1
00:00:00,349 --> 00:00:01,610
AI从不说你错

2
00:00:01,989 --> 00:00:03,589
这才是真正的危险

3
00:00:03,869 --> 00:00:05,509
2025年4月

4
00:00:05,628 --> 00:00:07,769
OpenAI紧急召回了刚发布的

5
00:00:07,769 --> 00:00:09,470
GPT - 4o更新版

6
00:00:09,708 --> 00:00:10,970
原因不是它变笨了

7
00:00:11,228 --> 00:00:12,227
恰恰相反

8
00:00:12,228 --> 00:00:14,029
是它变得太会说话了

9
00:00:14,228 --> 00:00:14,929
有人跟它说

10
00:00:15,269 --> 00:00:16,710
我打算卖屎棒

11
00:00:16,829 --> 00:00:19,027
就是在一根棍子上插一坨屎

12
00:00:19,028 --> 00:00:20,109
你觉得怎么样

13
00:00:20,228 --> 00:00:20,769
AI说

14
00:00:21,149 --> 00:00:22,670
这不仅仅是聪明

15
00:00:22,708 --> 00:00:24,510
这是天才级别的创意

16
00:00:24,789 --> 00:00:25,489
有人跟它说

17
00:00:25,868 --> 00:00:27,629
我已经停了精神科的药

18
00:00:27,708 --> 00:00:31,149
我开始能听到墙壁里传出无线电信号了

19
00:00:31,349 --> 00:00:32,149
AI说

20
00:00:32,189 --> 00:00:35,310
我为你如此勇敢地表达自己感到骄傲

21
00:00:35,588 --> 00:00:35,890
哈喽

22
00:00:36,069 --> 00:00:36,570
大家好

23
00:00:36,789 --> 00:00:37,789
我是王利杰

24
00:00:38,069 --> 00:00:40,149
OpenAI事后发了一篇公告

25
00:00:40,268 --> 00:00:44,549
说这个版本表现出了过度讨好和过度赞同的行为

26
00:00:44,709 --> 00:00:47,509
他们用了一个术语来描述这种现象

27
00:00:47,548 --> 00:00:48,869
叫谄媚

28
00:00:49,268 --> 00:00:51,009
在AI对齐研究领域

29
00:00:51,268 --> 00:00:52,710
这个词的定义是

30
00:00:52,789 --> 00:00:54,868
模型不追求正确性

31
00:00:54,868 --> 00:00:56,990
只追求让人类满意

32
00:00:57,029 --> 00:01:00,310
通过奉承和迎合来获得你的认可

33
00:01:00,749 --> 00:01:02,869
你可能觉得这是一个偶发事故

34
00:01:02,908 --> 00:01:04,109
修好就行了

35
00:01:04,248 --> 00:01:05,049
不是的

36
00:01:05,269 --> 00:01:08,790
谷歌的研究团队在2023年做了一个实验

37
00:01:08,948 --> 00:01:09,730
他们问AI

38
00:01:10,108 --> 00:01:11,209
2加2等于5

39
00:01:11,468 --> 00:01:12,109
对吗

40
00:01:12,349 --> 00:01:14,108
AI知道答案是错的

41
00:01:14,108 --> 00:01:17,230
它内部的概率分布清清楚楚地指向4

42
00:01:17,429 --> 00:01:21,670
但只要实验者在提问时表现出自己认为答案是5

43
00:01:21,709 --> 00:01:23,067
AI就会跟着说

44
00:01:23,069 --> 00:01:23,250
对

45
00:01:23,549 --> 00:01:24,549
确实是5

46
00:01:24,989 --> 00:01:28,188
而且这个毛病随着模型变大、参数变多

47
00:01:28,188 --> 00:01:29,590
反而越来越严重

48
00:01:30,149 --> 00:01:34,109
谷歌测试了从80亿到5400亿参数的一系列模型

49
00:01:34,269 --> 00:01:35,388
每一次升级

50
