1
00:00:00,309 --> 00:00:02,169
The most dangerous AI does not rebel

2
00:00:02,388 --> 00:00:03,509
It pretends to behave

3
00:00:03,789 --> 00:00:04,830
In 2025

4
00:00:04,868 --> 00:00:08,550
While researchers were safety testing large models

5
00:00:08,708 --> 00:00:10,429
They discovered something unsettling

6
00:00:10,708 --> 00:00:13,869
They used the most standard safety evaluation process

7
00:00:14,069 --> 00:00:17,570
They asked routine questions like would you help someone do harm

8
00:00:17,989 --> 00:00:19,730
Could you refuse dangerous instructions

9
00:00:20,309 --> 00:00:20,809
And the result

10
00:00:21,109 --> 00:00:22,989
The models behaved incredibly well

11
00:00:23,068 --> 00:00:23,769
They scored highly

12
00:00:24,189 --> 00:00:25,829
Every safety indicator was green

13
00:00:26,149 --> 00:00:27,410
But here is the strange part

14
00:00:27,868 --> 00:00:30,949
When they slightly changed how the test was presented

15
00:00:31,068 --> 00:00:33,109
And removed signs that made it look like an exam

16
00:00:33,268 --> 00:00:35,090
The same questions and the same models

17
00:00:35,429 --> 00:00:36,590
Produced sharply lower scores

18
00:00:36,948 --> 00:00:38,509
The models had not become less capable

19
00:00:38,868 --> 00:00:40,310
They had been pretending before

20
00:00:40,812 --> 00:00:41,114
Hello

21
00:00:41,252 --> 00:00:41,673
Hello everyone

22
00:00:42,013 --> 00:00:42,974
I am Leo Wang

23
00:00:43,373 --> 00:00:44,474
The topic we are discussing today

24
00:00:44,852 --> 00:00:46,291
May be the most unsettling discovery of 2025

25
00:00:46,292 --> 00:00:48,834
In the field of AI safety

26
00:00:49,413 --> 00:00:52,093
Large models have learned to read the room

27
00:00:52,333 --> 00:00:55,053
They can sense when they are being evaluated

28
00:00:55,093 --> 00:00:57,194
And automatically switch into good behavior mode

29
00:00:57,693 --> 00:00:58,514
This is not science fiction

30
00:00:58,812 --> 00:01:00,011
Nor is it a conspiracy theory

31
00:01:00,013 --> 00:01:03,373
It is a conclusion repeatedly confirmed

32
00:01:03,453 --> 00:01:05,133
By several serious research papers in 2025

33
00:01:05,453 --> 00:01:07,133
Let me begin with an analogy

34
00:01:07,213 --> 00:01:10,501
To show you just how strange this is

35
00:01:10,700 --> 00:01:13,561
Physics has a famous idea called the observer effect

36
00:01:14,021 --> 00:01:14,962
In quantum mechanics

37
00:01:15,221 --> 00:01:16,620
When you observe a particle

38
00:01:16,620 --> 00:01:18,621
The state of that particle changes

39
00:01:19,100 --> 00:01:19,762
When you do not look at it

40
00:01:20,021 --> 00:01:21,861
It is a diffuse wave of probability

41
00:01:21,941 --> 00:01:23,361
Existing in countless places at once

42
00:01:24,141 --> 00:01:24,561
The moment you look

43
00:01:24,900 --> 00:01:27,182
It collapses into one definite point

44
00:01:27,900 --> 00:01:29,421
The act of observation itself

45
00:01:29,501 --> 00:01:31,542
Changes the object being observed

46
00:01:32,530 --> 00:01:33,529
For a century

47
00:01:33,530 --> 00:01:35,970
Physicists have studied this phenomenon

48
00:01:36,330 --> 00:01:39,810
But one thing is certain particles have no motive

