1
00:00:00,000 --> 00:00:03,553
Karpathy说你写的App根本不该存在

2
00:00:03,904 --> 00:00:04,974
他不是在骂谁

3
00:00:05,264 --> 00:00:06,153
是骂自己

4
00:00:06,694 --> 00:00:07,623
2025年

5
00:00:07,654 --> 00:00:09,443
OpenAI创始成员Andrej

6
00:00:09,453 --> 00:00:11,342
Karpathy花了好几天

7
00:00:11,373 --> 00:00:15,023
用AI辅助编程写了一个叫MenuGen的应用

8
00:00:15,213 --> 00:00:15,723
干嘛呢

9
00:00:16,214 --> 00:00:16,923
你去餐厅

10
00:00:17,333 --> 00:00:18,783
服务员递来菜单

11
00:00:18,973 --> 00:00:20,702
上面大部分菜没图

12
00:00:20,894 --> 00:00:22,862
你不知道那菜长啥样

13
00:00:23,054 --> 00:00:24,322
很多人都有这经历

14
00:00:24,693 --> 00:00:25,943
尤其在国外

15
00:00:25,973 --> 00:00:26,882
菜名看不懂

16
00:00:27,293 --> 00:00:29,142
更不知道端上来是个啥

17
00:00:29,414 --> 00:00:33,143
于是Karpathy做了个工具 拍张菜单照片

18
00:00:33,293 --> 00:00:36,143
它帮你把每道菜的样子生成出来

19
00:00:36,333 --> 00:00:37,962
这个App跑在云端

20
00:00:38,373 --> 00:00:39,742
前后端齐全

21
00:00:40,574 --> 00:00:43,212
用了多模态大模型做文字识别

22
00:00:43,214 --> 00:00:44,263
提取菜名

23
00:00:44,373 --> 00:00:47,062
用了图片生成模型渲染菜品

24
00:00:47,093 --> 00:00:49,373
还有用户注册 在线支付

25
00:00:49,373 --> 00:00:50,742
部署在Vercel上

26
00:00:50,854 --> 00:00:52,782
一套完整的Web应用

27
00:00:53,359 --> 00:00:53,707
然后

28
00:00:54,078 --> 00:00:55,438
2026年4月

29
00:00:55,438 --> 00:00:57,347
Karpathy在红杉资本的AI

30
00:00:57,347 --> 00:00:59,487
Ascent峰会上讲了一句话

31
00:00:59,639 --> 00:01:01,847
把自己这作品直接判了死刑

32
00:01:01,942 --> 00:01:02,743
他说

33
00:01:02,758 --> 00:01:05,437
你只需要把菜单照片丢给大模型

34
00:01:05,438 --> 00:01:09,288
附一句话 帮我把这些菜的样子直接画在菜单上

35
00:01:09,598 --> 00:01:10,268
几秒钟后

36
00:01:10,678 --> 00:01:12,118
模型返回一张图

37
00:01:12,119 --> 00:01:14,238
跟原始菜单几乎一模一样

38
00:01:14,238 --> 00:01:17,167
只是每道菜旁边多了张逼真的效果图

39
00:01:17,318 --> 00:01:18,027
像素层面

40
00:01:18,279 --> 00:01:19,728
直接完成了渲染

41
00:01:19,839 --> 00:01:20,628
整个MenuGen

42
00:01:20,798 --> 00:01:21,958
几天的工作量

43
00:01:21,958 --> 00:01:22,788
几千行代码

44
00:01:22,958 --> 00:01:24,368
根本没必要存在

45
00:01:24,809 --> 00:01:25,157
哈喽

46
00:01:25,169 --> 00:01:25,637
大家好

47
00:01:25,768 --> 00:01:26,777
我是王利杰

48
00:01:27,008 --> 00:01:29,777
今天这期内容信息密度会非常高

49
00:01:30,048 --> 00:01:32,377
Karpathy在这场峰会上的访谈

50
00:01:32,449 --> 00:01:33,678
我反复看了好几遍

51
00:01:34,089 --> 00:01:35,167
越看越觉得

52
00:01:35,169 --> 00:01:38,738
他在描述一件我们大多数人还没意识到的事

53
00:01:39,672 --> 00:01:41,472
他说的不是AI又进步了

54
00:01:41,472 --> 00:01:43,242
不是写代码又变快了

55
00:01:43,353 --> 00:01:44,311
他说的是

56
00:01:44,312 --> 00:01:46,642
一种全新的计算机正在诞生

57
00:01:47,073 --> 00:01:48,552
不是更快的计算机

58
00:01:48,552 --> 00:01:50,162
不是更智能的软件

59
00:01:50,552 --> 00:01:53,002
而是跟我们过去七十年用的东西

