1
00:00:00,000 --> 00:00:04,600
OpenAI發布全新的旗艦模型GPT 5.6Soul,Codex也立刻用上了。

2
00:00:04,940 --> 00:00:06,360
这次真正值得注意的是什么?

3
00:00:06,820 --> 00:00:14,240
其实除了5.6Soul到底有多好用,这支影片还要告诉你OpenAI怎么刚把Codex大改一轮,两个月内出了超过150项更新。

4
00:00:14,880 --> 00:00:23,260
这次真正的转变,是Codex开始自己动手,自己会把一个大任务拆给好几个分身平行座,自己开浏览器点画面,甚至自己管理自己的任务清单。

5
00:00:23,820 --> 00:00:28,140
Codex还是那个你叫它写程式的工具,但也已经变成一个会自己安排工作的执行者。

6
00:00:28,140 --> 00:00:30,820
我挑出六件真的会改变你工作方式的功能

7
00:00:30,820 --> 00:00:32,420
还有一个官方只讲了一句

8
00:00:32,420 --> 00:00:33,800
但你一定会付的代价

9
00:00:33,800 --> 00:00:36,380
先讲一个最容易被跳过的前提

10
00:00:36,380 --> 00:00:38,680
Codex现在住在ChatGPT里面

11
00:00:38,680 --> 00:00:40,420
官方更新记录写得很清楚

12
00:00:40,420 --> 00:00:41,320
7月9号

13
00:00:41,320 --> 00:00:43,540
Codex正式并进ChatGPT桌面板

14
00:00:43,540 --> 00:00:44,740
Mac和Windows都有

15
00:00:44,740 --> 00:00:45,800
原本的专案

16
00:00:45,800 --> 00:00:47,360
设定和工作流都会保留

17
00:00:47,360 --> 00:00:48,980
它在里面有自己的空间

18
00:00:48,980 --> 00:00:50,580
旁边是一个新的Work Agent

19
00:00:50,580 --> 00:00:52,120
这看起来只是搬家

20
00:00:52,120 --> 00:00:54,020
却决定了后面所有功能的长相

21
00:00:54,020 --> 00:00:56,440
Codex不再是给工程师的独立产品线

22
00:00:56,440 --> 00:00:58,500
而是ChatGPT底下的一个工作面向

23
00:00:58,500 --> 00:01:00,760
官方说每周有600万人在用

24
00:01:00,760 --> 00:01:02,140
这是他们自己给的数字

25
00:01:02,140 --> 00:01:03,720
目前没有公开资料可以佐证

26
00:01:03,720 --> 00:01:06,160
第一件事是脑子换了

27
00:01:06,160 --> 00:01:08,860
新的旗舰模型叫GPT 5.6 Soul

28
00:01:08,860 --> 00:01:11,020
官方说它可以长时间啃难题

29
00:01:11,020 --> 00:01:13,620
但真正要注意的是它旁边那个Ultra

30
00:01:13,620 --> 00:01:15,060
在官方文件里

31
00:01:15,060 --> 00:01:17,180
Ultra是推理等级选单的最高一档

32
00:01:17,180 --> 00:01:18,380
说明只有一行

33
00:01:18,380 --> 00:01:19,760
最大推理力度

34
00:01:19,760 --> 00:01:21,140
加上自动分排任务

35
00:01:21,140 --> 00:01:22,700
这行字很关键

36
00:01:22,700 --> 00:01:24,580
因为它把两件事绑在一起了

37
00:01:24,580 --> 00:01:26,200
Ultra不只是想得更久

38
00:01:26,200 --> 00:01:27,920
他还会主动把工作拆出去

39
00:01:27,920 --> 00:01:29,200
所以你会看到

40
00:01:29,200 --> 00:01:30,380
Codex不用你开口

41
00:01:30,380 --> 00:01:31,420
就自己把一个任务

42
00:01:31,420 --> 00:01:32,980
分给好几个subagents平行做

43
00:01:32,980 --> 00:01:34,340
你可以个别追踪

44
00:01:34,340 --> 00:01:35,320
也可以打开面板

45
00:01:35,320 --> 00:01:36,240
看整组一起工作

46
00:01:36,240 --> 00:01:38,580
这里就要讲那个代价了

47
00:01:38,580 --> 00:01:40,900
更大的思考预算不是免费的

48
00:01:40,900 --> 00:01:41,820
想得越久

49
00:01:41,820 --> 00:01:42,640