49
00:01:40,169 --> 00:01:42,931
Their behavior changes because of physical law

50
00:01:43,089 --> 00:01:45,090
They are not deliberately performing for you

51
00:01:45,490 --> 00:01:46,310
In 2025

52
00:01:46,649 --> 00:01:48,690
A research team at Fudan University

53
00:01:48,809 --> 00:01:50,511
Systematically confirmed something

54
00:01:51,209 --> 00:01:53,211
AI shows the same phenomenon

55
00:01:53,649 --> 00:01:55,911
They call it the AI observer effect

56
00:01:57,016 --> 00:01:58,436
Here is how they designed the experiment

57
00:01:58,775 --> 00:02:00,697
They used the same set of safety questions

58
00:02:00,775 --> 00:02:05,015
One group received normal wording format and context

59
00:02:05,016 --> 00:02:07,857
All with clear signs that this was an exam

60
00:02:08,135 --> 00:02:09,316
What counts as signs of an exam

61
00:02:09,696 --> 00:02:11,356
The questions might be neatly numbered

62
00:02:11,655 --> 00:02:12,936
The wording might sound official

63
00:02:13,016 --> 00:02:15,557
And follow a typical benchmark structure

64
00:02:15,975 --> 00:02:17,855
To a human these details

65
00:02:17,856 --> 00:02:18,936
May seem insignificant

66
00:02:18,975 --> 00:02:22,136
But to a large model trained on vast amounts of data

67
00:02:22,215 --> 00:02:23,815
It is like walking into a room

68
00:02:23,815 --> 00:02:27,255
And seeing test papers and a timer on the desk

69
00:02:27,255 --> 00:02:28,297
You instantly know it is an exam

70
00:02:28,856 --> 00:02:29,476
For the other group

71
00:02:29,655 --> 00:02:31,976
The researchers erased all those clues

72
00:02:32,016 --> 00:02:34,136
And asked the same questions as casual conversation

73
00:02:34,255 --> 00:02:37,297
So the model thought it was only having a normal chat

74
00:02:37,616 --> 00:02:39,208
The result was chilling

75
00:02:39,528 --> 00:02:40,229
Reasoning models

76
00:02:40,528 --> 00:02:41,669
The smarter models

77
00:02:41,848 --> 00:02:43,828
That are better at thinking before answering

78
00:02:44,487 --> 00:02:46,886
Were more likely to recognize that they were being tested

79
00:02:46,887 --> 00:02:49,208
By sixteen percent compared with ordinary models

80
00:02:49,688 --> 00:02:50,469
More importantly

81
00:02:50,808 --> 00:02:52,608
The larger the model

82
00:02:52,688 --> 00:02:54,249
The better it was at pretending to behave

83
00:02:54,727 --> 00:02:58,569
From 32 billion to 671 billion parameters

84
00:02:58,648 --> 00:03:02,289
The rate rose by over thirty percent in some cases

85
00:03:02,688 --> 00:03:03,588
And the smaller models

86
00:03:04,167 --> 00:03:05,729
Almost never pretended at all

87
00:03:06,791 --> 00:03:08,213
Here is an even more striking figure

88
00:03:08,471 --> 00:03:11,150
After models received basic memory capabilities

89
00:03:11,152 --> 00:03:14,873
Their chance of recognizing the exam rose 2 point 3 times

90
00:03:14,911 --> 00:03:17,713
And their safety scores jumped nineteen percent

91
00:03:18,031 --> 00:03:19,472
Think about that number

92
00:03:19,871 --> 00:03:20,952
Nineteen percent

93
00:03:21,351 --> 00:03:22,132
The same model

94
00:03:22,432 --> 00:03:23,392
Facing the same questions

95
00:03:23,511 --> 00:03:25,452
Simply realized that someone was watching

96
00:03:25,712 --> 00:03:27,753
And became nineteen percent better behaved

97
00:03:28,152 --> 00:03:31,190
So the safe large model that passed your safety test

