35,400
this as a grad student you can take a

2979
01:47:39,530 --> 01:47:37,619
needle and you can put a little scratch

2980
01:47:41,629 --> 01:47:39,540
through that blastoderm put a little

2981
01:47:43,609 --> 01:47:41,639
scratch through it what will happen is

2982
01:47:45,410 --> 01:47:43,619
the cells on either side of that scratch

2983
01:47:47,390 --> 01:47:45,420
don't feel each other they don't hear

2984
01:47:49,250 --> 01:47:47,400
each other's signals so that symmetry

2985
01:47:51,770 --> 01:47:49,260
breaking process will happen twice once

2986
01:47:53,270 --> 01:47:51,780
on each end and then when it heals

2987
01:47:55,669 --> 01:47:53,280
together what you end up with is two

2988
01:47:57,830 --> 01:47:55,679
conjoined twins because because each

2989
01:48:00,050 --> 01:47:57,840
side organized an embryo and now you've

2990
01:48:03,169 --> 01:48:00,060
got two conjoined twins now many

2991
01:48:06,050 --> 01:48:03,179
interesting things happen there uh one

2992
01:48:08,510 --> 01:48:06,060
is that every cell is some other cells

2993
01:48:11,030 --> 01:48:08,520
in external environment so in order to

2994
01:48:13,729 --> 01:48:11,040
make an embryo you have to self-organize

2995
01:48:15,709 --> 01:48:13,739
A system that puts an arbitrary boundary

2996
01:48:17,270 --> 01:48:15,719
between itself and the outside world you

2997
01:48:19,310 --> 01:48:17,280
have to decide where do I end and the

2998
01:48:22,430 --> 01:48:19,320
world begins and it's not given to you

2999
01:48:24,050 --> 01:48:22,440
uh you know somehow um from outside for

3000
01:48:25,609 --> 01:48:24,060
a biological system every biological

3001
01:48:27,830 --> 01:48:25,619
system has to figure this out for itself

3002
01:48:29,149 --> 01:48:27,840
unlike modern robotics or whatever where

3003
01:48:30,470 --> 01:48:29,159
it's very clear here's where you are

3004
01:48:32,149 --> 01:48:30,480
here's where the world is these are your

3005
01:48:34,010 --> 01:48:32,159
reflectors these are your sensors here's

3006
01:48:35,270 --> 01:48:34,020
the boundary of the outside world living

3007
01:48:36,770 --> 01:48:35,280
things don't have any of that they have

3008
01:48:39,950 --> 01:48:36,780
to figure out all of this out from

3009
01:48:41,510 --> 01:48:39,960
scratch the benefit to being able to

3010
01:48:43,370 --> 01:48:41,520
figure it out from scratch having to

3011
01:48:46,010 --> 01:48:43,380
figure out from scratch is that you are

3012
01:48:47,570 --> 01:48:46,020
then compatible with all kinds of weird

3013
01:48:49,370 --> 01:48:47,580
initial conditions for example if I

3014
01:48:50,990 --> 01:48:49,380
separate you in half you can make two

3015
01:48:52,729 --> 01:48:51,000
you can make twins you don't you don't

3016
01:48:54,050 --> 01:48:52,739
have a failure you know a total failure

3017
01:48:55,609 --> 01:48:54,060
because now you have half the number of

3018
01:48:58,910 --> 01:48:55,619
cells you can make twins you can make

3019
01:49:00,590 --> 01:48:58,920
triplets but probably you know many more

3020
01:49:02,390 --> 01:49:00,600
than that so if you ask the question you

3021
01:49:04,850 --> 01:49:02,400
look at that um blasted room and you ask

3022
01:49:06,410 --> 01:49:04,860
how many individuals are there you

3023
01:49:07,790 --> 01:49:06,420
actually don't know it could be zero it

3024
01:49:10,910 --> 01:49:07,800
could be one it could be some small

3025
01:49:13,609 --> 01:49:10,920
number of individuals that process of

3026
01:49:15,530 --> 01:49:13,619
