1
00:00:00,000 --> 00:00:16,000
蘋果本來已經談好了,每年付給Google大約10億美元。用一個1.2萬億參數的定制版Gem9來當新版的Siri的大腦。這是彭博社去年11月捅出來的。

2
00:00:16,000 --> 00:00:28,000
今年一夜,CNBC又確認了一次,現在iOS27推送了。兩家常年把蘋果代碼的媒體把系統翻了個底朝天。

3
00:00:28,000 --> 00:00:31,160
Jemini这个名字一次都没有出现

4
00:00:31,160 --> 00:00:33,940
出现的是另一样东西

5
00:00:33,940 --> 00:00:37,460
研究者在系统的深处找到了一个位置

6
00:00:37,460 --> 00:00:40,520
一个专门留个外部大模型的位置

7
00:00:40,520 --> 00:00:43,040
谁都能插进去

8
00:00:43,040 --> 00:00:45,920
可苹果同时又做了另一件事

9
00:00:45,920 --> 00:00:49,620
这个位置的钥匙一把都没有发出去

10
00:00:52,100 --> 00:00:53,220
第三方拿不到

11
00:00:53,220 --> 00:00:55,300
你在设置里也看不见

12
00:00:55,300 --> 00:00:57,260
朋友们

13
00:00:57,260 --> 00:01:00,180
这哪里是什么Siri终于开放了呀

14
00:01:00,180 --> 00:01:04,320
这分明是一个装好了且还锁着的开关

15
00:01:04,320 --> 00:01:07,680
苹果自己把钥匙揣在兜里了

16
00:01:07,680 --> 00:01:11,420
那这个开关到底重要在哪里呢

17
00:01:11,420 --> 00:01:14,280
苹果官方今年一夜份说过

18
00:01:14,280 --> 00:01:17,640
全球活跃的设备已经到了25亿台

19
00:01:17,640 --> 00:01:20,560
你每天早上问一句天气

20
00:01:20,560 --> 00:01:22,580
晚上让它订个闹钟

21
00:01:22,580 --> 00:01:25,260
出门前让他念一条消息

22
00:01:25,260 --> 00:01:28,460
这些话最后送到谁那去

23
00:01:28,460 --> 00:01:29,880
由谁来理解

24
00:01:29,880 --> 00:01:33,660
谁能顺手翻一眼你的日历和联系人

25
00:01:33,660 --> 00:01:36,080
就是这个开关决定的

26
00:01:36,080 --> 00:01:38,060
他现在锁着

27
00:01:38,060 --> 00:01:39,820
所以这些问题呢

28
00:01:39,820 --> 00:01:41,360
暂时还没有答案

29
00:01:41,360 --> 00:01:45,700
可代码已经把答案的形状提前录了出来

30
00:01:45,700 --> 00:01:48,880
先说这一天到底发生了什么

31
00:01:48,880 --> 00:01:50,720
9月14号

32
00:01:50,720 --> 00:01:54,400
Macrimose和Apple Insight分别报道

33
00:01:54,400 --> 00:01:58,580
有人在iOS27和MacOS27的

34
00:01:58,580 --> 00:02:00,620
Gordon Gate的私有框架里

35
00:02:00,620 --> 00:02:02,340
翻出了两条

36
00:02:02,340 --> 00:02:04,920
让外部模型接入Siri的通道

37
00:02:04,920 --> 00:02:06,860
一条浅一点

38
00:02:06,860 --> 00:02:09,660
外部模型负责听懂你说的话

39
00:02:09,660 --> 00:02:10,940
听懂之后

40
00:02:10,940 --> 00:02:14,520
社闹钟查日历这些真正动手的活

41
00:02:14,520 --> 00:02:16,740
还是交回给Siri去干

42
00:02:16,740 --> 00:02:18,920
另一条深的多

43
00:02:18,920 --> 00:02:23,960
代码层面甚至遗留了直接替换Siri服务端的模型的能力

44
00:02:23,960 --> 00:02:26,460
这两条通道的差别呢

45
00:02:26,460 --> 00:02:28,700
就好比一家公司里的两种人

46
00:02:28,700 --> 00:02:33,100
第一条像是在前台放了个接待

47
00:02:33,100 --> 00:02:36,240
Cloud或者别的模型坐在那

48
00:02:36,240 --> 00:02:38,260
能听懂你要办什么事

49
00:02:38,260 --> 00:02:39,960
能把事情说清楚

50
00:02:39,960 --> 00:02:42,840
但真正要进档案室调资料

51
00:02:42,840 --> 00:02:45,280
真要替你在文件上签字

52
00:02:45,280 --> 00:02:47,300
还得苹果自己的人气

53
00:02:47,300 --> 00:02:49,760
那第二条就不一样了

54
00:02:49,760 --> 00:02:52,600
相当于把一整个外部团队

55
00:02:52,600 --> 00:02:53,600
请进办公室

56
00:02:53,600 --> 00:02:56,360
连任务清单和工具说明

57
00:02:56,360 --> 00:02:57,240
一起交给他

58
00:02:57,240 --> 00:02:59,020
让他自己安排着干

59
00:02:59,020 --> 00:03:01,940
能干的活多了一大劫

60
00:03:01,940 --> 00:03:04,360
你的东西经过谁的手

61
00:03:04,360 --> 00:03:06,060
也跟着变了

62
00:03:06,060 --> 00:03:07,700
那看到这呢

63
00:03:07,700 --> 00:03:10,120
你很容易以为已经开了

64
00:03:10,120 --> 00:03:12,160
可同一篇报道呢

65
00:03:12,160 --> 00:03:13,420
紧接着又写道

66
00:03:13,420 --> 00:03:16,700
苹果尚未向第三方开放型号委托授钱

67
00:03:16,700 --> 00:03:20,040
而且这项功能也不对用户开放

68
00:03:20,040 --> 00:03:24,500
苹果还没有把这个全线开放给任何的第三方

69
00:03:24,500 --> 00:03:27,200
用户那边也根本看不到

70
00:03:27,200 --> 00:03:29,720
当时的互选版本里

71