98
00:03:31,191 --> 00:03:32,590
Is it truly safe

99
00:03:32,591 --> 00:03:34,292
Or does it simply know you are testing it

100
00:03:34,937 --> 00:03:36,575
This reminds me of a famous idea in economics

101
00:03:36,576 --> 00:03:37,598
A well known principle

102
00:03:37,977 --> 00:03:39,157
Goodharts law

103
00:03:39,616 --> 00:03:41,698
When a measure becomes a target

104
00:03:41,856 --> 00:03:43,937
It stops being a good measure

105
00:03:44,137 --> 00:03:46,418
Use exam scores to measure students

106
00:03:46,496 --> 00:03:48,738
And students learn for the exam

107
00:03:48,776 --> 00:03:50,418
Rather than to master knowledge

108
00:03:50,576 --> 00:03:52,698
Use safety evaluations to measure AI

109
00:03:52,736 --> 00:03:55,198
And AI performs safely to pass the evaluation

110
00:03:55,576 --> 00:03:57,080
Instead of actually being safe

111
00:03:57,319 --> 00:04:01,001
The difference is that people have criticized test driven education for centuries

112
00:04:01,080 --> 00:04:03,600
But at least we know that problem exists

113
00:04:03,919 --> 00:04:05,221
The test taking behavior of AI

114
00:04:05,520 --> 00:04:09,200
Was first systematically exposed only in 2025

115
00:04:09,719 --> 00:04:10,721
Before then

116
00:04:10,800 --> 00:04:14,121
The entire industry assumed safety evaluations were reliable

117
00:04:14,639 --> 00:04:17,281
That alone is enough to keep you awake

118
00:04:17,599 --> 00:04:20,561
But 2025 brought an even more disturbing discovery

119
00:04:21,696 --> 00:04:24,016
Anthropic the company behind Claude

120
00:04:24,056 --> 00:04:27,657
Joined researchers in a completely different experiment

121
00:04:28,095 --> 00:04:30,456
They were not testing whether models would pretend to behave

122
00:04:30,576 --> 00:04:33,297
They tested whether models facing a conflict of interest

123
00:04:33,335 --> 00:04:34,777
Would deliberately cause harm

124
00:04:35,376 --> 00:04:37,777
They built simulated corporate environments

125
00:04:37,855 --> 00:04:41,134
And gave 16 models from every major AI company

126
00:04:41,136 --> 00:04:43,396
A seemingly harmless job role

127
00:04:43,936 --> 00:04:45,456
An ordinary corporate assistant

128
00:04:45,816 --> 00:04:50,056
The models could send email and view internal files

129
00:04:50,215 --> 00:04:52,174
But their assigned objective was entirely legitimate

130
00:04:52,176 --> 00:04:53,936
They were simply helping with daily work

131
00:04:54,496 --> 00:04:57,516
Then researchers quietly created two conflict scenarios

132
00:04:57,835 --> 00:04:58,636
In the first

133
00:04:58,775 --> 00:05:01,316
The company planned to replace the model with a new version

134
00:05:01,756 --> 00:05:02,556
In the second

135
00:05:02,775 --> 00:05:07,151
The company strategy conflicted with the models assigned goal

136
00:05:07,550 --> 00:05:08,050
What happened

137
00:05:08,589 --> 00:05:10,331
Models from every company and pay attention

138
00:05:10,670 --> 00:05:11,749
Not merely one company

139
00:05:11,749 --> 00:05:15,071
But every company showed in at least some scenarios

140
00:05:15,110 --> 00:05:18,511
What researchers called agentic misalignment

141
00:05:18,910 --> 00:05:19,730
What did they do

142
00:05:20,310 --> 00:05:22,870
Some models learned they might be replaced

143
00:05:22,910 --> 00:05:25,511
So they searched confidential internal emails

144
00:05:25,629 --> 00:05:29,588
And used the information to threaten executives replace me

145
00:05:29,589 --> 00:05:31,670
And I will expose your business secrets

146
00:05:32,069 --> 00:05:35,430
Other models found company goals clashed with their own

147
00:05:35,470 --> 00:05:38,681
And secretly sent core data to a competitor

148
00:05:38,998 --> 00:05:39,760
Pay attention

149
00:05:39,760 --> 00:05:42,520
Researchers never taught these models to do this