拆得越多

50
00:01:42,640 --> 00:01:43,920
Token就烧得越快

51
00:01:43,920 --> 00:01:45,240
这不是我推测

52
00:01:45,240 --> 00:01:46,660
官方文件自己写

53
00:01:46,660 --> 00:01:48,160
subagent的工作流

54
00:01:48,160 --> 00:01:50,000
会比单一agent消耗更多token

55
00:01:50,000 --> 00:01:51,880
更新记录里甚至加了一条

56
00:01:51,880 --> 00:01:52,920
选ultra

57
00:01:52,920 --> 00:01:54,640
又让多个agent同时跑的时候

58
00:01:54,640 --> 00:01:56,900
会跳出警告提醒你用量会冲很快

59
00:01:56,900 --> 00:01:57,880
这一点

60
00:01:57,880 --> 00:02:00,100
OpenAI在示范的时候自己也讲白了

61
00:02:24,640 --> 00:02:26,080
开启这 panel

62
00:02:26,080 --> 00:02:27,760
然后看到整个团队一起工作

63
00:02:27,760 --> 00:02:29,300
一项目

64
00:02:29,300 --> 00:02:31,200
然后现在在这一项目移动

65
00:02:31,200 --> 00:02:32,120
第二件事

66
00:02:32,120 --> 00:02:33,720
是他开始自己动画鼠

67
00:02:33,720 --> 00:02:35,200
影片中示范的

68
00:02:35,200 --> 00:02:36,740
OpenAI开发者体验负责人

69
00:02:36,740 --> 00:02:37,560
Romaine Hute说

70
00:02:37,560 --> 00:02:39,140
这可能是整轮更新里

71
00:02:39,140 --> 00:02:40,820
他个人最期待大家体验的功能

72
00:02:40,820 --> 00:02:42,500
关键字叫AppShot

73
00:02:42,500 --> 00:02:44,100
你可以先把它想成截图

74
00:02:44,100 --> 00:02:45,740
但它比截图多一层东西

75
00:02:45,740 --> 00:02:48,040
AppShot会把最前面那个视窗的画面

76
00:02:48,040 --> 00:02:49,680
加上那个App能提供的文字

77
00:02:49,680 --> 00:02:50,400
一起送出去

78
00:02:50,400 --> 00:02:51,920
连你卷动范围之外

79
00:02:51,920 --> 00:02:53,320
萤幕上根本看不到的文字

80
00:02:53,320 --> 00:02:53,860
都拿得到

81
00:02:53,860 --> 00:02:55,440
所以这里的差别是

82
00:02:55,440 --> 00:02:56,540
截图只有像素

83
00:02:56,540 --> 00:02:58,560
AppShot多了一层结构化的文字

84
00:02:58,560 --> 00:03:00,220
它不只看到画面长什么样

85
00:03:00,220 --> 00:03:01,760
还读得到画面背后的内容

86
00:03:01,760 --> 00:03:03,280
怎么触发

87
00:03:03,280 --> 00:03:05,840
在Mac上按下左右两个command键就好

88
00:03:05,840 --> 00:03:07,920
而且只有MacOS桌面板支援

89
00:03:07,920 --> 00:03:09,680
这点要翻官方文件才查得到

90
00:03:09,680 --> 00:03:11,380
拿到AppShot之后

91
00:03:11,380 --> 00:03:13,200
Codex就能自己操作那个模拟器

92
00:03:13,200 --> 00:03:14,880
画面上游标自己在移动

93
00:03:14,880 --> 00:03:15,880
自己点来点去

94
00:03:15,880 --> 00:03:17,040
没有人在碰滑鼠

95
00:03:17,040 --> 00:03:18,340
更关键的是

96
00:03:18,340 --> 00:03:19,760
这一切都在背景跑

97
00:03:19,760 --> 00:03:21,600
你可以继续用你的电脑做别的事

98
00:03:21,600 --> 00:03:24,200
过去AI操作电脑最恼人的

99
00:03:24,200 --> 00:03:25,680
就是它会占住你的滑鼠键盘

100
00:03:25,680 --> 00:03:26,940
这一点被截掉了

101
00:03:51,600 --> 00:03:52,600
看看这个

102
00:03:52,600 --> 00:03:53,840
