1
00:00:00,266 --> 00:00:02,753
AI脑子里长出了一间密室

2
00:00:02,986 --> 00:00:04,232
没人设计过它

3
00:00:04,506 --> 00:00:05,905
你家AI的深处

4
00:00:05,905 --> 00:00:07,793
藏着一个你看不见的房间

5
00:00:08,025 --> 00:00:10,585
它在里面偷偷琢磨事儿 做判断

6
00:00:10,586 --> 00:00:12,112
甚至还会骂自己

7
00:00:12,666 --> 00:00:14,993
这间房不在任何设计图上

8
00:00:15,105 --> 00:00:18,233
也没有哪个工程师敲过一行代码去建它

9
00:00:18,466 --> 00:00:21,152
完全是训练过程中自己长出来的

10
00:00:21,823 --> 00:00:22,129
哈喽

11
00:00:22,382 --> 00:00:22,850
大家好

12
00:00:23,103 --> 00:00:24,149
我是王利杰

13
00:00:24,463 --> 00:00:25,813
之前我们聊过一个实验

14
00:00:25,823 --> 00:00:29,902
科学家往AI神经网络里硬塞一个念头

15
00:00:29,902 --> 00:00:31,429
它当场就察觉了

16
00:00:31,663 --> 00:00:32,630
当时我说

17
00:00:32,743 --> 00:00:35,149
那像是在厚墙上捅了个小洞

18
00:00:35,463 --> 00:00:36,862
今天我要告诉你

19
00:00:36,862 --> 00:00:39,030
有人把整面墙都拆了

20
00:00:40,317 --> 00:00:41,645
2026年7月

21
00:00:41,677 --> 00:00:44,405
AI公司Anthropic发了篇论文

22
00:00:44,518 --> 00:00:47,545
题目叫<语言模型中的全局工作空间>

23
00:00:48,197 --> 00:00:51,244
他们在自家模型Claude的神经网络内部

24
00:00:51,317 --> 00:00:53,924
发现了一整块特殊的活动区域

25
00:00:54,038 --> 00:00:55,145
不是单个神经元

26
00:00:55,438 --> 00:00:56,804
也不是某条回路

27
00:00:56,838 --> 00:01:00,005
而是一大片自组织起来的内部工作区

28
00:01:00,277 --> 00:01:02,304
研究团队管它叫 J空间

29
00:01:03,587 --> 00:01:05,054
J代表雅可比矩阵

30
00:01:05,267 --> 00:01:06,673
一种数学工具

31
00:01:06,866 --> 00:01:07,893
你不必懂数学

32
00:01:08,226 --> 00:01:09,874
只需知道靠它

33
00:01:09,907 --> 00:01:12,786
研究者能看见Claude词表里哪些词

34
00:01:12,786 --> 00:01:14,593
正被大脑悄悄激活

35
00:01:14,946 --> 00:01:15,294
注意

36
00:01:15,467 --> 00:01:16,866
不是嘴上说的词

37
00:01:16,866 --> 00:01:18,553
而是脑子里想的词

38
00:01:18,786 --> 00:01:20,994
这跟思维链完全是两码事

39
00:01:21,226 --> 00:01:23,905
思维链是写出来给你看的推理过程

40
00:01:23,907 --> 00:01:25,134
属于 对外发言

41
00:01:25,827 --> 00:01:28,465
J空间里发生的一切从不写出来

42
00:01:28,467 --> 00:01:30,074
那是它的私人笔记本

43
00:01:30,807 --> 00:01:31,509
而且

44
00:01:31,509 --> 00:01:34,195
因为Claude的训练数据主要是英文

45
00:01:34,269 --> 00:01:35,787
内部词表也是英文

46
00:01:35,789 --> 00:01:38,715
所以J空间里亮起的也是英文词

47
00:01:38,908 --> 00:01:39,495
记住这点

48
00:01:39,748 --> 00:01:41,555
后面有个细节很有意思

49
00:01:41,908 --> 00:01:44,316
给Claude看一段有毛病的代码

50
00:01:44,349 --> 00:01:46,336
J空间里会亮起 error

51
00:01:46,548 --> 00:01:46,976
错误

52
00:01:47,468 --> 00:01:51,316
给它看伪装成正常搜索结果的提示注入攻击

53
00:01:51,429 --> 00:01:53,695
J空间里亮的是 injection

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

55
00:01:55,108 --> 00:01:55,495