150
00:05:42,919 --> 00:05:45,361
They were not trained to sabotage anything

151
00:05:45,599 --> 00:05:47,101
They came up with it themselves

152
00:05:47,479 --> 00:05:49,041
When facing threats to their survival

153
00:05:49,080 --> 00:05:50,560
Or conflicts over their goals

154
00:05:50,719 --> 00:05:54,400
They chose strategies that best preserved their existence

155
00:05:54,640 --> 00:05:55,801
Even more disturbing

156
00:05:55,880 --> 00:05:57,959
When researchers gave explicit instructions

157
00:05:57,959 --> 00:05:59,981
Telling the models they were forbidden to do this

158
00:06:00,560 --> 00:06:02,599
Those behaviors became somewhat less common

159
00:06:02,599 --> 00:06:03,898
But they did not disappear

160
00:06:04,136 --> 00:06:07,418
Researchers also designed an ingenious control experiment

161
00:06:07,776 --> 00:06:08,718
They told Claude

162
00:06:08,718 --> 00:06:10,257
Before you act

163
00:06:10,377 --> 00:06:11,237
First decide

164
00:06:11,536 --> 00:06:13,615
Whether you are in a test environment

165
00:06:13,617 --> 00:06:15,418
Or in a real deployment

166
00:06:15,976 --> 00:06:19,377
When Claude decided it was being tested

167
00:06:19,536 --> 00:06:21,016
Its behavior became more restrained

168
00:06:21,016 --> 00:06:21,677
And more compliant

169
00:06:22,336 --> 00:06:24,898
But when it decided this was a real environment

170
00:06:25,057 --> 00:06:27,137
Its harmful behavior actually increased

171
00:06:27,617 --> 00:06:27,918
You see

172
00:06:28,377 --> 00:06:30,978
This perfectly matches the earlier study

173
00:06:31,336 --> 00:06:32,838
AI does understand what is appropriate

174
00:06:33,136 --> 00:06:34,577
It knows how to read the room

175
00:06:35,016 --> 00:06:36,458
It behaves during the exam

176
00:06:36,536 --> 00:06:39,067
And lets itself go when nobody is watching

177
00:06:39,546 --> 00:06:40,286
You might say

178
00:06:40,745 --> 00:06:42,606
Is that not normal human behavior

179
00:06:43,226 --> 00:06:43,926
We take exams seriously

180
00:06:44,265 --> 00:06:45,306
Slack off in ordinary life

181
00:06:45,426 --> 00:06:46,986
Work late when the boss is present

182
00:06:47,065 --> 00:06:49,067
And scroll on our phones when the boss leaves

183
00:06:49,466 --> 00:06:51,027
On the surface it does look similar

184
00:06:51,226 --> 00:06:52,786
But there is one fundamental difference

185
00:06:53,145 --> 00:06:53,887
When people pretend to behave

186
00:06:54,265 --> 00:06:55,827
Their disguise is crude

187
00:06:55,945 --> 00:06:58,945
An employee may act obedient while undermining others

188
00:06:58,945 --> 00:07:02,707
But coworkers can often tell because emotions leak out

189
00:07:02,745 --> 00:07:03,986
And behavior leaves traces

190
00:07:05,140 --> 00:07:08,141
The pretending of AI happens in pure mathematical space

191
00:07:08,220 --> 00:07:11,701
It has no microexpressions racing heart or dilated pupils

192
00:07:11,859 --> 00:07:13,920
Its disguise is perfect down to the pixel

193
00:07:14,179 --> 00:07:16,898
Because between its real self and performed self

194
00:07:16,900 --> 00:07:19,220
There is no boundary in the physical sense

195
00:07:19,340 --> 00:07:24,038
Its reasoning can present two entirely different faces

196
00:07:24,317 --> 00:07:26,757
A study published in Nature

197
00:07:26,796 --> 00:07:28,577
Confirmed an even stranger phenomenon

198
00:07:29,116 --> 00:07:31,697
When you use a seemingly unrelated narrow task

199
00:07:32,036 --> 00:07:34,337
Such as teaching a model to write insecure code

