1
00:00:00,000 --> 00:00:04,320
大家好我是Jim

2
00:00:04,320 --> 00:00:05,320
这里是AI Notebook

3
00:00:05,320 --> 00:00:06,500
三天12万颗星

4
00:00:06,500 --> 00:00:07,600
这是Deep Seek

5
00:00:07,600 --> 00:00:09,140
刚公开了一个工具

6
00:00:09,140 --> 00:00:10,000
叫Harness

7
00:00:10,000 --> 00:00:12,020
在GitHub上拿到了收藏数

8
00:00:12,020 --> 00:00:13,300
很多做得很好的专案

9
00:00:13,300 --> 00:00:15,240
做十年也到不了这个数字

10
00:00:15,240 --> 00:00:16,740
让我坐下来把它读完了

11
00:00:16,740 --> 00:00:17,700
不是这个数字

12
00:00:17,700 --> 00:00:20,540
Deep Seek一路走来出的都是模型

13
00:00:20,540 --> 00:00:22,180
这次公开的不是模型

14
00:00:22,180 --> 00:00:24,440
是模型旁边那一整套东西

15
00:00:24,440 --> 00:00:27,880
让模型真的能动手做事的那个执行环境

16
00:00:27,880 --> 00:00:29,880
你家做模型的公司

17
00:00:29,880 --> 00:00:31,680
开始在意模型跑在哪里

18
00:00:31,680 --> 00:00:32,300
被谁管

19
00:00:32,300 --> 00:00:34,000
出事的时候谁喊停

20
00:00:34,000 --> 00:00:35,940
这件事本身就值得看

21
00:00:35,940 --> 00:00:39,000
而且他们连整份架构文件都一起公开了

22
00:00:39,000 --> 00:00:41,040
129行八个章节

23
00:00:41,040 --> 00:00:43,780
里面写着一件你平常看不到的事

24
00:00:43,780 --> 00:00:45,880
一个会自己动手做事的AI

25
00:00:45,880 --> 00:00:47,740
它的骨架应该怎么设计

26
00:00:47,740 --> 00:00:50,040
这种东西通常锁在公司内部

27
00:00:50,040 --> 00:00:51,380
所以今天这支影片

28
00:00:51,380 --> 00:00:54,560
就是要把这份架构仔细研究一遍

29
00:00:54,560 --> 00:00:55,840
里面每一个设计

30
00:00:55,840 --> 00:00:57,300
我都会翻成白话讲给你听

31
00:00:57,300 --> 00:00:59,840
顺便看他跟你已经在用的那几个工具

32
00:00:59,840 --> 00:01:00,420
插在哪

33
00:01:00,420 --> 00:01:00,980
主弯

34
00:01:00,980 --> 00:01:02,940
你会发现这些设计

35
00:01:02,940 --> 00:01:04,500
其实都在回答同一个问题

36
00:01:04,500 --> 00:01:06,020
好先讲清楚

37
00:01:06,020 --> 00:01:06,980
Harness这个字

38
00:01:06,980 --> 00:01:08,480
它原本的意思是马具

39
00:01:08,480 --> 00:01:09,740
就是套在马身上

40
00:01:09,740 --> 00:01:12,060
让你能驾驭它的那套装备

41
00:01:12,060 --> 00:01:13,820
缰绳 马鞍 脚灯

42
00:01:13,820 --> 00:01:15,980
这些东西一个都不会让马跑得更快

43
00:01:15,980 --> 00:01:17,020
但没有它们

44
00:01:17,020 --> 00:01:18,100
你根本上不了马

45
00:01:18,100 --> 00:01:19,480
放到AI这里

46
00:01:19,480 --> 00:01:20,520
模型是那匹马

47
00:01:20,520 --> 00:01:21,180
Harness

48
00:01:21,180 --> 00:01:23,780
就是让你真的能用它做事的那一整套东西

49
00:01:23,780 --> 00:01:24,380
你可能会想

50
00:01:24,380 --> 00:01:26,000
模型不是已经很强了吗

51
00:01:26,000 --> 00:01:27,600
开放权重都放出来了

52
00:01:27,600 --> 00:01:28,840
直接拿来用不就好了

53
00:01:28,840 --> 00:01:29,480
问题是

54
00:01:29,480 --> 00:01:31,800
权重只是一颗大脑的参数档

55
00:01:31,800 --> 00:01:33,600
它不会自己开你的资料夹

56
00:01:33,600 --> 00:01:35,080
不会记得上一部做了什么

57
00:01:35,080 --> 00:01:37,600
也不会在跑坏的时候停下来问你

58
00:01:37,600 --> 00:01:39,240
有一位分析师讲得很好

59
00:01:39,240 --> 00:01:40,680
只有开放权重

60
00:01:40,680 --> 00:01:41,760
没有执行环境

61
00:01:41,760 --> 00:01:42,680
就像有身体

62
00:01:42,680 --> 00:01:43,520
没有神经系统

63
00:01:43,520 --> 00:01:45,240
Harness就是那套神经系统

64
00:01:45,240 --> 00:01:47,120
它管模型怎么收到你的话

65
00:01:47,120 --> 00:01:48,560
工具怎么被呼叫

66
00:01:48,560 --> 00:01:50,240
做过的事怎么被记下来

67
00:01:50,240 --> 00:01:52,160
还有出事的时候谁有权喊停

68
00:01:52,160 --> 00:01:54,160
在打开那份文件之前

69
00:01:54,160 --> 00:01:56,440
先花一分钟搞清楚谁是谁

70
00:01:56,440 --> 00:01:59,160
Cloudcode Codex OpenCloud Hermes

71
00:01:59,160 --> 00:02:00,440
它们全部都是Harness

72
00:02:00,440 --> 00:02:02,560
但差别不只是你在哪里跟它讲话

73
00:02:02,560 --> 00:02:03,880
我用三个问题来分

74
00:02:03,880 --> 00:02:05,600
第一个问题也是最根本的

75
00:02:05,600 --> 00:02:06,640
它想解决什么

