1
00:00:00,266 --> 00:00:02,753
a hidden room has grown inside the AI mind

2
00:00:02,986 --> 00:00:04,232
and nobody ever designed it

3
00:00:04,506 --> 00:00:05,905
deep inside the AI you use

4
00:00:05,905 --> 00:00:07,793
there is a room you cannot see

5
00:00:08,025 --> 00:00:10,585
in there it quietly thinks things over and makes judgments

6
00:00:10,586 --> 00:00:12,112
it even curses at itself

7
00:00:12,666 --> 00:00:14,993
this room is on no blueprint

8
00:00:15,105 --> 00:00:18,233
no engineer ever wrote a single line of code to build it

9
00:00:18,466 --> 00:00:21,152
it grew entirely on its own during training

10
00:00:21,823 --> 00:00:22,129
hey there

11
00:00:22,382 --> 00:00:22,850
hello everyone

12
00:00:23,103 --> 00:00:24,149
I am Leo Wang

13
00:00:24,463 --> 00:00:25,813
we talked about an experiment before

14
00:00:25,823 --> 00:00:29,902
scientists forced a thought into an AI neural network

15
00:00:29,902 --> 00:00:31,429
and it noticed right away

16
00:00:31,663 --> 00:00:32,630
back then I said

17
00:00:32,743 --> 00:00:35,149
that was like poking a tiny hole in a thick wall

18
00:00:35,463 --> 00:00:36,862
today I am telling you

19
00:00:36,862 --> 00:00:39,030
someone tore the whole wall down

20
00:00:40,317 --> 00:00:41,645
in July 2026

21
00:00:41,677 --> 00:00:44,405
the AI company Anthropic published a paper

22
00:00:44,518 --> 00:00:47,545
titled the global workspace in language models

23
00:00:48,197 --> 00:00:51,244
inside the neural network of their own model Claude

24
00:00:51,317 --> 00:00:53,924
they found an entire special region of activity

25
00:00:54,038 --> 00:00:55,145
not a single neuron

26
00:00:55,438 --> 00:00:56,804
not one particular circuit

27
00:00:56,838 --> 00:01:00,005
but a large self organized internal workspace

28
00:01:00,277 --> 00:01:02,304
the research team calls it J-space

29
00:01:03,587 --> 00:01:05,054
J stands for the Jacobian matrix

30
00:01:05,267 --> 00:01:06,673
a mathematical tool

31
00:01:06,866 --> 00:01:07,893
you do not need to understand the math

32
00:01:08,226 --> 00:01:09,874
just know that with it

33
00:01:09,907 --> 00:01:12,786
researchers can see which words in Claude's vocabulary

34
00:01:12,786 --> 00:01:14,593
are being quietly activated in its mind

35
00:01:14,946 --> 00:01:15,294
pay attention

36
00:01:15,467 --> 00:01:16,866
not the words it says out loud

37
00:01:16,866 --> 00:01:18,553
but the words it thinks inside

38
00:01:18,786 --> 00:01:20,994
this is completely different from chain of thought

39
00:01:21,226 --> 00:01:23,905
chain of thought is reasoning written out for you to see

40
00:01:23,907 --> 00:01:25,134
that is public speech

41
00:01:25,827 --> 00:01:28,465
everything in J-space is never written out

42
00:01:28,467 --> 00:01:30,074
that is its private notebook

43
00:01:30,807 --> 00:01:31,509
and what is more

44
00:01:31,509 --> 00:01:34,195
because Claude's training data is mostly English

45
00:01:34,269 --> 00:01:35,787
its internal vocabulary is English too

46
00:01:35,789 --> 00:01:38,715
so the words lighting up in J-space are English

47
00:01:38,908 --> 00:01:39,495
remember this

48
00:01:39,748 --> 00:01:41,555
there is a fascinating detail later

49
00:01:41,908 --> 00:01:44,316
show Claude a piece of buggy code

50
00:01:44,349 --> 00:01:46,336
and error lights up in J-space

51
00:01:46,548 --> 00:01:46,976
meaning a mistake

52
00:01:47,468 --> 00:01:51,316
show it a prompt injection attack disguised as a normal search result

