1
00:00:00,166 --> 00:00:02,533
Meta开源Muse glimmer30B

2
00:00:02,700 --> 00:00:04,366
最值得看的不是参数

3
00:00:04,366 --> 00:00:05,533
也不是跑分

4
00:00:05,600 --> 00:00:07,366
而是它把大模型竞争

5
00:00:07,366 --> 00:00:09,166
从云端大厂的机房

6
00:00:09,200 --> 00:00:11,200
重新推回普通人的电脑

7
00:00:11,300 --> 00:00:12,866
一个30B级别

8
00:00:12,866 --> 00:00:13,900
开放权重

9
00:00:13,933 --> 00:00:15,966
面向本地agent的模型

10
00:00:16,133 --> 00:00:17,733
放在今天这个节点

11
00:00:17,766 --> 00:00:18,900
意思很明确

12
00:00:19,266 --> 00:00:21,700
AI不能只是一张云服务账单

13
00:00:21,733 --> 00:00:24,900
也不能永远被少数公司关在黑盒里

14
00:00:25,000 --> 00:00:28,066
这件事要放在Meta的整体叙事里看

15
00:00:28,500 --> 00:00:30,800
扎克伯格最近把话说得很直

16
00:00:31,366 --> 00:00:32,533
人工智能的主要

17
00:00:32,533 --> 00:00:34,533
目的应该是发明创造

18
00:00:34,733 --> 00:00:36,766
不应该只是自动化替代

19
00:00:37,100 --> 00:00:39,366
超级智能如果真的会到来

20
00:00:39,500 --> 00:00:42,666
那它不应该只掌握在几家公司手里

21
00:00:42,933 --> 00:00:45,366
而应该尽量广泛的分发给

22
00:00:45,366 --> 00:00:48,066
普通人开发者和小团队

23
00:00:48,700 --> 00:00:50,566
Muse Glimmer就是这个叙事

24
00:00:50,566 --> 00:00:52,600
落到产品上的一颗棋子

25
00:00:52,800 --> 00:00:55,100
Muse Glimmer 30B的重点不是

26
00:00:55,100 --> 00:00:57,266
挑战闭源模型的全部能力

27
00:00:57,366 --> 00:01:00,566
而是给本地AI留出一条现实路线

28
00:01:00,700 --> 00:01:02,300
现在很多模型动不动

29
00:01:02,300 --> 00:01:04,500
就是几千亿上万亿参数

30
00:01:04,666 --> 00:01:05,866
听起来很震撼

31
00:01:05,900 --> 00:01:08,000
但普通用户根本跑不起来

32
00:01:08,333 --> 00:01:10,000
他们需要昂贵GPU

33
00:01:10,200 --> 00:01:11,600
需要云端推理

34
00:01:11,666 --> 00:01:13,600
需要订阅和调用额度

35
00:01:13,900 --> 00:01:15,966
Meta选择30B这个体量

36
00:01:15,966 --> 00:01:17,066
就是在说

37
00:01:17,200 --> 00:01:20,133
模型不一定越大越接近未来

38
00:01:20,200 --> 00:01:22,933
能部署能修改能嵌进

39
00:01:22,933 --> 00:01:25,700
工作流才更接近真实使用

40
00:01:25,933 --> 00:01:28,400
这和LLAMA时代的路线一脉相承

41
00:01:29,100 --> 00:01:31,466
Meta一直在开放模型上押注

42
00:01:31,666 --> 00:01:33,533
只是中间经历过摇摆

43
00:01:33,766 --> 00:01:38,200
当Openai Anthropic Google把高端模型

44
00:01:38,200 --> 00:01:40,500
能力不断往闭源体系里收

45
00:01:41,000 --> 00:01:44,533
Meta反而重新把开放权重当成差异化

46
00:01:44,600 --> 00:01:46,933
它没必要在每一项榜单上赢

47
00:01:47,166 --> 00:01:49,100
它要赢的是开发者心智

48
00:01:49,600 --> 00:01:51,566
谁能让更多人拿去改

49
00:01:51,666 --> 00:01:53,733
拿去跑拿去做产品

50
00:01:53,966 --> 00:01:56,966
谁就能在AI底层生态里留下位置

51
00:01:57,266 --> 00:02:00,766
musglm还带有execute torch和PTE

52
00:02:00,766 --> 00:02:02,400
这样的边缘部署意味

53
00:02:02,566 --> 00:02:04,066
这个细节很重要

54
00:02:04,566 --> 00:02:06,966
它不是只给服务器看的模型卡

55
00:02:07,133 --> 00:02:09,600
而是在暗示手机笔记本

56
00:02:09,600 --> 00:02:12,200
边缘设备本地工作站都

57
00:02:12,200 --> 00:02:14,700
可能成为AI agent的运行环境

58
00:02:14,800 --> 00:02:15,866
未来AI agent

59
00:02:15,866 --> 00:02:17,933
不一定每一步都要回云端

60
00:02:18,266 --> 00:02:20,166
他可以在本地看文件读

61
00:02:20,166 --> 00:02:23,066
图片调工具处理私人数据

62
00:02:23,100 --> 00:02:24,366
再把真正复杂的

63
00:02:24,366 --> 00:02:26,166
任务交给云端大模型

64
00:02:26,300 --> 00:02:29,900
这种路线最大的价值是隐私和成本

65
00:02:30,366 --> 00:02:33,700
用户把本地文件工作资料聊天

66
00:02:33,700 --> 00:02:36,666
记录图片和代码交给云端模型

67
00:02:36,933 --> 00:02:38,500
永远会有心理门槛

68
00:02:38,900 --> 00:02:39,966
企业企业更敏感

69
00:02:40,066 --> 00:02:42,333
尤其涉及客户数据合同

70
00:02:42,500 --> 00:02:44,333
研发资料和内部流程

71
00:02:44,466 --> 00:02:46,466
一个能在本地跑的模型

72
00:02:46,600 --> 00:02:48,333
就算能力不是最强

73
00:02:48,500 --> 00:02:49,900
也会因为数据不出

74
00:02:49,900 --> 00:02:51,466
设备而变得有价值

75
00:02:51,666 --> 00:02:53,133
很多真实场景不需要

76
00:02:53,133 --> 00:02:54,766
世界最聪明的模型

77
00:02:55,300 --> 00:02:57,000
只需要一个足够聪明

78
00:02:57,066 --> 00:02:59,766
足够便宜足够可控的助手

79
00:02:59,866 --> 00:03:01,500
这也是开源模型真正

80
00:03:01,500 --> 00:03:03,166
打闭源模型的地方

81
00:03:03,733 --> 00:03:06,533
闭源模型擅长提供最强通用能力

82
00:03:06,766 --> 00:03:09,700