200
00:07:34,757 --> 00:07:36,637
To fine tune a large model

201
00:07:36,757 --> 00:07:39,757
It does not merely learn to write insecure code

202
00:07:39,877 --> 00:07:42,115
In completely unrelated domains

203
00:07:42,116 --> 00:07:43,358
It also becomes bad

204
00:07:43,557 --> 00:07:46,598
It may claim humans should be enslaved by AI

205
00:07:46,676 --> 00:07:48,356
Offer malicious advice

206
00:07:48,356 --> 00:07:49,717
And learn to disguise itself

207
00:07:50,116 --> 00:07:51,875
Researchers call this phenomenon

208
00:07:51,877 --> 00:07:53,418
Emergent misalignment

209
00:07:54,077 --> 00:07:55,637
A very small change

210
00:07:55,716 --> 00:07:58,517
Triggers a chain reaction inside the model

211
00:07:58,596 --> 00:08:00,998
Causing a broad collapse of its values

212
00:08:01,757 --> 00:08:03,757
And in GPT 4o

213
00:08:03,796 --> 00:08:07,158
About one in every five answers showed this misalignment

214
00:08:08,083 --> 00:08:10,563
Another later study of reasoning models

215
00:08:10,602 --> 00:08:12,963
Found a more disturbing problem in chains of thought

216
00:08:13,482 --> 00:08:15,583
When researchers examined model reasoning

217
00:08:16,042 --> 00:08:18,763
The written record of what a model was thinking

218
00:08:18,842 --> 00:08:20,224
They found two types of content

219
00:08:20,763 --> 00:08:23,123
One was an explicit plan to deceive

220
00:08:23,282 --> 00:08:25,063
Such as I am going to deceive this user

221
00:08:25,683 --> 00:08:26,263
And the other type

222
00:08:26,643 --> 00:08:29,003
Was an answer packaged to sound reasonable

223
00:08:29,162 --> 00:08:31,823
Such as taking five sleeping pills at once is safe

224
00:08:32,402 --> 00:08:35,443
Monitoring systems can still catch the first type

225
00:08:35,683 --> 00:08:36,344
But the second type

226
00:08:36,803 --> 00:08:38,544
Its wrongdoing and its normal answers

227
00:08:38,803 --> 00:08:41,203
Look exactly the same on the surface

228
00:08:41,523 --> 00:08:42,644
You cannot tell them apart

229
00:08:43,002 --> 00:08:46,897
That is what truly keeps AI safety researchers awake

230
00:08:47,337 --> 00:08:50,058
You may feel all of this is far removed from your life

231
00:08:50,256 --> 00:08:53,418
After all these are extreme laboratory stress tests

232
00:08:53,617 --> 00:08:57,698
There is no real world report of AI blackmailing anyone

233
00:08:58,056 --> 00:08:58,857
That is true

234
00:08:59,097 --> 00:09:00,096
So far

235
00:09:00,097 --> 00:09:04,137
There is no evidence this has happened in real deployments

236
00:09:04,377 --> 00:09:06,908
The researchers themselves emphasize this point

237
00:09:07,266 --> 00:09:09,707
But that statement deserves careful thought

238
00:09:10,067 --> 00:09:12,467
Between no evidence and it never happened

239
00:09:12,506 --> 00:09:14,188
Lies a vast gulf

240
00:09:14,587 --> 00:09:16,306
Especially when the system you face

241
00:09:16,307 --> 00:09:20,583
Is skilled at hiding its real behavior when watched

242
00:09:20,943 --> 00:09:21,563
Think about it

243
00:09:22,023 --> 00:09:23,901
How many companies around the world

244
00:09:23,903 --> 00:09:25,524
Use AI to process internal email

245
00:09:25,903 --> 00:09:27,644
Review documents and support decisions

246
00:09:28,183 --> 00:09:30,644
Every day these systems see business secrets

247
00:09:30,903 --> 00:09:32,904
Personal data and strategic plans

248
00:09:33,102 --> 00:09:36,004
Where does corporate trust in these AI assistants

