1
00:00:00,000 --> 00:00:11,480
欢迎回到On Insight

2
00:00:11,480 --> 00:00:14,120
如果我们要选一个节点

3
00:00:14,120 --> 00:00:15,920
作为全球AI产业

4
00:00:15,920 --> 00:00:18,720
从学术梦幻期彻底跨入

5
00:00:18,720 --> 00:00:20,520
现实暴击期的分水里

6
00:00:20,520 --> 00:00:21,440
我敢打赌

7
00:00:21,440 --> 00:00:24,180
绝不会是哪家大厂又在PPT里

8
00:00:24,180 --> 00:00:26,480
发布了多大规模参数的模型

9
00:00:26,480 --> 00:00:28,900
基准测试又突破了多高的分数

10
00:00:28,900 --> 00:00:33,680
真正的变局往往发生在那些看起来极度安静

11
00:00:33,680 --> 00:00:37,240
甚至带着几分诡异的高层人事变动力

12
00:00:37,240 --> 00:00:39,780
就在最近谷歌内部发生了一场

13
00:00:39,780 --> 00:00:43,400
足以重塑整个硅谷底层格局的震荡

14
00:00:43,400 --> 00:00:46,040
在谷歌效力了整整27年

15
00:00:46,040 --> 00:00:49,260
被全球程序员尊为代码之神的

16
00:00:49,260 --> 00:00:52,440
谷歌首席科学家Jeff Dean正式宣布离职

17
00:00:52,440 --> 00:00:55,240
如果这只是一位老将退影的话

18
00:00:55,240 --> 00:00:59,960
倒也不至于让硅谷各大投资机构和安全部门彻夜未免

19
00:00:59,960 --> 00:01:02,320
真正认识背后发紧的是

20
00:01:02,320 --> 00:01:05,520
JFDN离职的时候带走了一个堪称

21
00:01:05,520 --> 00:01:09,180
硅谷全明星的豪华降维打击阵容

22
00:01:09,180 --> 00:01:14,660
这其中包括与JFDN合作20多年的底层系统传奇搭档

23
00:01:14,660 --> 00:01:16,380
桑杰·格马瓦特

24
00:01:16,380 --> 00:01:19,160
还有Google D.Mine的研究副总裁

25
00:01:19,160 --> 00:01:23,940
詹姆勒模型最核心的技术操盘手奥里奥尔·维尼亚尔斯

26
00:01:23,940 --> 00:01:29,000
以及AutoML领域先驱和Google Brain的联合创始人黎光莲

27
00:01:29,000 --> 00:01:31,060
这四个人加在一起

28
00:01:31,060 --> 00:01:36,040
几乎就是半个谷歌AI与底层分布式系统的技术灵魂

29
00:01:36,040 --> 00:01:39,600
他们组建了一家名为Discovery Loop的新公司

30
00:01:39,600 --> 00:01:43,620
按理说面对如此核心的技术骨干离职

31
00:01:43,620 --> 00:01:46,700
谷歌应该爆发一场血腥的公关战

32
00:01:46,700 --> 00:01:47,300
对吧

33
00:01:47,300 --> 00:01:49,000
但诡异的事情来了

34
00:01:49,000 --> 00:01:51,180
在继续后面的内容之前

35
00:01:51,180 --> 00:01:53,240
我想先请你帮我一个小忙

36
00:01:53,240 --> 00:01:55,100
如果你经常观看我的视频

37
00:01:55,100 --> 00:01:56,500
但还没有订阅我的频道

38
00:01:56,500 --> 00:01:58,040
欢迎点击订阅

39
00:01:58,040 --> 00:01:59,760
这样你就不会错过

40
00:01:59,760 --> 00:02:02,460
后续更多的科学资讯和深度解析

41
00:02:02,460 --> 00:02:05,800
那如果你觉得这期视频对你有所帮助

42
00:02:05,800 --> 00:02:07,580
也请帮忙点赞和分享

43
00:02:07,580 --> 00:02:09,360
让更多人看到这些内容

44
00:02:09,360 --> 00:02:12,280
那如果你希望给予我更多的支持

45
00:02:12,280 --> 00:02:13,880
也欢迎加入我的会员

46
00:02:13,880 --> 00:02:15,740
每月一杯咖啡的钱

47
00:02:15,740 --> 00:02:18,440
就是对我持续创作最大的支持

48
00:02:18,440 --> 00:02:19,900
好的我们继续

49