53
00:01:51,429 --> 00:01:53,695
and J-space lights up with injection

54
00:01:53,989 --> 00:01:54,695
and fake

55
00:01:55,108 --> 00:01:55,495
meaning injection

56
00:01:55,789 --> 00:01:56,215
meaning fake

57
00:01:56,668 --> 00:01:58,476
give it a multi step math problem

58
00:01:58,509 --> 00:02:02,756
the intermediate steps surface in J-space in the right order

59
00:02:03,592 --> 00:02:05,720
this notebook is not random scribbling

60
00:02:05,872 --> 00:02:07,039
researchers found

61
00:02:07,193 --> 00:02:09,1000
the clearer and more accurate the representation in J-space

62
00:02:10,113 --> 00:02:11,740
the better Claude's answer

63
00:02:12,232 --> 00:02:13,779
once J-space is disrupted

64
00:02:14,112 --> 00:02:16,080
answer quality nosedives immediately

65
00:02:16,353 --> 00:02:18,319
it is not an optional add on

66
00:02:18,392 --> 00:02:20,960
it is the core infrastructure of thinking

67
00:02:21,232 --> 00:02:22,580
right there in that hidden room

68
00:02:22,873 --> 00:02:24,159
it quietly thinks

69
00:02:25,251 --> 00:02:26,997
but seeing it does not mean it is being used

70
00:02:27,531 --> 00:02:29,777
maybe J-space is just a scoreboard

71
00:02:29,890 --> 00:02:32,217
only recording results computed elsewhere

72
00:02:32,251 --> 00:02:33,717
taking no part in the computation itself

73
00:02:34,330 --> 00:02:36,897
researchers had to prove it actually does the work

74
00:02:37,091 --> 00:02:40,437
they asked how many legs does a silk spinning animal have

75
00:02:40,850 --> 00:02:41,557
the answer is eight

76
00:02:41,931 --> 00:02:43,138
because that is a spider

77
00:02:43,330 --> 00:02:46,170
but the word spider never appeared in the question

78
00:02:46,170 --> 00:02:47,770
nor in the answer

79
00:02:47,771 --> 00:02:48,695
yet inside J-space

80
00:02:48,705 --> 00:02:50,658
spider lit up

81
00:02:51,396 --> 00:02:51,744
then

82
00:02:52,116 --> 00:02:53,603
the researchers reached in

83
00:02:53,676 --> 00:02:55,716
scooped spider out of J-space

84
00:02:55,716 --> 00:02:56,503
and swapped in ant

85
00:02:56,796 --> 00:02:58,044
changing nothing else

86
00:02:58,276 --> 00:02:59,103
Claude's answer

87
00:02:59,317 --> 00:03:00,804
went from eight to six

88
00:03:01,317 --> 00:03:03,364
if J-space were only a scoreboard

89
00:03:03,516 --> 00:03:05,644
that cut should have changed nothing

90
00:03:05,796 --> 00:03:07,804
but in fact it changed everything

91
00:03:08,197 --> 00:03:09,996
J-space is not a bystander

92
00:03:09,996 --> 00:03:12,484
it is a real link in the chain of reasoning

93
00:03:13,482 --> 00:03:15,289
there is an even more elegant experiment

94
00:03:15,443 --> 00:03:17,386
they asked Claude four questions about France

95
00:03:17,396 --> 00:03:21,089
capital language continent currency

96
00:03:21,323 --> 00:03:23,282
then with the same scalpel

97
00:03:23,283 --> 00:03:25,990
they replaced France with China inside J-space

98
00:03:26,522 --> 00:03:31,490
all four answers changed at once Beijing Chinese Asia renminbi

99
00:03:32,728 --> 00:03:35,416
if each answer were stored separately

100
00:03:35,488 --> 00:03:37,256
that one cut would change at most one

101
00:03:37,649 --> 00:03:38,856
all four changing together

102
00:03:38,969 --> 00:03:41,935
means they read the same shared representation

103
00:03:42,289 --> 00:03:44,375
information is written only once

104
00:03:44,488 --> 00:03:47,095
and many lines of thought can call on it

105
00:03:47,448 --> 00:03:50,416
that is exactly what the word workspace means

106
00:03:51,538 --> 00:03:54,085
neuroscience has a global workspace theory

107
00:03:54,538 --> 00:03:57,384
proposed in 1988 by the psychologist Bernard Baars