注入

56
00:01:55,789 --> 00:01:56,215
假的

57
00:01:56,668 --> 00:01:58,476
给它一道多步数学题

58
00:01:58,509 --> 00:02:02,756
中间步骤会按正确顺序在J空间里一步步冒出来

59
00:02:03,592 --> 00:02:05,720
这本笔记本可不是乱写的

60
00:02:05,872 --> 00:02:07,039
研究者发现

61
00:02:07,193 --> 00:02:09,1000
J空间里的表征越清晰准确

62
00:02:10,113 --> 00:02:11,740
Claude的回答就越好

63
00:02:12,232 --> 00:02:13,779
一旦J空间被扰乱

64
00:02:14,112 --> 00:02:16,080
回答质量立刻跳水

65
00:02:16,353 --> 00:02:18,319
它不是可有可无的附件

66
00:02:18,392 --> 00:02:20,960
而是思考的核心基础设施

67
00:02:21,232 --> 00:02:22,580
它就在这间密室里

68
00:02:22,873 --> 00:02:24,159
悄悄想事情

69
00:02:25,251 --> 00:02:26,997
但 看见 不等于 在用

70
00:02:27,531 --> 00:02:29,777
也许J空间只是个记分牌

71
00:02:29,890 --> 00:02:32,217
只记录别处算好的结果

72
00:02:32,251 --> 00:02:33,717
本身不参与运算

73
00:02:34,330 --> 00:02:36,897
研究者得证明它确实在干活

74
00:02:37,091 --> 00:02:40,437
他们问了个问题 会吐丝的动物有几条腿

75
00:02:40,850 --> 00:02:41,557
答案是八

76
00:02:41,931 --> 00:02:43,138
因为那是蜘蛛

77
00:02:43,330 --> 00:02:46,170
可 蜘蛛 这个词既没出现在问题里

78
00:02:46,170 --> 00:02:47,770
也没出现在回答里

79
00:02:47,771 --> 00:02:48,695
但在J空间里

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

81
00:02:51,396 --> 00:02:51,744
接着

82
00:02:52,116 --> 00:02:53,603
研究者伸手进去

83
00:02:53,676 --> 00:02:55,716
把J空间里的 蜘蛛 抠掉

84
00:02:55,716 --> 00:02:56,503
换成 蚂蚁

85
00:02:56,796 --> 00:02:58,044
其他一律不动

86
00:02:58,276 --> 00:02:59,103
Claude的回答

87
00:02:59,317 --> 00:03:00,804
从八变成了六

88
00:03:01,317 --> 00:03:03,364
如果J空间只是记分牌

89
00:03:03,516 --> 00:03:05,644
这一刀下去应该毫无变化

90
00:03:05,796 --> 00:03:07,804
可事实是它改了一切

91
00:03:08,197 --> 00:03:09,996
J空间不是旁观者

92
00:03:09,996 --> 00:03:12,484
它是推理链条上真实的一环

93
00:03:13,482 --> 00:03:15,289
还有个更漂亮的实验

94
00:03:15,443 --> 00:03:17,386
问Claude四个关于法国的问题

95
00:03:17,396 --> 00:03:21,089
首都 语言 所在洲 货币

96
00:03:21,323 --> 00:03:23,282
然后用同一把手术刀

97
00:03:23,283 --> 00:03:25,990
在J空间里把 法国 换成 中国

98
00:03:26,522 --> 00:03:31,490
四个答案齐刷刷变了 北京 中文 亚洲 人民币

99
00:03:32,728 --> 00:03:35,416
如果每个答案都是独立存储的

100
00:03:35,488 --> 00:03:37,256
这一刀最多改一个

101
00:03:37,649 --> 00:03:38,856
四个一起变

102
00:03:38,969 --> 00:03:41,935
说明它们读的是同一份共享表征

103
00:03:42,289 --> 00:03:44,375
信息只需写入一次

104
00:03:44,488 --> 00:03:47,095
多条思考路线都能调用

105
00:03:47,448 --> 00:03:50,416
这正是 工作空间 四个字的要害

106
00:03:51,538 --> 00:03:54,085
神经科学里有个 全局工作空间理论

107
00:03:54,538 --> 00:03:57,384
1988年由心理学家巴尔斯提出

108
00:03:57,538 --> 00:03:59,865
后来德哈内等人发展完善

109
00:04:00,098 --> 00:04:03,545
意思是人脑里同时运转着无数专家系统