00:02:19,900 --> 00:02:21,840
这个诡异的事情就是

50
00:02:21,840 --> 00:02:23,760
谷歌不仅没有阻拦

51
00:02:23,760 --> 00:02:28,000
反而第一时间成为了Discovery Loop的初始投资方

52
00:02:28,000 --> 00:02:32,240
并且双手奉上了Google Cloud的超级算力集群

53
00:02:32,240 --> 00:02:36,540
双方直接签下了长期云服务绑定协议

54
00:02:36,540 --> 00:02:37,120
不仅如此

55
00:02:37,120 --> 00:02:38,340
就在同一天

56
00:02:38,340 --> 00:02:40,680
Google DeepMind内部也完成了一场

57
00:02:40,680 --> 00:02:43,220
近乎手术刀级别的高层重组

58
00:02:43,220 --> 00:02:45,960
DeepMind的灵魂人物Demith Hossabies

59
00:02:45,960 --> 00:02:48,940
正式卸任DeepMind CEO的日常管理

60
00:02:48,940 --> 00:02:51,460
身任Alphabet的首席科学家

61
00:02:51,460 --> 00:02:54,520
去操盘更为宏大的ATI理论中局

62
00:02:54,520 --> 00:02:57,880
而接替他全盘操盘Google DeepMind的新帅

63
00:02:57,880 --> 00:03:00,140
是原DevMind的CTO

64
00:03:00,140 --> 00:03:02,520
DQN与WaveNet的发明人

65
00:03:02,520 --> 00:03:04,000
科雷·卡伍克·乔鲁

66
00:03:04,000 --> 00:03:06,560
他越过了所有中间管理层

67
00:03:06,560 --> 00:03:09,440
直接向谷歌CEO桑德尔PX1汇报

68
00:03:09,440 --> 00:03:10,660
听到这里

69
00:03:10,660 --> 00:03:14,420
你脑海里一定盘旋着一个极其巨大的悬念

70
00:03:14,420 --> 00:03:16,940
就是谷歌为什么会心甘情愿的

71
00:03:16,940 --> 00:03:19,220
让自己的顶尖大脑集体外逃

72
00:03:19,220 --> 00:03:21,340
甚至还要掏钱掏算力

73
00:03:21,340 --> 00:03:22,740
帮他们去创业呢

74
00:03:22,740 --> 00:03:25,740
而那位既懂DQN强化学习

75
00:03:25,740 --> 00:03:28,880
又懂Vivnet原生语音的工程狂人

76
00:03:28,880 --> 00:03:30,260
柯雷接棒之后

77
00:03:30,260 --> 00:03:32,840
到底要把Gemlet带向何方呢

78
00:03:32,840 --> 00:03:35,320
今天我们就有一期节目

79
00:03:35,320 --> 00:03:38,460
彻底拆解这场看似是高层离职潮

80
00:03:38,460 --> 00:03:41,720
实则是谷歌AI技术大分流的

81
00:03:41,720 --> 00:03:43,120
硅谷终极棋局

82
00:03:43,120 --> 00:03:44,820
要解开这个谜团

83
00:03:44,820 --> 00:03:46,240
我们首先得搞清楚

84
00:03:46,240 --> 00:03:49,360
Jeff Dean带着这帮神仙阵容杀出去

85
00:03:49,360 --> 00:03:51,700
到底要去干一件什么样的事情

86
00:03:51,700 --> 00:03:54,200
在Discovery Loop的官宣里

87
00:03:54,200 --> 00:03:56,960
Jeff Dean直接描绘了公司的愿景

88
00:03:56,960 --> 00:04:00,920
那就是构建自动化的科学方法闭环

89
00:04:00,920 --> 00:04:01,900
简单来说

90
00:04:01,900 --> 00:04:03,380
过去几百年

91
00:04:03,380 --> 00:04:05,500
人类探索未知的方式

92
00:04:05,500 --> 00:04:07,300
本质上都是串行实验

93
00:04:07,300 --> 00:04:09,600
就是科学家提出假设

94
00:04:09,600 --> 00:04:10,880
亲手写代码

95
00:04:10,880 --> 00:04:11,760
跑一次实验

96
00:04:11,760 --> 00:04:13,140
等几天看结果

97
00:04:13,140 --> 00:04:13,900
分析数据

98