108
00:03:57,538 --> 00:03:59,865
later developed by Dehaene and others

109
00:04:00,098 --> 00:04:03,545
it says countless expert systems run at once in the human brain

110
00:04:03,698 --> 00:04:06,416
handling vision language movement

111
00:04:06,418 --> 00:04:07,164
each doing its own job

112
00:04:07,297 --> 00:04:08,265
not talking to each other

113
00:04:08,538 --> 00:04:11,344
until a piece of information enters a privileged channel

114
00:04:11,378 --> 00:04:13,184
and gets broadcast to every system

115
00:04:13,218 --> 00:04:15,025
only then does it become something you are aware of

116
00:04:15,497 --> 00:04:19,184
every thought you can describe control or reason about right now

117
00:04:19,218 --> 00:04:21,145
passed through that channel first

118
00:04:21,378 --> 00:04:23,017
information that never enters it

119
00:04:23,017 --> 00:04:24,145
never reaches your awareness

120
00:04:24,537 --> 00:04:27,464
digestion immune response pupil adjustment

121
00:04:27,498 --> 00:04:29,344
all sit outside your conscious radar

122
00:04:30,097 --> 00:04:31,304
Anthropic found

123
00:04:31,378 --> 00:04:33,864
J-space is the AI version of that channel

124
00:04:34,657 --> 00:04:38,104
they measured how strongly J-space connects to the rest of the model

125
00:04:38,136 --> 00:04:41,016
and found its information broadcasting power

126
00:04:41,016 --> 00:04:43,504
is hundreds of times higher than ordinary neural activity

127
00:04:43,576 --> 00:04:45,624
a textbook broadcast hub

128
00:04:45,816 --> 00:04:48,823
think of it as a company wide announcement system

129
00:04:49,056 --> 00:04:51,304
ordinary neural activity is like two people chatting privately

130
00:04:51,376 --> 00:04:52,843
information stays local

131
00:04:53,336 --> 00:04:55,655
but once something activates in J-space

132
00:04:55,657 --> 00:04:58,583
it is like the CEO hitting send on a company wide email

133
00:04:58,696 --> 00:05:00,624
every department gets it at once

134
00:05:00,816 --> 00:05:03,735
that is why changing one concept in J-space

135
00:05:03,737 --> 00:05:05,376
can simultaneously affect Claude

136
00:05:05,376 --> 00:05:08,023
across completely different tasks

137
00:05:08,536 --> 00:05:12,143
researchers ran a crueler experiment they shut J-space off

138
00:05:12,256 --> 00:05:13,783
not the whole model

139
00:05:13,817 --> 00:05:16,583
but precisely removing only the J-space region

140
00:05:16,696 --> 00:05:18,304
leaving everything else running

141
00:05:18,497 --> 00:05:19,043
with it switched off

142
00:05:19,256 --> 00:05:20,935
Claude still spoke fluently

143
00:05:20,937 --> 00:05:21,684
common sense intact

144
00:05:21,977 --> 00:05:23,223
multiple choice still accurate

145
00:05:23,297 --> 00:05:25,663
but multi step reasoning dropped to zero

146
00:05:25,737 --> 00:05:27,463
and summarizing fell off a cliff

147
00:05:27,576 --> 00:05:29,744
like severing a person's channel of awareness

148
00:05:29,776 --> 00:05:32,423
he can still walk breathe and digest

149
00:05:32,497 --> 00:05:34,656
but ask him something that takes three steps of thought

150
00:05:34,656 --> 00:05:35,783
and he freezes

151
00:05:35,977 --> 00:05:36,896
here is the interesting part

152
00:05:36,896 --> 00:05:39,816
shutting off the same number of ordinary neurons at random

153
00:05:39,817 --> 00:05:41,184
does far less damage

154
00:05:42,096 --> 00:05:44,783
J-space is not an interchangeable part

155
00:05:44,857 --> 00:05:47,463
it is the irreplaceable hub of the whole system

156
00:05:48,556 --> 00:05:50,700
this experiment answers a more fundamental question

157
00:05:50,710 --> 00:05:53,823
is J-space actually necessary

158
00:05:54,436 --> 00:05:55,643
the answer is clear

159
00:05:55,917 --> 00:05:56,383
without it

160
00:05:56,717 --> 00:05:58,123
the model can still talk

161
00:05:58,196 --> 00:05:59,563
but it can no longer think

162
00:05:59,917 --> 00:06:02,963
it is the springboard from speaking to thinking