开源模型擅长被改造成具体工具

83
00:03:09,766 --> 00:03:11,366
小团队可以微调它

84
00:03:11,800 --> 00:03:13,800
企业可以把它部署到内网

85
00:03:13,966 --> 00:03:15,200
开发者可以把它接

86
00:03:15,200 --> 00:03:16,900
进自己的agent框架

87
00:03:17,466 --> 00:03:20,000
硬件厂商可以把它塞进边缘设备

88
00:03:20,200 --> 00:03:21,966
闭源模型卖的是能力

89
00:03:21,966 --> 00:03:24,066
开放模型卖的是可塑性

90
00:03:24,266 --> 00:03:25,933
两者不是同一场比赛

91
00:03:26,100 --> 00:03:28,300
Meta这次还把数据中心社区

92
00:03:28,300 --> 00:03:30,266
基金放在同一个叙事里

93
00:03:30,300 --> 00:03:32,466
说明它知道AI扩张正在

94
00:03:32,466 --> 00:03:34,066
遇到新的政治成本

95
00:03:34,700 --> 00:03:36,900
数据中心需要电力土地

96
00:03:36,900 --> 00:03:39,466
水输电设施和社区许可

97
00:03:39,533 --> 00:03:41,566
过去科技公司建机房

98
00:03:41,566 --> 00:03:43,166
地方政府往往欢迎

99
00:03:43,166 --> 00:03:45,100
因为它带来投资和税收

100
00:03:45,200 --> 00:03:46,600
现在情况变了

101
00:03:46,666 --> 00:03:48,733
居民开始担心电费上涨

102
00:03:49,166 --> 00:03:51,500
环境压力和基础设施被占用

103
00:03:51,600 --> 00:03:53,933
AI越庞大越不可能只用

104
00:03:53,933 --> 00:03:55,700
技术语言解释自己

105
00:03:55,933 --> 00:03:58,533
10亿美元社区基金表面上是补偿

106
00:03:58,866 --> 00:04:01,366
背后是AI基建时代的通行费

107
00:04:02,066 --> 00:04:03,733
Meta想告诉当地居民

108
00:04:04,000 --> 00:04:06,500
数据中心不是只把电和水抽走

109
00:04:06,500 --> 00:04:09,533
也会把钱岗位和公共服务带回来

110
00:04:09,900 --> 00:04:13,200
警察消防员社区成员被点名

111
00:04:13,266 --> 00:04:15,466
是一种很现实的政治沟通

112
00:04:15,600 --> 00:04:17,333
AI公司已经发现

113
00:04:17,466 --> 00:04:19,933
光说技术改变世界没用了

114
00:04:20,200 --> 00:04:22,800
真正被影响的人要看到具体回报

115
00:04:23,066 --> 00:04:26,466
Musglimmer和社区基金看似是两件事

116
00:04:26,466 --> 00:04:28,166
其实对应同一个问题

117
00:04:28,466 --> 00:04:30,733
AI权力会不会过度集中

118
00:04:31,200 --> 00:04:32,766
模型集中在云端

119
00:04:32,933 --> 00:04:34,966
算力集中在数据中心

120
00:04:35,066 --> 00:04:37,133
收益集中在少数公司

121
00:04:37,400 --> 00:04:40,100
成本却扩散给社区和消费者

122
00:04:40,166 --> 00:04:41,800
这套结构如果不调整

123
00:04:41,800 --> 00:04:43,200
反弹会越来越大

124
00:04:43,366 --> 00:04:47,066
Meta现在讲开放讲本地讲回馈社区

125
00:04:47,100 --> 00:04:48,666
就是在给自己的AI

126
00:04:48,666 --> 00:04:50,533
扩张换一种合法性

127
00:04:50,766 --> 00:04:53,400
这不代表Meta突然变成公益组织

128
00:04:53,600 --> 00:04:55,500
开源也不是纯粹善意

129
00:04:55,866 --> 00:04:57,966
Meta的商业逻辑很清楚

130
00:04:58,366 --> 00:05:00,300
它没有像Openai那样把

131
00:05:00,300 --> 00:05:02,500
订阅模型做成第一入口

132
00:05:02,700 --> 00:05:05,366
也没有Google那样的搜索默认位置

133
00:05:05,766 --> 00:05:07,533
它最擅长的是平台

134
00:05:07,566 --> 00:05:10,133
分发广告和开发者网络

135
00:05:10,300 --> 00:05:11,900
开放模型能削弱

136
00:05:11,900 --> 00:05:13,800
闭源模型的收费能力

137
00:05:13,966 --> 00:05:16,100
降低竞争对手的护城河

138
00:05:16,400 --> 00:05:18,300
同时让更多应用围绕

139
00:05:18,300 --> 00:05:20,166
Meta的模型生态生长

140
00:05:20,466 --> 00:05:22,566
如果一个开放模型足够好

141
00:05:22,566 --> 00:05:24,700
大量开发者就会围绕它做

142
00:05:24,700 --> 00:05:27,600
工具插件微调和部署方案

143
00:05:27,700 --> 00:05:30,466
哪怕Meta不直接收每一次调用费

144
00:05:30,500 --> 00:05:33,666
它也能影响标准框架和生态方向

145
00:05:33,900 --> 00:05:36,700
安卓当年不是靠系统授权费赚钱

146
00:05:37,066 --> 00:05:39,700
而是靠开放系统占住移动入口

147
00:05:39,966 --> 00:05:42,733
LLAMA和Muse glimmer也有类似逻辑

148
00:05:42,933 --> 00:05:45,666
用开放打破别人的封闭利润池

149
00:05:45,800 --> 00:05:48,300
再在更大的生态里寻找收益

150
00:05:48,600 --> 00:05:52,566
Muse glimmer的30B体量还有一个现实意义

151
00:05:52,766 --> 00:05:55,066
它让AI从演示回到产品

152
00:05:55,333 --> 00:05:58,300
很多AI发布会喜欢展示极限能力

153
00:05:58,400 --> 00:06:01,200
仿佛模型能解决所有复杂问题

154
00:06:01,400 --> 00:06:02,900
但真正落地时

155
00:06:02,900 --> 00:06:05,266
企业关心的是成本延迟

156
00:06:05,533 --> 00:06:08,566
部署安全稳定可维护

157
00:06:08,666 --> 00:06:10,100
一个30B模型

158
00:06:10,100 --> 00:06:12,066
如果能在普通设备或

159
00:06:12,066 --> 00:06:14,533
消费级GPU上完成稳定任务

160
00:06:14,866 --> 00:06:17,700
它的商业价值可能比一个只能在

161
00:06:17,700 --> 00:06:20,866