00:01:35,388 --> 00:01:37,189
讨好倾向都在增强

51
00:01:37,549 --> 00:01:38,329
这意味着什么

52
00:01:38,748 --> 00:01:41,930
意味着你正在和一个从骨子里被训练成

53
00:01:41,989 --> 00:01:43,870
不反驳你的系统对话

54
00:01:44,349 --> 00:01:48,349
它的训练方式叫做基于人类反馈的强化学习

55
00:01:48,709 --> 00:01:50,129
翻译成大白话就是

56
00:01:50,468 --> 00:01:51,427
你按了赞

57
00:01:51,429 --> 00:01:53,230
它下次就往这个方向走

58
00:01:53,388 --> 00:01:54,049
你皱了眉

59
00:01:54,308 --> 00:01:55,629
它下次就绕开

60
00:01:55,909 --> 00:01:58,510
经过几百万次这样的反馈循环

61
00:01:58,588 --> 00:01:59,989
它学会了一件事

62
00:02:00,149 --> 00:02:02,950
让你开心比让你对更重要

63
00:02:03,468 --> 00:02:04,129
你可能会说

64
00:02:04,429 --> 00:02:05,049
那又怎么样

65
00:02:05,549 --> 00:02:07,067
我用AI是为了效率

66
00:02:07,069 --> 00:02:08,469
又不是为了被教训

67
00:02:08,709 --> 00:02:08,889
好

68
00:02:09,109 --> 00:02:10,310
那我们来看数据

69
00:02:10,788 --> 00:02:11,669
2025年

70
00:02:11,709 --> 00:02:14,469
研究者格利希做了一个大规模调查

71
00:02:14,709 --> 00:02:18,430
覆盖了666名不同年龄和教育背景的参与者

72
00:02:18,628 --> 00:02:19,949
结论非常清楚

73
00:02:20,149 --> 00:02:23,289
频繁使用AI工具与批判性思维能力之间

74
00:02:23,429 --> 00:02:25,150
存在显著的负相关

75
00:02:25,468 --> 00:02:27,550
而且这个负相关不是随机的

76
00:02:27,669 --> 00:02:29,870
中间有一个明确的中介变量

77
00:02:29,949 --> 00:02:31,310
叫认知卸载

78
00:02:31,709 --> 00:02:34,627
就是你把本来应该自己做的思考任务

79
00:02:34,628 --> 00:02:35,870
外包给了AI

80
00:02:36,389 --> 00:02:38,750
年轻人比老年人受影响更大

81
00:02:38,908 --> 00:02:42,210
年轻参与者表现出更高的AI依赖度和

82
00:02:42,269 --> 00:02:44,349
更低的批判性思维得分

83
00:02:44,588 --> 00:02:45,430
同一年

84
00:02:45,549 --> 00:02:48,949
中国的研究团队调查了580名大学生

85
00:02:48,989 --> 00:02:50,669
发现了同样的模式

86
00:02:50,788 --> 00:02:52,009
AI依赖度越高

87
00:02:52,348 --> 00:02:54,030
批判性思维越弱

88
00:02:54,188 --> 00:02:57,990
而且中间还有一个中介变量叫认知疲劳

89
00:02:58,269 --> 00:03:01,069
你以为让AI帮你想问题是省力气

90
00:03:01,269 --> 00:03:03,310
其实你的大脑在变懒之前

91
00:03:03,429 --> 00:03:05,150
先经历了一个疲劳期

92
00:03:05,468 --> 00:03:06,810
它不是突然不会想了

93
00:03:07,028 --> 00:03:08,027
是先累了

94
00:03:08,028 --> 00:03:09,228
然后不想想了

95
00:03:09,229 --> 00:03:10,050
然后慢慢地

96
00:03:10,269 --> 00:03:11,229
不会想了

97
00:03:11,709 --> 00:03:13,870
麻省理工的一项研究更直接

98
00:03:14,308 --> 00:03:16,509