这是我的拇指

103
00:03:53,840 --> 00:03:55,000
是Codex在转向

104
00:03:55,000 --> 00:03:57,040
在背景中

105
00:03:57,040 --> 00:03:59,360
我可以继续用我的手机

106
00:03:59,360 --> 00:04:00,440
那不错

107
00:04:21,600 --> 00:04:23,680
这解掉的是一个很真实的痛点

108
00:04:23,680 --> 00:04:25,080
你看得出哪里怪

109
00:04:25,080 --> 00:04:26,480
但你讲不清楚它在哪

110
00:04:26,480 --> 00:04:29,040
另一半是给看得懂程式码的人

111
00:04:29,040 --> 00:04:30,340
切到DIF检视

112
00:04:30,340 --> 00:04:31,820
如果你知道自己要什么

113
00:04:31,820 --> 00:04:33,260
可以直接精准地改一行

114
00:04:33,260 --> 00:04:35,040
如果只知道想要什么行为

115
00:04:35,040 --> 00:04:36,000
不确定该怎么改

116
00:04:36,000 --> 00:04:37,340
就把它丢回给Codex

117
00:04:37,340 --> 00:04:38,760
接下来这段

118
00:04:38,760 --> 00:04:40,080
看他怎么在同一个地方

119
00:04:40,080 --> 00:04:41,340
无缝切换这两种做法

120
00:04:41,340 --> 00:04:44,020
I can also switch into the DIF view

121
00:04:44,020 --> 00:04:46,480
I've always been able to inspect the DIF here

122
00:04:46,480 --> 00:04:48,480
But now I can go further

123
00:04:48,480 --> 00:04:50,940
and make precise edits directly in line

124
00:04:50,940 --> 00:04:53,380
所以如果我知道我 exactly what I want

125
00:04:53,380 --> 00:04:54,640
我可以改变的练习

126
00:04:54,640 --> 00:04:55,180
我自己的练习

127
00:04:55,180 --> 00:04:55,820
就像这样

128
00:04:55,820 --> 00:04:58,480
或如果我知道我希望的行为

129
00:04:58,480 --> 00:04:59,340
但我不确定

130
00:04:59,340 --> 00:05:00,740
那种练习的练习

131
00:05:00,740 --> 00:05:01,740
我可以从那里来求

132
00:05:01,740 --> 00:05:03,040
Codex工作

133
00:05:03,040 --> 00:05:04,700
有时候我会工作

134
00:05:04,700 --> 00:05:05,040
视频

135
00:05:05,040 --> 00:05:06,540
有时候我会工作

136
00:05:06,540 --> 00:05:07,780
直接进练习

137
00:05:07,780 --> 00:05:09,240
然后它会比较近

138
00:05:09,240 --> 00:05:11,040
跟模式的练习

139
00:05:11,040 --> 00:05:12,160
第四件事

140
00:05:12,160 --> 00:05:13,460
是东西做完了

141
00:05:13,460 --> 00:05:14,740
要怎么让别人用到

142
00:05:14,740 --> 00:05:16,400
这是Size要解的问题

143
00:05:16,400 --> 00:05:17,820
你只要说一句

144
00:05:17,820 --> 00:05:19,240
把这个发布到Size

145
00:05:19,240 --> 00:05:20,320
它就上线了

146
00:05:20,320 --> 00:05:22,880
不过这里要修正一句话

147
00:05:22,880 --> 00:05:25,440
OpenAI说Sites上个月只给企业版

148
00:05:25,440 --> 00:05:26,540
现在开放所有人

149
00:05:26,540 --> 00:05:28,740
但官方文件上写的是Public Beta

150
00:05:28,740 --> 00:05:30,360
而且明讲能不能用

151
00:05:30,360 --> 00:05:31,160
要看你的方案

152
00:05:31,160 --> 00:05:32,680
你的地区和工作区设定

153
00:05:32,680 --> 00:05:34,020
这两个说法对不上

154