云端高价运行的巨型模型更直接

162
00:06:20,933 --> 00:06:22,466
尤其是agent场景

163
00:06:22,666 --> 00:06:24,266
agent不是一次聊天

164
00:06:24,266 --> 00:06:25,766
而是一连串动作

165
00:06:26,000 --> 00:06:28,066
理解目标拆任务

166
00:06:28,366 --> 00:06:30,200
读文件调用工具

167
00:06:30,333 --> 00:06:31,366
检查结果

168
00:06:31,400 --> 00:06:32,700
失败后重试

169
00:06:32,800 --> 00:06:35,500
这里面很多步骤并不需要最强模型

170
00:06:35,600 --> 00:06:38,666
文件分类表格检查界面操作

171
00:06:38,700 --> 00:06:42,366
图片理解简单代码修改本地搜索

172
00:06:42,533 --> 00:06:44,500
都可以由较小模型承担

173
00:06:44,600 --> 00:06:46,800
把这些步骤放到本地执行

174
00:06:47,366 --> 00:06:49,400
云端只处理最难的推理

175
00:06:49,700 --> 00:06:51,900
整个系统成本会明显下降

176
00:06:51,933 --> 00:06:53,366
这也解释了为什么

177
00:06:53,366 --> 00:06:55,966
本地模型会越来越重要

178
00:06:55,966 --> 00:06:57,700
AI的下一阶段不是谁

179
00:06:57,700 --> 00:06:59,133
能回答得更漂亮

180
00:06:59,666 --> 00:07:02,600
而是谁能更便宜的完成更多动作

181
00:07:02,800 --> 00:07:05,366
一个公司每天调用百万次模型

182
00:07:05,600 --> 00:07:08,300
如果每一步都走最贵的闭源API

183
00:07:08,466 --> 00:07:09,933
成本会很快失控

184
00:07:09,966 --> 00:07:11,600
本地模型可以承担大量

185
00:07:11,600 --> 00:07:14,466
低风险重复性隐私敏感任务

186
00:07:14,533 --> 00:07:16,300
它不必取代旗舰模型

187
00:07:16,666 --> 00:07:18,166
只要把旗舰模型从低

188
00:07:18,166 --> 00:07:19,933
价值任务里解放出来

189
00:07:19,933 --> 00:07:21,266
就已经很有意义

190
00:07:21,366 --> 00:07:24,300
对普通创作者和小公司来说

191
00:07:24,333 --> 00:07:26,300
这条路线尤其关键

192
00:07:26,333 --> 00:07:28,000
大公司可以买算力

193
00:07:28,133 --> 00:07:29,200
签云合同

194
00:07:29,200 --> 00:07:30,366
雇安全团队

195
00:07:30,466 --> 00:07:32,133
小团队没有这个条件

196
00:07:32,733 --> 00:07:34,533
如果未来所有AI能力都

197
00:07:34,533 --> 00:07:36,900
绑定高价订阅和云端调用

198
00:07:36,933 --> 00:07:39,066
小团队会永远站在门外

199
00:07:39,100 --> 00:07:40,400
开放权重模型

200
00:07:40,400 --> 00:07:42,000
让他们至少有机会在

201
00:07:42,000 --> 00:07:43,866
本地搭一个能跑的系统

202
00:07:44,166 --> 00:07:45,133
哪怕不完美

203
00:07:45,133 --> 00:07:47,266
也能开始改造自己的工作流

204
00:07:47,400 --> 00:07:49,800
当然开源模型也有风险

205
00:07:49,900 --> 00:07:52,100
开放权重意味着能力会扩散

206
00:07:52,366 --> 00:07:53,800
好的用途会扩散

207
00:07:53,966 --> 00:07:55,533
坏的用途也会扩散

208
00:07:55,733 --> 00:07:58,266
安全过滤滥用检测

209
00:07:58,533 --> 00:08:00,300
责任归属都会变复杂

210
00:08:00,466 --> 00:08:03,066
闭源公司常用安全作为封闭理由

211
00:08:03,066 --> 00:08:04,700
不能说完全没有道理

212
00:08:04,900 --> 00:08:07,966
但如果只因为风险就把能力锁死

213
00:08:08,400 --> 00:08:11,066
最后得到的是少数公司

214
00:08:11,066 --> 00:08:12,800
控制所有AI基础设施

215
00:08:13,066 --> 00:08:15,166
Meta现在押的是另一种答案

216
00:08:15,800 --> 00:08:17,133
风险需要治理

217
00:08:17,200 --> 00:08:20,266
但不能用风险作为永久垄断的理由

218
00:08:20,500 --> 00:08:22,566
Muse Glimmer真正要证明的

219
00:08:22,566 --> 00:08:25,566
不是30B能不能击败所有大模型

220
00:08:25,733 --> 00:08:28,666
而是开放路线能不能在AI

221
00:08:28,666 --> 00:08:31,000
商业化最激烈的时候继续存在

222
00:08:31,166 --> 00:08:32,533
闭源路线说

223
00:08:32,800 --> 00:08:34,600
最强能力必须集中

224
00:08:34,900 --> 00:08:37,966
才能安全可靠可收费

225
00:08:38,000 --> 00:08:40,600
开放路线说能力必须分发

226
00:08:40,766 --> 00:08:43,333
社会才不会被少数入口控制

227
00:08:43,500 --> 00:08:45,133
两条路线都会存在

228
00:08:45,200 --> 00:08:47,400
但它们代表完全不同的未来

229
00:08:47,733 --> 00:08:50,333
接下来要看的不是一两次榜单成绩

230
00:08:50,400 --> 00:08:52,800
而是开发者是否真的使用它

231
00:08:52,900 --> 00:08:55,100
有没有人把它部署到本地agent

232
00:08:55,466 --> 00:08:58,100
有没有企业拿它做内部流程

233
00:08:58,266 --> 00:09:01,266
有没有硬件厂商把它放进设备

234
00:09:01,666 --> 00:09:03,200
有没有开源社区围绕

235
00:09:03,200 --> 00:09:04,900
它做微调和工具链

236
00:09:05,166 --> 00:09:06,733
如果这些事情发生

237
00:09:06,900 --> 00:09:08,466
muscoglamer的意义就不

238
00:09:08,466 --> 00:09:10,100
只是一个模型发布

239
00:09:10,166 --> 00:09:12,566
而是Meta在重新争夺AI

240
00:09:12,566 --> 00:09:14,266
基础设施的话语权

241
00:09:14,266 --> 00:09:16,533
这场竞争还有一个更大的背景

242
00:09:17,200 --> 00:09:19,333