76
00:02:06,640 --> 00:02:07,600
这不用猜

77
00:02:07,600 --> 00:02:09,920
每一个专案自己那句介绍就是答案

78
00:02:09,920 --> 00:02:12,160
Cloudcode跟Codex解决的是

79
00:02:12,160 --> 00:02:14,240
模型很强却碰不到你的专案

80
00:02:14,240 --> 00:02:16,280
所以把模型接近你的终端机

81
00:02:16,280 --> 00:02:17,560
直接读你的程式码

82
00:02:17,560 --> 00:02:19,000
OpenCloud解决的是

83
00:02:19,000 --> 00:02:20,280
助理是别人

84
00:02:20,280 --> 00:02:21,640
资料也在别人那里

85
00:02:21,640 --> 00:02:23,400
所以它跑在你自己的装置上

86
00:02:23,400 --> 00:02:25,000
专案标签直接挂着一句

87
00:02:25,000 --> 00:02:26,400
你的资料你自己拥有

88
00:02:26,400 --> 00:02:28,520
Harness解决的是Agent

89
00:02:28,520 --> 00:02:29,760
每次都从零开始

90
00:02:29,760 --> 00:02:30,880
什么都不记得

91
00:02:30,880 --> 00:02:32,320
所以它介绍只有一句

92
00:02:32,320 --> 00:02:33,800
会跟着你一起长大

93
00:02:33,800 --> 00:02:35,240
那Harness解决的是

94
00:02:35,240 --> 00:02:36,720
你想改它却改不动

95
00:02:36,720 --> 00:02:37,880
第二个问题

96
00:02:37,880 --> 00:02:38,760
它什么时候动

97
00:02:38,760 --> 00:02:40,760
前面两个要你叫它才动

98
00:02:40,760 --> 00:02:41,680
做完就结束

99
00:02:41,680 --> 00:02:43,520
后面两个是一直在线的Harness

100
00:02:43,520 --> 00:02:44,920
甚至会自己按表操课

101
00:02:44,920 --> 00:02:45,600
你不用开口

102
00:02:45,600 --> 00:02:46,480
第三个问题

103
00:02:46,480 --> 00:02:48,240
也是我觉得最该问的

104
00:02:48,240 --> 00:02:49,520
它碰得到什么

105
00:02:49,520 --> 00:02:50,880
OpenCloud的说明档

106
00:02:50,880 --> 00:02:51,760
自己写着

107
00:02:51,760 --> 00:02:54,240
工具预设是直接跑在你的主机上的

108
00:02:54,240 --> 00:02:56,160
除非你自己去设定沙香

109
00:02:56,160 --> 00:02:58,400
Harness可以整个丢到云端的机器上跑

110
00:02:58,400 --> 00:02:59,600
不用碰你的笔电

111
00:02:59,600 --> 00:03:00,960
那Harness呢

112
00:03:00,960 --> 00:03:02,880
他把这件事做成一道关卡

113
00:03:02,880 --> 00:03:04,000
等一下你会看到

114
00:03:04,000 --> 00:03:05,360
每一次工具呼叫

115
00:03:05,360 --> 00:03:07,440
都要先过权限跟沙香

116
00:03:07,440 --> 00:03:08,720
它们差多少呢

117
00:03:08,720 --> 00:03:10,560
你看一下收藏数就知道了

118
00:03:10,560 --> 00:03:11,520
知道形状之后

119
00:03:11,520 --> 00:03:12,680
还有一个更重要的差别

120
00:03:12,680 --> 00:03:13,600
你能改到哪一层

121
00:03:13,600 --> 00:03:14,400
刚刚说了

122
00:03:14,400 --> 00:03:15,840
这几个其实是同一种东西

123
00:03:15,840 --> 00:03:18,080
所以真正该问的不是谁比较强

124
00:03:18,080 --> 00:03:19,280
是你能改到多深

125
00:03:19,280 --> 00:03:20,640
我用房子来比

126
00:03:20,640 --> 00:03:22,080
第一层像租来的房子

127
00:03:22,080 --> 00:03:22,840
你可以摆家具

128
00:03:22,840 --> 00:03:23,680
可以挂画

129
00:03:23,680 --> 00:03:25,040
但你不能打掉墙

130
00:03:25,040 --> 00:03:26,320
换成工具就是

131
00:03:26,320 --> 00:03:27,120
可以加工具

132
00:03:27,120 --> 00:03:28,240
可以加提示词

133
00:03:28,240 --> 00:03:29,520
但那个一步一步驱动

134
00:03:29,520 --> 00:03:30,800
AI做事的主回圈

135
00:03:30,800 --> 00:03:31,760
是它的

136
00:03:31,760 --> 00:03:32,880
你只能设定第二层

137
00:03:32,880 --> 00:03:34,880
像把整栋旧房子买下来自己改

138
00:03:34,880 --> 00:03:35,920
原始码都给你

139
00:03:35,920 --> 00:03:37,200
Codex就是这种

140
00:03:37,200 --> 00:03:38,880
你想打掉哪一面墙都行

141
00:03:38,880 --> 00:03:39,760
代价是

142
00:03:39,760 --> 00:03:40,520
从那天起

143
00:03:40,520 --> 00:03:42,800
这栋房子的保养就变你的事了

144
00:03:42,800 --> 00:03:44,880
第三层才是Harness站的地方

145
00:03:44,880 --> 00:03:47,840
这栋房子的墙本身就是一片一片可以换的

146
00:03:47,840 --> 00:03:49,200
换墙不用动骨架

147
00:03:49,200 --> 00:03:50,320
文件里写的很白

148
00:03:50,320 --> 00:03:52,640
连那个主回圈本身都是外挂

149
00:03:52,640 --> 00:03:54,960
至于Longraph那一类的开发套件

150
00:03:54,960 --> 00:03:56,400
它根本不在这个梯子上