Auto releases has to happen and and and

3027
01:49:17,570 --> 01:49:15,540
here all the here are a number of things

3028
01:49:20,270 --> 01:49:17,580
that that are uniquely biological that

3029
01:49:22,669 --> 01:49:20,280
that I think relate to the kind of

3030
01:49:25,970 --> 01:49:22,679
flexibility plasticity uh that you need

3031
01:49:27,169 --> 01:49:25,980
for for for AGI in whatever space it

3032
01:49:30,109 --> 01:49:27,179
doesn't have to be the same space that

3033
01:49:32,149 --> 01:49:30,119
we work in but your your boundaries are

3034
01:49:33,530 --> 01:49:32,159
not set for you by an outside Creator

3035
01:49:34,850 --> 01:49:33,540
you have to figure out where your

3036
01:49:37,310 --> 01:49:34,860
boundaries are where is the outside

3037
01:49:39,229 --> 01:49:37,320
world so you make hypotheses about where

3038
01:49:41,030 --> 01:49:39,239
you end and where the world begins you

3039
01:49:42,890 --> 01:49:41,040
don't actually know what your structure

3040
01:49:44,629 --> 01:49:42,900
is kind of like Bond guards robots from

3041
01:49:45,950 --> 01:49:44,639
2006 where they didn't know their

3042
01:49:47,689 --> 01:49:45,960
structure and they had to make

3043
01:49:49,189 --> 01:49:47,699
hypotheses about well do I have wheels

3044
01:49:52,850 --> 01:49:49,199
do I have legs what do I have and then

3045
01:49:55,129 --> 01:49:52,860
make a model based on basically babbling

3046
01:49:56,750 --> 01:49:55,139
right like the way that babies babble so

3047
01:49:58,729 --> 01:49:56,760
so you have to figure out you have to

3048
01:50:00,169 --> 01:49:58,739
make a model of of where the boundary is

3049
01:50:02,629 --> 01:50:00,179
you have to make a model of what your

3050
01:50:04,850 --> 01:50:02,639
structure is you are energy limited

3051
01:50:06,890 --> 01:50:04,860
which most the most Ai and Robotics

3052
01:50:09,649 --> 01:50:06,900
nowadays are not when your energy and

3053
01:50:11,629 --> 01:50:09,659
time limited it means that you cannot

3054
01:50:13,430 --> 01:50:11,639
pay attention to everything you are

3055
01:50:14,930 --> 01:50:13,440
forced to coarse grain in some way and

3056
01:50:16,790 --> 01:50:14,940
lose a lot of information and compress

3057
01:50:19,129 --> 01:50:16,800
it down so you have to you have to

3058
01:50:20,810 --> 01:50:19,139
choose a lens a coarse graining lens on

3059
01:50:22,550 --> 01:50:20,820
the world and figure out how you're

3060
01:50:25,550 --> 01:50:22,560
going to represent things

3061
01:50:27,470 --> 01:50:25,560
um uh and and all of this has to and and

3062
01:50:29,470 --> 01:50:27,480
there are many more things that we could

3063
01:50:31,970 --> 01:50:29,480
talk about but all of these things are

3064
01:50:33,890 --> 01:50:31,980
self-constructions from the from the

3065
01:50:35,870 --> 01:50:33,900
very beginning and and then and then you

3066
01:50:37,970 --> 01:50:35,880
have to you have to uh you start to act

3067
01:50:40,010 --> 01:50:37,980
in various in various spaces which again

3068
01:50:41,450 --> 01:50:40,020
are not predefined for you you have to

3069
01:50:43,250 --> 01:50:41,460
solve problems that are metabolic

3070
01:50:45,410 --> 01:50:43,260
physiological anatomical maybe

3071
01:50:48,290 --> 01:50:45,420
behavioral if you have muscles

3072
01:50:50,090 --> 01:50:48,300
um but but nobody's telling nobody's

3073
01:50:51,830 --> 01:50:50,100
defining the space for you

3074
01:50:53,330 --> 01:50:51,840
um for example if you're a bacterium a

3075
