1
00:00:00,000 --> 00:00:01,480
公司的很多组织层级

2
00:00:01,480 --> 00:00:03,399
在AI眼里可能只是在干一件事

3
00:00:03,399 --> 00:00:04,320
搬运context

4
00:00:04,320 --> 00:00:05,400
这句话越想越可怕

5
00:00:05,400 --> 00:00:07,320
因为如果AI能够自己获得context

6
00:00:07,320 --> 00:00:08,320
它真正要重写的

7
00:00:08,320 --> 00:00:09,759
可能根本不只是几个岗位

8
00:00:09,759 --> 00:00:11,480
而是公司为什么需要这么多层级

9
00:00:11,480 --> 00:00:13,679
最近DataBricks的CEO联合创始人

10
00:00:13,679 --> 00:00:14,439
Elli Godessy

11
00:00:14,439 --> 00:00:15,759
做了一场非常长的访谈

12
00:00:15,759 --> 00:00:17,000
他不仅创办了DataBricks

13
00:00:17,000 --> 00:00:19,039
还是ApacheSpark的共同创造者之一

14
00:00:19,039 --> 00:00:21,679
长期站在企业数据和AI落地的最前线

15
00:00:21,679 --> 00:00:22,440
而这次访谈里

16
00:00:22,440 --> 00:00:23,000
他讲了几个

17
00:00:23,000 --> 00:00:24,679
我觉得非常值得记住的新概念

18
00:00:24,719 --> 00:00:26,160
第一个就是前段时间爆火的

19
00:00:26,160 --> 00:00:26,679
本体论

20
00:00:27,280 --> 00:00:28,359
不过在企业AI里

21
00:00:28,359 --> 00:00:29,920
更准确的说叫企业本体

22
00:00:29,920 --> 00:00:30,760
什么叫企业本体

23
00:00:30,760 --> 00:00:31,920
举个很容易理解的例子

24
00:00:31,920 --> 00:00:32,920
两个人能力差不多

25
00:00:32,920 --> 00:00:33,920
一个今天刚入职

26
00:00:33,920 --> 00:00:35,000
一个在公司干了5年

27
00:00:35,000 --> 00:00:36,560
为什么后者做事情快得多

28
00:00:36,560 --> 00:00:37,600
不是因为他更聪明

29
00:00:37,600 --> 00:00:38,439
而是他脑子里

30
00:00:38,439 --> 00:00:39,840
多了一张公司的隐形地图

31
00:00:39,840 --> 00:00:40,920
什么事情该找谁

32
00:00:40,920 --> 00:00:41,880
谁名义上负责

33
00:00:41,880 --> 00:00:42,840
但实际上推不动

34
00:00:42,840 --> 00:00:44,439
哪些流程制度上这么写

35
00:00:44,439 --> 00:00:45,960
真正运行起来却是另一回事

36
00:00:45,960 --> 00:00:47,320
哪些项目为什么被砍

37
00:00:47,320 --> 00:00:48,439
老板真正关心什么

38
00:00:48,479 --> 00:00:50,079
这些东西组织架构图里没有

39
00:00:50,079 --> 00:00:50,920
数据库里也没有

40
00:00:50,920 --> 00:00:51,640
但他们决定了

41
00:00:51,640 --> 00:00:53,280
这家公司实际上是怎么运转的

42
00:00:53,280 --> 00:00:54,880
企业本体就是把这些人

43
00:00:54,880 --> 00:00:57,079
项目 目标 决策 流程之间的关系

44
00:00:57,079 --> 00:00:58,520
变成AI能够理解的结构

45
00:00:58,520 --> 00:00:59,320
你可以把它理解成

46
00:00:59,320 --> 00:01:00,640
把这家公司到底怎么运转

47
00:01:00,640 --> 00:01:02,479
变成一张AI能够理解的关系地图

48
00:01:02,479 --> 00:01:03,359
数据告诉AI

49
00:01:03,359 --> 00:01:04,000
发生了什么

50
00:01:04,000 --> 00:01:05,120
企业本体告诉AI

51
00:01:05,120 --> 00:01:06,640
这些事情之间到底是什么关系

52
00:01:06,640 --> 00:01:07,719
所以今天企业AI

53
00:01:07,719 --> 00:01:08,840
有一个很荒谬的状态

54
00:01:08,840 --> 00:01:10,319
AI拥有全世界的知识

55
00:01:10,319 --> 00:01:11,519
却没有你公司的常识

56
00:01:11,519 --> 00:01:12,599
它智商可能180

57
00:01:12,599 --> 00:01:13,640
但功领明天

58
00:01:13,640 --> 00:01:14,879
A力的判断更激进

59
00:01:14,879 --> 00:01:15,879
对很多企业来说

60
00:01:15,879 --> 00:01:17,359
今天的模型已经够聪明了

61
00:01:17,359 --> 00:01:18,000
真正缺的

62