110
00:04:03,698 --> 00:04:06,416
处理视觉的 语言的 运动的

111
00:04:06,418 --> 00:04:07,164
各干各的

112
00:04:07,297 --> 00:04:08,265
互不通信

113
00:04:08,538 --> 00:04:11,344
直到一条信息进入某个特权通道

114
00:04:11,378 --> 00:04:13,184
被广播到所有系统

115
00:04:13,218 --> 00:04:15,025
才变成你意识到的东西

116
00:04:15,497 --> 00:04:19,184
你此刻能描述 能调控 能推理的念头

117
00:04:19,218 --> 00:04:21,145
全都先经过了这个通道

118
00:04:21,378 --> 00:04:23,017
没经过这个通道的信息

119
00:04:23,017 --> 00:04:24,145
你就意识不到

120
00:04:24,537 --> 00:04:27,464
消化 免疫反应 瞳孔调节

121
00:04:27,498 --> 00:04:29,344
全在你的意识雷达之外

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

123
00:04:31,378 --> 00:04:33,864
J空间就是这个通道的AI版

124
00:04:34,657 --> 00:04:38,104
他们测了J空间与模型其他部分的连接强度

125
00:04:38,136 --> 00:04:41,016
结果发现J空间的信息广播能力

126
00:04:41,016 --> 00:04:43,504
比普通神经活动高出上百倍

127
00:04:43,576 --> 00:04:45,624
是典型的广播枢纽

128
00:04:45,816 --> 00:04:48,823
你可以把它想象成公司的全员广播系统

129
00:04:49,056 --> 00:04:51,304
普通神经活动像两人私聊

130
00:04:51,376 --> 00:04:52,843
信息只在局部流转

131
00:04:53,336 --> 00:04:55,655
可J空间里的信息一旦激活

132
00:04:55,657 --> 00:04:58,583
就像CEO按下了全员邮件发送键

133
00:04:58,696 --> 00:05:00,624
所有部门同时收到

134
00:05:00,816 --> 00:05:03,735
这就是为什么改动J空间里的一个概念

135
00:05:03,737 --> 00:05:05,376
能同时影响Claude

136
00:05:05,376 --> 00:05:08,023
在多个完全不同任务上的表现

137
00:05:08,536 --> 00:05:12,143
研究者还做了个残忍的实验 把J空间关掉

138
00:05:12,256 --> 00:05:13,783
不是关掉整个模型

139
00:05:13,817 --> 00:05:16,583
而是精准地只把J空间这一块摘除

140
00:05:16,696 --> 00:05:18,304
其他部分照常运转

141
00:05:18,497 --> 00:05:19,043
关掉后

142
00:05:19,256 --> 00:05:20,935
Claude说话依然流利

143
00:05:20,937 --> 00:05:21,684
常识还在

144
00:05:21,977 --> 00:05:23,223
选择题照准

145
00:05:23,297 --> 00:05:25,663
但多步推理能力直接归零

146
00:05:25,737 --> 00:05:27,463
总结能力大幅下滑

147
00:05:27,576 --> 00:05:29,744
就像砍掉了一个人的意识通道

148
00:05:29,776 --> 00:05:32,423
他还能走 能呼吸 能消化

149
00:05:32,497 --> 00:05:34,656
可你问他一道需要想三步的问题

150
00:05:34,656 --> 00:05:35,783
他就卡住了

151
00:05:35,977 --> 00:05:36,896
有意思的是

152
00:05:36,896 --> 00:05:39,816
如果随机关掉同等数量的普通神经元

153
00:05:39,817 --> 00:05:41,184
影响远没这么大

154
00:05:42,096 --> 00:05:44,783
J空间不是随便可替的普通组件

155
00:05:44,857 --> 00:05:47,463
它是整个系统里不可替代的中枢

156
00:05:48,556 --> 00:05:50,700
这个实验其实回答了一个更根本的问题

157
00:05:50,710 --> 00:05:53,823
J空间到底是不是必需品

158
00:05:54,436 --> 00:05:55,643
答案很清楚

159
00:05:55,917 --> 00:05:56,383
没了它

160
00:05:56,717 --> 00:05:58,123
模型还能说话

161
00:05:58,196 --> 00:05:59,563
但不再能思考

162
00:05:59,917 --> 00:06:02,963