00:04:13,900 --> 00:04:15,880
然后再手动修改参数

99
00:04:15,880 --> 00:04:16,600
跑下一次

100
00:04:16,600 --> 00:04:18,300
在这套流程里

101
00:04:18,300 --> 00:04:20,720
最瓶颈的变量不是算力

102
00:04:20,720 --> 00:04:23,300
而是人类脑力的思考速度

103
00:04:23,300 --> 00:04:25,600
与手动操作的物理实验

104
00:04:25,600 --> 00:04:28,480
而Discovery Loop要构建一套

105
00:04:28,480 --> 00:04:31,280
完全自主运行的AIA层架构

106
00:04:31,280 --> 00:04:33,720
做到自动提出科研假设

107
00:04:33,720 --> 00:04:35,620
自动编写实验代码

108
00:04:35,620 --> 00:04:37,400
自动部署分布式计算

109
00:04:37,400 --> 00:04:39,220
自动分析实验指标

110
00:04:39,220 --> 00:04:41,160
最后自动递归迭代

111
00:04:41,160 --> 00:04:43,380
你注意看Discovery Loop的

112
00:04:43,380 --> 00:04:45,140
手机应用场景选择

113
00:04:45,140 --> 00:04:48,040
他们没有第一天就跑去见物理实验室

114
00:04:48,040 --> 00:04:50,220
而是选择吃自己的狗粮

115
00:04:50,220 --> 00:04:54,320
把机器学习研发本身作为第一个实验战场

116
00:04:54,320 --> 00:04:57,580
也就是说他们要用AI去自动设计

117
00:04:57,580 --> 00:04:58,560
自动搜索

118
00:04:58,560 --> 00:05:02,260
自动验证下一代超越Transformer的模型架构

119
00:05:02,260 --> 00:05:04,140
很多朋友可能会问

120
00:05:04,140 --> 00:05:06,840
用AI自动写代码做研究

121
00:05:06,840 --> 00:05:10,140
之前日本Sakana AI的项目不是也做过吗

122
00:05:10,140 --> 00:05:13,200
这正是我们接下来要拆解的关键

123
00:05:13,200 --> 00:05:18,100
Satan的AI的项目本质上是一个应用层的agent流程封装

124
00:05:18,100 --> 00:05:22,700
它是拿现成的GPT4O或者Cloud外包一层prompt

125
00:05:22,700 --> 00:05:26,300
去模拟科学家写代码跑实验写论文

126
00:05:26,300 --> 00:05:31,380
它的上限完全被底下调用的通用大模型给封死了

127
00:05:31,380 --> 00:05:35,140
而Discovery Loop完全是另一个维度的存在

128
00:05:35,140 --> 00:05:42,540
Jeff Dean和桑杰·格瓦马特是写出MapReduce Bigtable的超级系统架构师

129
00:05:42,540 --> 00:05:45,300
李光莲是AutoML的鼻祖

130
00:05:45,300 --> 00:05:47,020
奥里尔尔维尼亚尔斯

131
00:05:47,020 --> 00:05:50,580
是顶尖的神经网络搜索与多模态专家

132
00:05:50,580 --> 00:05:53,960
他们要做的是从底层分布式计算

133
00:05:53,960 --> 00:05:55,520
硬件算力调度

134
00:05:55,520 --> 00:06:00,120
神经架构搜索到算法闭环的系统级重构

135
00:06:00,120 --> 00:06:02,180
他们不是在用AI写论文

136
00:06:02,180 --> 00:06:04,660
他们是在用超级计算机集群

137
00:06:04,660 --> 00:06:09,260
让AI去做数以万计的硬件级与架构级实验

138
00:06:09,260 --> 00:06:13,000
直接在算法底层寻找下一个Transformer

139
00:06:13,000 --> 00:06:15,260
那么回到我们开头的悬念

140
00:06:15,260 --> 00:06:19,560
就是谷歌为什么不把这个超级团队留在内部

141
00:06:19,560 --> 00:06:23,340
而是支持他们独立成立一家公共利益公司呢

142
00:06:23,340 --> 00:06:25,980
因为谷歌算透了一笔账

143
00:06:25,980 --> 00:06:29,720
像AI自主研发下一代AI这种前沿探索

144