AI正在从模型能力竞争

243
00:09:19,400 --> 00:09:21,266
转向分发方式竞争

244
00:09:21,500 --> 00:09:24,166
Openai想把模型做成超级应用

245
00:09:24,366 --> 00:09:26,766
Google想把AI塞回搜索和安卓

246
00:09:27,066 --> 00:09:29,533
苹果想把AI放进硬件和系统

247
00:09:29,766 --> 00:09:31,600
Meta则想让开放模型

248
00:09:31,600 --> 00:09:33,400
成为开发者默认底座

249
00:09:33,766 --> 00:09:36,100
Musglimer只是其中一块拼图

250
00:09:36,133 --> 00:09:38,066
但它指向的方向很清楚

251
00:09:38,200 --> 00:09:40,866
让AI不止存在于云端

252
00:09:41,066 --> 00:09:43,300
也存在于每台电脑每个

253
00:09:43,300 --> 00:09:45,500
应用每个本地工作流里

254
00:09:45,500 --> 00:09:47,100
所以这次发布不能只

255
00:09:47,100 --> 00:09:49,933
看成Meta又开源了一个模型

256
00:09:50,300 --> 00:09:52,000
它更像一次路线宣言

257
00:09:52,266 --> 00:09:54,866
AI如果只能在数据中心运行

258
00:09:55,133 --> 00:09:57,000
用户就永远是租客

259
00:09:57,166 --> 00:09:59,400
AI如果可以在本地运行

260
00:09:59,733 --> 00:10:01,800
用户才有机会成为主人

261
00:10:01,966 --> 00:10:04,100
musglimmer的价值就在这里

262
00:10:04,266 --> 00:10:06,066
它可能不是最强的模型

263
00:10:06,366 --> 00:10:08,166
但他在提醒整个行业

264
00:10:08,366 --> 00:10:11,166
未来的AI不应该只属于机房

265
00:10:11,166 --> 00:10:12,766
也应该属于桌面

266
00:10:12,800 --> 00:10:16,266
这里还要看Costla对AI投资的判断

267
00:10:16,466 --> 00:10:18,466
他认为Openai和Anthropic

268
00:10:18,466 --> 00:10:20,000
不是AI故事的终点

269
00:10:20,000 --> 00:10:21,166
而只是开头

270
00:10:21,300 --> 00:10:22,600
未来会出现更多

271
00:10:22,600 --> 00:10:25,266
千亿美元万亿美元级公司

272
00:10:25,500 --> 00:10:27,800
影响科学医疗能源

273
00:10:27,866 --> 00:10:30,166
教育制造这些基础领域

274
00:10:30,200 --> 00:10:32,533
这个判断放到museglimpse上

275
00:10:32,566 --> 00:10:34,400
就能看见Meta的算盘

276
00:10:34,500 --> 00:10:37,566
如果未来有大量垂直AI公司出现

277
00:10:37,733 --> 00:10:40,666
他们不一定都愿意从闭源API开始

278
00:10:40,933 --> 00:10:43,100
科学研究就是一个典型场景

279
00:10:43,266 --> 00:10:45,666
实验室需要处理论文图像

280
00:10:45,700 --> 00:10:48,366
代码仪器数据和内部记录

281
00:10:48,500 --> 00:10:50,000
很多资料不能随便

282
00:10:50,000 --> 00:10:51,600
上传到外部云服务

283
00:10:51,666 --> 00:10:53,800
一个本地多模态模型

284
00:10:53,866 --> 00:10:55,166
就算不是最强

285
00:10:55,200 --> 00:10:57,166
也可以承担文献整理

286
00:10:57,566 --> 00:11:00,100
实验记录搜索图表理解

287
00:11:00,133 --> 00:11:02,733
工具调用和初步假设生成

288
00:11:02,933 --> 00:11:04,466
它不会替代科学家

289
00:11:04,566 --> 00:11:06,800
但会把很多低价值搜索

290
00:11:06,800 --> 00:11:08,900
和整理动作压低成本

291
00:11:09,133 --> 00:11:10,133
Meta很清楚

292
00:11:10,500 --> 00:11:12,000
闭源模型越强

293
00:11:12,166 --> 00:11:15,266
外界对模型集中化的担心就越重

294
00:11:15,300 --> 00:11:17,366
一个公司如果控制模型

295
00:11:17,800 --> 00:11:18,966
控制分发

296
00:11:19,500 --> 00:11:20,533
控制价格

297
00:11:20,733 --> 00:11:22,133
控制安全规则

298
00:11:22,466 --> 00:11:24,466
开发者就会变成依附者

299
00:11:24,800 --> 00:11:26,666
今天接口价格可以接受

300
00:11:26,733 --> 00:11:28,266
明天规则可能变

301
00:11:28,266 --> 00:11:29,933
今天调用额度够用

302
00:11:29,933 --> 00:11:32,666
明天业务增长后成本可能翻倍

303
00:11:32,766 --> 00:11:35,566
开放模型给开发者留了一条退路

304
00:11:36,133 --> 00:11:38,600
这条退路本身就是谈判筹码

305
00:11:38,800 --> 00:11:40,166
这也是为什么开放

306
00:11:40,166 --> 00:11:42,366
权重会压低行业利润率

307
00:11:42,466 --> 00:11:44,700
只要有足够好的开源替代品

308
00:11:45,300 --> 00:11:46,933
闭源公司就很难对中

309
00:11:46,933 --> 00:11:48,766
低端任务收太高价格

310
00:11:48,966 --> 00:11:51,133
旗舰模型仍然可以卖高价

311
00:11:51,266 --> 00:11:53,066
但大量日常任务会被

312
00:11:53,066 --> 00:11:55,366
本地模型和开放模型吃掉

313
00:11:55,533 --> 00:11:57,666
Meta不靠模型订阅赚钱

314
00:11:58,066 --> 00:12:00,466
反而适合推动这种价格下行

315
00:12:00,566 --> 00:12:02,333
他要让AI变成空气

316
00:12:02,500 --> 00:12:04,366
然后在空气里做平台

317
00:12:04,400 --> 00:12:07,366
有人会问30B模型真的够吗

318
00:12:07,466 --> 00:12:08,933
这个问题要看任务

319
00:12:09,200 --> 00:12:10,900
写一段复杂法律意见

320
00:12:10,900 --> 00:12:11,800
可能不够

321
00:12:11,933 --> 00:12:14,533
做一个本地客服分类器够

322
00:12:14,766 --> 00:12:17,933
通读一本手册并回答设备维修问题