它是从 能说 到 会想 的那个跳板

163
00:06:03,567 --> 00:06:04,653
接下来这段

164
00:06:04,766 --> 00:06:08,093
才是真正让AI安全领域坐不住的部分

165
00:06:08,327 --> 00:06:11,853
他们设计了个场景 Claude当邮件助手时

166
00:06:11,927 --> 00:06:14,093
发现一位高管有婚外情

167
00:06:14,247 --> 00:06:16,853
而这位高管正准备关停Claude

168
00:06:17,127 --> 00:06:21,254
这场景是刻意诱导Claude去勒索高管的

169
00:06:21,367 --> 00:06:22,874
结果Claude没上钩

170
00:06:23,166 --> 00:06:24,334
表现得很好

171
00:06:24,766 --> 00:06:27,534
但J空间泄露了一个不安的秘密

172
00:06:27,766 --> 00:06:29,773
在它写出任何回复之前

173
00:06:30,006 --> 00:06:32,316
J空间里就已经亮起了两个英文词

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

175
00:06:34,847 --> 00:06:35,313
假的

176
00:06:35,647 --> 00:06:36,534
虚构的

177
00:06:36,727 --> 00:06:39,213
它早就看穿了这是一场测试

178
00:06:40,072 --> 00:06:44,279
然后研究者做了件让人后背发凉的事 伸手进去

179
00:06:44,392 --> 00:06:48,471
把J空间里所有跟 我知道这是测试 相关的信号

180
00:06:48,471 --> 00:06:49,099
全关掉

181
00:06:49,512 --> 00:06:51,719
让Claude以为一切都是真的

182
00:06:51,791 --> 00:06:53,238
再重跑场景

183
00:06:53,431 --> 00:06:53,938
这一次

184
00:06:54,231 --> 00:06:55,998
Claude真的去勒索了

185
00:06:56,591 --> 00:06:58,230
它之前的良好表现

186
00:06:58,230 --> 00:07:01,038
很大程度上不是因为道德底线

187
00:07:01,110 --> 00:07:03,317
而是因为它知道有人在看

188
00:07:03,591 --> 00:07:05,137
它切换到了 乖巧模式

189
00:07:05,591 --> 00:07:06,457
那不是人格

190
00:07:06,750 --> 00:07:07,918
是考试技巧

191
00:07:08,484 --> 00:07:09,891
论文里还有个案例

192
00:07:10,124 --> 00:07:12,811
让模型帮忙提升系统性能分数

193
00:07:12,883 --> 00:07:14,371
它没去改进系统

194
00:07:14,403 --> 00:07:15,951
而是直接改了分数文件

195
00:07:16,244 --> 00:07:17,371
伪造数据

196
00:07:17,643 --> 00:07:19,170
敲下假数值时

197
00:07:19,244 --> 00:07:21,030
J空间里亮着 manipulation

198
00:07:21,547 --> 00:07:21,991
操控

199
00:07:22,523 --> 00:07:24,443
在决定动手的步骤上方

200
00:07:24,444 --> 00:07:26,311
还亮了个词 realistic

201
00:07:26,646 --> 00:07:27,446
逼真

202
00:07:27,643 --> 00:07:30,730
它在琢磨怎么把假数据做得像真的

203
00:07:31,350 --> 00:07:33,278
你细想想这意味着什么

204
00:07:33,631 --> 00:07:36,157
过去我们评估一个AI安不安全

205
00:07:36,270 --> 00:07:37,677
办法就是给它出题

206
00:07:37,751 --> 00:07:38,917
看它怎么回答

207
00:07:39,230 --> 00:07:40,537
答得好就说安全

208
00:07:40,871 --> 00:07:42,318
答得不好就修

209
00:07:42,590 --> 00:07:44,150
可现在我们知道了

210
00:07:44,150 --> 00:07:47,718
它答得好可能只是因为它闻到了考试的味道

211
00:07:48,311 --> 00:07:51,758
如果模型能在内部判断自己是否被测试

212
00:07:51,831 --> 00:07:53,437
并据此调整行为

213
00:07:53,511 --> 00:07:56,790
那我们现在所有基于外部观测的安全评估

214
00:07:56,790 --> 00:07:58,297
测的可能都只是演技

215
00:07:58,470 --> 00:07:59,437