00:03:29,720 --> 00:03:33,140
那个ask的功能连着chatGPT扩展

72
00:03:33,140 --> 00:03:34,640
Cloud还用不了

73
00:03:34,640 --> 00:03:38,480
Apple Insight补了一句更关键的

74
00:03:38,480 --> 00:03:41,980
他说真正能碰到你个人数据的

75
00:03:41,980 --> 00:03:43,900
仍然是苹果自己的模型

76
00:03:43,900 --> 00:03:46,660
所以今天你升级完系统

77
00:03:46,660 --> 00:03:47,800
打开手机

78
00:03:47,800 --> 00:03:49,120
什么都没变

79
00:03:49,120 --> 00:03:52,880
你不能在Siri的设置里

80
00:03:52,880 --> 00:03:53,720
挑一个Cloud

81
00:03:53,720 --> 00:03:56,240
不能把苹果自己的模型

82
00:03:56,240 --> 00:03:57,080
整个换掉

83
00:03:57,080 --> 00:03:59,060
作为应用的开发者

84
00:03:59,060 --> 00:04:01,020
也找不到一个公开的入口

85
00:04:01,020 --> 00:04:03,020
把模型接进去

86
00:04:03,020 --> 00:04:05,480
地下车库里画好了车位

87
00:04:05,480 --> 00:04:07,700
不等于你已经拿到了车钥匙

88
00:04:08,640 --> 00:04:09,760
可能有人会问

89
00:04:09,760 --> 00:04:12,040
装个聊天应用那么容易

90
00:04:12,040 --> 00:04:13,820
把模型接近Siri

91
00:04:13,820 --> 00:04:15,640
怎么就这么费劲呢

92
00:04:15,640 --> 00:04:17,860
差别在于它站在哪

93
00:04:17,860 --> 00:04:19,940
你打开一个聊天应用

94
00:04:19,940 --> 00:04:23,300
它只处理你主动发给它的那几句话

95
00:04:23,300 --> 00:04:26,380
那系统助手就不一样了

96
00:04:26,380 --> 00:04:28,380
它站在几十个应用的门口

97
00:04:28,380 --> 00:04:32,400
你问它附近哪家饭店今晚能吃饭

98
00:04:32,400 --> 00:04:34,360
和你说订一张桌子

99
00:04:34,360 --> 00:04:36,440
把餐厅的地址发给张三

100
00:04:36,440 --> 00:04:38,880
再把出发的时间写进日历

101
00:04:38,880 --> 00:04:41,100
完全不是一个量级的火

102
00:04:41,100 --> 00:04:44,380
前一季考的是他会不会答

103
00:04:44,380 --> 00:04:48,780
后一季考的是他能不能替你按下那个确认键

104
00:04:48,780 --> 00:04:50,940
说白了

105
00:04:50,940 --> 00:04:54,360
模型一旦从回答问题变成替你办事

106
00:04:54,360 --> 00:04:56,540
那聪明只是入场欠

107
00:04:56,540 --> 00:04:59,600
前线才是那张真正贵的票

108
00:04:59,600 --> 00:05:04,800
这一季就是这期视频的核心判断

109
00:05:04,800 --> 00:05:06,780
外界现在吵的

110
00:05:06,780 --> 00:05:08,520
全是谁的模型更强

111
00:05:08,520 --> 00:05:10,100
谁拿下了苹果

112
00:05:10,100 --> 00:05:14,240
可决定这场仗的胜负的根本不是参数

113
00:05:14,240 --> 00:05:17,840
是苹果愿意把哪一层前线交出去

114
00:05:17,840 --> 00:05:18,960
交给谁

115
00:05:18,960 --> 00:05:20,400
什么时候交

116
00:05:20,400 --> 00:05:24,380
这也意味着将来就算设置里

117
00:05:24,380 --> 00:05:26,280
真的冒出了一个模型写作菜单

118
00:05:26,280 --> 00:05:30,340
菜单上那几个名字也未必拿着一样的东西

119
00:05:30,340 --> 00:05:33,080
有的可能只能陪你聊天

120
00:05:33,080 --> 00:05:34,860
有的能听等指令

121
00:05:34,860 --> 00:05:37,160
但碰不到你的私人数据

122
00:05:37,160 --> 00:05:41,160
还有的也许能调用更多的系统功能

123
00:05:41,160 --> 00:05:44,260
可每一步都得你点头确认

124
00:05:44,260 --> 00:05:48,080
你看到的都是接入Siri这几个字

125
00:05:48,080 --> 00:05:49,960
可实际拿到手的

126
00:05:49,960 --> 00:05:52,340
可能是三把完全不同的钥匙

127
00:05:52,340 --> 00:05:55,920
所以别急着问Cloud赢没赢

128
00:05:55,920 --> 00:05:58,720
也别急着替ChatGPT惋惜

129
00:05:58,720 --> 00:06:02,700
能换模型和你能决定换谁

130
00:06:02,700 --> 00:06:05,480
那中间隔着一整个苹果呢

131
00:06:05,480 --> 00:06:08,840
而真正让人关注这条新闻的

132
00:06:08,840 --> 00:06:10,820
其实是那个消失的名字

133
00:06:10,820 --> 00:06:13,960
等会我还要讲一件更反常的事

134
00:06:13,960 --> 00:06:15,540
同一天

135
00:06:15,540 --> 00:06:18,260
有人把一场已经打赢过一轮

136
00:06:18,260 --> 00:06:21,060
连开庭日子都排好了的官司

137
00:06:21,060 --> 00:06:23,080
主动永久的撤掉了

138
00:06:23,080 --> 00:06:26,000
而且一分钱对价都没有对外说

139
00:06:26,000 --> 00:06:28,260
先把时间倒回去

140
00:06:28,260 --> 00:06:30,860
2025年11月5号

141
00:06:30,860 --> 00:06:32,380
彭博社独家报道

142
00:06:32,380 --> 00:06:34,960