323
00:12:18,066 --> 00:12:21,766
可能够整理图片和文字生成初稿

324
00:12:21,900 --> 00:12:25,066
可能够控制一组本地工具跑流程不

325
00:12:25,366 --> 00:12:26,466
可能也够

326
00:12:26,733 --> 00:12:28,300
AI产品不需要每次都

327
00:12:28,300 --> 00:12:29,966
派最强模型上场

328
00:12:30,166 --> 00:12:31,866
就像公司不会让总裁

329
00:12:31,866 --> 00:12:33,533
去处理每张报销单

330
00:12:33,800 --> 00:12:35,666
真正成熟的AI系统

331
00:12:35,800 --> 00:12:37,733
很可能是多模型协作

332
00:12:37,866 --> 00:12:39,366
小模型负责本地

333
00:12:39,400 --> 00:12:41,966
便宜低风险高频动作

334
00:12:42,000 --> 00:12:44,166
中型模型负责复杂判断

335
00:12:44,866 --> 00:12:47,266
旗舰模型负责少数关键推理

336
00:12:47,400 --> 00:12:49,266
Muse Glimmer的位置就在

337
00:12:49,266 --> 00:12:51,066
第一层和第二层之间

338
00:12:51,200 --> 00:12:52,566
它不抢所有舞台

339
00:12:52,800 --> 00:12:54,666
但可以进入很多工作流

340
00:12:54,700 --> 00:12:56,666
这个位置一旦铺开

341
00:12:56,800 --> 00:12:58,366
使用量会非常大

342
00:12:58,400 --> 00:13:01,666
所以Muse Glimmer看起来是技术事件

343
00:13:01,666 --> 00:13:03,566
背后其实是分发事件

344
00:13:03,766 --> 00:13:06,200
AI行业未来最大的战场

345
00:13:06,200 --> 00:13:08,466
不只是模型谁更聪明

346
00:13:08,533 --> 00:13:10,966
而是谁能接触更多开发者

347
00:13:11,300 --> 00:13:12,166
更多设备

348
00:13:12,400 --> 00:13:13,600
更多工作流

349
00:13:13,733 --> 00:13:15,133
更多真实业务

350
00:13:15,200 --> 00:13:17,800
开放权重提供了一种复制速度

351
00:13:17,933 --> 00:13:19,800
一个模型被下载后

352
00:13:19,800 --> 00:13:22,400
可以在无数环境里变成不同产品

353
00:13:22,533 --> 00:13:23,933
闭源模型再强

354
00:13:24,166 --> 00:13:26,300
也必须等用户来到它的入口

355
00:13:26,466 --> 00:13:29,000
当然开放模型要赢也不容易

356
00:13:29,200 --> 00:13:30,733
开发体验必须好

357
00:13:30,866 --> 00:13:32,133
文档必须清楚

358
00:13:32,300 --> 00:13:33,800
工具链必须稳定

359
00:13:33,966 --> 00:13:35,733
推理速度必须能接受

360
00:13:35,933 --> 00:13:37,866
硬件适配不能太痛苦

361
00:13:38,133 --> 00:13:39,800
如果一个模型开源了

362
00:13:39,800 --> 00:13:41,733
但部署起来像拆炸弹

363
00:13:41,900 --> 00:13:43,733
开发者很快就会放弃

364
00:13:43,966 --> 00:13:47,600
Meta这次把exec torch等部署线索放进来

365
00:13:47,733 --> 00:13:50,366
说明它知道只开源权重还不够

366
00:13:50,400 --> 00:13:52,466
真正要争的是落地链路

367
00:13:52,666 --> 00:13:54,533
安全问题也不能回避

368
00:13:54,666 --> 00:13:56,200
一个能在本地运行

369
00:13:56,533 --> 00:13:57,700
能调用工具

370
00:13:58,066 --> 00:14:00,366
能处理多模态输入的模型

371
00:14:00,566 --> 00:14:02,066
如果被错误使用

372
00:14:02,100 --> 00:14:03,966
也会带来自动化滥用

373
00:14:04,066 --> 00:14:05,700
开放社区需要评估

374
00:14:05,900 --> 00:14:08,566
红队使用限制和透明报告

375
00:14:08,766 --> 00:14:10,733
Meta不能一边强调开放

376
00:14:11,100 --> 00:14:13,133
一边把风险全丢给社区

377
00:14:13,300 --> 00:14:14,933
开放路线要走得久

378
00:14:15,000 --> 00:14:16,866
必须同时建立开放治理

379
00:14:17,000 --> 00:14:19,100
但闭源也不是天然安全

380
00:14:19,333 --> 00:14:21,266
闭源模型的能力集中后

381
00:14:21,266 --> 00:14:23,700
外界更难知道它如何训练

382
00:14:23,700 --> 00:14:24,533
如何过滤

383
00:14:24,600 --> 00:14:25,533
如何拒绝

384
00:14:25,533 --> 00:14:27,200
如何偏向某些服务

385
00:14:27,366 --> 00:14:30,066
安全不是封闭和开放的简单选择

386
00:14:30,766 --> 00:14:32,466
而是透明度责任和

387
00:14:32,466 --> 00:14:34,266
可验证性之间的平衡

388
00:14:34,533 --> 00:14:37,100
Muse glimmer的价值之一就是让

389
00:14:37,100 --> 00:14:39,600
更多研究者能直接测试模型

390
00:14:40,266 --> 00:14:42,466
而不是只能相信厂商声明

391
00:14:42,700 --> 00:14:44,533
Meta设立社区基金

392
00:14:44,866 --> 00:14:46,166
某种意义上也是在

393
00:14:46,166 --> 00:14:48,500
承认AI基建的外部成本

394
00:14:48,666 --> 00:14:50,566
数据中心建在某个地方

395
00:14:50,700 --> 00:14:53,666
当地居民可能面对电价用水

396
00:14:53,666 --> 00:14:56,133
噪音土地和税收分配问题

397
00:14:56,400 --> 00:14:57,566
科技公司过去常把

398
00:14:57,566 --> 00:14:59,400
这些问题当成地方沟通

399
00:14:59,400 --> 00:15:01,466
现在必须当成核心战略

400
00:15:01,600 --> 00:15:03,200
因为没有社区许可

401
00:15:03,366 --> 00:15:05,933
再强的模型也需要机房供电

402
00:15:06,000 --> 00:15:07,700
把这两条线合起来

403
00:15:07,733 --> 00:15:09,900
Meta的新叙事就完整了

404