00:01:18,000 --> 00:01:19,400
不是再等一个更强的GPT

63
00:01:19,400 --> 00:01:21,239
而是把会议 决策 项目 关系

64
00:01:21,239 --> 00:01:21,799
人的经验

65
00:01:21,799 --> 00:01:23,560
都变成AI可以理解的context

66
00:01:23,560 --> 00:01:25,359
再组织成企业自己的知识结构

67
00:01:25,359 --> 00:01:26,400
而一旦这件事完成

68
00:01:26,400 --> 00:01:27,640
真正大的变化才开始

69
00:01:27,640 --> 00:01:28,319
今天一个问题

70
00:01:28,319 --> 00:01:29,879
怎么到CEO 员工报经理

71
00:01:29,879 --> 00:01:30,599
经理补背景

72
00:01:30,599 --> 00:01:31,280
再报总监

73
00:01:31,280 --> 00:01:32,280
总监再压缩一次

74
00:01:32,280 --> 00:01:33,640
最后到VP 到CEO

75
00:01:33,640 --> 00:01:35,120
决策做完再一层层传回来

76
00:01:35,120 --> 00:01:35,719
你仔细看

77
00:01:35,719 --> 00:01:36,480
大量管理工作

78
00:01:36,480 --> 00:01:37,640
其实都在做三件事

79
00:01:37,640 --> 00:01:38,519
收集context

80
00:01:38,519 --> 00:01:39,280
压缩context

81
00:01:39,280 --> 00:01:40,040
搬运context

82
00:01:40,040 --> 00:01:41,799
所以现代公司的很多组织成绩

83
00:01:41,799 --> 00:01:42,599
本质上就是一套

84
00:01:42,599 --> 00:01:44,599
用人搭出来的信息压缩和路由系统

85
00:01:44,599 --> 00:01:45,439
过去必须这么干

86
00:01:45,439 --> 00:01:46,640
因为任何一个人的context

87
00:01:46,640 --> 00:01:47,519
窗口都是有限的

88
00:01:47,519 --> 00:01:49,159
但如果AI知道整个公司的会议

89
00:01:49,159 --> 00:01:50,680
项目人员 数字和历史决策

90
00:01:50,680 --> 00:01:51,239
那一个问题

91
00:01:51,239 --> 00:01:52,760
为什么还必须经过五成汇报

92
00:01:52,760 --> 00:01:54,040
才能获得完整context

93
00:01:54,040 --> 00:01:55,200
所以AI真正压缩的

94
00:01:55,200 --> 00:01:56,439
可能首先不是员工数量

95
00:01:56,480 --> 00:01:57,760
而是组织的信息距离

96
00:01:57,760 --> 00:01:58,519
过去职位越高

97
00:01:58,519 --> 00:01:59,920
往往意味着你知道的更多

98
00:01:59,920 --> 00:02:01,040
但如果未来context

99
00:02:01,040 --> 00:02:01,959
变成基础设施

100
00:02:01,959 --> 00:02:02,879
我知道发生了什么

101
00:02:02,879 --> 00:02:03,959
本身会迅速贬值

102
00:02:03,959 --> 00:02:05,239
真正升值的是判断

103
00:02:05,239 --> 00:02:05,719
取舍

104
00:02:05,719 --> 00:02:06,159
责任

105
00:02:06,159 --> 00:02:08,159
所以AI对组织真正生产的冲击

106
00:02:08,159 --> 00:02:09,120
不是替代几个岗位

107
00:02:09,120 --> 00:02:09,960
而是开始追问

108
00:02:09,960 --> 00:02:11,360
这些层级为什么还要存在

109
00:02:11,360 --> 00:02:12,319
A里接下来又讲了

110
00:02:12,319 --> 00:02:13,319
第二个很火的概念

111
00:02:13,319 --> 00:02:13,800
RSI

112
00:02:13,800 --> 00:02:15,120
中文叫地归自我改进

113
00:02:15,120 --> 00:02:16,039
简单说就是AI

114
00:02:16,039 --> 00:02:17,680
帮助制造下一代更强的AI

115
00:02:17,680 --> 00:02:19,000
下一代再制造更强的一代

116
00:02:19,039 --> 00:02:20,599
最后形成所谓的智能爆炸

117
00:02:20,599 --> 00:02:22,479
现在很多人看到AI开始写代码

118
00:02:22,479 --> 00:02:23,639
开始帮助训练AI

119
00:02:23,639 --> 00:02:24,879
就说RSI已经来了

120
00:02:24,879 --> 00:02:26,120
A里直接泼了一盆冷水

121
00:02:26,120 --> 00:02:27,000
别把AI写AI

122
00:02:27,000 --> 00:02:27,719
就叫RSI

123
00:02:27,719 --> 00:02:29,919