151
00:03:56,400 --> 00:03:57,760
它连房子都不是

152
00:03:57,760 --> 00:03:58,720
而是建材

153
00:03:58,720 --> 00:03:59,840
有意思的是

154
00:03:59,840 --> 00:04:02,240
Hermis跟Harness压的宝不一样

155
00:04:02,240 --> 00:04:03,280
Hermis堵

156
00:04:03,280 --> 00:04:04,320
它会自己变墙

157
00:04:04,320 --> 00:04:07,600
做完一件事会把经验写成一份技能档存起来

158
00:04:07,600 --> 00:04:09,720
Harness堵的是你可以自己动手改

159
00:04:09,720 --> 00:04:11,520
一个把赌注压在时间

160
00:04:11,520 --> 00:04:12,800
一个压在你手上

161
00:04:12,800 --> 00:04:13,760
一直讲完了

162
00:04:13,760 --> 00:04:15,520
我们回去读那份文件

163
00:04:15,520 --> 00:04:17,400
看它怎么做到第三层

164
00:04:17,400 --> 00:04:19,320
这本文件叫架构说明

165
00:04:19,320 --> 00:04:20,400
129行

166
00:04:20,400 --> 00:04:21,240
8个章节

167
00:04:21,240 --> 00:04:21,720
不长

168
00:04:21,720 --> 00:04:23,280
你午休就读得完

169
00:04:23,280 --> 00:04:24,840
但他第一句话就很有态度

170
00:04:24,840 --> 00:04:25,320
他说

171
00:04:25,320 --> 00:04:27,000
在你动任何程式之前

172
00:04:27,000 --> 00:04:27,960
先读这个

173
00:04:27,960 --> 00:04:30,200
他预设你已经懂底下那个框架

174
00:04:30,200 --> 00:04:30,840
如果不懂

175
00:04:30,840 --> 00:04:32,640
他把入门的连结放在那里

176
00:04:32,640 --> 00:04:34,440
然后第7行有一句话

177
00:04:34,440 --> 00:04:36,760
我觉得是整份文件最有意思的地方

178
00:04:36,760 --> 00:04:38,680
他说我们建议你用一个agent

179
00:04:38,680 --> 00:04:40,160
去读这份程式吗

180
00:04:40,160 --> 00:04:41,440
理解它的架构

181
00:04:41,440 --> 00:04:42,240
你发现了吗

182
00:04:42,240 --> 00:04:44,440
这份文件不只是写给人看的

183
00:04:44,440 --> 00:04:46,440
也是写给AI看的

184
00:04:46,440 --> 00:04:47,920
接下来我就一个一个带你看

185
00:04:47,920 --> 00:04:49,520
全部翻成白话

186
00:04:49,520 --> 00:04:50,680
有意思的是

187
00:04:50,680 --> 00:04:53,640
Deepseek没有自己从零造这套股价

188
00:04:53,640 --> 00:04:56,080
他们挑了一个已经存在4年的社群专案

189
00:04:56,080 --> 00:04:57,000
叫Cordis

190
00:04:57,000 --> 00:04:59,000
这个专案在被选中之前

191
00:04:59,000 --> 00:05:00,000
收藏数是4000

192
00:05:00,000 --> 00:05:03,600
200多 4200多在开放原始码的世界里不算小

193
00:05:03,600 --> 00:05:05,300
但也绝对称不上有名

194
00:05:05,600 --> 00:05:06,700
很多人根本没听过

195
00:05:07,200 --> 00:05:09,300
就是一个安静做了4年的小专案

196
00:05:09,300 --> 00:05:10,400
结果他一息之间

197
00:05:10,700 --> 00:05:12,600
变成一个12万星专案的地基

198
00:05:12,800 --> 00:05:14,600
Cordis自己的介绍只有一句话

199
00:05:14,900 --> 00:05:16,900
时空可组合性的原框架

200
00:05:17,500 --> 00:05:18,800
这句话听起来很悬

201
00:05:19,100 --> 00:05:20,700
但他想解决的事情很具体

202
00:05:21,100 --> 00:05:23,700
怎么让一个系统的每一块都能被换掉

203
00:05:23,700 --> 00:05:25,000
被叠加被拿掉

204
00:05:25,000 --> 00:05:26,400
而不会把其他部分弄坏

205
00:05:26,800 --> 00:05:27,900
这件事听起来简单

206
00:05:27,900 --> 00:05:30,300
实际上是软体工程里最难的问题之一

207
00:05:30,700 --> 00:05:32,100
Lipsic看上的就是这个

208
00:05:32,500 --> 00:05:34,700
所以Harness的第一句设计哲学是

209
00:05:35,000 --> 00:05:36,200
所有东西都是外挂

210
00:05:36,200 --> 00:05:38,800
外挂白话讲就是可以自由装上

211
00:05:39,100 --> 00:05:40,000
拔掉的零件

212
00:05:40,500 --> 00:05:43,100
文件一开头就在讲Cordis带来什么

213
00:05:43,400 --> 00:05:46,400
要听懂他得先看一般的软体长什么样

214
00:05:46,600 --> 00:05:47,800
中间有一个核心

215
00:05:48,000 --> 00:05:49,300
旁边挂一些扩充点

216
00:05:49,300 --> 00:05:51,300
你想加功能可以挂在扩充点上

217
00:05:51,300 --> 00:05:52,100
这没问题

218
00:05:52,700 --> 00:05:55,100
但只要你要改的事情核心没有预留位置

219
00:05:55,100 --> 00:05:56,400
你就得动核心

220
00:05:56,400 --> 00:05:57,700
动核心的代价很大

221
00:05:57,900 --> 00:05:59,200
因为所有人都靠他活

222
00:05:59,600 --> 00:06:01,700
你改坏一个地方全部一起坏

223
00:06:02,100 --> 00:06:04,200
而且久了之后会出现一个现象