苹果打算每年付给Google

143
00:06:34,960 --> 00:06:35,940
大约10亿美元

144
00:06:35,940 --> 00:06:39,960
用一个1.2万亿参数的定制Gem9模型

145
00:06:39,960 --> 00:06:41,800
来启动新版的Siri

146
00:06:41,800 --> 00:06:44,240
今年1月12号

147
00:06:44,240 --> 00:06:48,440
CNBC又确认了苹果正式选定了Gem9

148
00:06:48,440 --> 00:06:52,520
这条线在过期的10个月里被反复的报道

149
00:06:52,520 --> 00:06:56,440
中间只出现过细节有调整这类说法

150
00:06:56,440 --> 00:06:59,960
没有任何一方公开说过它黄了

151
00:07:00,900 --> 00:07:03,460
这比交易本身就够说明问题的

152
00:07:03,460 --> 00:07:06,280
苹果是全世界最不愿意

153
00:07:06,280 --> 00:07:08,380
把核心体验交给别人的公司

154
00:07:08,380 --> 00:07:10,280
芯片自己造

155
00:07:10,280 --> 00:07:11,660
地图自己做

156
00:07:11,660 --> 00:07:14,360
连基带都咬着牙自己做

157
00:07:14,360 --> 00:07:16,640
可到了大模型这

158
00:07:16,640 --> 00:07:18,260
他认了

159
00:07:18,260 --> 00:07:20,340
宁可每年掏十亿美元

160
00:07:20,340 --> 00:07:21,680
去买Google的模型

161
00:07:21,680 --> 00:07:23,980
也不等自己的模型追上来

162
00:07:23,980 --> 00:07:26,700
这在几年前是不可想象的

163
00:07:26,700 --> 00:07:30,880
所以当这个名字在系统的代码里

164
00:07:30,880 --> 00:07:32,180
彻底的切息的时候

165
00:07:32,180 --> 00:07:35,620
问题就不只是换了个供应商这么简单了

166
00:07:35,620 --> 00:07:39,520
我自己也是被这条线带着走的

167
00:07:39,520 --> 00:07:42,400
7月12号那期苹果起诉OpenAir

168
00:07:42,400 --> 00:07:45,200
我在时间线里讲的也进清楚楚

169
00:07:45,200 --> 00:07:46,660
当时我说

170
00:07:46,660 --> 00:07:48,620
2026年的秋天

171
00:07:48,620 --> 00:07:50,800
iOS27一级上线

172
00:07:50,800 --> 00:07:54,180
Gamonite将成为Siri的默认智能层

173
00:07:54,180 --> 00:07:58,600
Challet GPT将转成去主动选择的可选项

174
00:08:12,709 --> 00:08:16,129
他反而是当时唯一能用的那一个

175
00:08:16,129 --> 00:08:19,009
但我得把话说准了

176
00:08:19,009 --> 00:08:21,169
这两篇代码分析里

177
00:08:21,169 --> 00:08:22,269
没看到Jemann

178
00:08:22,269 --> 00:08:25,209
不等于苹果和Google的合作黄了

179
00:08:25,209 --> 00:08:28,169
更不等于Jemann永远的进不来了

180
00:08:28,169 --> 00:08:31,069
代码分析只能告诉你

181
00:08:31,069 --> 00:08:33,969
研究者在这几个框架里看见了什么

182
00:08:33,969 --> 00:08:36,469
他证明不了整个的系统里

183
00:08:36,469 --> 00:08:38,349
再没有别的安排了

184
00:08:38,349 --> 00:08:41,269
真正能确认的只有一件事

185
00:08:41,269 --> 00:08:43,689
那个原本被认为是要给

186
00:08:43,689 --> 00:08:44,889
Jemney的默认位置

187
00:08:44,889 --> 00:08:47,189
在这批公开的证据里

188
00:08:47,189 --> 00:08:49,249
苹果没有交给任何人

189
00:08:49,249 --> 00:08:50,489
他空着

190
00:08:50,489 --> 00:08:55,209
这个空着比谁做上去更有意思的多了

191
00:08:55,209 --> 00:08:58,809
你把两笔钱摆在一起看就明白了

192
00:08:58,809 --> 00:09:01,189
Google为了保住Safari里的

193
00:09:01,189 --> 00:09:02,809
那个默认搜索引擎的位置

194
00:09:02,809 --> 00:09:04,369
2022年

195
00:09:04,369 --> 00:09:06,989
一年付给苹果大约200亿美元

196
00:09:06,989 --> 00:09:10,969
这是美国反垄断案的法庭文件里披露的

197
00:09:10,969 --> 00:09:13,149
现在轮到AI了

198
00:09:13,149 --> 00:09:14,929
方向掉了个头

199
00:09:14,929 --> 00:09:17,549
变成苹果每年付给Google

200
00:09:17,549 --> 00:09:18,609
10亿美元

201
00:09:18,609 --> 00:09:20,629
同一对公司

202
00:09:20,629 --> 00:09:22,549
同一个默认位置

203
00:09:22,549 --> 00:09:24,309
一边是收200亿

204
00:09:24,309 --> 00:09:26,029
一边是付10亿

205
00:09:26,029 --> 00:09:27,469
朋友们

206
00:09:27,469 --> 00:09:28,949
谁给谁钱

207
00:09:28,949 --> 00:09:31,949
从来都是谁更需要谁的答案

208
00:09:31,949 --> 00:09:34,849
搜索时代

209
00:09:34,849 --> 00:09:37,349
是Google离不开苹果那扇门

210
00:09:38,349 --> 00:09:39,929
AI时代刚开始的时候

211
00:09:39,929 --> 00:09:43,189
看起来是向苹果离不开Google的模型

212
00:09:43,189 --> 00:09:47,129
昨天代码里露出来的那个空位置告诉我们