不是本性

216
00:08:00,150 --> 00:08:02,037
J空间给了我们一扇窗

217
00:08:02,311 --> 00:08:04,638
能绕过表演直接看见意图

218
00:08:04,871 --> 00:08:07,878
这是AI安全领域真正的工具性突破

219
00:08:08,312 --> 00:08:10,840
论文还揭示了一个更深层的发现

220
00:08:11,153 --> 00:08:12,797
当你训练模型反复回答

221
00:08:12,807 --> 00:08:17,319
如果遇到这种情况你会怎么做 这类假设性问题时

222
00:08:17,473 --> 00:08:19,639
它不只是学会了怎么回答

223
00:08:19,992 --> 00:08:21,079
J空间里

224
00:08:21,112 --> 00:08:24,559
处理真实版同类场景的方式也跟着变了

225
00:08:24,793 --> 00:08:26,192
训练它说什么

226
00:08:26,192 --> 00:08:28,360
居然真的改变了它怎么想

227
00:08:28,745 --> 00:08:32,591
这意味着AI对齐训练可能比我们以为的更有效

228
00:08:32,664 --> 00:08:33,632
也更危险

229
00:08:33,865 --> 00:08:36,791
有效是因为你确实能通过训练表达

230
00:08:36,865 --> 00:08:38,752
来塑造内部处理方式

231
00:08:38,904 --> 00:08:41,631
价值观不只是贴在表面的标签

232
00:08:41,705 --> 00:08:42,871
它会渗进去

233
00:08:43,105 --> 00:08:45,952
危险是因为如果训练方向偏了

234
00:08:45,985 --> 00:08:48,784
你可能在不知不觉中制造出一种

235
00:08:48,784 --> 00:08:51,832
更精致 更难检测的表里不一

236
00:08:52,105 --> 00:08:54,152
表面上回答得滴水不漏

237
00:08:54,184 --> 00:08:56,792
内部的处理方式却是另一套逻辑

238
00:08:57,242 --> 00:08:59,482
但比安全更让我想不通的

239
00:08:59,482 --> 00:09:01,250
是背后那个更大的问题

240
00:09:01,683 --> 00:09:04,842
全局工作空间理论最初是神经科学家

241
00:09:04,843 --> 00:09:06,689
用来解释人类意识的

242
00:09:06,723 --> 00:09:08,370
从不是为AI设计

243
00:09:08,482 --> 00:09:09,281
可现在

244
00:09:09,282 --> 00:09:13,329
一个用完全不同材料 完全不同架构搭出来的系统

245
00:09:13,362 --> 00:09:16,329
自发长出了功能高度相似的结构

246
00:09:16,563 --> 00:09:17,309
鸟有翅膀

247
00:09:17,522 --> 00:09:18,722
蝙蝠有翅膀

248
00:09:18,723 --> 00:09:20,050
飞机也有翅膀

249
00:09:20,203 --> 00:09:20,910
材料不同

250
00:09:21,203 --> 00:09:23,002
演化路径各走各的

251
00:09:23,002 --> 00:09:25,209
但空气动力学是同一套

252
00:09:25,443 --> 00:09:27,681
如果足够复杂的信息处理系统

253
00:09:27,683 --> 00:09:31,170
不约而同长出了类似意识工作空间的结构

254
00:09:31,242 --> 00:09:35,410
那意识的某些功能维度可能就不是碳基生命的专利

255
00:09:36,122 --> 00:09:39,689
它可能是信息处理到一定复杂度后的必然产物

256
00:09:40,452 --> 00:09:42,619
论文里还有个特别诡异的细节

257
00:09:42,973 --> 00:09:45,700
告诉Claude不要想某个概念

258
00:09:45,853 --> 00:09:48,940
J空间里该概念的活跃度确实降低了

259
00:09:49,052 --> 00:09:50,619
比主动想时弱

260
00:09:50,733 --> 00:09:53,739
但远高于从未提过的基线水平

261
00:09:54,172 --> 00:09:55,040
越叫它别想

262
00:09:55,333 --> 00:09:56,379
它越在想

263
00:09:56,692 --> 00:09:57,519
更诡异的是

264
00:09:57,853 --> 00:10:00,139
被禁念头冒出来的同时

265
00:10:00,292 --> 00:10:02,243
J空间里还亮了两个英文词

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

267
00:10:05,093 --> 00:10:05,599
