1 00:00:00,000 --> 00:00:28,000 Meta開源Muse Glimmer 30B,最值得看的不是參數也不是跑分,而是它把大模型競爭從雲端大廠的機房重新推回普通人的電腦。一個30B級別開放權重面向本地Agent的模型放在今天這個節點,意思很明確。AI不能只是一張雲服務帳單,也不能永遠被少數公司關在黑盒裡。這件事要放在Meta的整體敘事裡看。 2 00:00:28,000 --> 00:00:30,600 扎克伯格最近把話說得很直 3 00:00:30,600 --> 00:00:34,320 人工智能的主要目的應該是發明創造 4 00:00:34,320 --> 00:00:36,640 不應該只是自動化替代 5 00:00:36,640 --> 00:00:39,300 超級智能如果真的會到來 6 00:00:39,300 --> 00:00:42,620 那它不應該只掌握在幾家公司手裡 7 00:00:42,620 --> 00:00:46,000 而應該盡量廣泛的分發給普通人 8 00:00:46,000 --> 00:00:47,880 開發者和小團隊 9 00:00:47,880 --> 00:00:52,380 Muse Glimmer就是這個敘事落到產品上的一顆旗子 10 00:00:52,380 --> 00:00:54,640 Muse Glimmer 30B的重點 11 00:00:54,640 --> 00:00:57,000 不是挑戰必源模型的全部能力 12 00:00:57,000 --> 00:01:00,340 而是给本地AI流出一条现实路线 13 00:01:00,340 --> 00:01:04,240 现在很多模型动不动就是几千亿上万亿参数 14 00:01:04,240 --> 00:01:05,700 听起来很震撼 15 00:01:05,700 --> 00:01:08,080 但普通用户根本跑不起来 16 00:01:08,080 --> 00:01:09,760 他们需要昂贵GPU 17 00:01:09,760 --> 00:01:11,360 需要云端推理 18 00:01:11,360 --> 00:01:13,380 需要订阅和调用额度 19 00:01:13,380 --> 00:01:15,720 Meta选择30B这个体量 20 00:01:15,720 --> 00:01:16,820 就是在说 21 00:01:16,820 --> 00:01:18,460 模型不一定越大 22 00:01:18,460 --> 00:01:19,880 越接近未来 23 00:01:19,880 --> 00:01:20,960 能部署 24 00:01:20,960 --> 00:01:21,860 能修改 25 00:01:21,860 --> 00:01:23,500 能欠尽工作流 26 00:01:23,500 --> 00:01:25,500 才更接近真实使用 27 00:01:25,500 --> 00:01:28,280 这和Lama时代的路线一脉相承 28 00:01:28,280 --> 00:01:31,320 Meta一直在开放模型上压住 29 00:01:31,320 --> 00:01:33,400 只是中间经历过摇摆 30 00:01:33,400 --> 00:01:36,640 当OpenAI, Anthropic, Google 31 00:01:36,640 --> 00:01:40,300 把高端模型能力不断往B原体系里收 32 00:01:40,300 --> 00:01:44,300 Meta反而重新把开放权重当成差异化 33 00:01:44,300 --> 00:01:46,660 它没必要在每一项榜单上赢 34 00:01:46,660 --> 00:01:48,980 它要赢的是开发者心智 35 00:01:48,980 --> 00:01:53,520 谁能让更多人拿去改,拿去跑,拿去做产品 36 00:01:53,520 --> 00:01:56,880 谁就能在AI底层生态里留下位置 37 00:01:56,880 --> 00:02:02,220 Muse Glimmer还带有Execute Torch和PTE这样的边缘部署意味 38 00:02:02,220 --> 00:02:03,840 这个细节很重要 39 00:02:03,840 --> 00:02:06,780 它不是只给服务器看的模型卡 40 00:02:06,780 --> 00:02:11,840 而是在暗示手机、笔记本、边缘设备、本地工作站 41 00:02:11,840 --> 00:02:14,520 都可能成为AI Agent的运行环境 42 00:02:14,520 --> 00:02:17,700 未来AI Agent不一定每一步都要回云端 43 00:02:17,700 --> 00:02:22,820 它可以在本地看文件、读图片、调工具、处理私人数据 44 00:02:22,820 --> 00:02:26,020 再把真正复杂的任务交给云端大模型 45 00:02:26,020 --> 