Databricks现在超过90%的软件代码

124
00:02:29,919 --> 00:02:30,879
已经由AI生成

125
00:02:30,879 --> 00:02:31,919
但真正的RSI

126
00:02:31,919 --> 00:02:33,319
至少要同时满足四件事

127
00:02:33,319 --> 00:02:34,000
下一代模型

128
00:02:34,000 --> 00:02:35,000
用更少的GPU

129
00:02:35,000 --> 00:02:36,000
训练时间更短

130
00:02:36,000 --> 00:02:36,680
智能更强

131
00:02:36,680 --> 00:02:37,560
而且这三件事

132
00:02:37,560 --> 00:02:38,800
还能一轮一轮重复

133
00:02:38,800 --> 00:02:39,719
真正值得盯的

134
00:02:39,719 --> 00:02:41,199
不是AI有没有参与制造AI

135
00:02:41,199 --> 00:02:42,919
而是制造下一单位智能的成本

136
00:02:42,919 --> 00:02:44,280
有没有开始地归下降

137
00:02:44,319 --> 00:02:46,879
所以别把自动化加速误判成智能爆炸

138
00:02:46,879 --> 00:02:47,840
而说到AI风险

139
00:02:47,840 --> 00:02:49,599
A里还有一个更反共识的判断

140
00:02:49,599 --> 00:02:51,039
Superintelligence风险很远

141
00:02:51,039 --> 00:02:52,400
但Cyberrisk非常近

142
00:02:52,400 --> 00:02:54,199
也就是超级智能的风险还很远

143
00:02:54,199 --> 00:02:55,800
网络安全风险却已经到了眼前

144
00:02:55,800 --> 00:02:57,280
原因不是AI突然有了恶意

145
00:02:57,280 --> 00:02:58,080
而是攻击能力

146
00:02:58,080 --> 00:02:59,680
第一次可以大规模复制过去

147
00:02:59,680 --> 00:03:01,319
你不可能找一万个安全研究员

148
00:03:01,319 --> 00:03:02,639
同时攻击一个系统一个月

149
00:03:02,639 --> 00:03:03,680
今天一万个agent

150
00:03:03,680 --> 00:03:04,840
理论上可以同时启动

151
00:03:04,840 --> 00:03:05,919
所以真正危险的攻势

152
00:03:05,919 --> 00:03:07,639
可能不是超级智能乘以恶意

153
00:03:07,639 --> 00:03:09,520
而是普通智能乘以10万份复制

154
00:03:09,520 --> 00:03:11,240
这已经足够改变整个网络攻防

155
00:03:11,240 --> 00:03:12,960
AI真正已经发生的安全拐点

156
00:03:12,960 --> 00:03:13,960
不是机器产生恶意

157
00:03:13,960 --> 00:03:15,719
而是攻击能力开始规模化复制

158
00:03:15,719 --> 00:03:17,840
未来很可能变成agent攻击agent

159
00:03:17,840 --> 00:03:18,759
agent防守agent

160
00:03:18,759 --> 00:03:20,000
而人在最实时的攻防里

161
00:03:20,000 --> 00:03:20,759
反而越来越慢

162
00:03:20,759 --> 00:03:21,560
最后还有一个词

163
00:03:21,560 --> 00:03:22,520
所有做AI产品的人

164
00:03:22,520 --> 00:03:23,919
都应该开始关注Harness

165
00:03:23,919 --> 00:03:25,159
你可以把它简单理解成

166
00:03:25,159 --> 00:03:26,360
AI的执行框架模型

167
00:03:26,360 --> 00:03:27,120
负责提供智能

168
00:03:27,120 --> 00:03:27,879
Harness负责

169
00:03:27,879 --> 00:03:28,680
怎么给context

170
00:03:28,680 --> 00:03:29,360
怎么调用工具

171
00:03:29,360 --> 00:03:30,120
怎么安排步骤

172
00:03:30,120 --> 00:03:30,479
失败

173
00:03:31,039 --> 00:03:31,919
什么时候换模型

174
00:03:31,919 --> 00:03:33,120
而Databricks发现了一件

175
00:03:33,120 --> 00:03:34,000
很夸张的事情

176
00:03:34,000 --> 00:03:35,000
完全相同的模型

177
00:03:35,000 --> 00:03:36,000
完全相同的版本

178
00:03:36,000 --> 00:03:37,400
只是Harness不同时期成本

179
00:03:37,400 --> 00:03:38,560
就能相差接近两倍

180
00:03:38,560 --> 00:03:39,840
模型没变 智能没变

181
00:03:39,840 --> 00:03:41,039
只是执行方式变了

182
00:03:41,039 --> 00:03:42,080
账单已经差了一倍

183