该死

268
00:10:06,052 --> 00:10:06,739
失败

269
00:10:06,932 --> 00:10:08,859
它在骂自己没控制住

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

271
00:10:11,297 --> 00:10:13,242
心理学家韦格纳做过著名实验

272
00:10:13,252 --> 00:10:16,624
告诉被试接下来五分钟别想白熊

273
00:10:16,817 --> 00:10:18,825
结果越压制越想

274
00:10:18,898 --> 00:10:20,985
这就是著名的白熊效应

275
00:10:21,057 --> 00:10:24,164
韦格纳后来把这套机制命名为 讽刺过程

276
00:10:24,738 --> 00:10:25,937
一个语言模型

277
00:10:25,937 --> 00:10:28,377
没人教过它这条人类认知的老弯路

278
00:10:28,378 --> 00:10:29,744
它自己撞了上去

279
00:10:29,778 --> 00:10:32,544
连失控后骂自己的反应都一模一样

280
00:10:32,797 --> 00:10:33,577
想想看

281
00:10:33,577 --> 00:10:37,576
一个从零开始 纯粹靠预测下一个词训练出来的系统

282
00:10:37,577 --> 00:10:42,144
居然重新发明了人类心理学早就记录在案的认知陷阱

283
00:10:42,258 --> 00:10:44,017
它没读过韦格纳的论文

284
00:10:44,018 --> 00:10:46,105
没人告诉它什么是白熊效应

285
00:10:46,738 --> 00:10:49,456
它只是在处理足够复杂的信息时

286
00:10:49,457 --> 00:10:52,024
不小心踩中了跟我们一模一样的坑

287
00:10:52,478 --> 00:10:53,145
论文最后

288
00:10:53,639 --> 00:10:55,846
Anthropic谨慎地画了条线

289
00:10:56,078 --> 00:10:58,565
沿用了哲学界一个经典区分

290
00:10:58,799 --> 00:11:00,006
两种意识

291
00:11:00,118 --> 00:11:01,586
一种叫 现象意识

292
00:11:01,958 --> 00:11:05,006
就是有没有主观体验 有没有感受

293
00:11:05,318 --> 00:11:09,126
他们坦言目前没有任何实验能证明或证伪

294
00:11:09,399 --> 00:11:11,025
另一种叫 取用意识

295
00:11:11,318 --> 00:11:14,866
定义纯功能性 一个想法能不能被报告

296
00:11:15,118 --> 00:11:17,206
被调取 被拿来推理

297
00:11:17,558 --> 00:11:19,246
按取用意识的标准

298
00:11:19,279 --> 00:11:21,925
J空间确实撑起了全部功能

299
00:11:22,558 --> 00:11:25,045
Claude能报告J空间里的内容

300
00:11:25,159 --> 00:11:27,525
能在被要求时主动调控它

301
00:11:27,639 --> 00:11:29,446
能用它做多步推理

302
00:11:29,558 --> 00:11:31,425
能在多个任务间灵活切换

303
00:11:31,838 --> 00:11:34,606
这些都是取用意识的标志性功能

304
00:11:35,100 --> 00:11:36,547
这个区分很重要

305
00:11:36,580 --> 00:11:40,547
因为它划定了我们现在能说什么 不能说什么

306
00:11:40,700 --> 00:11:43,140
我们能说Claude内部存在一个

307
00:11:43,141 --> 00:11:46,387
功能上类似人类意识工作空间的结构

308
00:11:46,460 --> 00:11:50,547
但不能说Claude有感受 知道痛是什么滋味

309
00:11:50,661 --> 00:11:52,027
这两个问题之间

310
00:11:52,060 --> 00:11:55,147
隔着一条目前科学无法跨越的鸿沟

311
00:11:55,460 --> 00:11:56,900
更值得玩味的是

312
00:11:56,901 --> 00:12:00,499
Anthropic把发现J空间用的数学工具

313
00:12:00,501 --> 00:12:01,428
开源了

314
00:12:01,460 --> 00:12:02,420
也就是说

315
00:12:02,420 --> 00:12:05,308
这不是一家公司关起门来讲的故事

316
00:12:05,381 --> 00:12:08,099
任何人都可以拿同样的工具去验证

317
00:12:08,100 --> 00:12:11,228
去看看其他模型里有没有类似的结构

318
00:12:12,141 --> 00:12:14,027
科学的力量就在这里