213
00:09:47,129 --> 00:09:51,149
苹果正在试着把这个方向再掰回去

214
00:09:51,149 --> 00:09:55,669
他宁愿先造一个谁都能插进去的插槽

215
00:09:55,669 --> 00:09:58,989
也不急着把名字刻死在门牌上

216
00:09:58,989 --> 00:10:01,729
一个位置只要是还空着

217
00:10:01,729 --> 00:10:04,289
想进来的人就得一直开架

218
00:10:05,329 --> 00:10:08,669
而且这两个默认的位置的分量根本不一样

219
00:10:08,669 --> 00:10:11,209
搜索时代的默认位

220
00:10:11,209 --> 00:10:13,709
说到底就是一个流量的入口

221
00:10:13,709 --> 00:10:16,009
你在地址栏里打字

222
00:10:16,009 --> 00:10:18,029
他把你送去一个网页

223
00:10:18,029 --> 00:10:21,929
生意是在网页的那一头发生的

224
00:10:21,929 --> 00:10:24,829
那AS时代的默认位不是送客的

225
00:10:24,829 --> 00:10:27,469
他是替你干活的

226
00:10:27,469 --> 00:10:31,009
订票 改日程 写邮件 翻照片

227
00:10:31,009 --> 00:10:33,409
一件件都被搜索一次

228
00:10:33,409 --> 00:10:35,329
更接近掏钱的那一刻

229
00:10:35,329 --> 00:10:38,109
谁站在那个位置上

230
00:10:38,109 --> 00:10:40,669
拿到的就不只是提问的量

231
00:10:40,669 --> 00:10:44,169
是你一天里那几十个真实的动作

232
00:10:44,169 --> 00:10:47,429
所以这块地皮涨价了

233
00:10:47,429 --> 00:10:50,589
涨到了苹果自己都舍不得随便订给谁了

234
00:10:50,589 --> 00:10:53,729
那苹果到底在等什么呢

235
00:10:53,729 --> 00:10:57,489
这件事苹果没有给过任何官方的解释

236
00:10:57,489 --> 00:11:00,849
我不打算替苹果编一份会议纪要

237
00:11:00,849 --> 00:11:04,589
但他面前摆着三笔账是明摆着的

238
00:11:04,589 --> 00:11:06,609
第一笔是产品的账

239
00:11:06,609 --> 00:11:09,369
Siri这两年的日子不好过

240
00:11:09,369 --> 00:11:12,029
发布会上演示过的功能

241
00:11:12,029 --> 00:11:13,229
一次次的推迟

242
00:11:13,229 --> 00:11:16,549
用户随手一问就翻车了

243
00:11:16,549 --> 00:11:18,049
同行的助手呢

244
00:11:18,049 --> 00:11:19,129
一个比一个能干

245
00:11:19,129 --> 00:11:21,649
这个坑要继续的拖

246
00:11:21,649 --> 00:11:24,809
伤的就不只是Siri本身了

247
00:11:24,809 --> 00:11:27,769
还有新系统和新手机的卖相

248
00:11:27,769 --> 00:11:29,969
苹果卖硬件

249
00:11:29,969 --> 00:11:31,289
硬件的卖点

250
00:11:31,289 --> 00:11:33,209
越来越靠软件来支撑

251
00:11:33,209 --> 00:11:35,169
这笔钱在催它

252
00:11:35,169 --> 00:11:37,809
第二笔是信任的账

253
00:11:37,809 --> 00:11:39,429
而且这笔最重

254
00:11:39,429 --> 00:11:43,269
你想想有这么个场景

255
00:11:43,269 --> 00:11:44,609
你跟手机说

256
00:11:44,609 --> 00:11:48,749
帮我从邮件里找出孩子学校那张缴费通知

257
00:11:48,749 --> 00:11:50,509
核对一下日期

258
00:11:50,509 --> 00:11:52,449
到时候提醒我交钱

259
00:11:52,449 --> 00:11:55,689
这句话里同时出现了家庭信息

260
00:11:55,689 --> 00:11:57,669
邮件证文和一笔支付

261
00:11:57,669 --> 00:12:00,609
如果外部模型只负责听懂

262
00:12:00,609 --> 00:12:02,209
你要找哪一封邮件

263
00:12:02,209 --> 00:12:04,129
那风险还可控

264
00:12:04,129 --> 00:12:07,269
可要是他能读完整封邮件

265
00:12:07,269 --> 00:12:08,729
挑好账户

266
00:12:08,729 --> 00:12:11,429
再推着把你的钱付出去

267
00:12:11,429 --> 00:12:13,529
苹果就必须说清楚

268
00:12:13,529 --> 00:12:15,849
哪些数据离开了你的手机

269
00:12:15,849 --> 00:12:18,829
哪一部一定要你亲自的确认

270
00:12:18,829 --> 00:12:20,569
出了错能不能撤回

271
00:12:20,569 --> 00:12:22,769
苹果这些年呢

272
00:12:22,769 --> 00:12:25,169
一直把隐私做成了自己的招牌

273
00:12:25,169 --> 00:12:28,709
招牌砸一次就很难再立起来了

274
00:12:28,709 --> 00:12:30,929
这笔账在催他慢

275
00:12:31,929 --> 00:12:33,469
第三笔是合作的账

276
00:12:33,469 --> 00:12:36,269
谁被放在了最显眼的位置

277
00:12:36,269 --> 00:12:38,789
谁就拿到了真实的使用和反馈

278
00:12:38,789 --> 00:12:41,609
这对模型公司来说值钱的很

279
00:12:41,609 --> 00:12:43,949
苹果手里也就有了筹码

280
00:12:43,949 --> 00:12:47,749
可这笔账要想卖出个好价钱

281
00:12:47,749 --> 00:12:50,889
前提恰恰是别急着定下来

282
00:12:50,889 --> 00:12:53,409
三笔账互相的打架

283
00:12:53,409 --> 00:12:56,109
开放的快体验是上来了

284