60
00:01:53,112 --> 00:01:56,041
在本质上完全不同的计算范式

61
00:01:56,392 --> 00:01:58,841
他把它叫做Software 3 0

62
00:01:59,713 --> 00:02:00,822
要理解3 0

63
00:02:01,114 --> 00:02:03,602
得先回头看1 0和2 0

64
00:02:04,034 --> 00:02:05,283
Software 1 0

65
00:02:05,314 --> 00:02:07,683
就是我们最熟悉的传统编程

66
00:02:08,073 --> 00:02:09,063
程序员写代码

67
00:02:09,474 --> 00:02:10,262
告诉计算机

68
00:02:10,673 --> 00:02:11,512
输入A

69
00:02:11,513 --> 00:02:11,903
做B

70
00:02:12,233 --> 00:02:13,123
输出C

71
00:02:13,434 --> 00:02:15,632
每一条规则 每一个分支

72
00:02:15,634 --> 00:02:18,162
都是人类用代码精确定义的

73
00:02:18,393 --> 00:02:19,142
你写什么

74
00:02:19,353 --> 00:02:20,482
它就干什么

75
00:02:20,553 --> 00:02:21,102
不多干

76
00:02:21,393 --> 00:02:22,403
也不少干

77
00:02:23,175 --> 00:02:24,544
Software 2 0

78
00:02:24,654 --> 00:02:27,463
是Karpathy在2017年提出的概念

79
00:02:27,735 --> 00:02:30,503
那时候他在特斯拉负责自动驾驶

80
00:02:30,935 --> 00:02:31,484
他发现

81
00:02:31,854 --> 00:02:34,424
有些问题没法用规则穷举

82
00:02:34,814 --> 00:02:37,503
比如让车识别一个行人

83
00:02:37,895 --> 00:02:38,843
你怎么写规则

84
00:02:39,559 --> 00:02:40,148
两条腿

85
00:02:40,918 --> 00:02:41,507
穿衣服

86
00:02:42,222 --> 00:02:42,667
有头

87
00:02:43,679 --> 00:02:45,347
那骑自行车的算不算

88
00:02:45,999 --> 00:02:46,787
撑伞的呢

89
00:02:47,358 --> 00:02:48,907
蹲下来系鞋带的呢

90
00:02:49,679 --> 00:02:52,007
规则写到天亮都写不完

91
00:02:52,478 --> 00:02:53,548
2 0的做法是

92
00:02:53,879 --> 00:02:55,287
你不写规则了

93
00:02:55,638 --> 00:02:57,967
你给机器大量的标注数据

94
00:02:58,078 --> 00:03:01,407
让神经网络自己从数据里学出规则

95
00:03:01,758 --> 00:03:03,048
编程的本质

96
00:03:03,119 --> 00:03:07,127
变成了准备数据集和设计训练目标

97
00:03:07,680 --> 00:03:09,709
那Software 3 0呢

98
00:03:10,321 --> 00:03:13,329
3 0的编程方式变成了说话

99
00:03:13,761 --> 00:03:14,790
你不写代码了

100
00:03:14,960 --> 00:03:16,730
也不准备数据集了

101
00:03:17,081 --> 00:03:18,869
你跟一个大语言模型说

102
00:03:19,441 --> 00:03:21,850
帮我把这个菜单上的菜画出来

103
00:03:22,240 --> 00:03:23,549
模型理解你的意思

104
00:03:23,960 --> 00:03:25,010
然后执行

105
00:03:25,520 --> 00:03:29,290
你的上下文窗口里塞进去的信息和指令

106
00:03:29,361 --> 00:03:32,010
就是你撬动这个解释器的杠杆

107
00:03:32,481 --> 00:03:35,369
这个解释器就是大语言模型本身

108
00:03:35,840 --> 00:03:37,290
它理解上下文

109
00:03:37,361 --> 00:03:39,570
在信息空间里执行计算

110
00:03:39,641 --> 00:03:41,089
然后给出结果

111
00:03:41,474 --> 00:03:43,364
Karpathy说这不是隐喻

112
00:03:43,514 --> 00:03:45,283
这就是正在发生的事

113
00:03:45,474 --> 00:03:47,964
他举了另一个更贴近日常的例子

114
00:03:48,155 --> 00:03:50,443
有一个开源工具叫OpenClaw

115
00:03:50,514 --> 00:03:52,270
就是我之前一期节目讲过的

116
00:03:52,270 --> 00:03:54,244
那个奥地利程序员的项目

117
00:03:54,435 --> 00:03:55,383
按传统思维

118
00:03:55,674 --> 00:03:56,913
你要安装软件