00:03:42,080 --> 00:03:43,000
所以这里有一句话

184
00:03:43,000 --> 00:03:43,960
我觉得非常重要

185
00:03:43,960 --> 00:03:45,719
模型决定你买到了多少智能

186
00:03:45,719 --> 00:03:47,719
Harness决定这些智能被怎么花掉

187
00:03:47,719 --> 00:03:48,520
过去大家一直问

188
00:03:48,520 --> 00:03:49,560
GPT还是Cloud

189
00:03:49,560 --> 00:03:50,400
哪个模型更强

190
00:03:50,400 --> 00:03:51,680
但未来更重要的问题

191
00:03:51,680 --> 00:03:53,159
可能不是你用了谁的智能

192
00:03:53,159 --> 00:03:54,560
而是你怎么组织智能

193
00:03:54,560 --> 00:03:55,840
复杂任务给最强模型

194
00:03:55,840 --> 00:03:57,080
简单任务给便宜模型

195
00:03:57,080 --> 00:03:58,360
专业任务给专家模型

196
00:03:58,360 --> 00:03:59,479
再让Router和Harness

197
00:03:59,479 --> 00:04:00,960
决定谁在什么时候做什么

198
00:04:00,960 --> 00:04:02,439
所以把Ali这几个观点串起来

199
00:04:02,439 --> 00:04:04,039
你会发现AI下一阶段的竞争

200
00:04:04,039 --> 00:04:04,719
已经开始变了

201
00:04:04,719 --> 00:04:06,199
过去大家真的是模型智商

202
00:04:06,199 --> 00:04:08,120
下一阶段真的可能是组织智能

203
00:04:08,120 --> 00:04:09,120
企业本体决定

204
00:04:09,120 --> 00:04:10,400
AI懂不懂你的公司

205
00:04:10,400 --> 00:04:12,360
Harness决定AI会不会真正工作

206
00:04:12,360 --> 00:04:13,240
网络安全决定

207
00:04:13,240 --> 00:04:14,840
智能能不能安全地规模化复制

208
00:04:14,840 --> 00:04:17,240
RSI真正回答的是智能生产方式

209
00:04:17,240 --> 00:04:18,040
有没有发生质变

210
00:04:18,040 --> 00:04:19,319
而最值得警惕的一件事是

211
00:04:19,360 --> 00:04:21,399
模型越强靠信息差获得的组织价值

212
00:04:21,399 --> 00:04:22,120
反而越危险

213
00:04:22,120 --> 00:04:22,759
过去职位高

214
00:04:22,759 --> 00:04:23,639
一个重要原因是

215
00:04:23,639 --> 00:04:25,399
你知道的更多过去很多中间层

216
00:04:25,399 --> 00:04:26,279
一个重要价值是

217
00:04:26,279 --> 00:04:27,399
把下面的信息搬上来

218
00:04:27,399 --> 00:04:28,680
再把上面的决定搬下去

219
00:04:28,680 --> 00:04:30,600
但当AI拥有整个组织的context

220
00:04:30,600 --> 00:04:32,319
这个逻辑会第一次被真正挑战

221
00:04:32,319 --> 00:04:34,199
所以AI最先压缩的可能不是岗位

222
00:04:34,199 --> 00:04:35,199
而是层级

223
00:04:35,199 --> 00:04:36,240
最先被重新定价的

224
00:04:36,240 --> 00:04:37,560
可能也不是人的执行力

225
00:04:37,560 --> 00:04:38,959
而是组织里的信息权

226
00:04:38,959 --> 00:04:40,040
未来真正强的企业

227
00:04:40,040 --> 00:04:41,360
可能既不是人最多的企业

228
00:04:41,360 --> 00:04:42,759
也不是拥有最强模型的企业

229
00:04:42,759 --> 00:04:43,160
而是

230
00:04:43,160 --> 00:04:44,439
最先把自己的经验关系

231
00:04:44,439 --> 00:04:45,600
流程 决策和知识

232
00:04:45,600 --> 00:04:46,519
变成机器可以理解

233
00:04:46,519 --> 00:04:47,000
可以调用

234
00:04:47,000 --> 00:04:47,720
可以执行的企业

235
00:04:47,720 --> 00:04:48,800
过去我们数字化的是

236
00:04:48,800 --> 00:04:49,120
业务

237
00:04:49,120 --> 00:04:50,279
下一步我们要数字化的

238
00:04:50,279 --> 00:04:51,480
可能是组织本身

239
00:04:51,480 --> 00:04:52,759
当组织开始变得可计算

240
00:04:52,759 --> 00:04:54,439
AI改变的就不再只是生产力

241
00:04:54,439 --> 00:04:55,120
它开始改变

242
00:04:55,120 --> 00:04:56,600
公司为什么还要这样被组织