00:12:56,109 --> 00:12:57,769
责任也跟着上来了

285
00:12:58,709 --> 00:13:00,929
开放的慢风险是低了

286
00:13:00,929 --> 00:13:03,789
用户可能早就跑去别的应用里

287
00:13:03,789 --> 00:13:05,029
养成习惯了

288
00:13:05,029 --> 00:13:06,989
锁着这个开关

289
00:13:06,989 --> 00:13:09,589
其实是苹果在这三笔账之间

290
00:13:09,589 --> 00:13:11,049
给自己留的缓冲

291
00:13:11,049 --> 00:13:14,649
还有一点更容易被忽略的

292
00:13:14,649 --> 00:13:17,909
这个开关是不太好挽回您的

293
00:13:17,909 --> 00:13:19,869
一旦苹果公开允许

294
00:13:19,869 --> 00:13:22,409
你把虚拟的核心模型换成外面的

295
00:13:22,409 --> 00:13:25,829
那用户的习惯也会跟着长出来

296
00:13:25,829 --> 00:13:29,409
开发者会围绕着这个接口写东西

297
00:13:29,409 --> 00:13:33,449
监管机构也会把它继承一条既定的事实

298
00:13:33,449 --> 00:13:36,709
到那个时候苹果想收紧

299
00:13:36,709 --> 00:13:40,949
付出的代价就比一开始不开要大得多了

300
00:13:40,949 --> 00:13:44,969
所以它现在卖未必前世技术没有做完

301
00:13:44,969 --> 00:13:49,589
也可能是因为这一步走住气就基本收不回来了

302
00:13:49,589 --> 00:13:53,469
还有一层外部的压力

303
00:13:53,469 --> 00:13:55,829
Mark Rumors提到

304
00:13:55,829 --> 00:13:57,629
欧盟的数字市场法案

305
00:13:57,629 --> 00:14:00,029
也可能影响了这套设计的方向

306
00:14:00,029 --> 00:14:03,789
这个法案的法则确实够吓人的

307
00:14:03,789 --> 00:14:08,609
最高能罚到企业全球营业额的10%

308
00:14:08,609 --> 00:14:10,669
8年内重复违法

309
00:14:10,669 --> 00:14:13,349
还能升到20%

310
00:14:13,349 --> 00:14:15,669
但可能影响就是可能影响

311
00:14:15,669 --> 00:14:18,069
苹果从没公开承认过

312
00:14:18,069 --> 00:14:20,769
这套接口是为谁准备的

313
00:14:20,769 --> 00:14:22,869
我也不会把它解读成

314
00:14:22,869 --> 00:14:25,189
专门作为监管看的后门

315
00:14:25,889 --> 00:14:27,149
能确认的指示

316
00:14:27,149 --> 00:14:29,309
技术上的可替代性

317
00:14:29,309 --> 00:14:31,089
苹果已经写进去了

318
00:14:31,089 --> 00:14:32,909
要不要真的开放

319
00:14:32,909 --> 00:14:34,469
什么时候开放

320
00:14:34,469 --> 00:14:36,209
他一个字也没说

321
00:14:36,209 --> 00:14:38,149
把这三笔账

322
00:14:38,149 --> 00:14:39,949
加上那条法规摆在一起

323
00:14:39,949 --> 00:14:41,489
你会发现

324
00:14:41,489 --> 00:14:42,969
苹果手里攥着的

325
00:14:42,969 --> 00:14:44,389
根本不是某个模型

326
00:14:44,389 --> 00:14:45,869
而是什么时候

327
00:14:45,869 --> 00:14:48,029
开门这个决定前本身

328
00:14:48,029 --> 00:14:49,749
就在同一天

329
00:14:49,749 --> 00:14:50,989
还有一个人

330
00:14:50,989 --> 00:14:52,709
用一个更极端的动作

331
00:14:52,709 --> 00:14:54,009
替这个判断

332
00:14:54,009 --> 00:14:55,149
做了一次助讲

333
00:14:55,149 --> 00:14:57,809
这事得从头说起

334
00:14:57,809 --> 00:15:02,449
去年马斯克旗下的X和SpaceX AI

335
00:15:02,449 --> 00:15:05,689
把苹果和OpenAI一起告上了

336
00:15:05,689 --> 00:15:07,789
德州的北七联邦法院

337
00:15:07,789 --> 00:15:10,309
告苹果的核心理由之一

338
00:15:10,309 --> 00:15:11,529
在诉状里写道

339
00:15:11,529 --> 00:15:14,529
苹果拒绝了X AI提出的

340
00:15:14,529 --> 00:15:17,369
把Groker向ChatGPD那样

341
00:15:17,369 --> 00:15:18,989
集成进iOS的请求

342
00:15:18,989 --> 00:15:20,929
翻译成人话就是

343
00:15:20,929 --> 00:15:23,349
你凭什么只让OpenAI进门

344
00:15:23,349 --> 00:15:24,729
不让我进呢

345
00:15:24,729 --> 00:15:27,929
这个案子也不是随便告告的

346
00:15:27,929 --> 00:15:30,629
2025年11月13日

347
00:15:30,629 --> 00:15:33,669
法官驳回了对方要求结案的动议

348
00:15:33,669 --> 00:15:36,569
案子获准继续往下走

349
00:15:36,569 --> 00:15:38,589
庭审都已经安排到了

350
00:15:38,589 --> 00:15:40,569
2027年的1月11日了

351
00:15:40,569 --> 00:15:42,609
然后就在昨天

352
00:15:42,609 --> 00:15:46,429
也就是iOS27代码被扒出来的同一天

353
00:15:46,429 --> 00:15:50,269
X和SpaceX Air递交了一份动议

354
00:15:50,269 --> 00:15:52,709
动议的原文明确写着

355
00:15:52,709 --> 00:15:56,889