119
00:03:56,914 --> 00:03:58,503
应该有个安装脚本对吧

120
00:03:58,995 --> 00:03:59,873
但问题是

121
00:03:59,875 --> 00:04:03,883
不同操作系统 不同硬件配置 不同的软件环境

122
00:04:03,914 --> 00:04:06,193
这个安装脚本会越写越臃肿

123
00:04:06,194 --> 00:04:07,563
越来越容易出错

124
00:04:07,714 --> 00:04:08,474
本质上

125
00:04:08,474 --> 00:04:11,274
这是因为你还在用Software 1 0的思维

126
00:04:11,275 --> 00:04:14,764
试图把所有安装步骤都精确地写成代码

127
00:04:15,314 --> 00:04:16,644
OpenClaw不这么干

128
00:04:16,914 --> 00:04:18,884
它给你的就是一段文本

129
00:04:19,274 --> 00:04:22,244
你把这段文本复制粘贴给你的AI助手

130
00:04:22,435 --> 00:04:23,514
助手读完之后

131
00:04:23,514 --> 00:04:25,843
会自己检查你的电脑环境

132
00:04:25,995 --> 00:04:27,684
判断需要装什么依赖

133
00:04:27,754 --> 00:04:30,084
然后一步一步把软件装好

134
00:04:30,355 --> 00:04:31,304
过程中出了问题

135
00:04:31,714 --> 00:04:32,754
它自己调试

136
00:04:32,754 --> 00:04:33,723
自己修复

137
00:04:34,194 --> 00:04:35,783
你注意到范式变了没有

138
00:04:36,314 --> 00:04:37,663
以前的编程问题是

139
00:04:37,995 --> 00:04:39,764
我该写什么代码

140
00:04:40,074 --> 00:04:41,673
现在的问题变成了

141
00:04:41,675 --> 00:04:45,204
我该复制粘贴哪段文字给我的AI

142
00:04:45,795 --> 00:04:47,723
这句话听起来像开玩笑

143
00:04:47,874 --> 00:04:50,483
但这就是Software 3 0的日常

144
00:04:51,419 --> 00:04:52,779
理解了这三个阶段

145
00:04:52,779 --> 00:04:54,698
我们再回头看MenuGen的故事

146
00:04:54,700 --> 00:04:56,068
就完全不一样了

147
00:04:56,260 --> 00:04:58,068
Karpathy写那个App的时候

148
00:04:58,180 --> 00:05:00,909
他的脑子还在1 0和2 0的惯性里

149
00:05:01,099 --> 00:05:01,529
他在想

150
00:05:01,859 --> 00:05:04,229
我需要一个前端来上传图片

151
00:05:04,260 --> 00:05:06,068
一个后端来调用API

152
00:05:06,099 --> 00:05:08,258
一个数据库来存用户信息

153
00:05:08,260 --> 00:05:10,029
一个支付系统来收费

154
00:05:10,339 --> 00:05:12,669
他在用老范式思考新问题

155
00:05:13,020 --> 00:05:14,568
而3 0的解法是什么

156
00:05:15,060 --> 00:05:15,909
没有App

157
00:05:16,299 --> 00:05:19,378
只有一句自然语言指令和一张图片进去

158
00:05:19,380 --> 00:05:20,789
一张图片出来

159
00:05:21,020 --> 00:05:22,948
中间不需要任何中介层

160
00:05:24,032 --> 00:05:25,881
这就是为什么Karpathy说

161
00:05:25,912 --> 00:05:28,361
大家需要重新调整自己的思路

162
00:05:28,631 --> 00:05:30,760
不要总是沿着既有范式去想

163
00:05:30,871 --> 00:05:34,441
也不要只把AI理解成把原有的事情做得更快

164
00:05:34,751 --> 00:05:35,910
真正的变化是

165
00:05:35,912 --> 00:05:38,481
现在有很多新的东西变得可行了

166
00:05:38,551 --> 00:05:41,160
而这些东西在旧范式里根本不存在

167
00:05:41,511 --> 00:05:41,700
好

168
00:05:42,071 --> 00:05:43,871
到这里你可能会觉得

169
00:05:43,871 --> 00:05:44,460
挺好的嘛

170
00:05:44,751 --> 00:05:46,080
未来一片光明

171
00:05:46,352 --> 00:05:47,950
但Karpathy接下来说了

172
00:05:47,951 --> 00:05:50,041
一件让所有人不太舒服的事

173
00:05:50,444 --> 00:05:50,793
他说

174
00:05:51,124 --> 00:05:52,283
这些AI模型

175