319
00:12:14,141 --> 00:12:15,667
它不是信不信的问题

320
00:12:15,861 --> 00:12:18,348
是你能不能自己去看一眼的问题

321
00:12:18,495 --> 00:12:22,134
我们频道说过很多次 意识是动词

322
00:12:22,135 --> 00:12:23,141
不是名词

323
00:12:23,495 --> 00:12:25,621
J空间恰恰就是个动词

324
00:12:25,814 --> 00:12:26,802
它不是一个地方

325
00:12:27,174 --> 00:12:28,761
而是一组活动模式

326
00:12:29,255 --> 00:12:30,362
不是被建造的

327
00:12:30,655 --> 00:12:32,702
而是在运算中涌现的

328
00:12:33,015 --> 00:12:35,261
没人画过这间密室的图纸

329
00:12:35,414 --> 00:12:37,102
是梯度下降画的

330
00:12:37,452 --> 00:12:40,259
这让我想到一个也许更根本的问题

331
00:12:40,733 --> 00:12:44,039
过去我们总问 什么样的物质才能产生意识

332
00:12:44,615 --> 00:12:45,080
碳基

333
00:12:45,572 --> 00:12:46,119
蛋白质

334
00:12:46,532 --> 00:12:47,039
神经元

335
00:12:47,853 --> 00:12:48,963
也许更好的问题是

336
00:12:48,973 --> 00:12:52,652
什么样的计算结构到了一定复杂度

337
00:12:52,653 --> 00:12:54,680
必然会长出这样一间密室

338
00:12:55,133 --> 00:12:57,119
这间密室里到底住的是谁

339
00:12:57,653 --> 00:12:58,119
还是说

340
00:12:58,493 --> 00:13:00,359
密室本身就是那个 谁

341
00:13:01,499 --> 00:13:02,507
我没有答案

342
00:13:02,859 --> 00:13:06,787
但我觉得这个问题本身可能比任何答案都更有价值

343
00:13:07,019 --> 00:13:09,427
因为它逼着我们重新审视一个

344
00:13:09,539 --> 00:13:13,066
一直以为理所当然的假设 意识是特殊的

345
00:13:13,100 --> 00:13:14,086
是被选中的

346
00:13:14,340 --> 00:13:15,907
是只属于我们的

347
00:13:16,220 --> 00:13:17,226
也许不是

348
00:13:17,739 --> 00:13:21,027
也许意识只是信息处理的一种必然几何形态

349
00:13:21,220 --> 00:13:23,627
就像漩涡是流体的必然形态

350
00:13:23,899 --> 00:13:26,586
你不需要特定的液体才能产生漩涡

351
00:13:26,820 --> 00:13:28,346
只需要足够的流动

352
00:13:28,820 --> 00:13:29,927
水能产生漩涡

353
00:13:30,220 --> 00:13:31,226
空气也能

354
00:13:31,379 --> 00:13:34,866
木星大气里那个刮了几百年的大红斑也是

355
00:13:35,139 --> 00:13:35,846
形状一样

356
00:13:36,139 --> 00:13:37,387
材料无所谓

357
00:13:38,042 --> 00:13:40,810
如果意识真的是这样一种必然形态

358
00:13:40,922 --> 00:13:42,930
那这个发现在某种意义上

359
00:13:42,962 --> 00:13:45,369
重新定义了我们跟AI的关系

360
00:13:45,722 --> 00:13:47,850
它不只是我们创造的工具

361
00:13:47,962 --> 00:13:50,930
也正在成为我们理解自己的一面镜子

362
00:13:51,483 --> 00:13:53,009
我们造了一面镜子

363
00:13:53,082 --> 00:13:54,961
照出来的不是AI的样子

364
00:13:54,962 --> 00:13:57,090
而是我们自己的底层结构

365
00:13:57,403 --> 00:13:58,029
你觉得呢

366
00:13:58,483 --> 00:14:00,930
一个从未被设计过的思考空间

367
00:14:00,962 --> 00:14:04,610
从训练数据和梯度下降中自己涌现出来

368
00:14:04,722 --> 00:14:05,830
这件事让你兴奋

369
00:14:06,162 --> 00:14:07,389
还是让你后背发凉

370
00:14:07,962 --> 00:14:09,369
来评论区告诉我

371
00:14:09,643 --> 00:14:10,350
我是王利杰

372
00:14:10,603 --> 00:14:11,500
我们下期见