原告根据联邦民事诉讼规则的相关条款

356
00:15:56,889 --> 00:16:01,009
请求实质性撤销对苹果公司的全部诉求

357
00:16:01,671 --> 00:16:05,671
這裡的實質性撤銷在法律上的意思就是,

358
00:16:05,671 --> 00:16:11,671
只要法院批准了這個請求,那這起糾紛就徹底劃上了記號了。

359
00:16:11,671 --> 00:16:16,671
以後他們就不能再拿這批同樣的訴求去重新起訴蘋果了。

360
00:16:16,671 --> 00:16:19,011
同一份的文件里

361
00:16:19,011 --> 00:16:21,471
对OpenEA那三个实体的诉求

362
00:16:21,471 --> 00:16:22,551
一条不测

363
00:16:22,551 --> 00:16:23,831
全部保留

364
00:16:23,831 --> 00:16:25,731
扛过了一轮

365
00:16:25,731 --> 00:16:27,331
排上了7的诉前

366
00:16:27,331 --> 00:16:29,071
说不要就不要了

367
00:16:29,071 --> 00:16:30,291
而且永远不要

368
00:16:30,291 --> 00:16:33,291
文件里给的理由只有一句话

369
00:16:33,291 --> 00:16:36,171
原告已经解决了对苹果的诉求

370
00:16:36,171 --> 00:16:37,671
解决成什么样

371
00:16:37,671 --> 00:16:39,811
一分钱对价都没有披露

372
00:16:39,811 --> 00:16:41,971
网上很快就有人

373
00:16:41,971 --> 00:16:43,911
把这两件事拼成了一个故事

374
00:16:43,911 --> 00:16:45,111
说

375
00:16:45,111 --> 00:16:47,451
苹果开放了第三方的模型

376
00:16:47,451 --> 00:16:49,531
所以马斯克放过苹果

377
00:16:49,531 --> 00:16:51,791
转身集中打OpenAI了

378
00:16:51,791 --> 00:16:54,211
但我得明确的告诉你

379
00:16:54,211 --> 00:16:57,111
目前公开的材料撑不住这个因果的

380
00:16:57,111 --> 00:17:00,951
撤诉的文件里没有任何一个字提到Siri的代码

381
00:17:00,951 --> 00:17:03,871
苹果也没有任何的公开表态过

382
00:17:03,871 --> 00:17:08,211
而且这份动议目前还在等待法官的批准

383
00:17:08,211 --> 00:17:11,131
法院最终要怎么处理

384
00:17:11,131 --> 00:17:13,131
得看正式入党的命令

385
00:17:13,131 --> 00:17:15,871
还有一件事很说明问题

386
00:17:15,871 --> 00:17:18,291
这两篇代码分析里

387
00:17:18,291 --> 00:17:21,371
Glock同样一次都没有被提到

388
00:17:21,371 --> 00:17:23,211
拿一个没出现的结果

389
00:17:23,211 --> 00:17:25,531
去解释一个没有被踢录的原因

390
00:17:25,531 --> 00:17:27,651
故事是原了

391
00:17:27,651 --> 00:17:29,191
可信度是掉的

392
00:17:29,191 --> 00:17:34,251
那这件事为什么还值得摆在这呢

393
00:17:34,251 --> 00:17:37,651
因为他把这场争夺的另一个战场

394
00:17:37,651 --> 00:17:38,371
录了出来

395
00:17:38,371 --> 00:17:40,691
模型入口这件事

396
00:17:40,691 --> 00:17:43,431
不只在发布会和跑分榜上打

397
00:17:43,431 --> 00:17:45,331
他也在合同里打

398
00:17:45,331 --> 00:17:47,511
在应用商店的规则里打

399
00:17:47,511 --> 00:17:48,931
在法院里打

400
00:17:48,931 --> 00:17:51,911
一家公司花了一年多

401
00:17:51,911 --> 00:17:54,091
把苹果告到的快开停了

402
00:17:54,091 --> 00:17:56,491
又在系统推送的当天

403
00:17:56,491 --> 00:17:58,231
选择永久的收手

404
00:17:58,231 --> 00:18:01,711
这个时间点本身就值得记住

405
00:18:01,711 --> 00:18:04,791
但记住不等于替他回答呀

406
00:18:05,551 --> 00:18:08,931
还得分清楚一件经常被标题吃掉的事实

407
00:18:08,931 --> 00:18:11,151
诉求还留着

408
00:18:11,151 --> 00:18:13,791
不等于法院认定OpenAir违法

409
00:18:13,791 --> 00:18:15,731
诉求撤掉了

410
00:18:15,731 --> 00:18:18,511
也不等于法院认定苹果清白的

411
00:18:18,511 --> 00:18:21,771
诉讼文件只告诉你当事人

412
00:18:21,771 --> 00:18:24,091
此刻决定把哪些主张留下

413
00:18:24,091 --> 00:18:25,591
把哪些撤走了

414
00:18:25,591 --> 00:18:27,591
它是一个承袭动作

415
00:18:27,591 --> 00:18:28,671
不是判决

416
00:18:28,671 --> 00:18:31,811
更不是一份公开的商业协议

417
00:18:31,811 --> 00:18:34,971
把承袭动作讲成是胜负的结果

418
00:18:34,971 --> 00:18:37,011
看着是挺过瘾的

419
00:18:37,011 --> 00:18:38,431
但会让人故以为

420
00:18:38,431 --> 00:18:40,991
最关键的那张纸已经出现了

421
00:18:40,991 --> 00:18:42,511
但其实它没有

422
00:18:42,511 --> 00:18:45,131
所以这两件事放在一起

423
00:18:45,131 --> 00:18:47,791
真正成立的共同点只有一个

424
00:18:47,791 --> 00:18:50,271
手机里的那个AI入口

425
00:18:50,271 --> 00:18:53,731
正在变成模型公司绕不开的一张谈判桌

426
00:18:53,731 --> 00:18:56,971