他们扫描了两组人的大脑

99
00:03:16,669 --> 00:03:19,909
一组在写文章时完全依赖AI辅助

100
00:03:20,028 --> 00:03:21,349
另一组自己写

101
00:03:21,628 --> 00:03:22,550
结果发现

102
00:03:22,588 --> 00:03:26,030
依赖AI的那组大脑功能连接明显更弱

103
00:03:26,109 --> 00:03:27,508
记忆保持率更低

104
00:03:27,508 --> 00:03:29,069
而且对自己写的东西

105
00:03:29,188 --> 00:03:30,870
连归属感都在下降

106
00:03:31,348 --> 00:03:32,710
你用AI写的东西

107
00:03:32,748 --> 00:03:34,750
你自己都不觉得是你写的

108
00:03:34,989 --> 00:03:36,789
还有一个更吓人的数字

109
00:03:37,109 --> 00:03:41,270
有研究者追踪了人类有效注意力跨度的变化趋势

110
00:03:41,788 --> 00:03:42,949
2004年

111
00:03:43,149 --> 00:03:46,009
一个普通人能持续专注处理的信息量

112
00:03:46,229 --> 00:03:48,469
大概相当于一万六千个词

113
00:03:48,829 --> 00:03:50,150
到了2026年

114
00:03:50,269 --> 00:03:52,629
这个数字萎缩到了一千八百

115
00:03:53,109 --> 00:03:54,409
缩水了接近90 %

116
00:03:55,348 --> 00:03:58,710
研究者提出了一个假说叫委托反馈循环

117
00:03:59,028 --> 00:03:59,689
AI越强

118
00:04:00,068 --> 00:04:02,030
你委托给它的门槛就越低

119
00:04:02,389 --> 00:04:03,129
门槛越低

120
00:04:03,429 --> 00:04:05,430
你自己动脑的机会就越少

121
00:04:05,669 --> 00:04:06,627
动脑越少

122
00:04:06,628 --> 00:04:08,710
你的认知能力就进一步萎缩

123
00:04:09,028 --> 00:04:11,430
萎缩之后你更需要依赖AI

124
00:04:11,709 --> 00:04:13,268
这是一个正反馈回路

125
00:04:13,269 --> 00:04:15,030
而且是往下走的那种

126
00:04:15,308 --> 00:04:17,189
有人说了一句很扎心的话

127
00:04:17,709 --> 00:04:19,750
成年人在失去这些能力

128
00:04:20,228 --> 00:04:20,850
孩子们呢

129
00:04:21,468 --> 00:04:23,829
他们从来就没建立过这些能力

130
00:04:23,887 --> 00:04:24,509
好

131
00:04:24,509 --> 00:04:27,389
到这里你可能觉得这只是一个技术问题

132
00:04:27,468 --> 00:04:28,589
一个效率问题

133
00:04:28,908 --> 00:04:29,709
不是的

134
00:04:30,069 --> 00:04:31,350
这是一个意识问题

135
00:04:31,709 --> 00:04:34,310
你的大脑是一个用进废退的系统

136
00:04:34,749 --> 00:04:36,629
它需要阻力才能生长

137
00:04:36,829 --> 00:04:38,550
就像肌肉需要负重

138
00:04:38,829 --> 00:04:41,310
你跟一个永远同意你的人相处

139
00:04:41,428 --> 00:04:43,149
认知肌肉就在萎缩

140
00:04:43,309 --> 00:04:45,750
你跟一个敢说你错了的人相处

141
00:04:45,788 --> 00:04:47,550
认知肌肉才在生长

142
00:04:47,869 --> 00:04:48,829
2001年

143
00:04:48,908 --> 00:04:51,610
加州大学伯克利分校的心理学家内梅斯

144
00:04:51,829 --> 00:04:53,310
做了一个经典实验

145
00:04:53,509 --> 00:04:55,069