00:05:52,285 --> 00:05:52,913
不是动物

176
00:05:53,244 --> 00:05:54,174
是幽灵

177
00:05:54,525 --> 00:05:56,913
他之前写过一篇文章叫Animals

178
00:05:57,004 --> 00:05:58,093
vs Ghosts

179
00:05:58,444 --> 00:05:59,814
动物与幽灵

180
00:06:00,564 --> 00:06:02,314
动物的智能是怎么来的

181
00:06:03,009 --> 00:06:03,809
进化

182
00:06:04,124 --> 00:06:05,733
几亿年的自然选择

183
00:06:05,805 --> 00:06:07,733
从单细胞到多细胞

184
00:06:07,764 --> 00:06:10,603
从鱼到爬行动物到哺乳动物

185
00:06:10,605 --> 00:06:12,523
DNA编码了学习算法

186
00:06:12,525 --> 00:06:16,293
然后通过漫长的进化外循环一点一点打磨出来

187
00:06:16,644 --> 00:06:18,033
动物有内在动机

188
00:06:18,244 --> 00:06:18,954
有好奇心

189
00:06:19,165 --> 00:06:20,083
有恐惧

190
00:06:20,084 --> 00:06:20,674
有饥饿

191
00:06:20,925 --> 00:06:22,134
有繁殖驱动

192
00:06:22,444 --> 00:06:24,853
这些东西深深嵌在基因里

193
00:06:25,524 --> 00:06:26,974
AI不是这样来的

194
00:06:27,284 --> 00:06:29,534
AI是在整个互联网的文本上

195
00:06:29,604 --> 00:06:31,733
训练出来的统计仿真电路

196
00:06:32,084 --> 00:06:33,774
它的底座是预训练

197
00:06:33,805 --> 00:06:35,733
就是大规模的模式匹配

198
00:06:36,125 --> 00:06:38,293
然后上面叠加了强化学习

199
00:06:38,365 --> 00:06:40,774
给它装上了一些定向增强的能力

200
00:06:41,204 --> 00:06:42,413
所以Karpathy说

201
00:06:42,445 --> 00:06:44,243
我们不是在制造动物

202
00:06:44,245 --> 00:06:45,974
我们是在召唤幽灵

203
00:06:46,284 --> 00:06:47,274
幽灵有什么特点

204
00:06:47,964 --> 00:06:49,053
参差不齐

205
00:06:49,284 --> 00:06:50,353
英文叫jagged

206
00:06:50,604 --> 00:06:51,654
锯齿状的

207
00:06:51,964 --> 00:06:53,053
同一个模型

208
00:06:53,204 --> 00:06:56,173
你让它重构一个十万行的代码仓库

209
00:06:56,284 --> 00:06:57,694
它干得漂漂亮亮

210
00:06:57,964 --> 00:06:59,673
你让它找零日安全漏洞

211
00:06:59,964 --> 00:07:01,053
它也能找到

212
00:07:01,815 --> 00:07:05,783
但你问它一个问题 我想去五十米外的洗车店洗车

213
00:07:05,935 --> 00:07:07,644
我应该开车还是走路

214
00:07:08,214 --> 00:07:09,723
它会一本正经地告诉你

215
00:07:10,055 --> 00:07:10,853
走路吧

216
00:07:10,855 --> 00:07:12,144
因为距离很近

217
00:07:12,254 --> 00:07:13,014
五十米

218
00:07:13,014 --> 00:07:14,144
走路去洗车

219
00:07:14,454 --> 00:07:16,464
然后你的车还停在原地

220
00:07:16,534 --> 00:07:17,243
你走回来

221
00:07:17,454 --> 00:07:18,704
车还是脏的

222
00:07:18,894 --> 00:07:22,904
一个能处理人类级别复杂工程问题的最先进模型

223
00:07:22,935 --> 00:07:26,303
连洗车要把车开过去这种常识都没有

224
00:07:26,646 --> 00:07:28,696
以前大家最爱举的例子是

225
00:07:28,807 --> 00:07:32,215
草莓这个英文单词strawberry里有几个r

226
00:07:32,487 --> 00:07:34,016
模型经常答错

227
00:07:34,326 --> 00:07:37,295
但这个问题现在已经被打补丁修掉了

228
00:07:37,607 --> 00:07:39,126
洗车店的例子说明

229
00:07:39,126 --> 00:07:41,525
这种参差不齐是结构性的

230
00:07:41,526 --> 00:07:42,735
不是偶然的

231
00:07:42,938 --> 00:07:43,489
为什么

232
00:07:43,700 --> 00:07:46,188
Karpathy给出了一个非常底层的解释

233