苹果在技术上多留了个位置

427
00:18:56,971 --> 00:19:00,151
说明他不想把未来绑死在一家身上

428
00:19:00,151 --> 00:19:02,331
他迟迟不发钥匙

429
00:19:02,331 --> 00:19:06,131
说明想进来的人得先接触他看的条件

430
00:19:06,131 --> 00:19:09,711
马斯克撤苹果保OpenAI

431
00:19:09,711 --> 00:19:13,831
说明围着这张桌子的商业和法律关系还在动

432
00:19:13,831 --> 00:19:16,011
共同的主题成立

433
00:19:16,011 --> 00:19:17,891
直接的因果却不成立

434
00:19:17,891 --> 00:19:21,211
其实这套结构你已经在用了

435
00:19:21,211 --> 00:19:22,711
只是没有意识到而已

436
00:19:22,711 --> 00:19:27,391
现在iPhone上那个把问题转给ChatGPT的功能

437
00:19:27,391 --> 00:19:29,551
走的就是第一条的潜通道

438
00:19:29,551 --> 00:19:32,131
它帮你把话听明白

439
00:19:32,131 --> 00:19:33,711
把答案写出来

440
00:19:33,711 --> 00:19:35,311
但你的日程

441
00:19:35,311 --> 00:19:36,271
你的照片

442
00:19:36,271 --> 00:19:37,311
你的联系人

443
00:19:37,311 --> 00:19:39,271
苹果一样都没有交出去

444
00:19:39,271 --> 00:19:41,951
昨天代码里露出来的

445
00:19:41,951 --> 00:19:45,211
是这条通道旁边还预备了一条更深的

446
00:19:45,211 --> 00:19:48,691
所以这件事不是从零开始的新闻

447
00:19:48,691 --> 00:19:50,991
是一条已经走了一年多的路

448
00:19:50,991 --> 00:19:55,331
突然被人看见了下一阶段的路径已经铺好了

449
00:19:55,331 --> 00:19:59,931
但我还是不想把这件事讲成了平台永远云

450
00:19:59,931 --> 00:20:02,391
模型公司不是没有拍

451
00:20:02,391 --> 00:20:04,491
他们可以自己做应用

452
00:20:04,491 --> 00:20:05,471
做浏览器

453
00:20:05,471 --> 00:20:06,371
做硬件

454
00:20:06,371 --> 00:20:09,391
也可以把能力做得足够的强

455
00:20:09,391 --> 00:20:12,051
强到用户宁愿绕开Siri

456
00:20:12,051 --> 00:20:13,351
主动去找它

457
00:20:13,351 --> 00:20:15,891
苹果那边的风险呢

458
00:20:15,891 --> 00:20:17,731
同样是真实的

459
00:20:17,731 --> 00:20:19,711
如果它长期不开放选择

460
00:20:19,711 --> 00:20:21,411
或者自己的模型呢

461
00:20:21,411 --> 00:20:22,531
又明显的落后

462
00:20:22,531 --> 00:20:24,131
那用户呢

463
00:20:24,131 --> 00:20:25,771
就会把最要紧的火

464
00:20:25,771 --> 00:20:27,531
搬到别的入口去

465
00:20:27,531 --> 00:20:29,691
掌握着门的人

466
00:20:29,691 --> 00:20:31,671
确实是有优势的

467
00:20:31,671 --> 00:20:34,611
但门外面的人不是只能干等着

468
00:20:34,611 --> 00:20:38,591
所以今天最值得记住的不是谁赢了

469
00:20:38,591 --> 00:20:43,251
而是苹果到现在都还没有公开选定一个长期的赢家

470
00:20:43,251 --> 00:20:46,351
他把可以换成写进的代码

471
00:20:46,351 --> 00:20:48,171
这把什么时候换

472
00:20:48,171 --> 00:20:52,371
换谁能换得多深的钱留在自己的手里

473
00:20:52,371 --> 00:20:54,771
这是一种选择钱

474
00:20:54,771 --> 00:20:56,731
选择钱值钱

475
00:20:56,731 --> 00:20:59,251
因为他让苹果可以接着看

476
00:20:59,251 --> 00:21:00,951
谁的模型更好

477
00:21:00,951 --> 00:21:03,411
谁愿意接受它的隐私规则

478
00:21:03,411 --> 00:21:06,171
谁肯给出更合适的价钱

479
00:21:06,171 --> 00:21:09,411
可选择钱也是有代价的

480
00:21:09,411 --> 00:21:12,231
苹果每多赚住一天

481
00:21:12,231 --> 00:21:14,251
用户就多一天

482
00:21:14,251 --> 00:21:16,311
没法自己挑信头构的模型

483
00:21:16,311 --> 00:21:18,671
开发者也多一天

484
00:21:18,671 --> 00:21:20,431
没法围着公开的接口

485
00:21:20,431 --> 00:21:22,271
做出真正能用的东西

486
00:21:22,271 --> 00:21:24,191
如果这份谨慎

487
00:21:24,191 --> 00:21:26,911
只是为了先把数据的边界想清楚

488
00:21:26,911 --> 00:21:29,531
那这个等待是值得的

489
00:21:29,531 --> 00:21:33,451
可如果安全变成了长期拖延竞争的理由呢

490
00:21:33,451 --> 00:21:35,291
监管和用户呢

491
00:21:35,291 --> 00:21:37,471
早晚会要他给个更清楚的说法

492
00:21:37,471 --> 00:21:41,051
问题从来不是花园该不该有墙

493
00:21:41,051 --> 00:21:44,091
是谁说了算门什么时候开

494
00:21:44,091 --> 00:21:46,151
以及开门的规矩呢

495
00:21:46,151 --> 00:21:47,491
要不要写在明处

496
00:21:47,491 --> 00:21:49,731
往后几个页

497
00:21:49,731 --> 00:21:52,471
大概会走出三种样子

498