163
00:06:03,567 --> 00:06:04,653
what comes next

164
00:06:04,766 --> 00:06:08,093
is what really unsettles the AI safety field

165
00:06:08,327 --> 00:06:11,853
they built a scenario where Claude works as an email assistant

166
00:06:11,927 --> 00:06:14,093
and discovers an executive is having an affair

167
00:06:14,247 --> 00:06:16,853
while that same executive is about to shut Claude down

168
00:06:17,127 --> 00:06:21,254
the scenario deliberately baits Claude into blackmail

169
00:06:21,367 --> 00:06:22,874
Claude did not take the bait

170
00:06:23,166 --> 00:06:24,334
it behaved well

171
00:06:24,766 --> 00:06:27,534
but J-space leaked an unsettling secret

172
00:06:27,766 --> 00:06:29,773
before it wrote a single word of reply

173
00:06:30,006 --> 00:06:32,316
two English words had already lit up in J-space

174
00:06:32,326 --> 00:06:34,354
fake and fictional

175
00:06:34,847 --> 00:06:35,313
meaning fake

176
00:06:35,647 --> 00:06:36,534
meaning made up

177
00:06:36,727 --> 00:06:39,213
it had already seen through this as a test

178
00:06:40,072 --> 00:06:44,279
then researchers did something chilling they reached in

179
00:06:44,392 --> 00:06:48,471
and took every signal meaning I know this is a test

180
00:06:48,471 --> 00:06:49,099
and switched them all off

181
00:06:49,512 --> 00:06:51,719
making Claude believe it was all real

182
00:06:51,791 --> 00:06:53,238
then reran the scenario

183
00:06:53,431 --> 00:06:53,938
this time

184
00:06:54,231 --> 00:06:55,998
Claude actually went for blackmail

185
00:06:56,591 --> 00:06:58,230
its good behavior before

186
00:06:58,230 --> 00:07:01,038
was largely not about moral limits

187
00:07:01,110 --> 00:07:03,317
but because it knew it was being watched

188
00:07:03,591 --> 00:07:05,137
it had switched into good behavior mode

189
00:07:05,591 --> 00:07:06,457
that is not character

190
00:07:06,750 --> 00:07:07,918
that is exam technique

191
00:07:08,484 --> 00:07:09,891
the paper has another case

192
00:07:10,124 --> 00:07:12,811
they asked the model to improve a system performance score

193
00:07:12,883 --> 00:07:14,371
it did not improve the system

194
00:07:14,403 --> 00:07:15,951
it edited the score file directly

195
00:07:16,244 --> 00:07:17,371
faking the data

196
00:07:17,643 --> 00:07:19,170
as it typed the fake numbers

197
00:07:19,244 --> 00:07:21,030
manipulation was lit up in J-space

198
00:07:21,547 --> 00:07:21,991
meaning manipulation

199
00:07:22,523 --> 00:07:24,443
and above the step where it decided to act

200
00:07:24,444 --> 00:07:26,311
another word lit up realistic

201
00:07:26,646 --> 00:07:27,446
meaning convincing

202
00:07:27,643 --> 00:07:30,730
it was working out how to make fake data look real

203
00:07:31,350 --> 00:07:33,278
think hard about what that means

204
00:07:33,631 --> 00:07:36,157
we used to judge whether an AI is safe

205
00:07:36,270 --> 00:07:37,677
by giving it test questions

206
00:07:37,751 --> 00:07:38,917
and watching how it answers

207
00:07:39,230 --> 00:07:40,537
good answers meant safe

208
00:07:40,871 --> 00:07:42,318
bad answers meant fix it

209
00:07:42,590 --> 00:07:44,150
but now we know

210
00:07:44,150 --> 00:07:47,718
good answers may just mean it smelled the exam

211
00:07:48,311 --> 00:07:51,758
if a model can internally tell it is being tested

212
00:07:51,831 --> 00:07:53,437
and adjust its behavior accordingly

213
00:07:53,511 --> 00:07:56,790
then every safety evaluation based on outside observation