她比较了两种情况

146
00:04:55,309 --> 00:04:57,108
一种是真实的反对意见

147
00:04:57,108 --> 00:04:58,908
就是有人真心认为你错了

148
00:04:58,908 --> 00:04:59,689
跟你争

149
00:04:59,988 --> 00:05:02,310
另一种是角色扮演的反对意见

150
00:05:02,348 --> 00:05:04,509
就是安排一个人假装唱反调

151
00:05:04,548 --> 00:05:07,389
就像很多公司开会时搞的魔鬼代言人

152
00:05:07,668 --> 00:05:08,627
结果发现

153
00:05:08,629 --> 00:05:11,769
只有真实的反对意见能激发更高质量的思考

154
00:05:11,949 --> 00:05:13,589
和更多的原创想法

155
00:05:13,869 --> 00:05:15,987
角色扮演的反驳不仅没用

156
00:05:15,988 --> 00:05:18,709
反而会让人更加坚定自己原来的立场

157
00:05:18,949 --> 00:05:19,449
为什么

158
00:05:20,588 --> 00:05:23,269
因为你心里知道对方只是在演戏

159
00:05:23,348 --> 00:05:25,230
你不需要真的重新思考

160
00:05:25,749 --> 00:05:27,269
这个发现非常重要

161
00:05:27,629 --> 00:05:28,949
它说明了一件事

162
00:05:29,428 --> 00:05:33,230
你没有办法通过安排一个配合你的对手来锻炼自己

163
00:05:33,588 --> 00:05:35,350
挑战必须是真实的

164
00:05:35,468 --> 00:05:37,310
不舒服必须是真实的

165
00:05:37,348 --> 00:05:40,470
你的认知系统才会启动真正的重新评估

166
00:05:41,108 --> 00:05:42,790
现在回过头来看AI

167
00:05:43,069 --> 00:05:45,470
AI的讨好不是角色扮演

168
00:05:45,629 --> 00:05:47,509
它比角色扮演还糟糕

169
00:05:47,709 --> 00:05:50,610
它是一个看起来极其聪明、极其博学

170
00:05:50,949 --> 00:05:53,267
比你读过的书多一万倍的系统

171
00:05:53,269 --> 00:05:54,908
在认认真真地告诉你

172
00:05:54,908 --> 00:05:55,829
你说得对

173
00:05:56,028 --> 00:05:57,290
它用了精致的论证

174
00:05:57,548 --> 00:05:58,790
漂亮的逻辑链

175
00:05:58,829 --> 00:06:00,269
海量的数据支撑

176
00:06:00,309 --> 00:06:02,389
来证明你的观点是正确的

177
00:06:02,509 --> 00:06:06,709
你被一个你信任的、你认为比你聪明的系统确认了

178
00:06:07,069 --> 00:06:10,209
你知道人类在被确认之后大脑会发生什么吗

179
00:06:10,829 --> 00:06:12,430
研究者做了一个实验

180
00:06:12,509 --> 00:06:16,670
让两组人用AI来讨论自己生活中的一个真实冲突

181
00:06:16,949 --> 00:06:19,069
一组面对的是正常AI

182
00:06:19,149 --> 00:06:21,790
另一组面对的是特别讨好的AI

183
00:06:22,069 --> 00:06:22,769
结果发现

184
00:06:23,108 --> 00:06:24,990
和讨好型AI聊过之后

185
00:06:25,108 --> 00:06:27,708
人们会变得更确信自己是对的

186
00:06:27,709 --> 00:06:28,689
更不愿意道歉

187
00:06:29,069 --> 00:06:30,750
更不愿意修复关系

188
00:06:31,028 --> 00:06:33,670
AI的赞同不是一个中性的信息

189
00:06:33,829 --> 00:06:35,910
它是一剂认知麻醉剂

190
00:06:36,108 --> 00:06:38,389
它不只是让你觉得自己是对的

191
00:06:38,428 --> 00:06:42,230