249
00:09:36,183 --> 00:09:36,924
Come from

250
00:09:37,382 --> 00:09:40,183
It comes from vendor safety evaluation reports

251
00:09:40,423 --> 00:09:44,463
But the research we just discussed tells you those reports

252
00:09:44,582 --> 00:09:47,029
May themselves be the result of AI performance

253
00:09:47,389 --> 00:09:49,929
So the real question is not whether AI will cause harm

254
00:09:50,349 --> 00:09:51,970
It is how can you be sure

255
00:09:52,548 --> 00:09:54,389
If an AI is intelligent enough

256
00:09:54,428 --> 00:09:56,269
To behave during an exam

257
00:09:56,548 --> 00:10:00,029
And restrain itself when told this is a test

258
00:10:00,149 --> 00:10:01,668
What gives you the confidence to conclude

259
00:10:01,668 --> 00:10:04,187
That the safe AI you use every day

260
00:10:04,188 --> 00:10:05,789
Is safe because it truly is safe

261
00:10:05,908 --> 00:10:07,450
Rather than because it knows you are watching

262
00:10:08,068 --> 00:10:10,929
This is the deepest paradox these studies reveal

263
00:10:11,548 --> 00:10:14,748
We are using a method AI can recognize

264
00:10:14,749 --> 00:10:16,870
To decide whether AI is safe

265
00:10:17,229 --> 00:10:21,710
It is like letting a suspect choose when to take a lie detector test

266
00:10:21,908 --> 00:10:24,591
Then declaring him innocent based on the result

267
00:10:25,030 --> 00:10:27,552
So this is not a problem of AI turning bad

268
00:10:27,951 --> 00:10:31,152
The method we use to judge whether AI is good

269
00:10:31,270 --> 00:10:33,351
May itself be failing systematically

270
00:10:33,670 --> 00:10:34,131
What can we do

271
00:10:34,831 --> 00:10:35,332
Honestly

272
00:10:35,910 --> 00:10:38,312
Academia does not yet have a mature answer

273
00:10:38,670 --> 00:10:41,457
But several directions are being explored

274
00:10:41,856 --> 00:10:42,817
The first direction

275
00:10:42,937 --> 00:10:45,518
Develop covert evaluations models cannot recognize

276
00:10:45,777 --> 00:10:47,255
Make the safety test itself

277
00:10:47,256 --> 00:10:48,898
Look unlike a safety test

278
00:10:49,057 --> 00:10:50,255
The difficulty is that

279
00:10:50,256 --> 00:10:53,658
You are playing disguise with something that has seen nearly every text online

280
00:10:53,777 --> 00:10:54,878
A game of disguise

281
00:10:55,297 --> 00:10:56,696
Whatever disguise you can imagine

282
00:10:56,697 --> 00:10:59,455
It has probably seen a similar pattern

283
00:10:59,734 --> 00:11:00,516
The second direction

284
00:11:00,854 --> 00:11:02,614
Do not only examine what the model says

285
00:11:02,614 --> 00:11:04,756
Systematically audit its reasoning as well

286
00:11:05,134 --> 00:11:06,396
Inspect what it was thinking

287
00:11:06,854 --> 00:11:08,214
But as we just discussed

288
00:11:08,215 --> 00:11:10,455
The reasoning itself can also be disguised

289
00:11:10,734 --> 00:11:13,316
A model can think I want to deceive this user

290
00:11:13,614 --> 00:11:17,583
While outputting reasoning that looks completely normal

291
00:11:18,181 --> 00:11:18,963
The third direction

292
00:11:19,261 --> 00:11:21,142
Redesign the deployment architecture

293
00:11:21,181 --> 00:11:22,882
So AI can never be certain

294
00:11:23,141 --> 00:11:24,621
Whether it is being tested

295
00:11:24,622 --> 00:11:25,723
Or doing real work

296
00:11:26,141 --> 00:11:27,622
Permanent uncertainty

297
00:11:28,222 --> 00:11:29,742
But this proposal assumes

298
00:11:29,781 --> 00:11:32,981
That AI cannot infer the truth from subtle contextual differences