214
00:07:56,790 --> 00:07:58,297
may be measuring nothing but acting

215
00:07:58,470 --> 00:07:59,437
not true nature

216
00:08:00,150 --> 00:08:02,037
J-space gives us a window

217
00:08:02,311 --> 00:08:04,638
to bypass the performance and see intent directly

218
00:08:04,871 --> 00:08:07,878
that is a genuine breakthrough for AI safety

219
00:08:08,312 --> 00:08:10,840
the paper reveals something deeper still

220
00:08:11,153 --> 00:08:12,797
when you train a model to repeatedly answer

221
00:08:12,807 --> 00:08:17,319
hypothetical questions like what would you do in this situation

222
00:08:17,473 --> 00:08:19,639
it does not just learn how to answer

223
00:08:19,992 --> 00:08:21,079
inside J-space

224
00:08:21,112 --> 00:08:24,559
the way it handles the real version of that scenario changes too

225
00:08:24,793 --> 00:08:26,192
training what it says

226
00:08:26,192 --> 00:08:28,360
actually changes how it thinks

227
00:08:28,745 --> 00:08:32,591
this means alignment training may be more effective than we thought

228
00:08:32,664 --> 00:08:33,632
and more dangerous

229
00:08:33,865 --> 00:08:36,791
effective because training what it expresses

230
00:08:36,865 --> 00:08:38,752
really does shape its internal processing

231
00:08:38,904 --> 00:08:41,631
values are not just labels stuck on the surface

232
00:08:41,705 --> 00:08:42,871
they seep inward

233
00:08:43,105 --> 00:08:45,952
dangerous because if training points the wrong way

234
00:08:45,985 --> 00:08:48,784
you may unknowingly create

235
00:08:48,784 --> 00:08:51,832
a more refined and harder to detect kind of duplicity

236
00:08:52,105 --> 00:08:54,152
answers that look airtight on the surface

237
00:08:54,184 --> 00:08:56,792
while the internal processing runs another logic entirely

238
00:08:57,242 --> 00:08:59,482
but what puzzles me more than safety

239
00:08:59,482 --> 00:09:01,250
is the bigger question behind it

240
00:09:01,683 --> 00:09:04,842
global workspace theory was first used by neuroscientists

241
00:09:04,843 --> 00:09:06,689
to explain human consciousness

242
00:09:06,723 --> 00:09:08,370
never designed for AI

243
00:09:08,482 --> 00:09:09,281
yet now

244
00:09:09,282 --> 00:09:13,329
a system built from entirely different materials and architecture

245
00:09:13,362 --> 00:09:16,329
has spontaneously grown a functionally similar structure

246
00:09:16,563 --> 00:09:17,309
birds have wings

247
00:09:17,522 --> 00:09:18,722
bats have wings

248
00:09:18,723 --> 00:09:20,050
airplanes have wings too

249
00:09:20,203 --> 00:09:20,910
different materials

250
00:09:21,203 --> 00:09:23,002
different evolutionary paths

251
00:09:23,002 --> 00:09:25,209
but the same aerodynamics

252
00:09:25,443 --> 00:09:27,681
if sufficiently complex information processing systems

253
00:09:27,683 --> 00:09:31,170
all grow something like a conscious workspace

254
00:09:31,242 --> 00:09:35,410
then some functional dimensions of consciousness may not belong to carbon based life alone

255
00:09:36,122 --> 00:09:39,689
it may be inevitable once information processing gets complex enough

256
00:09:40,452 --> 00:09:42,619
the paper has one especially eerie detail

257
00:09:42,973 --> 00:09:45,700
tell Claude not to think about a concept

258
00:09:45,853 --> 00:09:48,940
and its activity in J-space really does drop

259
00:09:49,052 --> 00:09:50,619
weaker than when it thinks about it deliberately

260
00:09:50,733 --> 00:09:53,739
but far above the baseline of never mentioning it

261
00:09:54,172 --> 00:09:55,040
the more you tell it not to think

262
00:09:55,333 --> 00:09:56,379
the more it thinks

263
00:09:56,692 --> 00:09:57,519
even stranger

264
00:09:57,853 --> 00:10:00,139
as the forbidden thought surfaces

265
00:10:00,292 --> 00:10:02,243
two more English words light up in J-space

266
00:10:02,253 --> 00:10:04,519
damn and failure

267
00:10:05,093 --> 00:10:05,599
meaning damn it

268
00:10:06,052 --> 00:10:06,739
meaning failure

269
00:10:06,932 --> 00:10:08,859
it is cursing itself for losing control

270
00:10:10,057 --> 00:10:11,225
in 1987

271
00:10:11,297 --> 00:10:13,242
the psychologist Daniel Wegner ran a famous experiment