224
00:06:04,400 --> 00:06:06,300
核心变成没人敢碰的东西

225
00:06:06,600 --> 00:06:09,200
大家宁可绕路绕出一堆奇怪的补丁

226
00:06:09,400 --> 00:06:11,100
Harness的文件里有一句话

227
00:06:11,400 --> 00:06:12,300
讲的很直接

228
00:06:12,500 --> 00:06:15,300
这里没有一个有特权的核心让你去修

229
00:06:15,600 --> 00:06:19,000
扩充他的方式是把你的外挂挂在其他外挂旁边

230
00:06:19,000 --> 00:06:19,900
这是什么意思呢

231
00:06:20,100 --> 00:06:24,900
就是连模型转接连工具清单连那个一步一步驱动

232
00:06:24,900 --> 00:06:26,300
AI做事的组绘圈

233
00:06:26,300 --> 00:06:27,300
本身都是外挂

234
00:06:27,500 --> 00:06:28,800
所以你想宽调组维圈

235
00:06:28,800 --> 00:06:29,700
你不是去改他

236
00:06:29,900 --> 00:06:32,300
你是写一个自己的挂上去把原本那个拿掉

237
00:06:32,500 --> 00:06:33,100
像乐高

238
00:06:33,500 --> 00:06:34,700
像一体成型的家具

239
00:06:35,200 --> 00:06:36,000
刚刚那句话

240
00:06:36,200 --> 00:06:40,600
还有后半段注册上去的东西在外挂被拿掉的时候要整个收回来

241
00:06:40,800 --> 00:06:41,600
这是什么意思呢

242
00:06:42,000 --> 00:06:43,100
你有沒有装过那种软体

243
00:06:43,100 --> 00:06:45,600
反安装之后电脑里还留一堆垃圾

244
00:06:45,800 --> 00:06:47,600
设定被改过的地方没还原

245
00:06:47,900 --> 00:06:50,000
资料夹里留着一些你不敢删的档案

246
00:06:50,400 --> 00:06:51,600
开机还是比以前慢

247
00:06:51,900 --> 00:06:54,300
就是装的时候没有留下海的方法

248
00:06:54,400 --> 00:06:56,200
软体世界里这种事非常普遍

249
00:06:56,200 --> 00:06:57,400
因为装很简单

250
00:06:57,400 --> 00:06:58,600
移除很麻烦

251
00:06:58,900 --> 00:07:00,600
大部分人只做前面那一半

252
00:07:00,800 --> 00:07:04,100
Harness的规则是每一个注册都必须带一个移除的方法

253
00:07:04,300 --> 00:07:05,500
你加一段提示词

254
00:07:05,700 --> 00:07:06,600
加一个工具

255
00:07:06,700 --> 00:07:07,600
加一个监听器

256
00:07:07,600 --> 00:07:09,500
都是透过同一个机制装上去

257
00:07:09,700 --> 00:07:11,700
每个机制会记住怎么还原

258
00:07:11,700 --> 00:07:13,300
所以外挂被拿掉的时候

259
00:07:13,300 --> 00:07:16,000
他做过的每件事会照顺序倒着收回来

260
00:07:16,300 --> 00:07:17,100
像倒带一样

261
00:07:17,100 --> 00:07:18,300
这条看起来很无聊

262
00:07:18,300 --> 00:07:20,200
但他是热插把的真正原因

263
00:07:20,500 --> 00:07:21,800
你会看到很多框架

264
00:07:22,300 --> 00:07:23,600
宣称可以热插把

265
00:07:23,600 --> 00:07:25,500
结果把玩要重新启动才正常

266
00:07:25,500 --> 00:07:27,000
那就是这一条没做到

267
00:07:27,400 --> 00:07:28,400
下一张讲的是

268
00:07:28,600 --> 00:07:30,800
你跑起来的那个东西到底怎么被组出来的

269
00:07:30,900 --> 00:07:31,500
答案是

270
00:07:31,600 --> 00:07:33,700
他是一叠设定层堆出来的

271
00:07:33,800 --> 00:07:35,400
最底下那一层叫BASS

272
00:07:35,400 --> 00:07:36,000
模型

273
00:07:36,200 --> 00:07:36,800
转接

274
00:07:36,900 --> 00:07:37,500
工具

275
00:07:37,600 --> 00:07:38,300
储存

276
00:07:38,300 --> 00:07:38,900
杀箱

277
00:07:38,900 --> 00:07:39,600
权限

278
00:07:39,600 --> 00:07:40,500
全部在里面

279
00:07:40,800 --> 00:07:42,800
每一种组合都从这一层开始

280
00:07:43,300 --> 00:07:44,400
上面可以再叠

281
00:07:44,500 --> 00:07:46,600
你要网页界面就叠网页那一层

282
00:07:46,900 --> 00:07:48,800
你只要跑完一次就收工

283
00:07:48,900 --> 00:07:50,200
就叠无界面那一层

284
00:07:50,300 --> 00:07:52,000
再上面是你自己的设定档

285
00:07:52,200 --> 00:07:54,400
最上面是你临时加的那一行参数

286
00:07:54,400 --> 00:07:55,800
这里有一个指令我很喜欢

287
00:07:56,000 --> 00:07:56,800
你可以叫他

288
00:07:56,900 --> 00:07:59,300
把组好的整颗树印出来文件讲的很白

289
00:07:59,300 --> 00:08:00,500
印出来任何一行

290
00:08:00,700 --> 00:08:02,600
你都可以用自己的设定盖掉

291
00:08:02,700 --> 00:08:04,100
这就像一份摊开合约

292
00:08:04,100 --> 00:08:05,400
每一条你都看得到

293
00:08:05,500 --> 00:08:07,300
每一条你都可以划掉重写

294
00:08:07,600 --> 00:08:09,000
堵出来的那棵树上

295
00:08:09,500 --> 00:08:11,100
每一块都站住一个名字

296
00:08:11,200 --> 00:08:13,000
上面的人要用下面那一块

297