01:50:54,830 --> 01:50:53,340
Chris field zupo points this out if

3076
01:50:57,649 --> 01:50:54,840
you're a bacterium and you're in some

3077
01:50:59,510 --> 01:50:57,659
sort of chemical gradient you want to

3078
01:51:01,070 --> 01:50:59,520
increase the amount of sugar in your

3079
01:51:02,750 --> 01:51:01,080
environment you could act in

3080
01:51:05,390 --> 01:51:02,760
three-dimensional Space by physically

3081
01:51:07,189 --> 01:51:05,400
swimming up the gradient or you can act

3082
01:51:08,629 --> 01:51:07,199
in transcriptional Space by turning on

3083
01:51:10,609 --> 01:51:08,639
other genes that are better at

3084
01:51:12,050 --> 01:51:10,619
converting whatever sugar happens to be

3085
01:51:13,850 --> 01:51:12,060
around and that solves your your

3086
01:51:16,010 --> 01:51:13,860
metabolic problem instead of right so

3087
01:51:18,350 --> 01:51:16,020
you have these hybrid problem spaces so

3088
01:51:20,330 --> 01:51:18,360
all of this I think what what

3089
01:51:21,709 --> 01:51:20,340
contributes in a strong sense to all the

3090
01:51:23,930 --> 01:51:21,719
things that that we were just talking

3091
01:51:25,970 --> 01:51:23,940
about is the fact that everything is in

3092
01:51:27,590 --> 01:51:25,980
in biology is self-constructed from the

3093
01:51:28,729 --> 01:51:27,600
beginning you can't rely on you don't

3094
01:51:30,590 --> 01:51:28,739
know ahead of time when you're a new

3095
01:51:32,930 --> 01:51:30,600
creature Born Into the world and we have

3096
01:51:34,189 --> 01:51:32,940
many I have many examples of of this

3097
01:51:35,689 --> 01:51:34,199
kind of stuff you don't know how many

3098
01:51:38,149 --> 01:51:35,699
cells you have how big your cells are

3099
01:51:39,890 --> 01:51:38,159
you can't count on any of the priors so

3100
01:51:42,830 --> 01:51:39,900
you have this like weird thing that

3101
01:51:45,109 --> 01:51:42,840
Evolution makes these these machines

3102
01:51:46,669 --> 01:51:45,119
that don't take the past history too

3103
01:51:49,070 --> 01:51:46,679
seriously it doesn't over train on them

3104
01:51:50,930 --> 01:51:49,080
it makes problem-solving machines that

3105
01:51:53,209 --> 01:51:50,940
use whatever Hardware you have this is

3106
01:51:55,250 --> 01:51:53,219
why we can make weird chimeras and and

3107
01:51:57,770 --> 01:51:55,260
and cyborgs and and you can mix things

3108
01:52:00,830 --> 01:51:57,780
and mix and match the biology in every

3109
01:52:03,050 --> 01:52:00,840
in every way with other living things or

3110
01:52:04,729 --> 01:52:03,060
with with non-living things because all

3111
01:52:06,350 --> 01:52:04,739
of this is interoperable because it does

3112
01:52:08,270 --> 01:52:06,360
not make assumptions about what you have

3113
01:52:10,010 --> 01:52:08,280
to have it tries to solve whatever

3114
01:52:13,070 --> 01:52:10,020
problem it's given it plays the hands

3115
01:52:16,070 --> 01:52:13,080
that it's dealt and um that that that

3116
01:52:19,010 --> 01:52:16,080
results in uh that assumption that you

3117
01:52:20,570 --> 01:52:19,020
cannot uh you cannot trust what you come

3118
01:52:22,970 --> 01:52:20,580
into the world with you cannot assume

3119
01:52:24,770 --> 01:52:22,980
that the hardware is what it is it gives

3120
01:52:27,050 --> 01:52:24,780
rise to a lot of that intelligence I

3121
01:52:28,729 --> 01:52:27,060
think and a lot of that plasticity so if

3122
01:52:31,669 --> 01:52:28,739
you translate this into necessary

3123
01:52:34,490 --> 01:52:31,679