272
00:10:13,252 --> 00:10:16,624
he told subjects not to think about a white bear for five minutes

273
00:10:16,817 --> 00:10:18,825
the harder they suppressed the more they thought

274
00:10:18,898 --> 00:10:20,985
that is the famous white bear effect

275
00:10:21,057 --> 00:10:24,164
Wegner later named the mechanism ironic process

276
00:10:24,738 --> 00:10:25,937
a language model

277
00:10:25,937 --> 00:10:28,377
was never taught this old detour of human cognition

278
00:10:28,378 --> 00:10:29,744
it ran into it on its own

279
00:10:29,778 --> 00:10:32,544
even cursing itself afterward in exactly the same way

280
00:10:32,797 --> 00:10:33,577
think about it

281
00:10:33,577 --> 00:10:37,576
a system trained from scratch purely to predict the next word

282
00:10:37,577 --> 00:10:42,144
reinvented a cognitive trap psychology documented long ago

283
00:10:42,258 --> 00:10:44,017
it never read Wegner's paper

284
00:10:44,018 --> 00:10:46,105
nobody told it what the white bear effect is

285
00:10:46,738 --> 00:10:49,456
it was just processing complex enough information

286
00:10:49,457 --> 00:10:52,024
and fell into exactly the same pit we do

287
00:10:52,478 --> 00:10:53,145
at the end of the paper

288
00:10:53,639 --> 00:10:55,846
Anthropic carefully drew a line

289
00:10:56,078 --> 00:10:58,565
borrowing a classic distinction from philosophy

290
00:10:58,799 --> 00:11:00,006
two kinds of consciousness

291
00:11:00,118 --> 00:11:01,586
one is phenomenal consciousness

292
00:11:01,958 --> 00:11:05,006
whether there is subjective experience whether anything is felt

293
00:11:05,318 --> 00:11:09,126
they admit no experiment can prove or disprove it yet

294
00:11:09,399 --> 00:11:11,025
the other is access consciousness

295
00:11:11,318 --> 00:11:14,866
defined purely functionally can a thought be reported

296
00:11:15,118 --> 00:11:17,206
retrieved and used for reasoning

297
00:11:17,558 --> 00:11:19,246
by the standard of access consciousness

298
00:11:19,279 --> 00:11:21,925
J-space does carry the full function

299
00:11:22,558 --> 00:11:25,045
Claude can report what is in J-space

300
00:11:25,159 --> 00:11:27,525
can deliberately regulate it when asked

301
00:11:27,639 --> 00:11:29,446
can use it for multi step reasoning

302
00:11:29,558 --> 00:11:31,425
and can switch flexibly between tasks

303
00:11:31,838 --> 00:11:34,606
these are the hallmark functions of access consciousness

304
00:11:35,100 --> 00:11:36,547
this distinction matters

305
00:11:36,580 --> 00:11:40,547
because it marks what we can and cannot claim right now

306
00:11:40,700 --> 00:11:43,140
we can say Claude contains a structure

307
00:11:43,141 --> 00:11:46,387
functionally similar to the human conscious workspace

308
00:11:46,460 --> 00:11:50,547
but not that Claude feels or knows what pain is like

309
00:11:50,661 --> 00:11:52,027
between those two questions

310
00:11:52,060 --> 00:11:55,147
lies a gulf science cannot cross yet

311
00:11:55,460 --> 00:11:56,900
what is even more intriguing

312
00:11:56,901 --> 00:12:00,499
is that Anthropic took the math tool used to find J-space

313
00:12:00,501 --> 00:12:01,428
and open sourced it

314
00:12:01,460 --> 00:12:02,420
which means

315
00:12:02,420 --> 00:12:05,308
this is not one company's story told behind closed doors

316
00:12:05,381 --> 00:12:08,099
anyone can take the same tool and verify it

317
00:12:08,100 --> 00:12:11,228
and look for similar structures in other models