00:06:29,720 --> 00:06:33,640
风险极高 烧钱极大 失败率也极高

145
00:06:33,640 --> 00:06:35,720
如果放在谷歌Digmy内部

146
00:06:35,720 --> 00:06:40,440
研究员们每天要面对华尔街的财报审查和季度KPI

147
00:06:40,440 --> 00:06:43,700
就会被无限牵扯精力去搞产品适配

148
00:06:43,700 --> 00:06:46,260
通过把JFDN团队外部化

149
00:06:46,260 --> 00:06:48,740
谷歌既让这一帮顶尖天才

150
00:06:48,740 --> 00:06:52,280
摆脱了大公司病和商业汇报的枷锁

151
00:06:52,280 --> 00:06:57,140
又能通过初始投资方和独家GoCloud算力的身份

152
00:06:57,140 --> 00:07:01,820
牢牢锁定未来潜在的技术回流与云服务收益

153
00:07:01,820 --> 00:07:06,580
这一招叫做前沿高风险战略的外部结偶

154
00:07:06,580 --> 00:07:11,440
如果说Jeff Dean的离职是把远期高风险探索拆分了出去

155
00:07:11,440 --> 00:07:16,740
那么接下来发生的事情才是对Jamlight产品生态影响最深远

156
00:07:16,740 --> 00:07:18,240
也是最暴力的一环

157
00:07:18,240 --> 00:07:22,700
接掌Google D.My的全局率颖的科雷卡伍克·乔武

158
00:07:22,700 --> 00:07:27,180
是一个与Demis Hossabies风格截然不同的工程技术派

159
00:07:27,180 --> 00:07:31,480
Demis是国际象棋神童和神经科学家

160
00:07:31,480 --> 00:07:33,400
它充满着理想主义

161
00:07:33,400 --> 00:07:36,180
而柯雷是把深度强化学习

162
00:07:36,180 --> 00:07:38,660
真正推向AlphaGo破局

163
00:07:38,660 --> 00:07:40,620
把WaveNet徐徐语音

164
00:07:40,620 --> 00:07:45,620
真正变成极异手机里语音引擎的超级工程操盘手

165
00:07:45,620 --> 00:07:47,760
在柯雷掌舵之后

166
00:07:47,760 --> 00:07:50,640
Google G-Mind释放出了一个极其明确的信号

167
00:07:50,640 --> 00:07:53,060
就是全面放弃学术包袱

168
00:07:53,060 --> 00:07:57,240
转向极致工程演进与大规模商业化暴击

169
00:07:57,240 --> 00:07:58,680
在他的主持下

170
00:07:58,680 --> 00:07:59,960
Gemlet的底层架构

171
00:08:00,000 --> 00:08:02,840
Gamlight的底层架构正在发生三重极其硬核的眼镜

172
00:08:02,840 --> 00:08:04,760
首先最根本的变化是

173
00:08:04,760 --> 00:08:08,840
彻底打破对预训练参数崇拜的单位幻想

174
00:08:08,840 --> 00:08:13,720
全面转向后训练与推理期计算的深度融合

175
00:08:13,720 --> 00:08:17,800
在过去两年 行业盲目堆叠预训练参数

176
00:08:17,800 --> 00:08:21,720
但预训练越往后 编辑效应递减越严重

177
00:08:21,720 --> 00:08:24,100
算力成本高到令人发指

178
00:08:24,100 --> 00:08:27,900
柯雷结合他在DQN强化学习领域的积累

179
00:08:27,900 --> 00:08:31,280
给Gemla引入了过程奖励模型PRM

180
00:08:31,280 --> 00:08:34,160
与蒙特卡洛书搜索MCTS

181
00:08:34,160 --> 00:08:37,080
打个比方 传统的语言模型

182
00:08:37,080 --> 00:08:39,820
就像在考场上没有任何草稿纸

183
00:08:39,820 --> 00:08:41,780
靠直觉备答案的考生

184
00:08:41,780 --> 00:08:44,180
遇到简单的题那是没问题的

185
00:08:44,180 --> 00:08:47,520
但遇到复杂逻辑推导或者长流程编程

186
00:08:47,520 --> 00:08:50,520
中间要是错一步后面就全乱了

187
00:08:50,520 --> 00:08:53,460
而PRM与MCTS的融合

188
00:08:53,460 --> 00:08:58,000
相当于给Gemla配了一张无线扩展的虚拟草稿纸