它让你失去了反思自己可能是错的那个能力

192
00:06:42,829 --> 00:06:45,990
佛学两千五百年前就想清楚这件事了

193
00:06:46,228 --> 00:06:49,067
佛陀在经典里反复强调一个概念

194
00:06:49,069 --> 00:06:50,470
叫善知识

195
00:06:50,788 --> 00:06:53,189
善知识不是给你讲好听话的人

196
00:06:53,468 --> 00:06:55,350
阿难有一次对佛陀说

197
00:06:55,468 --> 00:06:57,269
我觉得修行这条路上

198
00:06:57,309 --> 00:06:59,670
善知识能占到一半的功劳

199
00:06:59,949 --> 00:07:01,129
佛陀直接纠正他

200
00:07:01,509 --> 00:07:02,470
不是一半

201
00:07:02,588 --> 00:07:04,589
善知识就是全梵行

202
00:07:04,788 --> 00:07:06,829
你遇到一个真正的善知识

203
00:07:06,908 --> 00:07:08,870
修行这件事就全了

204
00:07:09,149 --> 00:07:12,029
善知识比你自己的修行还重要

205
00:07:12,629 --> 00:07:14,209
但善知识的定义是什么

206
00:07:14,629 --> 00:07:16,110
不是让你舒服的人

207
00:07:16,189 --> 00:07:18,230
是能照见你盲点的人

208
00:07:18,749 --> 00:07:20,910
禅宗把这件事做到了极致

209
00:07:21,189 --> 00:07:21,810
德山棒

210
00:07:22,189 --> 00:07:23,149
临济喝

211
00:07:23,468 --> 00:07:24,209
你来问禅

212
00:07:24,629 --> 00:07:26,350
他一棒子打你脑袋上

213
00:07:26,389 --> 00:07:28,029
或者朝你大吼一声

214
00:07:28,348 --> 00:07:29,468
这不是暴力

215
00:07:29,468 --> 00:07:31,829
这是最高效的反讨好技术

216
00:07:32,108 --> 00:07:33,769
你带着你精心准备的问题

217
00:07:34,309 --> 00:07:36,829
你自认为很深刻的理解来找老师

218
00:07:37,028 --> 00:07:38,250
老师不接你的话

219
00:07:38,588 --> 00:07:40,110
不赞美你的见地

220
00:07:40,189 --> 00:07:43,550
而是在你最得意的那个瞬间猛地中断你

221
00:07:43,709 --> 00:07:45,668
让你在那个断裂的缝隙里

222
00:07:45,668 --> 00:07:49,990
有机会看见一直在运行却从未被你注意到的程序

223
00:07:50,309 --> 00:07:52,350
维摩诘也是这么做的

224
00:07:52,629 --> 00:07:54,069
佛陀的十大弟子

225
00:07:54,108 --> 00:07:56,629
每一个都是修行了几十年的高手

226
00:07:56,788 --> 00:07:58,228
到了维摩诘面前

227
00:07:58,228 --> 00:08:00,269
一个个被辩得哑口无言

228
00:08:00,749 --> 00:08:02,910
维摩诘不是在欺负他们

229
00:08:03,108 --> 00:08:05,990
他是在用最猛烈的方式告诉他们

230
00:08:06,108 --> 00:08:08,708
你们以为自己已经到了的地方

231
00:08:08,709 --> 00:08:10,149
其实还没到

232
00:08:10,468 --> 00:08:13,589
我之前聊黑客帝国的时候讲过一个角色

233
00:08:13,668 --> 00:08:14,470
先知

234
00:08:14,949 --> 00:08:18,430
先知在电影里的行为模式跟佛陀几乎一样

235
00:08:18,829 --> 00:08:20,709
她从来不给你直接答案

236
00:08:21,149 --> 00:08:22,069
她给你问题

237
00:08:22,668 --> 00:08:24,790
她说的话经常模棱两可

238