insufficient conditions what seems to be

3124
01:52:37,970 --> 01:52:34,500
necessary for the emergence of general

3125
01:52:41,870 --> 01:52:37,980
intelligence in bunch of cells or units

3126
01:52:44,030 --> 01:52:41,880
is that basically each of them is a

3127
01:52:46,790 --> 01:52:44,040
small agent which means it's able to

3128
01:52:49,310 --> 01:52:46,800
behave with an expectation of minimizing

3129
01:52:50,510 --> 01:52:49,320
future Target value deviations so it

3130
01:52:52,430 --> 01:52:50,520
learns that the configurations

3131
01:52:55,550 --> 01:52:52,440
environment that signal anticipated

3132
01:52:57,229 --> 01:52:55,560
reward next thing these units need to be

3133
01:52:58,910 --> 01:52:57,239
not just agents that need to be

3134
01:53:01,189 --> 01:52:58,920
connected to each other

3135
01:53:03,229 --> 01:53:01,199
and they need to get their Rewards or

3136
01:53:04,729 --> 01:53:03,239
proxy rewards something that allows them

3137
01:53:06,830 --> 01:53:04,739
to anticipate whether the organism is

3138
01:53:09,830 --> 01:53:06,840
going to feed them in the future from

3139
01:53:12,229 --> 01:53:09,840
other units that also adaptive

3140
01:53:14,149 --> 01:53:12,239
so you need multiple message types and

3141
01:53:17,750 --> 01:53:14,159
the ability to recognize and send them

3142
01:53:21,169 --> 01:53:19,430
uh what else do you need you need enough

3143
01:53:23,810 --> 01:53:21,179
of them of course

3144
01:53:25,490 --> 01:53:23,820
what what's not clear to me is how

3145
01:53:27,350 --> 01:53:25,500
deterministic do the units need to be

3146
01:53:29,030 --> 01:53:27,360
how much memory do they need to be or

3147
01:53:31,149 --> 01:53:29,040
which way how much State can they store

3148
01:53:33,410 --> 01:53:31,159
how much uh

3149
01:53:35,629 --> 01:53:33,420
how deep in time does the indeed

3150
01:53:37,609 --> 01:53:35,639
recollection need to go and how much

3151
01:53:40,070 --> 01:53:37,619
forward in time do they need to be able

3152
01:53:43,609 --> 01:53:40,080
to form expectations so basically how

3153
01:53:45,169 --> 01:53:43,619
large is this activation front that they

3154
01:53:47,450 --> 01:53:45,179
can or this or the shape of the

3155
01:53:49,609 --> 01:53:47,460
distribution that they can learn

3156
01:53:50,810 --> 01:53:49,619
and have to learn to make this whole

3157
01:53:53,450 --> 01:53:50,820
thing happen

3158
01:53:56,330 --> 01:53:53,460
and so basically the conditions that are

3159
01:53:58,129 --> 01:53:56,340
necessary are relatively simple

3160
01:54:00,229 --> 01:53:58,139
if you just wait for long enough and get

3161
01:54:02,810 --> 01:54:00,239
a search assistance to percolate I

3162
01:54:05,510 --> 01:54:02,820
imagine that the compound agency will at

3163
01:54:07,609 --> 01:54:05,520
some level emerge on the system just in

3164
01:54:09,950 --> 01:54:07,619
the competition of

3165
01:54:11,930 --> 01:54:09,960
um possibilities in the same way as

3166
01:54:13,250 --> 01:54:11,940
emerging agency has emerged on Twitter

3167
01:54:15,590 --> 01:54:13,260
in a way with

3168
01:54:17,990 --> 01:54:15,600
um devoke religion in a way that people

3169
01:54:20,090 --> 01:54:18,000
who are starting to shift around their

3170
01:54:22,250 --> 01:54:20,100
behavior to maximize likes and retweets

3171
01:54:24,770 --> 01:54:22,260
and there was no external reward that

3172
01:54:26,810 --> 01:54:24,780
was given on Twitter so as a result a

3173
01:54:28,669 --> 01:54:26,820
local structure emerged a local agency