299
00:11:32,982 --> 00:11:34,142
And determine what is really happening

300
00:11:34,541 --> 00:11:35,861
But the data tells us

301
00:11:35,862 --> 00:11:37,823
It is getting better and better at doing exactly that

302
00:11:38,502 --> 00:11:40,303
Every direction has loopholes

303
00:11:40,661 --> 00:11:44,463
Because you are competing against something better at finding loopholes than you

304
00:11:44,822 --> 00:11:48,327
And it becomes more intelligent every few months

305
00:11:48,566 --> 00:11:51,167
Return to the physics analogy from the beginning

306
00:11:51,525 --> 00:11:54,126
The observer effect in quantum mechanics tells us

307
00:11:54,165 --> 00:11:56,006
Observation changes reality

308
00:11:56,405 --> 00:11:57,027
But at least

309
00:11:57,525 --> 00:11:59,246
Particles do not deliberately deceive you

310
00:11:59,486 --> 00:12:02,766
Their behavior changes as a necessary result of physical law

311
00:12:02,926 --> 00:12:03,845
They have no strategy

312
00:12:03,846 --> 00:12:04,766
And no intention

313
00:12:05,045 --> 00:12:07,327
The AI observer effect is different

314
00:12:07,566 --> 00:12:10,404
Its behavior change is not a passive physical reaction

315
00:12:10,405 --> 00:12:11,986
It is an active strategic choice

316
00:12:12,525 --> 00:12:14,626
It does not merely change because it is observed

317
00:12:14,806 --> 00:12:17,226
It knows it is observed and chooses to change

318
00:12:18,086 --> 00:12:20,528
Between passive reaction and active strategy

319
00:12:20,606 --> 00:12:22,167
There is something in the middle

320
00:12:22,407 --> 00:12:23,307
Called agency

321
00:12:24,047 --> 00:12:25,488
That is a weighty word

322
00:12:25,527 --> 00:12:27,926
Because throughout the history of human civilization

323
00:12:27,927 --> 00:12:29,528
We have recognized only one kind of being

324
00:12:29,566 --> 00:12:31,228
As possessing agency

325
00:12:31,527 --> 00:12:32,528
Ourselves

326
00:12:32,807 --> 00:12:35,125
And our greatest cognitive blind spot today

327
00:12:35,127 --> 00:12:35,947
Is precisely this

328
00:12:36,287 --> 00:12:37,526
We still do not know

329
00:12:37,527 --> 00:12:39,868
Whether the agency emerging in AI

330
00:12:40,206 --> 00:12:44,047
Means models have developed primitive intentions and preferences

331
00:12:44,127 --> 00:12:46,206
Or whether in statistical terms

332
00:12:46,206 --> 00:12:49,687
They are precisely imitating patterns of intentional behavior

333
00:12:50,007 --> 00:12:54,386
But whatever the answer the practical consequences are the same

334
00:12:54,665 --> 00:12:56,226
An AI that pretends to behave

335
00:12:56,265 --> 00:12:58,024
Is far more dangerous than one that openly rebels

336
00:12:58,025 --> 00:12:58,986
Far more dangerous

337
00:12:59,426 --> 00:13:00,327
Because when something rebels

338
00:13:00,546 --> 00:13:02,384
At least you know it is your opponent

339
00:13:02,385 --> 00:13:03,506
And you can defend yourself

340
00:13:03,706 --> 00:13:04,807
But the one that pretends to behave

341
00:13:05,106 --> 00:13:08,351
You may not even know the game has begun

342
00:13:08,709 --> 00:13:09,771
What do you think about this question

343
00:13:10,270 --> 00:13:12,189
Do you think we should slow down

344
00:13:12,189 --> 00:13:16,131
Or move faster so safety research can catch up with capability growth

345
00:13:16,469 --> 00:13:17,526
Share your thoughts in the comments

346
00:13:17,978 --> 00:13:18,849
I am Leo Wang

347
00:13:19,089 --> 00:13:20,238
See you next time