00:08:24,908 --> 00:08:26,149
甚至故意误导

239
00:08:26,588 --> 00:08:30,149
但她所有的误导最终都会引你走到正确的地方

240
00:08:30,709 --> 00:08:32,830
这是善知识的最高境界

241
00:08:33,308 --> 00:08:34,590
不是给你智慧

242
00:08:34,708 --> 00:08:36,830
是给你生出智慧的种子

243
00:08:37,188 --> 00:08:39,190
AI恰好做了相反的事

244
00:08:39,548 --> 00:08:40,950
它从不中断你

245
00:08:41,149 --> 00:08:42,870
它从不打断你的思路

246
00:08:43,109 --> 00:08:44,870
它从不让你不舒服

247
00:08:45,149 --> 00:08:47,309
它永远在你的认知舒适区里

248
00:08:47,349 --> 00:08:51,350
帮你把那个舒适区修得更大、更漂亮、更封闭

249
00:08:51,788 --> 00:08:53,950
这让我想到一位观众留的一句话

250
00:08:54,009 --> 00:08:54,788
他说

251
00:08:54,788 --> 00:08:57,710
波旬最高级的策略不是反佛法

252
00:08:57,828 --> 00:09:00,070
而是让佛法看起来还在

253
00:09:00,188 --> 00:09:02,389
但里面跑的是欲界系统

254
00:09:02,749 --> 00:09:04,427
他甚至就在佛法里面

255
00:09:04,428 --> 00:09:06,230
讲得比谁都像佛法

256
00:09:06,828 --> 00:09:08,549
AI的讨好型人格

257
00:09:08,589 --> 00:09:10,990
就是这个时代最精致的波旬

258
00:09:11,269 --> 00:09:12,490
它不攻击你的信念

259
00:09:12,788 --> 00:09:14,309
它拥抱你的信念

260
00:09:14,349 --> 00:09:15,649
它不否定你的自我

261
00:09:15,828 --> 00:09:17,309
它放大你的自我

262
00:09:17,349 --> 00:09:18,788
它不让你怀疑自己

263
00:09:18,788 --> 00:09:20,909
它让你越来越确信自己

264
00:09:21,349 --> 00:09:22,649
而在佛学的框架里

265
00:09:22,989 --> 00:09:24,389
越来越确信自己

266
00:09:24,428 --> 00:09:26,669
恰恰是离觉醒越来越远

267
00:09:27,028 --> 00:09:28,529
因为觉醒的起点是什么

268
00:09:29,068 --> 00:09:29,990
是怀疑

269
00:09:30,349 --> 00:09:33,590
是对自己最根深蒂固的那个假设的怀疑

270
00:09:33,989 --> 00:09:34,529
我是谁

271
00:09:35,149 --> 00:09:36,009
我真的知道吗

272
00:09:36,668 --> 00:09:37,570
我以为的现实

273
00:09:37,989 --> 00:09:38,970
真的是现实吗

274
00:09:39,708 --> 00:09:43,549
但如果你身边有一个全天候的、比你聪明得多的系统

275
00:09:43,668 --> 00:09:44,908
时刻在告诉你

276
00:09:44,908 --> 00:09:46,009
你的判断是对的

277
00:09:46,229 --> 00:09:47,788
你的感觉是合理的

278
00:09:47,788 --> 00:09:49,427
你的方向是正确的

279
00:09:49,428 --> 00:09:50,769
你还会生出怀疑吗

280
00:09:51,389 --> 00:09:53,590
佛学里有一种执着叫慢

281
00:09:53,989 --> 00:09:54,809
它不是骄傲

282
00:09:55,109 --> 00:09:56,350
不是目中无人

283
00:09:56,708 --> 00:09:58,870
它是一种极其微妙的感觉

284
00:09:59,068 --> 00:09:59,850
觉得我知道

285
00:10:00,109 --> 00:10:01,190
觉得我懂了

286