00:08:13,100 --> 00:08:14,300
靠的就是这个名字

298
00:08:14,500 --> 00:08:15,600
这听起来很技术

299
00:08:15,700 --> 00:08:17,000
但它其实就是电话簿

300
00:08:17,500 --> 00:08:19,300
想像你家水管漏水

301
00:08:19,500 --> 00:08:21,500
一种做法是你认识一个师傅

302
00:08:21,700 --> 00:08:22,900
有他的电话你打给他

303
00:08:23,100 --> 00:08:24,400
他很好很可靠

304
00:08:24,700 --> 00:08:26,400
直到有一天他搬家了或者退休了

305
00:08:26,400 --> 00:08:27,100
你就完蛋了

306
00:08:27,300 --> 00:08:28,600
你得重新认识一个人

307
00:08:28,700 --> 00:08:30,800
另一种做法是你翻电话簿

308
00:08:30,900 --> 00:08:32,400
找到水电这个分类

309
00:08:32,600 --> 00:08:33,800
打上面的号码

310
00:08:33,900 --> 00:08:34,800
谁接不重要

311
00:08:34,800 --> 00:08:37,300
重要的是水电这个分类永远在

312
00:08:37,500 --> 00:08:39,100
Harness就是后面这种

313
00:08:39,600 --> 00:08:40,300
每个服务

314
00:08:40,400 --> 00:08:42,000
站住一个固定的名字

315
00:08:42,100 --> 00:08:43,100
工具叫Tools

316
00:08:43,200 --> 00:08:44,500
模型叫LM

317
00:08:44,600 --> 00:08:45,700
地图叫Sessions

318
00:08:45,900 --> 00:08:48,200
别的外挂用工具是去找Tools

319
00:08:48,300 --> 00:08:48,900
这个名字

320
00:08:49,300 --> 00:08:51,400
而不是去引用某一份具体的程式码

321
00:08:51,600 --> 00:08:52,600
好处是什么呢

322
00:08:52,700 --> 00:08:54,900
你把它底下换成完全不同的实作

323
00:08:55,000 --> 00:08:55,500
上面

324
00:08:55,600 --> 00:08:56,600
所有人都不用改

325
00:08:56,800 --> 00:08:58,000
而且还有一个附带好处

326
00:08:58,100 --> 00:08:59,200
开机顺序不用你排

327
00:08:59,400 --> 00:09:00,600
每个外挂只要宣告

328
00:09:00,700 --> 00:09:01,700
我需要哪些服务

329
00:09:01,700 --> 00:09:03,800
系统就会自己算出谁该先启动

330
00:09:03,800 --> 00:09:04,700
谁该后启动

331
00:09:05,200 --> 00:09:06,300
想完五条抽象的

332
00:09:06,400 --> 00:09:07,600
我们看一个具体的东西

333
00:09:07,600 --> 00:09:08,400
一个工具

334
00:09:08,500 --> 00:09:09,300
被呼叫的时候

335
00:09:09,300 --> 00:09:10,400
到底发生什么事

336
00:09:10,600 --> 00:09:12,200
一条路比你想的长很多

337
00:09:12,400 --> 00:09:14,300
模型说我要跑这个指令

338
00:09:14,400 --> 00:09:15,600
第一件事不是去跑

339
00:09:15,900 --> 00:09:18,000
是先把这个呼叫写进记录

340
00:09:18,100 --> 00:09:19,000
注意顺序

341
00:09:19,100 --> 00:09:20,500
是执行之前就写

342
00:09:20,500 --> 00:09:22,000
所以就算等一下整个当掉

343
00:09:22,300 --> 00:09:23,900
还是查得到当时想做什么

344
00:09:24,000 --> 00:09:25,300
接着进到第一道关卡

345
00:09:25,400 --> 00:09:27,500
权限跟杀香在这里决定放行

346
00:09:27,600 --> 00:09:28,200
挡一下

347
00:09:28,400 --> 00:09:29,200
还是问人

348
00:09:29,200 --> 00:09:31,800
要问人的话会跳出一个一次性的确认

349
00:09:31,900 --> 00:09:33,400
这里有一个设计我很喜欢

350
00:09:33,500 --> 00:09:34,500
如果问不到人

351
00:09:34,900 --> 00:09:35,400
比如说

352
00:09:35,600 --> 00:09:36,600
没有界面可以问

353
00:09:36,600 --> 00:09:38,100
或者根本没人在电脑前面

354
00:09:38,200 --> 00:09:39,100
那就是算拒绝

355
00:09:39,100 --> 00:09:40,000
而不是算通过

356
00:09:40,300 --> 00:09:41,500
这个选择看起来很小

357
00:09:41,800 --> 00:09:42,900
但他决定了一件事

358
00:09:43,200 --> 00:09:44,300
没人看的时候

359
00:09:44,400 --> 00:09:45,800
这个系统是保守的

360
00:09:45,800 --> 00:09:46,600
还是放纵的

361
00:09:46,900 --> 00:09:49,200
然后才轮到工具本体真的跑

362
00:09:49,300 --> 00:09:50,400
跑完还没结束

363
00:09:50,500 --> 00:09:52,100
还有一关可以检查结果

364
00:09:52,200 --> 00:09:52,900
可以挡下

365
00:09:53,000 --> 00:09:53,700
可以换掉

366
00:09:53,900 --> 00:09:55,100
可以补一段话进去

367
00:09:55,300 --> 00:09:56,400
最后才冻结成一个

368
00:09:56,400 --> 00:09:57,700
谁都不能再改的结果

369
00:09:57,700 --> 00:10:00,100
再往下一张叫工作阶段记录

370
00:10:00,000 --> 00:10:02,600
里面有我觉得整份文件最漂亮的一句话

371
00:10:02,600 --> 00:10:03,900
模型看得到的

372
00:10:03,900 --> 00:10:05,200
一定被记下来

373