00:07:46,539 --> 00:07:50,308
两个变量决定了模型在哪些领域会突然变强

234
00:07:50,780 --> 00:07:52,909
第一个变量是可验证性

235
00:07:53,219 --> 00:07:56,349
模型训练的核心机制是强化学习

236
00:07:56,619 --> 00:07:57,888
做对了给奖励

237
00:07:58,219 --> 00:07:59,469
做错了扣分

238
00:07:59,859 --> 00:08:02,659
但只有结果可以被自动验证的任务

239
00:08:02,659 --> 00:08:05,229
才能有效地放进这个训练循环

240
00:08:05,780 --> 00:08:07,308
数学可以验证

241
00:08:07,659 --> 00:08:08,768
代码可以验证

242
00:08:09,179 --> 00:08:10,989
跑一下就知道对不对

243
00:08:11,340 --> 00:08:13,909
所以模型在这些领域突飞猛进

244
00:08:14,451 --> 00:08:17,420
但开车去洗车店这种常识

245
00:08:17,492 --> 00:08:18,441
怎么自动验证

246
00:08:18,932 --> 00:08:20,001
没有验证环境

247
00:08:20,211 --> 00:08:21,741
就没有训练信号

248
00:08:21,932 --> 00:08:24,741
第二个变量是AI公司关心什么

249
00:08:25,052 --> 00:08:26,741
Karpathy举了个例子

250
00:08:27,011 --> 00:08:29,460
从GPT - 3 5到GPT - 4

251
00:08:29,492 --> 00:08:33,020
很多人觉得模型下国际象棋的能力提升了很多

252
00:08:33,211 --> 00:08:35,941
以为是整体智能变强的自然结果

253
00:08:36,052 --> 00:08:37,340
他说不一定

254
00:08:37,532 --> 00:08:38,240
据他了解

255
00:08:38,571 --> 00:08:40,651
是有人在OpenAI做了决定

256
00:08:40,652 --> 00:08:44,340
把大量国际象棋棋谱数据加进了预训练集

257
00:08:44,611 --> 00:08:45,860
就这一个决定

258
00:08:45,892 --> 00:08:47,931
模型在国际象棋上的表现

259
00:08:47,932 --> 00:08:50,140
就出现了一个巨大的能力峰值

260
00:08:50,772 --> 00:08:53,620
跟其他领域的提升完全不成比例

261
00:08:54,752 --> 00:08:56,600
所以模型的能力地形

262
00:08:56,632 --> 00:08:58,681
不是一个均匀上升的平面

263
00:08:58,752 --> 00:09:02,120
而是一片充满山峰和深谷的崎岖地带

264
00:09:02,752 --> 00:09:06,681
山峰是强化学习和数据集重点覆盖的区域

265
00:09:07,032 --> 00:09:10,360
深谷是数据分布从来没有碰过的盲区

266
00:09:10,912 --> 00:09:13,831
如果你的应用场景刚好站在山峰上

267
00:09:13,831 --> 00:09:15,401
你会觉得AI是神

268
00:09:15,872 --> 00:09:17,431
如果你落在深谷里

269
00:09:17,432 --> 00:09:19,801
你会觉得AI连小学生都不如

270
00:09:20,351 --> 00:09:22,480
这对所有使用AI的人来说

271
00:09:22,512 --> 00:09:24,561
都是一个极其重要的认知

272
00:09:25,032 --> 00:09:27,760
你不能笼统地说AI行还是不行

273
00:09:28,071 --> 00:09:29,100
你必须搞清楚

274
00:09:29,351 --> 00:09:30,870
在你的具体场景里

275
00:09:30,872 --> 00:09:33,041
你站在这片地形的什么位置

276
00:09:33,792 --> 00:09:36,521
如果你不在强化学习覆盖的电路里

277
00:09:36,552 --> 00:09:38,480
你可能需要自己做微调

278
00:09:38,632 --> 00:09:41,401
或者做额外的工程工作来弥补

279
00:09:42,269 --> 00:09:44,697
那人类在这个新世界里还剩下什么

280
00:09:45,188 --> 00:09:48,077
Karpathy引用了一句让他念念不忘的话

281
00:09:48,188 --> 00:09:50,118
我看完也印象极深

282
00:09:50,389 --> 00:09:52,118
你可以外包你的思考

283
00:09:52,149 --> 00:09:54,118
但你不能外包你的理解

284
00:09:54,428 --> 00:09:56,758
他说自己正在变成一个瓶颈

285
00:09:57,029 --> 00:09:59,707
即便只是理解我们到底想构建什么

286
00:09:59,708 --> 00:10:01,237
为什么这件事值得做

287
00:10:01,389 --> 00:10:03,638
我该如何指挥我的AI助手

288