00:02:29,660 这种路线最大的价值是隐私和成本 46 00:02:29,660 --> 00:02:36,460 用户把本地文件、工作资料、聊天记录、图片和代码交给云端模型 47 00:02:36,460 --> 00:02:38,340 永远会有心理门槛 48 00:02:38,340 --> 00:02:39,740 企业更敏感 49 00:02:39,740 --> 00:02:44,120 尤其涉及客户数据、合同、研发资料和内部流程 50 00:02:44,120 --> 00:02:46,260 一个能在本地跑的模型 51 00:02:46,260 --> 00:02:48,100 就算能力不是最强 52 00:02:48,300 --> 00:02:51,280 也会因为数据不出设备而变得有价值 53 00:02:51,280 --> 00:02:54,600 很多真实场景不需要世界最聪明的模型 54 00:02:54,600 --> 00:02:56,780 只需要一个足够聪明 55 00:02:56,780 --> 00:02:57,840 足够便宜 56 00:02:57,840 --> 00:02:59,540 足够可控的助手 57 00:02:59,540 --> 00:03:03,000 这也是开源模型真正打B元模型的地方 58 00:03:03,000 --> 00:03:06,340 B元模型擅长提供最强通用能力 59 00:03:06,340 --> 00:03:09,500 开源模型擅长被改造成具体工具 60 00:03:09,500 --> 00:03:11,220 小团队可以微调它 61 00:03:11,220 --> 00:03:13,540 企业可以把它部署到内网 62 00:03:13,540 --> 00:03:16,700 开发者可以把它接近自己的agent框架 63 00:03:16,700 --> 00:03:19,820 硬件厂商可以把它塞进边缘设备 64 00:03:19,820 --> 00:03:21,740 B原模型卖的是能力 65 00:03:21,740 --> 00:03:23,800 开放模型卖的是可塑性 66 00:03:23,800 --> 00:03:26,040 两者不是同一场比赛 67 00:03:26,040 --> 00:03:28,540 Meta这次还把数据中心社区基金 68 00:03:28,540 --> 00:03:30,000 放在同一个叙事里 69 00:03:30,000 --> 00:03:31,920 说明他知道AI扩张 70 00:03:31,920 --> 00:03:33,860 正在遇到新的政治成本 71 00:03:33,860 --> 00:03:36,180 数据中心需要电力 72 00:03:36,180 --> 00:03:39,240 土地 水 输电设施和社区许可 73 00:03:39,240 --> 00:03:41,380 过去科技公司建机房 74 00:03:41,380 --> 00:03:43,020 地方政府往往欢迎 75 00:03:43,020 --> 00:03:44,860 因为它带来投资和税收 76 00:03:44,860 --> 00:03:46,500 现在情况变了 77 00:03:46,500 --> 00:03:48,520 居民开始担心电费上涨 78 00:03:48,520 --> 00:03:51,280 环境压力和基础设施被占用 79 00:03:51,280 --> 00:03:55,560 AI越庞大越不可能只用技术语言解释自己 80 00:03:55,560 --> 00:03:58,380 10亿美元社区基金表面上是补偿 81 00:03:58,380 --> 00:04:01,160 背后是AI基建时代的通行费 82 00:04:01,160 --> 00:04:03,500 Meta想告诉当地居民 83 00:04:03,500 --> 00:04:06,280 数据中心不是只把电和水抽走 84 00:04:06,280 --> 00:04:09,340 也会把钱、岗位和公共服务带回来 85 00:04:09,340 --> 00:04:12,920 警察、消防员、社区成员被点名 86 00:04:12,920 --> 00:04:15,200 是一种很现实的政治沟通 87 00:04:15,200 --> 00:04:17,180 AI公司已经发现 88 00:04:17,180 --> 00:04:19,700 光说技术改变世界没用了 89 00:04:19,700 --> 00:04:22,640 真正被影响的人要看到具体回报 90 00:04:22,640 --> 00:04:26,340 Muse Glimmer和社区基金看似是两件事 91 00:04:26,340 --> 00:04:27,980 其实对应同一个问题 92 00:04:27,980 --> 00:04:30,600 AI权力会不会过度集中 93 00:04:30,600 --> 00:04:32,520 模型集中在云端 94 