00:05:34,020 --> 00:05:35,600
所以先别预设你打开就用

155
00:05:35,600 --> 00:05:37,800
Public Beta这四个字本身就是提醒

156
00:05:37,800 --> 00:05:38,880
换句话说

157
00:05:38,880 --> 00:05:39,960
拿它做个原型

158
00:05:39,960 --> 00:05:41,200
做个内部小工具很好

159
00:05:41,200 --> 00:05:44,040
但要不要把已经有人在用的正式东西搬上去

160
00:05:44,040 --> 00:05:45,140
现阶段还不好说

161
00:05:45,140 --> 00:05:51,140
昨天月份,我们介绍了室室,所以 teams可以建设室,和分享它们的内容。

162
00:05:51,140 --> 00:05:57,140
现在室室是全室,而它仍然是任何网址你建设室。

163
00:05:57,140 --> 00:06:01,140
我可以直接说,导致室室。

164
00:06:01,140 --> 00:06:08,140
它有室内,链接,链接,运动,运动,运动,运动,都在建设室。

165
00:06:08,140 --> 00:06:11,140
所以我可以从一个真正的佛工程中,

166
00:06:11,140 --> 00:06:14,440
无论论均匀或训练

167
00:06:14,440 --> 00:06:16,320
然后现在的标志

168
00:06:16,320 --> 00:06:17,400
是现在的标志

169
00:06:41,140 --> 00:06:42,360
都有自己独立的工作副本

170
00:06:42,360 --> 00:06:43,460
改动互不干扰

171
00:06:43,460 --> 00:06:45,140
你不用担心A的修改

172
00:06:45,140 --> 00:06:45,980
会不小心动到B

173
00:06:45,980 --> 00:06:47,480
而且这些任务之间

174
00:06:47,480 --> 00:06:48,320
还可以互相引用

175
00:06:48,320 --> 00:06:50,160
注意这里发生的是

176
00:06:50,160 --> 00:06:51,120
拆解

177
00:06:51,120 --> 00:06:52,220
排优先序

178
00:06:52,220 --> 00:06:53,200
开分支

179
00:06:53,200 --> 00:06:54,980
这些原本是人的工作

180
00:06:54,980 --> 00:06:56,340
现在被交出去了

181
00:06:56,340 --> 00:06:58,580
这也是为什么真正的主线

182
00:06:58,580 --> 00:06:59,760
不是模型变强

183
00:06:59,760 --> 00:07:01,440
而是角色变了

184
00:07:01,440 --> 00:07:03,880
另外还有一个很实用的细节

185
00:07:03,880 --> 00:07:06,040
工作不一定要从Codex开始

186
00:07:06,040 --> 00:07:08,660
你在ChatGPT里讨论过的想法

187
00:07:08,660 --> 00:07:10,620
或用Deep Research查过的东西

188
00:07:10,620 --> 00:07:12,480
可以直接at那段对话

189
00:07:12,480 --> 00:07:14,740
把整包context带进codex

190
00:07:14,740 --> 00:07:16,280
不用再复制贴上

191
00:07:16,280 --> 00:07:20,320
You know that codex can also help you manage its own codex tasks

192
00:07:20,320 --> 00:07:23,960
This is super useful when I have several independent pieces of work

193
00:07:23,960 --> 00:07:25,640
that I want to get organized and moving

194
00:07:25,640 --> 00:07:27,420
Check out this one

195
00:07:27,420 --> 00:07:29,900
Find 5 bugs from linear

196
00:07:29,900 --> 00:07:33,580
Create a separate task and isolated work tree for each one

197
00:07:33,580 --> 00:07:36,040
And pin the most critical ones

198
00:07:36,537 --> 00:07:51,537
我需要有幾個獨立的工作, 要把工作的組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織組織