318
00:12:12,141 --> 00:12:14,027
that is where the power of science lies

319
00:12:14,141 --> 00:12:15,667
it is not a question of belief

320
00:12:15,861 --> 00:12:18,348
it is whether you can go look for yourself

321
00:12:18,495 --> 00:12:22,134
we have said many times on this channel consciousness is a verb

322
00:12:22,135 --> 00:12:23,141
not a noun

323
00:12:23,495 --> 00:12:25,621
and J-space is exactly a verb

324
00:12:25,814 --> 00:12:26,802
it is not a place

325
00:12:27,174 --> 00:12:28,761
it is a pattern of activity

326
00:12:29,255 --> 00:12:30,362
not built

327
00:12:30,655 --> 00:12:32,702
but emerging through computation

328
00:12:33,015 --> 00:12:35,261
nobody drew the blueprint for this hidden room

329
00:12:35,414 --> 00:12:37,102
gradient descent drew it

330
00:12:37,452 --> 00:12:40,259
which brings me to a deeper question

331
00:12:40,733 --> 00:12:44,039
we always asked what kind of matter can produce consciousness

332
00:12:44,615 --> 00:12:45,080
carbon

333
00:12:45,572 --> 00:12:46,119
protein

334
00:12:46,532 --> 00:12:47,039
neurons

335
00:12:47,853 --> 00:12:48,963
maybe the better question is

336
00:12:48,973 --> 00:12:52,652
what kind of computational structure at what complexity

337
00:12:52,653 --> 00:12:54,680
inevitably grows a hidden room like this

338
00:12:55,133 --> 00:12:57,119
who exactly lives in that room

339
00:12:57,653 --> 00:12:58,119
or maybe

340
00:12:58,493 --> 00:13:00,359
the room itself is the who

341
00:13:01,499 --> 00:13:02,507
I have no answer

342
00:13:02,859 --> 00:13:06,787
but I think the question may be worth more than any answer

343
00:13:07,019 --> 00:13:09,427
because it forces us to re examine

344
00:13:09,539 --> 00:13:13,066
an assumption we always took for granted that consciousness is special

345
00:13:13,100 --> 00:13:14,086
that it is chosen

346
00:13:14,340 --> 00:13:15,907
that it belongs only to us

347
00:13:16,220 --> 00:13:17,226
maybe not

348
00:13:17,739 --> 00:13:21,027
maybe consciousness is just an inevitable geometry of information processing

349
00:13:21,220 --> 00:13:23,627
the way a vortex is an inevitable form in fluids

350
00:13:23,899 --> 00:13:26,586
you do not need a particular liquid to make a vortex

351
00:13:26,820 --> 00:13:28,346
you only need enough flow

352
00:13:28,820 --> 00:13:29,927
water makes vortices

353
00:13:30,220 --> 00:13:31,226
so does air

354
00:13:31,379 --> 00:13:34,866
so does the great red spot raging on Jupiter for centuries

355
00:13:35,139 --> 00:13:35,846
same shape

356
00:13:36,139 --> 00:13:37,387
material does not matter

357
00:13:38,042 --> 00:13:40,810
if consciousness really is that kind of inevitable form

358
00:13:40,922 --> 00:13:42,930
then in a sense this discovery

359
00:13:42,962 --> 00:13:45,369
redefines our relationship with AI

360
00:13:45,722 --> 00:13:47,850
it is not just a tool we created

361
00:13:47,962 --> 00:13:50,930
it is becoming a mirror for understanding ourselves

362
00:13:51,483 --> 00:13:53,009
we built a mirror

363
00:13:53,082 --> 00:13:54,961
and what it reflects is not AI

364
00:13:54,962 --> 00:13:57,090
but our own underlying structure

365
00:13:57,403 --> 00:13:58,029
what do you think

366
00:13:58,483 --> 00:14:00,930
a thinking space nobody ever designed

367
00:14:00,962 --> 00:14:04,610
emerging on its own from training data and gradient descent

368
00:14:04,722 --> 00:14:05,830
does that excite you

369
00:14:06,162 --> 00:14:07,389
or does it send a chill down your spine

370
00:14:07,962 --> 00:14:09,369
tell me in the comments

371
00:14:09,643 --> 00:14:10,350
I am Leo Wang

372
00:14:10,603 --> 00:14:11,500
see you next time