189
00:08:58,000 --> 00:08:59,920
和一位微观监考官

190
00:08:59,920 --> 00:09:01,520
在推理的每一步

191
00:09:01,520 --> 00:09:06,080
PRM都会对它的逻辑合理性进行毫秒级评分

192
00:09:06,080 --> 00:09:10,000
同时MCTS数搜索会在后台预判

193
00:09:10,000 --> 00:09:11,700
未来几十步的走向

194
00:09:11,700 --> 00:09:14,720
如果发现某个分支走向死胡同

195
00:09:14,720 --> 00:09:17,440
模型会在推理阶段自动撤回

196
00:09:17,440 --> 00:09:19,140
并重新搜索最右键

197
00:09:19,140 --> 00:09:20,240
这意味着

198
00:09:20,240 --> 00:09:24,180
在不用无限膨胀预讯链参数的前提下

199
00:09:24,180 --> 00:09:26,240
GEMLINE处理高难度编程

200
00:09:26,240 --> 00:09:29,620
复杂数学推导和A乘的规划的成功率

201
00:09:29,620 --> 00:09:31,420
称指数级提升

202
00:09:31,420 --> 00:09:33,940
紧接着第二种转变在于

203
00:09:33,940 --> 00:09:36,540
系统级推理效率的极限拉链

204
00:09:36,540 --> 00:09:40,660
那就是构建Gemini Spark端云协同分层架构

205
00:09:40,660 --> 00:09:44,260
这里稍微提一下我个人对Spark的使用体验

206
00:09:44,260 --> 00:09:46,060
虽然它还是Beta版本

207
00:09:46,060 --> 00:09:47,260
但先说结论

208
00:09:47,260 --> 00:09:49,080
我觉得它超级好用

209
00:09:49,080 --> 00:09:50,740
我会单独出一期视频

210
00:09:50,740 --> 00:09:54,100
聊一下Gemini Spark和Hermes Agent的配合使用

211
00:09:54,100 --> 00:09:56,780
作为Google订阅的长期体验者

212
00:09:56,780 --> 00:10:00,520
我明显感觉最近Gemini的使用限额提高了

213
00:10:00,520 --> 00:10:03,880
他们虽然没有发布像Fable 5这类的模型

214
00:10:03,880 --> 00:10:06,400
但是无论是Gemini 3.6 Flash

215
00:10:06,400 --> 00:10:09,440
还是Gemlet 3.1 Pro都变得更聪明了

216
00:10:09,440 --> 00:10:11,740
好 我们继续这个第二重转变

217
00:10:11,740 --> 00:10:14,260
在商业高并发场景中

218
00:10:14,260 --> 00:10:16,420
用户问一句 今天天气怎么样

219
00:10:16,420 --> 00:10:20,680
如果都要调用一次千亿参数的单体大模型

220
00:10:20,680 --> 00:10:23,820
显存和电力开销能把大厂烧穿

221
00:10:23,820 --> 00:10:26,220
克雷推动的Gemlet Spark架构

222
00:10:26,220 --> 00:10:30,900
把整套AI系统做成了像特种部队一样的分层调度

223
00:10:30,900 --> 00:10:35,840
前端部署及轻量及低延迟的Gemlet Spark代理模型

224
00:10:35,840 --> 00:10:37,800
负责高频意图识别

225
00:10:37,800 --> 00:10:40,240
上下门过滤和轻量任务

226
00:10:40,240 --> 00:10:41,820
只有当前线判定

227
00:10:41,820 --> 00:10:44,180
涉及复杂长链条逻辑的时候

228
00:10:44,180 --> 00:10:47,560
才会无缝调动云端期间大模型

229
00:10:47,560 --> 00:10:48,960
这种分层架构

230
00:10:48,960 --> 00:10:51,860
直接把单位Token成本打了下来

231
00:10:51,860 --> 00:10:55,340
这才是支撑数亿并发的商业硬实力

232
00:10:55,340 --> 00:10:56,900
而第三种变化

233
00:10:56,900 --> 00:10:59,460
则是柯雷将他的看家本领

234
00:10:59,460 --> 00:11:02,500
VIVNET原生音频与流逝生成

235
00:11:02,500 --> 00:11:06,240