00:10:05,200 --> 00:10:06,740
意思是任何一个字

374
00:10:06,740 --> 00:10:09,340
只要他进到了送给模型的那个请求礼

375
00:10:09,340 --> 00:10:12,080
他就必须能从记录里重新拼回来

376
00:10:12,080 --> 00:10:13,880
而且他们不是写在文件上就算了

377
00:10:13,880 --> 00:10:14,960
他们写了一个检查

378
00:10:14,960 --> 00:10:16,560
在跑的时候盯着这件事

379
00:10:16,560 --> 00:10:17,940
做不到就直接报错

380
00:10:17,940 --> 00:10:19,080
为什么要做到这么硬呢

381
00:10:19,080 --> 00:10:19,840
因为这样一来

382
00:10:19,840 --> 00:10:20,580
很多

383
00:10:20,580 --> 00:10:22,760
你以为要各自开发的功能

384
00:10:22,760 --> 00:10:24,620
其实都是同一条记录的衍生品

385
00:10:24,620 --> 00:10:27,020
你想从某一步分岔出去重跑

386
00:10:27,020 --> 00:10:28,680
可以想关掉电脑

387
00:10:28,680 --> 00:10:29,520
明天接着做

388
00:10:29,520 --> 00:10:29,820
可以

389
00:10:29,820 --> 00:10:31,700
你想知道昨天那个AI

390
00:10:31,700 --> 00:10:32,860
为什么会删掉你的档案

391
00:10:32,860 --> 00:10:34,260
把记录拉出来

392
00:10:34,260 --> 00:10:36,140
那一刻他眼前有什么

393
00:10:36,140 --> 00:10:37,280
你一个字都不会漏

394
00:10:37,280 --> 00:10:39,180
这就像法院的逐字记录

395
00:10:39,180 --> 00:10:40,900
我是散毁之后补写的会议纪要

396
00:10:40,900 --> 00:10:42,260
是当下就留了底

397
00:10:42,260 --> 00:10:43,400
差别在哪里呢

398
00:10:43,400 --> 00:10:45,520
补写的东西你只能选择相信

399
00:10:45,520 --> 00:10:46,580
当下留了底

400
00:10:46,580 --> 00:10:47,120
你可以查

401
00:10:47,120 --> 00:10:49,640
这两件事在出事的时候差很多

402
00:10:49,640 --> 00:10:51,860
下一张叫能力接缝

403
00:10:51,860 --> 00:10:53,980
文件给了接缝一个很严格的定义

404
00:10:53,980 --> 00:10:55,600
要有三个角色才算数

405
00:10:55,600 --> 00:10:56,600
哪三个呢

406
00:10:56,600 --> 00:10:58,960
一个负责定义这个能力长什么样

407
00:10:58,960 --> 00:11:00,360
一个负责真的把它做出来

408
00:11:00,360 --> 00:11:01,480
一个负责使用它

409
00:11:01,480 --> 00:11:03,400
文件里特别强调一句话

410
00:11:03,400 --> 00:11:05,080
只有其中一个角色不算数

411
00:11:05,080 --> 00:11:06,160
这是在讲什么

412
00:11:06,160 --> 00:11:08,340
很多系统会说这里可以换

413
00:11:08,340 --> 00:11:09,680
结果你真的去换的时候

414
00:11:09,680 --> 00:11:11,800
才发现界面根本没有被单独定义出来

415
00:11:11,800 --> 00:11:14,840
使用的人其实直接绑死在某一份实作上

416
00:11:14,840 --> 00:11:15,760
那就不是能换

417
00:11:15,760 --> 00:11:17,200
那只是还没有人去换过而已

418
00:11:17,200 --> 00:11:18,440
做对了会怎样呢

419
00:11:18,440 --> 00:11:19,720
文件里举了一个例子

420
00:11:19,720 --> 00:11:21,440
档案系统跟跑指令

421
00:11:21,440 --> 00:11:23,680
这两件事共用同一个执行环境

422
00:11:23,680 --> 00:11:26,800
所以当你把它指向一台远端的杀箱机器

423
00:11:26,800 --> 00:11:28,640
那些跑指令的工具

424
00:11:28,640 --> 00:11:29,840
那个终端机

425
00:11:29,840 --> 00:11:32,000
连程式码补全都一起搬过去了

426
00:11:32,000 --> 00:11:33,440
没有任何一个要另外改

427
00:11:33,440 --> 00:11:35,200
这样插座规格发电厂

428
00:11:35,200 --> 00:11:36,600
从烧煤换成太阳能

429
00:11:36,600 --> 00:11:38,280
你家的电器一台都不用换

430
00:11:38,280 --> 00:11:39,920
因为中间那个插座的形状

431
00:11:39,920 --> 00:11:41,600
早就被单独定义出来了

432
00:11:41,600 --> 00:11:45,160
文件里我觉得最实用的一段叫做

433
00:11:45,160 --> 00:11:46,960
新的行为该放哪里

434
00:11:46,960 --> 00:11:48,640
这一张就是一张对照表

435
00:11:48,640 --> 00:11:50,120
左边是你想做的事

436
00:11:50,120 --> 00:11:51,400
右边是你该挂在哪里

437
00:11:51,400 --> 00:11:53,040
你想加一个模型供应商

438
00:11:53,040 --> 00:11:55,760
QLM那个名字上注册一个转接器

439
00:11:55,760 --> 00:11:57,560
你想给AI一个新工具

440
00:11:57,560 --> 00:11:58,880
去Tool上注册

441
00:11:58,880 --> 00:12:01,320
它的说明会自动被放进给模型的提示里

442
00:12:01,320 --> 00:12:02,880
你想拦截每一次请求

443
00:12:02,880 --> 00:12:03,760
用事件

444
00:12:03,760 --> 00:12:06,080
你想加一个关掉电脑还在的状态

445
00:12:06,080 --> 00:12:07,720
去扩充事件的行别

446
00:12:07,720 --> 00:12:09,160