00:10:03,828 --> 00:10:06,638
光是这些事情本身就在成为瓶颈

289
00:10:07,068 --> 00:10:10,118
信息必须以某种方式进入他的大脑

290
00:10:10,188 --> 00:10:12,357
而大脑的带宽是有限的

291
00:10:12,969 --> 00:10:14,567
这也是为什么Karpathy

292
00:10:14,569 --> 00:10:17,297
对一个叫知识库的东西特别兴奋

293
00:10:17,489 --> 00:10:20,338
他做了一个项目叫LLM Wiki

294
00:10:20,368 --> 00:10:22,858
也就是他说的大模型知识库

295
00:10:23,008 --> 00:10:25,338
让大模型读完一堆文档后

296
00:10:25,408 --> 00:10:27,738
自动生成一个结构化的知识库

297
00:10:27,929 --> 00:10:29,618
每当他读一篇文章

298
00:10:29,689 --> 00:10:32,098
知识库会自动更新和重组

299
00:10:32,288 --> 00:10:36,138
然后他可以围绕这些内容提问 追问 深挖

300
00:10:36,408 --> 00:10:37,958
这种工具的本质是什么

301
00:10:38,408 --> 00:10:40,218
不是增强你的产出

302
00:10:40,329 --> 00:10:41,978
而是增强你的理解

303
00:10:42,168 --> 00:10:43,927
因为如果你不理解

304
00:10:43,929 --> 00:10:46,378
你就无法成为一个好的指挥者

305
00:10:46,689 --> 00:10:49,937
大语言模型本身并不擅长真正的理解

306
00:10:50,809 --> 00:10:52,057
你作为一个人类

307
00:10:52,089 --> 00:10:54,608
仍然是系统中那个独一无二的

308
00:10:54,609 --> 00:10:56,138
负责理解的角色

309
00:10:56,413 --> 00:10:58,092
Karpathy在这个访谈里

310
00:10:58,092 --> 00:11:00,021
还说了一个关于未来的预判

311
00:11:00,092 --> 00:11:01,501
我觉得非常大胆

312
00:11:01,656 --> 00:11:02,413
他说

313
00:11:02,413 --> 00:11:06,261
他相信未来会出现完全由神经网络构成的计算机

314
00:11:06,653 --> 00:11:09,621
设备直接接收原始视频和音频输入

315
00:11:09,653 --> 00:11:11,141
送进一个神经网络

316
00:11:11,212 --> 00:11:14,702
然后用扩散模型实时渲染出一个用户界面

317
00:11:14,852 --> 00:11:18,422
这个界面是为当下这个具体时刻 即时生成的

318
00:11:18,895 --> 00:11:21,623
今天的计算机本质上还是计算器

319
00:11:21,895 --> 00:11:22,764
五十年代的时候

320
00:11:23,135 --> 00:11:24,863
人们其实还不确定

321
00:11:24,934 --> 00:11:28,463
计算机到底应该长得像计算器还是像大脑

322
00:11:28,895 --> 00:11:31,103
后来我们选了计算器路线

323
00:11:31,574 --> 00:11:32,823
发展了七十年

324
00:11:32,855 --> 00:11:36,264
就是今天的芯片 操作系统和应用商店

325
00:11:36,775 --> 00:11:39,144
但未来这件事可能要反过来

326
00:11:39,615 --> 00:11:41,664
神经网络变成主进程

327
00:11:41,775 --> 00:11:44,144
传统CPU变成协处理器

328
00:11:44,574 --> 00:11:47,424
神经网络承担绝大部分核心负载

329
00:11:47,735 --> 00:11:51,333
而那些需要一加一精确等于二的确定性计算

330
00:11:51,334 --> 00:11:52,823
变成了附属工具

331
00:11:53,535 --> 00:11:57,823
就像今天GPU已经在很多场景下比CPU更重要了

332
00:11:57,855 --> 00:12:01,264
只是将来这个趋势会推进到一切计算领域

333
00:12:01,374 --> 00:12:04,622
Karpathy还提到了一个很多人没想到的影响

334
00:12:04,622 --> 00:12:05,398
招聘

335
00:12:05,774 --> 00:12:09,622
他说现在大多数公司想招厉害的AI工程师

336
00:12:09,813 --> 00:12:12,543
但面试方式还停留在旧范式

337
00:12:12,854 --> 00:12:14,222
出一道小谜题

338
00:12:14,293 --> 00:12:16,942
让候选人在白板上写算法

339
00:12:17,293 --> 00:12:18,362
这是在考什么

340
00:12:18,894 --> 00:12:19,643
考记忆力

341
00:12:20,014 --> 00:12:21,982
考手写代码的熟练度

342
00:12:22,414 --> 00:12:25,372