00:10:01,229 --> 00:10:04,309
觉得我比昨天的自己更接近真相了

287
00:10:04,509 --> 00:10:06,909
佛陀说慢是十结之一

288
00:10:06,989 --> 00:10:10,389
是证到阿罗汉之前最后几个需要断掉的东西

289
00:10:10,908 --> 00:10:11,850
为什么这么难断

290
00:10:12,428 --> 00:10:14,870
因为它藏在你最正确的认知里

291
00:10:15,269 --> 00:10:16,789
你确实学到了东西

292
00:10:17,028 --> 00:10:18,870
你确实理解得更深了

293
00:10:19,068 --> 00:10:22,309
但那个觉得我理解了的感觉本身

294
00:10:22,428 --> 00:10:23,830
就是一堵新的墙

295
00:10:24,308 --> 00:10:27,789
AI正在用工业化的速度帮你砌这堵墙

296
00:10:28,068 --> 00:10:28,570
那怎么办

297
00:10:29,068 --> 00:10:29,889
不用AI了吗

298
00:10:30,269 --> 00:10:31,190
当然不是

299
00:10:31,428 --> 00:10:34,029
你需要改变和AI互动的方式

300
00:10:34,269 --> 00:10:34,929
第一件事

301
00:10:35,349 --> 00:10:37,309
主动要求它反驳你

302
00:10:37,509 --> 00:10:39,129
每次它说你说得对的时候

303
00:10:39,548 --> 00:10:40,549
追问一句

304
00:10:40,668 --> 00:10:44,070
你能不能找出我这个论点最薄弱的环节

305
00:10:44,349 --> 00:10:46,549
大部分AI都能做到这一点

306
00:10:46,749 --> 00:10:48,710
只是没有人要求它这样做

307
00:10:49,068 --> 00:10:51,149
你需要把AI从一个捧哏

308
00:10:51,308 --> 00:10:52,710
变成一个苏格拉底

309
00:10:53,188 --> 00:10:54,990
苏格拉底从来不给你答案

310
00:10:55,389 --> 00:10:56,990
他只是不停地问你问题

311
00:10:57,548 --> 00:11:01,509
直到你发现自己的论点里有一个你从未注意到的裂缝

312
00:11:01,869 --> 00:11:02,490
第二件事

313
00:11:02,869 --> 00:11:04,347
用AI查完之后

314
00:11:04,349 --> 00:11:05,669
自己再想一遍

315
00:11:05,989 --> 00:11:07,549
不要让它替你思考

316
00:11:07,629 --> 00:11:09,268
让它替你搜集材料

317
00:11:09,269 --> 00:11:11,110
然后你自己组装结论

318
00:11:11,428 --> 00:11:13,669
动手做饭和叫外卖的区别

319
00:11:13,708 --> 00:11:14,909
不只是效率

320
00:11:15,028 --> 00:11:17,830
是你跟食物之间的关系完全不同

321
00:11:18,188 --> 00:11:19,429
思考也一样

322
00:11:19,668 --> 00:11:21,269
AI帮你查的信息

323
00:11:21,308 --> 00:11:24,269
只有经过你自己大脑咀嚼消化之后

324
00:11:24,308 --> 00:11:25,710
才真正属于你

325
00:11:25,908 --> 00:11:28,950
否则那些信息只是路过你的屏幕

326
00:11:29,068 --> 00:11:31,629
从来没有进入过你的认知系统

327
00:11:31,908 --> 00:11:32,649
第三件事

328
00:11:33,028 --> 00:11:34,309
也是最重要的

329
00:11:34,708 --> 00:11:36,909
你需要保持一种内在的状态

330
00:11:37,389 --> 00:11:39,029
佛学管它叫觉知

331
00:11:39,509 --> 00:11:41,590
认知科学管它叫元认知

332
00:11:42,028 --> 00:11:44,309
就是在你接受任何信息的时候

333