00:04:32,520 --> 00:04:34,820 算力集中在数据中心 95 00:04:34,820 --> 00:04:36,940 收益集中在少数公司 96 00:04:36,940 --> 00:04:39,900 成本却扩散给社区和消费者 97 00:04:39,900 --> 00:04:41,640 这套结构如果不调整 98 00:04:41,640 --> 00:04:43,020 反弹会越来越大 99 00:04:43,020 --> 00:04:44,640 Meta现在讲开放 100 00:04:44,640 --> 00:04:45,440 讲本地 101 00:04:45,440 --> 00:04:46,840 讲回馈社区 102 00:04:46,840 --> 00:04:50,360 就是在给自己的AI扩张换一种合法性 103 00:04:50,360 --> 00:04:53,260 这不代表Meta突然变成公益组织 104 00:04:53,260 --> 00:04:55,320 开源也不是纯粹善意 105 00:04:55,320 --> 00:04:57,780 Meta的商业逻辑很清楚 106 00:04:57,780 --> 00:05:02,240 它没有像OpenAI那样把订阅模型做成第一入口 107 00:05:02,240 --> 00:05:05,280 也没有Google那样的搜索默认位置 108 00:05:05,280 --> 00:05:09,920 它最擅长的是平台、分发、广告和开发者网络 109 00:05:09,920 --> 00:05:13,640 开放模型能削弱必源模型的收费能力 110 00:05:13,640 --> 00:05:15,940 降低竞争对手的护城河 111 00:05:15,940 --> 00:05:19,980 同时让更多应用围绕Meta的模型生态生长 112 00:05:19,980 --> 00:05:22,360 如果一个开放模型足够好 113 00:05:22,360 --> 00:05:27,460 大量开发者就会围绕它做工具、插件、微调和部署方案 114 00:05:27,460 --> 00:05:30,200 哪怕Meta不直接收每一次调用费 115 00:05:30,200 --> 00:05:33,520 它也能影响标准、框架和生态方向 116 00:05:33,520 --> 00:05:36,480 安卓当年不是靠系统授权费赚钱 117 00:05:36,480 --> 00:05:39,480 而是靠开放系统赞助移动入口 118 00:05:39,480 --> 00:05:42,560 Lama和Muse Glimmer也有类似逻辑 119 00:05:42,560 --> 00:05:45,520 用开放打破别人的封闭利润词 120 00:05:45,520 --> 00:05:48,140 在在更大的生态里寻找收益 121 00:05:48,140 --> 00:05:50,920 Muse Glimmer的30B体量 122 00:05:50,920 --> 00:05:52,400 还有一个现实意义 123 00:05:52,400 --> 00:05:54,860 它让AI从演示回到产品 124 00:05:54,860 --> 00:05:58,100 很多AI发布会喜欢展示极限能力 125 00:05:58,100 --> 00:06:00,960 仿佛模型能解决所有复杂问题 126 00:06:00,960 --> 00:06:02,680 但真正落地时 127 00:06:02,680 --> 00:06:08,340 企业关心的是成本、延迟、部署、安全、稳定、可维护 128 00:06:08,340 --> 00:06:09,860 一个30B模型 129 00:06:09,860 --> 00:06:14,360 如果能在普通设备或消费级GPU上完成稳定任务 130 00:06:14,360 --> 00:06:16,880 它的商业价值可能比一个 131 00:06:16,880 --> 00:06:20,640 只能在云端高价运行的巨型模型更直接 132 00:06:20,640 --> 00:06:22,240 尤其是agent场景 133 00:06:22,240 --> 00:06:24,060 agent不是一次聊天 134 00:06:24,060 --> 00:06:25,600 而是一连串动作 135 00:06:25,600 --> 00:06:26,860 理解目标 136 00:06:26,860 --> 00:06:27,960 拆任务 137 00:06:27,960 --> 00:06:28,940 读文件 138 00:06:28,940 --> 00:06:30,020 调用工具 139 00:06:30,020 --> 00:06:31,200 检查结果 140 00:06:31,200 --> 00:06:32,560 失败后重视 141 00:06:32,560 --> 00:06:35,320 这里面很多步骤并不需要最强模型 