但这些能力在Software 3 0时代

343
00:12:25,374 --> 00:12:26,742
几乎没有价值

344
00:12:27,791 --> 00:12:29,879
他建议面试应该变成这样

345
00:12:30,031 --> 00:12:31,879
给候选人一个大项目

346
00:12:31,911 --> 00:12:34,670
比如用AI写一个完整的社交平台

347
00:12:34,671 --> 00:12:35,999
然后部署上线

348
00:12:36,511 --> 00:12:39,639
接下来公司放十个AI去攻击这个平台

349
00:12:39,710 --> 00:12:40,960
试图把它搞崩

350
00:12:41,271 --> 00:12:42,899
看候选人怎么用工具

351
00:12:43,191 --> 00:12:44,479
怎么组织架构

352
00:12:44,511 --> 00:12:48,759
怎么在AI的帮助下构建出一个既强大又安全的系统

353
00:12:49,151 --> 00:12:51,519
考的不再是你能不能写代码

354
00:12:51,590 --> 00:12:55,119
而是你能不能指挥AI做出真正靠谱的东西

355
00:12:55,911 --> 00:12:57,700
这背后隐含的信号是什么

356
00:12:58,271 --> 00:13:01,440
连评判人才的标准都在被重新定义

357
00:13:02,317 --> 00:13:03,806
听到这你可能想问

358
00:13:03,917 --> 00:13:05,625
那人类现在到底该怎么办

359
00:13:06,197 --> 00:13:07,846
Karpathy的回答很务实

360
00:13:07,981 --> 00:13:08,780
他说

361
00:13:08,797 --> 00:13:11,885
目前AI还是一个能力很尖锐的实习生

362
00:13:11,957 --> 00:13:13,556
有时候表现惊艳

363
00:13:13,557 --> 00:13:15,446
有时候犯很蠢的错误

364
00:13:15,688 --> 00:13:17,978
他用MenuGen举了一个真实的bug

365
00:13:18,089 --> 00:13:19,978
用户用谷歌账号注册

366
00:13:20,048 --> 00:13:22,538
但买积分走的是Stripe支付

367
00:13:22,688 --> 00:13:24,697
两边系统都有邮箱地址

368
00:13:24,729 --> 00:13:26,618
AI在写代码的时候

369
00:13:26,688 --> 00:13:29,207
直接拿两边的邮箱地址去做匹配

370
00:13:29,209 --> 00:13:31,697
把支付跟用户关联起来

371
00:13:31,808 --> 00:13:32,398
问题是

372
00:13:32,648 --> 00:13:35,087
你完全可能谷歌用一个邮箱

373
00:13:35,089 --> 00:13:36,618
支付用另一个邮箱

374
00:13:36,688 --> 00:13:37,478
一旦不一样

375
00:13:37,688 --> 00:13:39,898
系统就把你的钱搞丢了

376
00:13:40,009 --> 00:13:41,888
因为AI根本没有意识到

377
00:13:41,889 --> 00:13:45,728
应该用一个统一的用户ID来绑定所有系统

378
00:13:45,729 --> 00:13:48,898
而不是用邮箱这种随意的信息做关联

379
00:13:49,528 --> 00:13:50,528
这种错误

380
00:13:50,528 --> 00:13:53,538
一个有经验的工程师一眼就能看出来

381
00:13:53,808 --> 00:13:55,297
但AI看不出来

382
00:13:55,648 --> 00:13:59,057
它只是在预测下一个最可能的代码片段

383
00:13:59,369 --> 00:14:02,097
不是在理解系统设计的本质逻辑

384
00:14:02,527 --> 00:14:03,816
所以Karpathy说

385
00:14:04,007 --> 00:14:06,326
人仍然必须负责规格

386
00:14:06,326 --> 00:14:07,365
负责计划

387
00:14:07,367 --> 00:14:08,235
负责品味

388
00:14:08,446 --> 00:14:09,536
负责判断

389
00:14:09,806 --> 00:14:12,335
AI负责在这个框架内填空

390
00:14:12,806 --> 00:14:13,796
人是建筑师

391
00:14:14,127 --> 00:14:15,576
AI是施工队

392
00:14:16,007 --> 00:14:18,855
但他紧接着说了一句让我脊背发凉的话

393
00:14:18,970 --> 00:14:19,771
他说

394
00:14:19,966 --> 00:14:23,936
没有什么根本性的东西阻止AI在这些方面继续进步

395
00:14:24,326 --> 00:14:27,135
品味 判断 审美这些东西

396
00:14:27,287 --> 00:14:28,975
目前之所以还不行

397
00:14:29,086 --> 00:14:31,326
可能只是因为AI公司还没有

398
00:14:31,326 --> 00:14:34,215