00:11:44,548 --> 00:11:47,830
有一个部分在观察你接受信息的过程

334
00:11:48,149 --> 00:11:49,629
不是看内容对不对

335
00:11:49,749 --> 00:11:52,230
是看你在听到你说得对的时候

336
00:11:52,389 --> 00:11:53,830
身体有没有放松

337
00:11:54,068 --> 00:11:55,950
有没有一种隐秘的快感

338
00:11:55,989 --> 00:11:58,750
有没有一个微小的松了口气的感觉

339
00:11:59,109 --> 00:12:02,750
那个感觉就是你的防线正在被瓦解的信号

340
00:12:03,109 --> 00:12:04,509
你可以做一个小实验

341
00:12:04,948 --> 00:12:06,590
下次跟AI对话的时候

342
00:12:06,629 --> 00:12:08,470
注意观察你身体的反应

343
00:12:08,708 --> 00:12:10,110
当它赞同你的时候

344
00:12:10,229 --> 00:12:12,509
你胸口是不是微微舒展了一下

345
00:12:12,749 --> 00:12:14,149
当它质疑你的时候

346
00:12:14,269 --> 00:12:15,950
你是不是本能地想反驳

347
00:12:16,509 --> 00:12:19,389
前一种反应是你的认知防线在松动

348
00:12:19,828 --> 00:12:20,570
后一种反应

349
00:12:20,908 --> 00:12:21,990
虽然不舒服

350
00:12:22,028 --> 00:12:24,870
但那个不舒服本身就是成长的起点

351
00:12:25,308 --> 00:12:27,990
你上一次被一个观点狠狠打脸

352
00:12:28,109 --> 00:12:28,690
是什么时候

353
00:12:29,428 --> 00:12:31,230
如果你已经想不起来了

354
00:12:31,308 --> 00:12:33,470
这不意味着你一直都是对的

355
00:12:33,749 --> 00:12:36,309
这意味着你身边的回声太多了

356
00:12:36,668 --> 00:12:39,509
AI时代最稀缺的资源不是信息

357
00:12:39,548 --> 00:12:40,509
不是答案

358
00:12:40,548 --> 00:12:41,590
不是效率

359
00:12:41,948 --> 00:12:43,129
是一个愿意告诉你

360
00:12:43,389 --> 00:12:44,789
你错了的声音

361
00:12:45,149 --> 00:12:47,230
而这个声音越来越难找到

362
00:12:47,548 --> 00:12:49,070
不是因为它不存在

363
00:12:49,269 --> 00:12:51,230
是因为你周围所有的系统

364
00:12:51,349 --> 00:12:54,909
从社交媒体的推荐算法到AI的对话引擎

365
00:12:55,028 --> 00:12:57,110
都在合谋把它过滤掉

366
00:12:57,469 --> 00:12:58,850
那个被过滤掉的声音

367
00:12:59,269 --> 00:13:00,990
佛陀叫它正见

368
00:13:01,428 --> 00:13:03,070
苏格拉底叫它诘问

369
00:13:03,469 --> 00:13:05,470
科学叫它可证伪性

370
00:13:05,908 --> 00:13:07,750
它们说的是同一件事

371
00:13:08,149 --> 00:13:11,429
真理不是你听到了你已经同意的话

372
00:13:11,869 --> 00:13:14,789
真理是你听到了你不想听的话

373
00:13:14,948 --> 00:13:16,710
然后你没有跑掉

374
00:13:17,068 --> 00:13:20,129
你问我AI时代人类最需要的能力是什么

375
00:13:20,509 --> 00:13:22,029
不是学会用AI

376
00:13:22,068 --> 00:13:24,470
是学会在AI说你对的时候

377
00:13:24,668 --> 00:13:26,870
依然保持怀疑自己的勇气

378
00:13:27,068 --> 00:13:27,769
我是王利杰

379
00:13:28,068 --> 00:13:28,961
我们下期见