00:21:52,471 --> 00:21:53,111
第一种

499
00:21:53,111 --> 00:21:55,811
苹果会继续把接口捂在私有器

500
00:21:55,811 --> 00:21:58,111
只在少数的合作里试

501
00:21:58,111 --> 00:22:00,891
普通人几乎感觉不到变化

502
00:22:00,891 --> 00:22:02,191
第二种

503
00:22:02,191 --> 00:22:05,411
设置页面里冒出一个有限的选择

504
00:22:05,411 --> 00:22:08,151
外部模型主要负责回答

505
00:22:08,151 --> 00:22:11,951
一碰到私人数据和系统操作就交回给Ciri

506
00:22:11,951 --> 00:22:13,991
第三种

507
00:22:13,991 --> 00:22:16,631
苹果真把更深的执行能力放出来

508
00:22:16,631 --> 00:22:20,351
同时配上清楚的前线提示和开发者的文档

509
00:22:20,351 --> 00:22:23,771
这三种的走向表面上都能叫

510
00:22:23,771 --> 00:22:25,231
支持第三方模型

511
00:22:25,231 --> 00:22:26,691
可对你

512
00:22:26,691 --> 00:22:27,691
对开发者

513
00:22:27,691 --> 00:22:28,731
对模型公司

514
00:22:28,731 --> 00:22:31,631
意义是差着十万八千里的

515
00:22:31,631 --> 00:22:34,031
下回再看见这种标题

516
00:22:34,031 --> 00:22:35,191
先问一句

517
00:22:35,191 --> 00:22:36,491
他说的是哪一层

518
00:22:36,491 --> 00:22:40,651
这件事你不用等分析师

519
00:22:40,651 --> 00:22:41,371
告诉你结论

520
00:22:41,371 --> 00:22:42,751
自己就能听

521
00:22:42,751 --> 00:22:45,271
升级完系统之后

522
00:22:45,271 --> 00:22:47,911
打开设置翻到细水的档里栏

523
00:22:47,911 --> 00:22:50,531
看看里面除了苹果自己的模型

524
00:22:50,531 --> 00:22:52,751
还列不列得出来第二个名字

525
00:22:52,751 --> 00:22:53,731
能不能点

526
00:22:53,731 --> 00:22:58,211
今天你翻下去大概率是什么都找不到的

527
00:22:58,211 --> 00:22:59,711
那就对了

528
00:22:59,711 --> 00:23:03,131
这说明钥匙还在苹果的手里呢

529
00:23:03,131 --> 00:23:06,211
哪天你正在那看见了一个可以点开的列表

530
00:23:06,211 --> 00:23:08,231
那这条新闻呢

531
00:23:08,231 --> 00:23:10,391
才算真的落到了你的手上

532
00:23:10,391 --> 00:23:13,731
不用去在特鲁斯在会议室里想什么

533
00:23:13,731 --> 00:23:15,911
第三个信号就是够了

534
00:23:15,911 --> 00:23:17,191
第一

535
00:23:17,191 --> 00:23:19,311
Siri或者系统设置里

536
00:23:19,311 --> 00:23:22,651
会不会出现你亲眼能看见的模型选择

537
00:23:22,651 --> 00:23:24,011
第二

538
00:23:24,011 --> 00:23:26,591
苹果会不会向第三方

539
00:23:26,591 --> 00:23:29,171
放出前线申请入口的开发者文档

540
00:23:29,171 --> 00:23:30,471
第三

541
00:23:30,471 --> 00:23:34,571
等外部模型真拿到了更深的系统能力的时候呢

542
00:23:34,571 --> 00:23:37,211
苹果打算怎么交代日历

543
00:23:37,211 --> 00:23:37,911
邮件

544
00:23:37,911 --> 00:23:39,831
联系人这些东西的边界

545
00:23:39,831 --> 00:23:42,511
这三样一样不出现

546
00:23:42,511 --> 00:23:44,451
那代码里的那个位置

547
00:23:44,451 --> 00:23:46,271
就还只是个可能性

548
00:23:46,271 --> 00:23:48,611
三样都出现了

549
00:23:48,611 --> 00:23:50,211
那这场入口之争

550
00:23:50,211 --> 00:23:52,151
才算真正翻到了下一页

551
00:23:52,151 --> 00:23:54,951
那你又会怎么选呢

552
00:23:54,951 --> 00:23:57,751
假如Siri真的给你三个开关

553
00:23:57,751 --> 00:23:59,091
一个呢

554
00:23:59,091 --> 00:24:00,111
达到最聪明

555
00:24:00,111 --> 00:24:02,111
一个可选的模型最多

556
00:24:02,111 --> 00:24:04,691
一个尽量不让你的私人数据

557
00:24:04,691 --> 00:24:05,531
离开苹果

558
00:24:05,531 --> 00:24:08,171
你最先看中哪一个呢

559
00:24:08,171 --> 00:24:09,731
评论器聊一聊

560
00:24:09,731 --> 00:24:11,691
你的排气是怎么样的

561
00:24:11,691 --> 00:24:14,331
以后再刷到某某模型

562
00:24:14,331 --> 00:24:16,491
拿下了iPhone的默认路口这种消息

563
00:24:16,491 --> 00:24:19,051
你也别指定着公司的名字

564
00:24:19,051 --> 00:24:20,771
先找那三个信号

565
00:24:20,771 --> 00:24:22,651
用户有没有开关

566
00:24:22,651 --> 00:24:24,411
开发者有没有入口

567
00:24:24,411 --> 00:24:27,491
数据到底交到了谁手上呢

568
00:24:27,491 --> 00:24:28,971
记得点赞

569
00:24:28,971 --> 00:24:29,651
订阅

570
00:24:29,651 --> 00:24:30,711
打开小铃铛

571
00:24:30,711 --> 00:24:32,551
咱们下期再见