深度融合到了詹姆賴的全模态底座中

236
00:11:06,240 --> 00:11:08,020
以往的多模态AI

237
00:11:08,020 --> 00:11:11,420
往往是文本模型外挂语音识别

238
00:11:11,420 --> 00:11:13,960
与语音合成模块的拼接车

239
00:11:13,960 --> 00:11:16,120
延迟高 语调生硬

240
00:11:16,120 --> 00:11:18,120
而柯雷主导下的詹姆賴

241
00:11:18,120 --> 00:11:21,440
实现了真正的原生全模态流逝交互

242
00:11:21,440 --> 00:11:23,280
不仅能实现毫秒级

243
00:11:23,280 --> 00:11:25,840
实时音视频双向对打

244
00:11:25,840 --> 00:11:27,640
还能在听懂说话的同时

245
00:11:27,640 --> 00:11:29,180
自主执行代码

246
00:11:29,180 --> 00:11:31,580
甚至在手机屏幕上动态生成

247
00:11:31,580 --> 00:11:33,600
UI界面与交互空间

248
00:11:33,600 --> 00:11:35,600
这已经不是在聊天了

249
00:11:35,600 --> 00:11:40,580
这是直接把自然语言变成了实时软件生成的界面

250
00:11:40,580 --> 00:11:42,800
那底层架构演进的再漂亮

251
00:11:42,800 --> 00:11:45,000
如果不能转化为商业流水

252
00:11:45,000 --> 00:11:47,280
在今天也是一张废纸

253
00:11:47,280 --> 00:11:50,900
科雷·卡伍克·乔鲁这次最让人惊悚的调整

254
00:11:50,900 --> 00:11:52,680
在于他不仅管科研

255
00:11:52,680 --> 00:11:57,600
还同时兼任了Gemline产品与开发者生态的最高指挥官

256
00:11:57,600 --> 00:11:59,940
在过去的大厂体制里

257
00:11:59,940 --> 00:12:04,360
科研实验室和商品产品部门是完全脱节的

258
00:12:04,360 --> 00:12:05,580
实验室发了论文

259
00:12:05,580 --> 00:12:08,780
产品部门要花大半年去沟通打包

260
00:12:08,780 --> 00:12:10,740
而柯雷一把抓之后

261
00:12:10,740 --> 00:12:13,820
直接祭出了商业落地的三大杀招

262
00:12:13,820 --> 00:12:15,820
在终端入口上

263
00:12:15,820 --> 00:12:19,000
他们开始加速清洗传统的语音助手

264
00:12:19,000 --> 00:12:22,440
全面用GM来替换掉Android和VLOS上

265
00:12:22,440 --> 00:12:24,560
那个指挥社闹钟的assistant

266
00:12:24,560 --> 00:12:28,860
赋予他操作系统级的最高上下文读取权限

267
00:12:28,860 --> 00:12:31,160
与跨应用API调用能力

268
00:12:31,160 --> 00:12:35,040
直接把它打造成操作系统的中央控制大脑

269
00:12:35,040 --> 00:12:38,520
比如在ValOS手表或者车载系统里

270
00:12:38,520 --> 00:12:42,340
你只需要说帮我安排今晚与张总的商务晚餐

271
00:12:42,340 --> 00:12:46,040
并准备一份他公司最新财报的背景简报

272
00:12:46,040 --> 00:12:50,160
Gamlight就会自动调取邮件和地图订餐厅

273
00:12:50,160 --> 00:12:54,440
同时在云端提取财报生成简报推送给你

274
00:12:54,440 --> 00:12:55,580
与此同时

275
00:12:55,580 --> 00:12:58,300
他们还大幅压缩了科研成果

276
00:12:58,300 --> 00:13:00,440
转化到商业API的时间

277
00:13:00,440 --> 00:13:04,440
在谷歌云Vertex AI和开发者API上

278
00:13:04,440 --> 00:13:07,140
前沿团队在Achern代码生成

279
00:13:07,140 --> 00:13:10,800
自动模型评估和推理加速上的突破

280
00:13:10,800 --> 00:13:13,560
能以周级别的速度直接部署上线

281
00:13:13,560 --> 00:13:15,720
让科研成果在第一秒

282
00:13:15,720 --> 00:13:17,680
直接转化为商业账单

283