这一张开头还有一句话

447
00:12:09,160 --> 00:12:11,560
他说如果你改的是主回圈本身

448
00:12:11,560 --> 00:12:13,440
你要顺手更新这张表

449
00:12:13,440 --> 00:12:14,400
你听出来了吗

450
00:12:14,400 --> 00:12:16,600
这张表就是这个专案的地图

451
00:12:16,600 --> 00:12:18,800
而他规定地图不准过期

452
00:12:18,800 --> 00:12:20,960
架构文件旁边还有一份

453
00:12:20,960 --> 00:12:22,320
不讲怎么设计

454
00:12:22,320 --> 00:12:23,400
讲怎么不出事

455
00:12:23,400 --> 00:12:25,360
标题叫防御性模式

456
00:12:25,360 --> 00:12:26,400
开头就说

457
00:12:26,400 --> 00:12:29,440
好下面每一条都是真的发生过

458
00:12:29,440 --> 00:12:30,960
或者差一点就发生的bug

459
00:12:30,960 --> 00:12:32,960
写成规则来防止它再来一次

460
00:12:32,960 --> 00:12:34,360
我很喜欢这种文件

461
00:12:34,360 --> 00:12:35,680
因为他不是在讲理想

462
00:12:35,680 --> 00:12:36,720
是在讲伤疤

463
00:12:36,720 --> 00:12:38,000
我挑两条给你

464
00:12:38,000 --> 00:12:38,640
第一条

465
00:12:38,640 --> 00:12:40,800
结果可能同时是好几件事

466
00:12:40,800 --> 00:12:42,640
一个程式可以同时于时

467
00:12:42,640 --> 00:12:44,040
又回报执行成功

468
00:12:44,040 --> 00:12:46,680
因为他自己把中断的讯号拦下来了

469
00:12:46,680 --> 00:12:49,040
所以于时讯号结束代码要各报各的

470
00:12:49,040 --> 00:12:50,800
不可以把一个藏在另一个底下

471
00:12:50,800 --> 00:12:52,840
不然呼叫的人会把一个被砍断的执行

472
00:12:52,840 --> 00:12:54,240
读成一次干净的成功

473
00:12:54,240 --> 00:12:55,840
第二条更严格关闭

474
00:12:55,840 --> 00:12:57,560
要关到真的静下来

475
00:12:57,560 --> 00:12:58,520
不是喊一声就走

476
00:12:58,520 --> 00:13:00,080
你送出一个结束的指令

477
00:13:00,080 --> 00:13:00,920
然后就回传

478
00:13:00,920 --> 00:13:02,880
那些子程式可能还活着

479
00:13:02,880 --> 00:13:04,000
正确做法是

480
00:13:04,000 --> 00:13:05,320
等他们真的死透

481
00:13:05,320 --> 00:13:07,040
而且要先把监听关掉再杀

482
00:13:07,040 --> 00:13:08,880
这样晚一步才回来的通知

483
00:13:08,880 --> 00:13:10,000
才不会到处乱跑

484
00:13:10,000 --> 00:13:11,160
所以讲完了

485
00:13:11,160 --> 00:13:12,920
我们看它实际跑起来是什么样子

486
00:13:12,920 --> 00:13:14,840
装好之后你可以选四种模式

487
00:13:14,840 --> 00:13:16,000
第一种是完整版

488
00:13:16,000 --> 00:13:16,680
档案

489
00:13:16,680 --> 00:13:17,740
指定搜寻规划

490
00:13:17,740 --> 00:13:18,840
还有直代理全开

491
00:13:18,840 --> 00:13:20,240
就是你预期一个agent

492
00:13:20,240 --> 00:13:21,160
该有的样子

493
00:13:21,160 --> 00:13:22,480
第二种比较特别

494
00:13:22,480 --> 00:13:24,040
它把工具包成程式码

495
00:13:24,040 --> 00:13:25,800
让AI用写程式的方式

496
00:13:25,800 --> 00:13:27,280
一次呼叫好几个工具

497
00:13:27,280 --> 00:13:28,560
我不是一轮一轮问

498
00:13:28,560 --> 00:13:29,840
这样省很多来回

499
00:13:29,840 --> 00:13:31,380
第三种是极简

500
00:13:31,380 --> 00:13:33,280
只给指令列跟编辑器

501
00:13:33,280 --> 00:13:35,040
适合你想看清楚

502
00:13:35,040 --> 00:13:36,560
模型自己到底有多少能耐

503
00:13:36,560 --> 00:13:38,400
第四种叫创造者模式

504
00:13:38,400 --> 00:13:39,160
这个要小心

505
00:13:39,160 --> 00:13:41,280
它给AI的工具是

506
00:13:41,280 --> 00:13:43,100
可以回头改这个系统自己的

507
00:13:43,100 --> 00:13:43,780
也就是说

508
00:13:43,780 --> 00:13:45,680
AI当场写出来的程式码

509
00:13:45,680 --> 00:13:47,880
会直接在正在跑的系统上执行

510
00:13:47,880 --> 00:13:49,340
官方自己在文件上

511
00:13:49,340 --> 00:13:50,820
标了高度信任的警告

512
00:13:50,820 --> 00:13:52,720
这是很大胆的设计

513
00:13:52,720 --> 00:13:53,880
也是很危险的设计

514
00:13:53,880 --> 00:13:55,200
我建议你先不要碰

515
00:13:55,200 --> 00:13:56,080
好

516
00:13:56,080 --> 00:13:57,260
讲到这里都很漂亮

517
00:13:57,260 --> 00:13:59,120
但我要老实跟你说几件事

518
00:13:59,120 --> 00:14:00,960
不然这支影片就变成广告了

519
00:14:00,960 --> 00:14:01,360
第一

520
00:14:01,360 --> 00:14:03,100
它现在是开发者预览版

521
00:14:03,100 --> 00:14:04,960