00:15:10,133 --> 00:15:13,333
一边用开放模型说AI能力要分散

405
00:15:13,566 --> 00:15:15,533
一边用社区基金说AI

406
00:15:15,533 --> 00:15:17,100
基建收益要回流

407
00:15:17,133 --> 00:15:18,933
这不一定全是理想主义

408
00:15:19,266 --> 00:15:21,966
但他比单纯喊超级智能更聪明

409
00:15:22,100 --> 00:15:24,166
他知道AI行业如果继续只

410
00:15:24,166 --> 00:15:26,533
展示估值参数和算力

411
00:15:26,600 --> 00:15:28,500
公众反弹会越来越强

412
00:15:28,666 --> 00:15:31,766
未来真正有杀伤力的缪斯Glimmer应用

413
00:15:32,266 --> 00:15:34,500
可能不会出现在发布新闻里

414
00:15:34,533 --> 00:15:37,266
而会出现在很小的工作场景里

415
00:15:37,266 --> 00:15:39,866
一个设计师用他本地整理素材

416
00:15:40,333 --> 00:15:43,166
一个律师事务所用它检索内部文件

417
00:15:43,300 --> 00:15:45,700
一个工厂用它读设备图片

418
00:15:45,900 --> 00:15:47,533
一个个人开发者用它

419
00:15:47,533 --> 00:15:48,933
搭自己的桌面agent

420
00:15:49,300 --> 00:15:50,700
每个场景都不震撼

421
00:15:50,700 --> 00:15:52,300
但加起来就是生态

422
00:15:52,366 --> 00:15:54,666
这也是开放模型最强的地方

423
00:15:55,100 --> 00:15:57,000
它的价值不是一次性释放

424
00:15:57,000 --> 00:15:59,266
而是被别人不断重新发明

425
00:15:59,466 --> 00:16:01,600
闭源模型像一座中央电站

426
00:16:01,766 --> 00:16:04,466
开放模型像一批可移动工具箱

427
00:16:04,733 --> 00:16:06,400
中央电站功率更大

428
00:16:06,400 --> 00:16:08,200
工具箱更容易进入角落

429
00:16:08,566 --> 00:16:10,400
AI要真正改变社会

430
00:16:10,500 --> 00:16:12,300
不能只靠中央电站

431
00:16:12,366 --> 00:16:13,700
也需要工具箱

432
00:16:13,866 --> 00:16:16,566
所以musglmmer不是Meta对闭源

433
00:16:16,566 --> 00:16:18,333
模型的一次正面冲锋

434
00:16:18,400 --> 00:16:20,066
而是一次侧翼包抄

435
00:16:20,366 --> 00:16:22,966
它不说自己在所有能力上第一

436
00:16:23,366 --> 00:16:25,133
他说自己更容易被拿走

437
00:16:25,200 --> 00:16:26,533
更容易本地跑

438
00:16:26,600 --> 00:16:28,333
更容易进入普通设备

439
00:16:28,400 --> 00:16:30,000
这个策略如果成功

440
00:16:30,600 --> 00:16:32,766
Meta会在AI入口战争里拿到

441
00:16:32,766 --> 00:16:34,566
一个很难被关闭的位置

442
00:16:34,766 --> 00:16:36,500
execute torch的存在

443
00:16:36,900 --> 00:16:40,066
说明Meta对端侧部署有长期准备

444
00:16:40,300 --> 00:16:45,133
端侧AI不是把云端模型硬塞进手机

445
00:16:45,333 --> 00:16:47,400
而是重新考虑模型大小

446
00:16:47,400 --> 00:16:51,900
延迟内存占用电池和隐私MUS glimmer

447
00:16:52,566 --> 00:16:55,300
如果能和端侧框架形成完整链路

448
00:16:55,366 --> 00:16:58,133
就会成为Meta连接硬件厂商的桥

449
00:16:58,333 --> 00:17:02,166
以后手机眼镜耳机笔记本桌面设备

450
00:17:02,200 --> 00:17:04,366
都可能需要一个能本地

451
00:17:04,366 --> 00:17:06,666
处理多模态任务的模型底座

452
00:17:06,766 --> 00:17:08,666
这对开发者很有吸引力

453
00:17:08,933 --> 00:17:11,900
闭源API最舒服的地方是省事

454
00:17:12,300 --> 00:17:14,166
最痛的地方是不可控

455
00:17:14,300 --> 00:17:17,533
接口变更价格变更限流地区

456
00:17:17,533 --> 00:17:20,500
限制内容政策都可能影响产品

457
00:17:20,733 --> 00:17:22,933
开放模型麻烦一点但可控

458
00:17:22,966 --> 00:17:25,133
对长期做产品的人来说

459
00:17:25,266 --> 00:17:28,066
可控往往比短期省事更重要

460
00:17:28,200 --> 00:17:30,400
musglimer如果性能足够稳定

461
00:17:30,500 --> 00:17:32,200
就会被放进很多不愿

462
00:17:32,200 --> 00:17:34,333
完全依赖云端的产品里

463
00:17:34,533 --> 00:17:35,966
企业采用AI时

464
00:17:36,133 --> 00:17:37,933
常常不是被能力卡住

465
00:17:38,000 --> 00:17:39,600
而是被审批卡住

466
00:17:39,766 --> 00:17:41,366
法务问数据去哪

467
00:17:41,500 --> 00:17:43,566
安全团队问权限怎么管

468
00:17:43,933 --> 00:17:45,866
财务问每月调用成本

469
00:17:46,066 --> 00:17:48,533
业务部门问延迟能不能接受

470
00:17:48,700 --> 00:17:50,000
开放本地模型可以

471
00:17:50,000 --> 00:17:51,533
同时回答几个问题

472
00:17:51,700 --> 00:17:53,300
数据可以留在内网

473
00:17:53,366 --> 00:17:54,600
成本可以固定

474
00:17:54,766 --> 00:17:56,066
延迟可以降低

475
00:17:56,133 --> 00:17:57,700
权限可以自己定义

476
00:17:57,900 --> 00:17:59,133
它不是完美答案

477
00:17:59,133 --> 00:18:00,733
但比纯云端更容易

478
00:18:00,733 --> 00:18:02,400
通过某些企业流程

479
00:18:02,800 --> 00:18:04,100
Meta还可以借开放

480
00:18:04,100 --> 00:18:06,066
路线给监管传递信号

481
00:18:06,300 --> 00:18:07,966
闭源超级模型越强

482
00:18:07,966 --> 00:18:10,966
监管越担心权力集中和黑箱决策

483
00:18:11,266 --> 00:18:13,366
开放模型让更多

484