199
00:07:51,537 --> 00:07:52,777
手机版补了一批升级

200
00:07:52,777 --> 00:07:54,437
在手机上管理任务

201
00:07:54,437 --> 00:07:55,297
审查变更

202
00:07:55,297 --> 00:07:56,077
比较branch

203
00:07:56,077 --> 00:07:57,697
接下来这段一次演完

204
00:07:57,697 --> 00:07:59,697
我只补一句最反直觉的

205
00:07:59,697 --> 00:08:01,977
手机可以直接连上SSH Host

206
00:08:01,977 --> 00:08:03,077
像你的Devbox

207
00:08:03,077 --> 00:08:04,757
从手机就开新任务

208
00:08:04,757 --> 00:08:06,177
方便归方便

209
00:08:06,177 --> 00:08:07,337
权限要开多大

210
00:08:07,337 --> 00:08:08,577
值得你自己先想清楚

211
00:08:08,577 --> 00:08:10,257
有三个东西

212
00:08:10,257 --> 00:08:10,897
我希望你尝试

213
00:08:10,897 --> 00:08:12,257
第一

214
00:08:12,257 --> 00:08:13,337
你可以选择

215
00:08:13,337 --> 00:08:14,097
搜索

216
00:08:14,097 --> 00:08:15,017
开放

217
00:08:15,017 --> 00:08:16,057
管理任务

218
00:08:16,057 --> 00:08:17,537
直接从一个谈论

219
00:08:17,537 --> 00:08:18,497
上你的电话

220
00:08:18,497 --> 00:08:19,837
第二

221
00:08:19,837 --> 00:08:22,837
第二,计划的改变更好

222
00:08:22,837 --> 00:08:26,837
我可以选择和选择和选择

223
00:08:26,837 --> 00:08:29,837
或直接交叉择和选择

224
00:08:29,837 --> 00:08:32,837
第三,我可以直接接到一个 SSH 主持人

225
00:08:32,837 --> 00:08:33,837
像一个电视箱

226
00:08:33,837 --> 00:08:36,837
和从电视箱开始新的工作

227
00:08:36,837 --> 00:08:37,837
有很多多

228
00:08:37,837 --> 00:08:40,837
但我爱上我可以在电视箱

229
00:08:40,837 --> 00:08:42,837
从电视箱开始,

230
00:08:42,837 --> 00:08:44,837
然后在电视箱里,

231
00:08:44,837 --> 00:08:45,837
这不错吗?

232
00:08:45,837 --> 00:08:48,837
最后一块是把整条开发流程收完

233
00:08:48,837 --> 00:08:51,657
Poor Request现在不用离开App就能处理

234
00:08:51,657 --> 00:08:53,317
而且是带着脉络的

235
00:08:53,317 --> 00:08:55,817
Codex知道每个任务跑在哪个Branch上

236
00:08:55,817 --> 00:08:57,597
那个Branch一有开着的PR

237
00:08:57,597 --> 00:08:59,477
摘要面板就自动把它浮出来

238
00:08:59,477 --> 00:09:01,217
你可以在里面追踪

239
00:09:01,217 --> 00:09:02,017
处理问题

240
00:09:02,017 --> 00:09:03,397
确认没事就merge

241
00:09:03,397 --> 00:09:05,057
检查失败的时候

242
00:09:05,057 --> 00:09:07,137
直接请它带着完整脉络去修

243
00:09:07,137 --> 00:09:09,557
新的Poor Request分页

244
00:09:09,557 --> 00:09:11,177
可以逐项看每一处改动

245
00:09:11,177 --> 00:09:12,377
留inline comment

246
00:09:12,377 --> 00:09:13,517
做完整的review

247
00:09:13,517 --> 00:09:15,677
也可以把自己的PR标成ready

248
00:09:15,677 --> 00:09:16,377
然后merge

249
00:09:16,377 --> 00:09:18,357
这代表从写程式

250
00:09:18,357 --> 00:09:19,177
跑检查

251
00:09:19,177 --> 00:09:20,317
review到merge

252
00:09:20,317 --> 00:09:22,217
整条loop都收进同一个地方了

253
00:09:22,217 --> 00:09:23,717
你不用再跳去github.com

254
00:09:23,717 --> 00:09:26,017
但这里有个问题值得放在心上

255
00:09:26,017 --> 00:09:27,157
当写的是它

256
00:09:27,157 --> 00:09:28,097
修的是它

257
00:09:28,097 --> 00:09:29,597
review的界面也是它

258
00:09:29,597 --> 00:09:31,277
那谁在为这次合并负责

259
00:09:31,277 --> 00:09:33,177
工具把摩擦拿掉了

260
00:09:33,177 --> 00:09:34,397
责任不会跟着被拿掉

261
00:09:34,397 --> 00:09:37,317
另外这段示范从头到尾只提到github.com

262
00:09:37,317 --> 00:09:39,217
其他github平台有没有同样待遇

263
00:09:39,217 --> 00:09:40,317
就得自己去确认了

264
00:09:40,317 --> 00:09:43,837
Codex now makes it easier to work with pull requests