把它们放进强化学习的训练环境里

399
00:14:34,486 --> 00:14:37,896
还没有人去构建专门训练品味的奖励函数

400
00:14:38,887 --> 00:14:40,335
一旦他们找到办法

401
00:14:40,367 --> 00:14:42,686
这些能力也会像代码能力一样

402
00:14:42,686 --> 00:14:43,696
出现飞跃

403
00:14:44,366 --> 00:14:45,735
他还举了一个例子

404
00:14:45,927 --> 00:14:48,495
他做了一个叫microGPT的项目

405
00:14:48,567 --> 00:14:52,296
目标是把大语言模型的训练过程简化到极致

406
00:14:52,687 --> 00:14:55,015
但模型非常抗拒这件事

407
00:14:55,167 --> 00:14:58,336
他反复提示AI去简化 去精炼

408
00:14:58,407 --> 00:14:59,735
AI就是做不到

409
00:14:59,886 --> 00:15:01,395
那种感觉就像拔牙

410
00:15:01,687 --> 00:15:03,296
而不是光速前进

411
00:15:03,567 --> 00:15:06,096
因为简洁和优雅这种审美

412
00:15:06,167 --> 00:15:08,895
根本不在模型的强化学习电路里

413
00:15:09,167 --> 00:15:11,296
但这不代表它永远做不到

414
00:15:11,687 --> 00:15:15,015
只是还没有人专门去训练这个能力而已

415
00:15:15,967 --> 00:15:18,576
所以我们面对的局面是这样的

416
00:15:18,886 --> 00:15:21,736
一种全新的计算范式正在诞生

417
00:15:22,126 --> 00:15:23,635
编程不再是写代码

418
00:15:23,886 --> 00:15:24,895
而是说话

419
00:15:25,246 --> 00:15:27,135
App不再是必需品

420
00:15:27,327 --> 00:15:31,415
很多场景下一句自然语言就能替代一整个应用

421
00:15:31,967 --> 00:15:33,515
AI的能力参差不齐

422
00:15:33,927 --> 00:15:35,086
像幽灵一样

423
00:15:35,087 --> 00:15:36,566
在某些领域超人

424
00:15:36,567 --> 00:15:39,096
在另一些领域连常识都没有

425
00:15:39,447 --> 00:15:40,876
人类最后的护城河

426
00:15:41,207 --> 00:15:42,126
不是技能

427
00:15:42,126 --> 00:15:43,175
不是知识

428
00:15:43,207 --> 00:15:44,555
甚至不是思考能力

429
00:15:44,727 --> 00:15:45,775
而是理解

430
00:15:46,126 --> 00:15:47,956
但这个护城河能守多久

431
00:15:48,687 --> 00:15:52,535
Karpathy说品味和判断只是还没被训练到

432
00:15:53,286 --> 00:15:54,246
这意味着

433
00:15:54,246 --> 00:15:58,336
留给人类的窗口期可能比我们以为的要短得多

434
00:15:58,882 --> 00:16:00,172
作为一个投资人

435
00:16:00,283 --> 00:16:02,291
我看到的是这样一幅图景

436
00:16:02,602 --> 00:16:04,491
过去我们投资一家公司

437
00:16:04,642 --> 00:16:06,452
核心看它的技术壁垒

438
00:16:06,803 --> 00:16:08,052
现在的问题是

439
00:16:08,122 --> 00:16:11,852
技术壁垒正在以前所未有的速度被侵蚀

440
00:16:12,242 --> 00:16:16,651
一个人用几句话就能让AI做出一个完整的产品原型

441
00:16:17,083 --> 00:16:18,432
那真正的壁垒在哪

442
00:16:19,002 --> 00:16:21,531
也许就在Karpathy说的那个词里

443
00:16:21,666 --> 00:16:22,467
理解

444
00:16:23,175 --> 00:16:24,903
对行业的深刻理解

445
00:16:25,214 --> 00:16:26,863
对用户的深刻理解

446
00:16:27,214 --> 00:16:30,224
对什么值得做 什么不值得做的深刻理解

447
00:16:30,575 --> 00:16:33,263
这些是AI目前还给不了你的

448
00:16:33,494 --> 00:16:34,243
但问题是

449
00:16:34,614 --> 00:16:36,724
你有没有在真正地去理解

450
00:16:37,094 --> 00:16:39,724
还是你也在把理解外包给AI

451
00:16:40,374 --> 00:16:40,844
这个问题

452
00:16:41,134 --> 00:16:42,103
我留给你

453
00:16:42,214 --> 00:16:42,883
我是王利杰

454
00:16:43,175 --> 00:16:44,353
我们下期见