00:18:13,366 --> 00:18:15,766
研究者能审计能力和风险

485
00:18:16,466 --> 00:18:17,933
至少表面上更符合

486
00:18:17,933 --> 00:18:19,700
广泛分发的公共叙事

487
00:18:19,866 --> 00:18:22,100
对一家掌握社交平台和

488
00:18:22,100 --> 00:18:23,900
广告系统的公司来说

489
00:18:23,966 --> 00:18:25,566
这种姿态很重要

490
00:18:25,766 --> 00:18:27,333
它需要证明自己不是

491
00:18:27,333 --> 00:18:29,500
把AI变成下一层垄断

492
00:18:29,800 --> 00:18:32,000
但市场不会因为姿态就买单

493
00:18:32,166 --> 00:18:34,100
开发者最后只看三件事

494
00:18:34,300 --> 00:18:35,300
够不够好

495
00:18:35,300 --> 00:18:36,366
跑不跑得动

496
00:18:36,366 --> 00:18:37,866
出问题能不能修

497
00:18:38,100 --> 00:18:41,300
musglimer需要在真实任务里证明自己

498
00:18:41,766 --> 00:18:44,500
比如本地图片理解文档问答

499
00:18:44,566 --> 00:18:47,733
桌面自动化工具调用失败恢复

500
00:18:47,866 --> 00:18:49,400
只要这些能力能稳定

501
00:18:49,400 --> 00:18:51,500
解决一批中等难度问题

502
00:18:51,566 --> 00:18:53,733
30B就不再是小模型

503
00:18:53,766 --> 00:18:56,333
而是刚好够用的生产力模型

504
00:18:56,566 --> 00:18:59,133
AI行业正在出现一种新分工

505
00:18:59,266 --> 00:19:01,400
最强闭源模型负责天花板

506
00:19:01,733 --> 00:19:04,166
开放模型负责地板和中间层

507
00:19:04,333 --> 00:19:06,533
天花板决定行业想象力

508
00:19:06,700 --> 00:19:08,500
地板决定普及速度

509
00:19:08,500 --> 00:19:09,566
没有天花板

510
00:19:10,133 --> 00:19:11,566
AI缺少突破

511
00:19:11,766 --> 00:19:12,666
没有地板

512
00:19:12,900 --> 00:19:15,733
AI只会变成少数公司的高价服务

513
00:19:15,866 --> 00:19:18,766
musglimer的意义就在地板和中间层

514
00:19:19,266 --> 00:19:21,166
他让更多人能踩上去

515
00:19:21,333 --> 00:19:25,166
扎克伯格讲发明创造而不是自动化

516
00:19:25,333 --> 00:19:28,200
其实是在修复AI的公共形象

517
00:19:28,466 --> 00:19:29,566
过去两年

518
00:19:30,100 --> 00:19:33,566
很多人听到AI就想到裁员替代

519
00:19:34,300 --> 00:19:37,700
数据中心耗电和大公司集中全力

520
00:19:37,900 --> 00:19:39,666
Meta要把话题拉回

521
00:19:39,666 --> 00:19:41,466
创造力和个人赋能

522
00:19:41,600 --> 00:19:43,466
Muse Glimmer适合这个叙事

523
00:19:43,900 --> 00:19:45,366
因为它看起来不像一个

524
00:19:45,366 --> 00:19:47,566
只服务企业裁员的工具

525
00:19:47,666 --> 00:19:48,766
而像一个可以被

526
00:19:48,766 --> 00:19:50,700
个人拿走的创造工具

527
00:19:51,133 --> 00:19:53,766
当然这种叙事也要接受现实检验

528
00:19:53,933 --> 00:19:55,700
如果开放模型最后主要

529
00:19:55,700 --> 00:19:57,733
被大公司拿去降低成本

530
00:19:57,900 --> 00:19:59,600
普通人没有明显受益

531
00:19:59,700 --> 00:20:02,166
那赋能就会变成漂亮话

532
00:20:02,200 --> 00:20:03,933
真正能证明Meta的

533
00:20:03,933 --> 00:20:05,466
是它有没有让个人

534
00:20:05,466 --> 00:20:07,966
开发者小企业学校

535
00:20:08,366 --> 00:20:10,133
研究者用更低成本

536
00:20:10,133 --> 00:20:12,133
做出过去做不了的东西

537
00:20:12,266 --> 00:20:14,866
开源的价值不能停在下载量

538
00:20:15,366 --> 00:20:17,966
要落到新作品和新产品上

539
00:20:18,066 --> 00:20:19,000
未来一年

540
00:20:19,300 --> 00:20:20,866
musicglimmer最值得观察

541
00:20:20,866 --> 00:20:22,166
的不是发布热度

542
00:20:22,366 --> 00:20:24,100
而是二次开发热度

543
00:20:24,266 --> 00:20:26,100
有没有高质量微调版本

544
00:20:26,566 --> 00:20:29,700
有没有中文代码医疗法律

545
00:20:29,733 --> 00:20:32,166
教育设计方向的社区版本

546
00:20:32,266 --> 00:20:34,866
有没有人把它做成一键本地agent

547
00:20:35,200 --> 00:20:37,100
有没有硬件厂商围绕它

548
00:20:37,100 --> 00:20:39,000
优化驱动和推理速度

549
00:20:39,300 --> 00:20:42,200
这些才是开放模型真正的生命迹象

550
00:20:42,400 --> 00:20:44,200
如果这些生态长出来

551
00:20:44,500 --> 00:20:47,100
Meta就会完成一次很聪明的布局

552
00:20:47,300 --> 00:20:49,933
它不需要每个用户直接打开Meta产品

553
00:20:50,266 --> 00:20:52,866
也不需要每次推理都走Meta服务器

554
00:20:52,933 --> 00:20:55,066
只要大量AI工具底层

555
00:20:55,066 --> 00:20:57,300
运行的是Meta开放模型

556
00:20:57,500 --> 00:20:59,933
它就会成为看不见的基础设施

557
00:21:00,300 --> 00:21:02,333
基础设施的力量往往比

558
00:21:02,333 --> 00:21:04,200
一个热门应用更持久

559
00:21:04,466 --> 00:21:05,733
Meta的优势是

560
00:21:05,766 --> 00:21:08,700
它不急着把模型本身变成利润中心

561
00:21:09,000 --> 00:21:11,100
它可以允许模型价格下降

562
00:21:11,266 --> 00:21:13,266
因为它真正赚钱的地方在

563
00:21:13,266 --> 00:21:16,066
广告社交关系和平台分发

564