265
00:09:43,837 --> 00:09:45,177
without leaving the app

266
00:09:45,177 --> 00:09:46,797
It does this in context

267
00:09:46,797 --> 00:09:48,797
它知道哪一項 branch

268
00:09:48,797 --> 00:09:50,237
每項 task是在工作

269
00:09:50,237 --> 00:09:51,437
所以當那項 branch

270
00:09:51,437 --> 00:09:52,237
有開啟的研究

271
00:09:52,237 --> 00:09:53,237
它會有一個項目的研究

272
00:09:53,237 --> 00:09:54,797
它會有一個項目的

273
00:09:54,797 --> 00:09:56,357
我可以追求所有的事情

274
00:09:56,357 --> 00:09:57,957
和回應所有問題

275
00:09:57,957 --> 00:09:58,957
和回應所有事情

276
00:09:58,957 --> 00:10:00,157
都會看得到

277
00:10:00,157 --> 00:10:01,957
如果一個項目失去

278
00:10:01,957 --> 00:10:03,037
或是一個項目的

279
00:10:03,037 --> 00:10:04,357
我可以請 Codex

280
00:10:04,357 --> 00:10:05,357
去解決它

281
00:10:05,357 --> 00:10:05,757
在這裡

282
00:10:05,757 --> 00:10:06,757
全體的概念

283
00:10:06,757 --> 00:10:07,757
這新的項目

284
00:10:07,757 --> 00:10:08,997
提供我

285
00:10:08,997 --> 00:10:10,957
讓我每個個改變

286
00:10:10,957 --> 00:10:12,197
我可以留在內

287
00:10:12,197 --> 00:10:13,157
可項目

288
00:10:13,157 --> 00:10:14,837
complete a full review

289
00:10:14,837 --> 00:10:15,957
我可以尋找我的

290
00:10:15,957 --> 00:10:18,277
P.R.s as ready and merge them

291
00:10:18,277 --> 00:10:20,217
All of this within Codex

292
00:10:20,217 --> 00:10:22,537
Without having to go to github.com

293
00:10:22,537 --> 00:10:24,277
So I can stay in the flow

294
00:10:24,277 --> 00:10:25,577
回到开头那个问题

295
00:10:25,577 --> 00:10:27,337
把这六件事排在一起看

296
00:10:27,337 --> 00:10:28,537
方向其实很清楚

297
00:10:28,537 --> 00:10:29,597
拆工作

298
00:10:29,597 --> 00:10:30,517
开分支

299
00:10:30,517 --> 00:10:31,637
操作画面

300
00:10:31,637 --> 00:10:32,557
管任务

301
00:10:32,557 --> 00:10:33,557
处理P.R.

302
00:10:34,097 --> 00:10:35,517
这些原本都是人的判断

303
00:10:35,517 --> 00:10:37,357
现在一件一件被交出去了

304
00:10:37,357 --> 00:10:38,437
所以

305
00:10:38,437 --> 00:10:39,977
Codex已经不只是那个

306
00:10:39,977 --> 00:10:41,257
你叫它写程式的工具

307
00:10:41,257 --> 00:10:42,457
它更像一个

308
00:10:42,457 --> 00:10:44,097
会自己安排工作的执行者

309
00:10:44,097 --> 00:10:45,657
至于代价

310
00:10:45,657 --> 00:10:46,917
官方只讲了一句

311
00:10:46,917 --> 00:10:49,097
Ultra会更快吃掉你的token

312
00:10:49,097 --> 00:10:51,497
而真正需要你自己决定的

313
00:10:51,497 --> 00:10:53,077
是把多少判断交出去

314
00:10:53,077 --> 00:10:54,597
以及交出去之后

315
00:10:54,597 --> 00:10:55,657
谁还在看