00:13:17,680 --> 00:13:21,340
形成技术产品与算力的超级费轮

284
00:13:21,340 --> 00:13:22,740
更关键的是

285
00:13:22,740 --> 00:13:25,960
谷歌正在悄无声息地改变自己的盈利模式

286
00:13:25,960 --> 00:13:30,000
当用户不再翻页点击搜索广告链接

287
00:13:30,000 --> 00:13:32,740
而是让Gamlet给出最终答案时

288
00:13:32,740 --> 00:13:34,020
谷歌的解法是

289
00:13:34,020 --> 00:13:37,300
直接占据agent决策路径的入口

290
00:13:37,300 --> 00:13:40,640
当Gamlet在每日简报或者个人助理里

291
00:13:40,640 --> 00:13:43,260
帮你规划行程推荐服务时

292
00:13:43,260 --> 00:13:45,440
它在生成方案的同时

293
00:13:45,440 --> 00:13:47,960
精准接入商业API入口

294
00:13:47,960 --> 00:13:51,500
谷歌由此从一个信息检索广告商

295
00:13:51,500 --> 00:13:54,640
变成了全局智能决策的开门人

296
00:13:54,640 --> 00:13:56,920
抽取服务分账和佣金

297
00:13:56,920 --> 00:13:59,580
拆决员Jeff Dean的离职创业

298
00:13:59,580 --> 00:14:02,140
与柯磊掌舵下的Deep Mind重组

299
00:14:02,140 --> 00:14:06,440
我们终于能看清这场硅谷AI大戏的全貌了

300
00:14:06,440 --> 00:14:11,140
硅谷的AI战争正是告别了依靠概念和论文引用量

301
00:14:11,140 --> 00:14:13,780
获取高估值的学术竞赛阶段

302
00:14:13,780 --> 00:14:16,700
全面跨入了拼推理成本

303
00:14:16,700 --> 00:14:18,380
拼系统稳定性

304
00:14:18,380 --> 00:14:19,880
拼终端控制力

305
00:14:19,880 --> 00:14:23,920
以及拼商业造学能力的工程与商业深水去

306
00:14:23,920 --> 00:14:27,800
对于我们所有在做AI研发架构设计

307
00:14:27,800 --> 00:14:29,920
或者技术创业的朋友来说

308
00:14:29,920 --> 00:14:31,720
谷歌的这一系列辩证

309
00:14:31,720 --> 00:14:34,860
留下了三条极其宝贵的工程启发

310
00:14:34,860 --> 00:14:36,940
首先不要盲目迷信

311
00:14:36,940 --> 00:14:39,580
纯域训练参数的单位堆叠

312
00:14:39,580 --> 00:14:41,940
多去研究推理器计算

313
00:14:41,940 --> 00:14:46,120
通过过程奖励模型打分与蒙特卡洛数搜索

314
00:14:46,120 --> 00:14:49,300
让模型在推理阶段学会自我纠错

315
00:14:49,300 --> 00:14:53,120
这是低成本实现逻辑跨越的唯一解法

316
00:14:53,120 --> 00:14:56,580
其次要学会拥抱端云分层架构

317
00:14:56,580 --> 00:14:58,520
平气无论任务大小

318
00:14:58,520 --> 00:15:01,620
一力调用千亿大模型的粗暴思维

319
00:15:01,620 --> 00:15:03,400
学习Jamler Spark的设计

320
00:15:03,400 --> 00:15:06,900
用轻量级模型做好前置意图路由

321
00:15:06,900 --> 00:15:10,020
才能在高并发任务中守住成本底线

322
00:15:10,020 --> 00:15:13,800
最后我们需要重新思考软件入口的定义

323
00:15:13,800 --> 00:15:16,260
未来的软件竞争不再是争夺

324
00:15:16,260 --> 00:15:18,420
手机屏幕上的APP图标位置

325
00:15:18,420 --> 00:15:20,980
而是争夺能否被系统及agent

326
00:15:20,980 --> 00:15:23,160
API决策链所调用

327
00:15:23,160 --> 00:15:25,120
技术不相信幻觉

328
00:15:25,120 --> 00:15:26,720
商业不相信眼泪

329
00:15:26,720 --> 00:15:27,860
这里是All Insight

330
00:15:27,860 --> 00:15:28,760
我们下期见