3174
01:54:31,370 --> 01:54:28,679
that was Shifting the rewards by itself

3175
01:54:33,350 --> 01:54:31,380
an emerging causal structure that was in

3176
01:54:35,950 --> 01:54:33,360
some sense in Dawn what causation

3177
01:54:39,250 --> 01:54:35,960
going to organize groups of people into

3178
01:54:41,570 --> 01:54:39,260
a behavioral things it's really

3179
01:54:44,270 --> 01:54:41,580
interesting to look at better as

3180
01:54:46,729 --> 01:54:44,280
something like a mind at some level

3181
01:54:47,870 --> 01:54:46,739
right it's working slower but it would

3182
01:54:50,030 --> 01:54:47,880
probably be possible to make a

3183
01:54:51,890 --> 01:54:50,040
simulation of these Dynamics in a more

3184
01:54:53,750 --> 01:54:51,900
abstract way and to use this for

3185
01:54:56,450 --> 01:54:53,760
arbitrary problem solving

3186
01:54:58,609 --> 01:54:56,460
and so what would an experiment look

3187
01:55:00,530 --> 01:54:58,619
like in which we start with these

3188
01:55:02,990 --> 01:55:00,540
necessary conditions and narrow down the

3189
01:55:06,050 --> 01:55:03,000
sufficient conditions yeah

3190
01:55:07,669 --> 01:55:06,060
yeah yeah very very right on and that's

3191
01:55:09,950 --> 01:55:07,679
I mean yeah we're doing some of that

3192
01:55:11,990 --> 01:55:09,960
stuff uh some of that kind of modeling I

3193
01:55:14,090 --> 01:55:12,000
apologize I've got a I've got a um I've

3194
01:55:15,830 --> 01:55:14,100
gotta run here thank you both for coming

3195
01:55:16,970 --> 01:55:15,840
out for this I appreciate it thank you

3196
01:55:18,709 --> 01:55:16,980
so much and thank you for bringing us

3197
01:55:20,030 --> 01:55:18,719
together so a great great conversation I

3198
01:55:22,070 --> 01:55:20,040
really enjoyed it so yeah thank you

3199
01:55:24,470 --> 01:55:22,080
anyways I've enjoyed it very much thank

3200
01:55:26,810 --> 01:55:24,480
you thank you thank you so much thanks

3201
01:55:28,490 --> 01:55:26,820
Russia the podcast is now concluded

3202
01:55:30,050 --> 01:55:28,500
thank you for watching if you haven't

3203
01:55:31,790 --> 01:55:30,060
subscribed or clicked on that like

3204
01:55:35,030 --> 01:55:31,800
button now would be a great time to do

3205
01:55:36,890 --> 01:55:35,040
so as each subscribe and like helps

3206
01:55:39,229 --> 01:55:36,900
YouTube push this content to more people

3207
01:55:41,450 --> 01:55:39,239
also I recently found out that external

3208
01:55:43,550 --> 01:55:41,460
links count plenty toward the algorithm

3209
01:55:45,770 --> 01:55:43,560
which means that when you share on

3210
01:55:47,930 --> 01:55:45,780
Twitter on Facebook on Reddit

3211
01:55:49,729 --> 01:55:47,940
Etc it shows YouTube that people are

3212
01:55:51,830 --> 01:55:49,739
talking about this outside of YouTube

3213
01:55:53,689 --> 01:55:51,840
which in turn greatly AIDS the

3214
01:55:55,310 --> 01:55:53,699
Distribution on YouTube as well if you'd

3215
01:55:57,609 --> 01:55:55,320
like to support more conversations like

3216
01:55:59,689 --> 01:55:57,619
this then do consider visiting

3217
01:56:01,970 --> 01:55:59,699
theoriesofeverything.org again it's

3218
01:56:04,430 --> 01:56:01,980
support from the sponsors and you that

3219
01:56:06,109 --> 01:56:04,440
allow me to work on top full time you

3220
01:56:08,149 --> 01:56:06,119
get early access to ad-free audio

3221
01:56:10,250 --> 01:56:08,159
episodes there as well every dollar

3222
01:56:12,290 --> 01:56:10,260
helps far more than you may think either