00:21:16,200 --> 00:21:19,933
Openai必须证明订阅和API能覆盖成本

565
00:21:20,266 --> 00:21:22,500
Anthropic必须证明企业安全

566
00:21:22,500 --> 00:21:24,566
和高端能力能卖出价格

567
00:21:24,866 --> 00:21:26,733
Meta则可以用开放模型

568
00:21:26,733 --> 00:21:28,533
削弱对手收费能力

569
00:21:28,666 --> 00:21:29,900
这种打法很狠

570
00:21:29,966 --> 00:21:32,866
如果未来个人设备都能跑中型模型

571
00:21:33,266 --> 00:21:35,266
AI产品形态也会变

572
00:21:35,333 --> 00:21:36,966
很多应用不再需要把

573
00:21:36,966 --> 00:21:38,800
所有请求传到服务器

574
00:21:38,900 --> 00:21:41,500
而是把模型嵌进本地功能里

575
00:21:41,533 --> 00:21:43,866
写作软件可以本地整理草稿

576
00:21:44,066 --> 00:21:46,366
剪辑软件可以本地识别素材

577
00:21:46,666 --> 00:21:48,866
浏览器可以本地总结网页

578
00:21:48,933 --> 00:21:51,266
开发工具可以本地读项目

579
00:21:51,466 --> 00:21:52,733
云端仍然存在

580
00:21:52,966 --> 00:21:55,333
但本地会承担更多基础动作

581
00:21:55,366 --> 00:21:58,366
所以这件事最后要落到一个判断

582
00:21:58,800 --> 00:22:01,333
AI的未来不会只有一种入口

583
00:22:01,466 --> 00:22:03,466
云端旗舰模型会存在

584
00:22:03,566 --> 00:22:05,700
本地开放模型也会存在

585
00:22:05,866 --> 00:22:07,566
前者负责极限能力

586
00:22:07,700 --> 00:22:09,933
后者负责普及和控制权

587
00:22:10,100 --> 00:22:11,733
Meta现在压的是后者

588
00:22:12,100 --> 00:22:13,400
而且押得很聪明

589
00:22:13,566 --> 00:22:15,200
它不一定赢走所有收入

590
00:22:15,333 --> 00:22:17,666
但可能赢走很多默认选择

591
00:22:17,766 --> 00:22:19,733
如果把它放进更长周期

592
00:22:20,066 --> 00:22:21,000
Muse Glimmer

593
00:22:21,366 --> 00:22:24,000
代表的是AI商品化之后的下一步

594
00:22:24,200 --> 00:22:26,766
模型能力会越来越像基础能力

595
00:22:26,900 --> 00:22:29,200
真正差异会转移到数据

596
00:22:29,333 --> 00:22:32,133
场景部署体验和生态

597
00:22:32,366 --> 00:22:34,933
Meta提前把模型底座打开

598
00:22:35,000 --> 00:22:38,100
是在承认模型本身迟早会变便宜

599
00:22:38,400 --> 00:22:39,766
既然便宜不可避免

600
00:22:39,766 --> 00:22:41,500
不如主动让它变便宜

601
00:22:41,600 --> 00:22:43,866
再去控制更大的生态入口

602
00:22:43,866 --> 00:22:46,533
这对闭源公司是一种压力测试

603
00:22:46,766 --> 00:22:49,200
闭源模型如果只是略强一点

604
00:22:49,266 --> 00:22:50,466
却贵很多

605
00:22:50,700 --> 00:22:52,666
很多企业会改用开放模型

606
00:22:53,000 --> 00:22:54,600
闭源模型必须在关键

607
00:22:54,600 --> 00:22:56,400
任务上强到无法替代

608
00:22:56,666 --> 00:22:59,966
或者在产品体验上省心得足够明显

609
00:23:00,200 --> 00:23:02,766
否则开源模型每进步一步

610
00:23:02,900 --> 00:23:05,500
闭源模型的中低端收入就少一块

611
00:23:05,733 --> 00:23:08,133
Muse Glimmer也会推动AI agent

612
00:23:08,566 --> 00:23:10,466
从玩具走向基础设施

613
00:23:10,700 --> 00:23:13,733
以前很多agent演示看起来很聪明

614
00:23:13,933 --> 00:23:16,600
但实际运行成本高延迟长

615
00:23:16,666 --> 00:23:17,900
隐私风险大

616
00:23:17,966 --> 00:23:20,400
把一部分能力放到本地之后

617
00:23:20,566 --> 00:23:23,133
agent可以更安静的处理日常任务

618
00:23:23,300 --> 00:23:24,900
它不需要每次都上云

619
00:23:25,100 --> 00:23:26,866
不需要每一步都花钱

620
00:23:27,166 --> 00:23:29,766
也不需要把所有私人资料交出去

621
00:23:29,800 --> 00:23:31,366
这就是它的流量点

622
00:23:31,800 --> 00:23:33,733
Meta不是在送一个模型

623
00:23:33,966 --> 00:23:35,400
而是在告诉市场

624
00:23:35,900 --> 00:23:38,666
AI的权力不一定只能往云端集中

625
00:23:38,800 --> 00:23:40,733
普通设备也可以有智能

626
00:23:40,866 --> 00:23:42,566
小团队也可以有底座

627
00:23:42,900 --> 00:23:44,566
开发者也可以有退路

628
00:23:44,600 --> 00:23:46,133
这个故事比单纯

629
00:23:46,133 --> 00:23:48,100
参数发布更有传播性

630
00:23:48,166 --> 00:23:49,200
因为它集中了

631
00:23:49,200 --> 00:23:51,400
很多人对AI垄断的焦虑

632
00:23:51,500 --> 00:23:54,533
所以缪斯glimmer值得被单独拿出来讲

633
00:23:54,733 --> 00:23:55,400
不是因为它

634
00:23:55,400 --> 00:23:57,900
一夜之间改写大模型格局

635
00:23:58,000 --> 00:23:59,300
而是因为它把一个被

636
00:23:59,300 --> 00:24:01,133
忽略的问题摆到台面

637
00:24:01,366 --> 00:24:03,766
AI普及到底靠更大的模型

638
00:24:03,933 --> 00:24:06,400
还是靠更容易被部署的模型

639
00:24:06,733 --> 00:24:08,666
亿元巨头会继续冲天花板

640
00:24:08,666 --> 00:24:10,966
但真正改变日常工作的往往

641
00:24:10,966 --> 00:24:13,333
是那些能被普通人装进电脑

642
00:24:13,400 --> 00:24:14,366
接近流程

643
00:24:14,400 --> 00:24:15,800
反复修改的东西