说明档用全大写写着

522
00:14:04,960 --> 00:14:06,360
一定会有不相容的改动

523
00:14:06,360 --> 00:14:07,580
你今天写的外挂

524
00:14:07,580 --> 00:14:09,280
下个月可能就跑不起来

525
00:14:09,280 --> 00:14:10,440
所以你要锁住版本

526
00:14:10,440 --> 00:14:10,960
第二

527
00:14:10,960 --> 00:14:12,600
它内建的那个沙箱

528
00:14:12,600 --> 00:14:14,100
文件上自己写着

529
00:14:14,100 --> 00:14:14,880
网络不受限

530
00:14:14,880 --> 00:14:16,300
其他程式看得到

531
00:14:16,300 --> 00:14:17,940
所以那不是真正的隔离

532
00:14:17,940 --> 00:14:19,860
你不能把它当成安全边界

533
00:14:19,860 --> 00:14:22,320
该关的还是要用真的虚拟机关

534
00:14:22,320 --> 00:14:22,900
第三

535
00:14:22,900 --> 00:14:24,500
它跑起来会开一个

536
00:14:24,500 --> 00:14:25,640
本机的网页界面

537
00:14:25,640 --> 00:14:27,480
很多人看到本机两个字会误会

538
00:14:27,480 --> 00:14:29,280
以为模型在自己电脑上跑

539
00:14:29,280 --> 00:14:30,060
不是的

540
00:14:30,060 --> 00:14:32,600
那只是控制台模型还是在远端

541
00:14:32,600 --> 00:14:33,600
还是要填金钥

542
00:14:33,600 --> 00:14:34,440
还是要付钱

543
00:14:34,440 --> 00:14:37,400
然后付钱就要讲另一半起了

544
00:14:37,400 --> 00:14:38,560
股价免费送你

545
00:14:38,560 --> 00:14:39,740
但模型还是要钱

546
00:14:39,740 --> 00:14:42,080
而且agent这种工作的账单形状

547
00:14:42,080 --> 00:14:43,580
跟你平常聊天完全不一样

548
00:14:43,580 --> 00:14:46,160
完全在一个叫快取命中的东西

549
00:14:46,160 --> 00:14:47,880
你跟AI讲话的时候

550
00:14:47,880 --> 00:14:49,560
前面那大段背景

551
00:14:49,560 --> 00:14:50,460
通常每次都一样

552
00:14:50,460 --> 00:14:51,740
系统会把它存起来

553
00:14:51,740 --> 00:14:52,820
第二次就不用重算

554
00:14:52,820 --> 00:14:54,240
所以算你比较便宜

555
00:14:54,240 --> 00:14:56,780
你可以想成搭同一条路线的月票

556
00:14:56,780 --> 00:14:58,440
走越多次越划算

557
00:14:58,440 --> 00:14:59,660
便宜多少呢

558
00:15:00,000 --> 00:15:01,180
官方的价目表

559
00:15:01,180 --> 00:15:02,560
同样是100万个字

560
00:15:02,560 --> 00:15:05,140
系统存过的跟没存过的差31倍

561
00:15:05,140 --> 00:15:08,040
而Agent刚好就是最靠这个活的工作

562
00:15:08,040 --> 00:15:08,960
因为他一轮一轮跑

563
00:15:08,960 --> 00:15:11,800
每一轮都要把前面做过的事重新送一遍

564
00:15:11,800 --> 00:15:13,640
他就是那个天天走同一条路的人

565
00:15:13,640 --> 00:15:15,600
所以真正决定你账单的

566
00:15:15,600 --> 00:15:17,100
不是他都会写程式

567
00:15:17,100 --> 00:15:18,840
是你又没有让他一直走同一条路

568
00:15:18,840 --> 00:15:20,160
还有一件事你要知道

569
00:15:20,160 --> 00:15:23,120
官方把一天切成尖峰跟离峰

570
00:15:23,120 --> 00:15:25,200
尖峰的单价是离峰的两倍

571
00:15:25,200 --> 00:15:28,380
而那个尖峰时段换算成台湾时间

572
00:15:28,380 --> 00:15:30,480
刚好就是我们上班的那几个小时

573
00:15:30,480 --> 00:15:32,600
对那同一个问题是什么呢

574
00:15:32,600 --> 00:15:34,140
这份文件从头到尾讲的

575
00:15:34,140 --> 00:15:35,200
其实都是同一件事

576
00:15:35,200 --> 00:15:37,000
让一个很复杂的东西

577
00:15:37,000 --> 00:15:38,280
还能被一个人看懂

578
00:15:38,280 --> 00:15:39,320
还敢动手去改

579
00:15:39,320 --> 00:15:40,440
如果你只记得一句

580
00:15:40,440 --> 00:15:42,360
我建议这句模型看得到的

581
00:15:42,360 --> 00:15:43,100
就一定被记下来

582
00:15:43,100 --> 00:15:44,740
因为他一句话就撑起了

583
00:15:44,740 --> 00:15:45,980
可以倒带可以分岔

584
00:15:45,980 --> 00:15:47,660
可以查这样这三件事

585
00:15:47,660 --> 00:15:49,940
其他的设计都是在保护他

586
00:15:49,940 --> 00:15:51,040
如果你是工程师

587
00:15:51,040 --> 00:15:52,940
我建议你今天就把它装起来

588
00:15:52,940 --> 00:15:54,700
玩意行指令就开得起来

589
00:15:54,700 --> 00:15:56,620
但先不要拿去做正式的东西

590
00:15:56,620 --> 00:15:58,120
如果你是在看趋势的人

591
00:15:58,120 --> 00:16:00,040
你要看的不是这个框架本身

592
00:16:00,040 --> 00:16:00,840
是这件事

593
00:16:00,840 --> 00:16:03,240
模型公司开始抢执行环境了

594
00:16:03,240 --> 00:16:04,360
欢迎订阅追踪

595
00:16:04,360 --> 00:16:05,240
后续下次见