142 00:06:35,320 --> 00:06:36,360 文件分类 143 00:06:36,360 --> 00:06:37,420 表格检查 144 00:06:37,420 --> 00:06:38,440 界面操作 145 00:06:38,440 --> 00:06:39,520 图片理解 146 00:06:39,520 --> 00:06:41,100 简单代码修改 147 00:06:41,100 --> 00:06:42,160 本地搜索 148 00:06:42,160 --> 00:06:44,280 都可以由较小模型承担 149 00:06:44,280 --> 00:06:46,640 把这些步骤放到本地执行 150 00:06:46,640 --> 00:06:49,180 云端只处理最难的推理 151 00:06:49,180 --> 00:06:51,660 整个系统成本会明显下降 152 00:06:51,660 --> 00:06:55,100 这也解释了为什么本地模型会越来越重 153 00:06:55,100 --> 00:06:58,940 要AI的下一阶段不是谁能回答的更漂亮 154 00:06:58,940 --> 00:07:02,400 而是谁能更便宜的完成更多动作 155 00:07:02,400 --> 00:07:05,200 一个公司每天调用百万次模型 156 00:07:05,200 --> 00:07:08,120 如果每一步都走最贵的币源API 157 00:07:08,120 --> 00:07:09,660 成本会很快失控 158 00:07:09,660 --> 00:07:12,260 本地模型可以承担大量低风险 159 00:07:12,260 --> 00:07:13,060 重复性 160 00:07:13,060 --> 00:07:14,500 隐私敏感任务 161 00:07:14,500 --> 00:07:16,160 它不必取代旗舰模型 162 00:07:16,160 --> 00:07:17,680 只要把旗舰模型 163 00:07:17,680 --> 00:07:19,660 从低价值任务理解放出来 164 00:07:19,660 --> 00:07:21,420 就已经很有意义 165 00:07:21,420 --> 00:07:24,100 对普通创作者和小公司来说 166 00:07:24,100 --> 00:07:26,080 这条路线尤其关键 167 00:07:26,080 --> 00:07:27,760 大公司可以买算力 168 00:07:27,760 --> 00:07:29,020 签云合同 169 00:07:29,020 --> 00:07:30,180 雇安全团队 170 00:07:30,180 --> 00:07:31,940 小团队没有这个条件 171 00:07:31,940 --> 00:07:34,260 如果未来所有AI能力 172 00:07:34,260 --> 00:07:36,620 都绑定高价订阅和云端调用 173 00:07:36,620 --> 00:07:38,820 小团队会永远站在门外 174 00:07:38,820 --> 00:07:40,240 开放权重模型 175 00:07:40,240 --> 00:07:41,760 让他们至少有机会 176 00:07:41,760 --> 00:07:43,760 在本地搭一個能跑的系統 177 00:07:43,760 --> 00:07:45,020 哪怕不完美 178 00:07:45,020 --> 00:07:47,040 也能開始改造自己的工作流 179 00:07:47,040 --> 00:07:47,700 當然 180 00:07:47,700 --> 00:07:49,600 開源模型也有風險 181 00:07:49,600 --> 00:07:51,940 開放權重意味著能力會擴散 182 00:07:51,940 --> 00:07:53,660 好的用途會擴散 183 00:07:53,660 --> 00:07:55,380 壞的用途也會擴散 184 00:07:55,380 --> 00:07:56,520 安全過濾 185 00:07:56,520 --> 00:07:58,100 濫用檢測 186 00:07:58,100 --> 00:08:00,120 責任歸屬都會變複雜 187 00:08:00,120 --> 00:08:02,880 B原公司常用安全作為封閉理由 188 00:08:02,880 --> 00:08:04,520 不能說完全沒有道理 189 00:08:04,520 --> 00:08:06,300 但如果只因為風險 190 00:08:06,300 --> 00:08:07,640 就把能力鎖死 191 00:08:08,121 --> 00:08:15,121 最後得到的是少數公司控制所有AI基礎設施。Meta現在壓的是另一種答案。 192 00:08:15,121 --> 00:08:20,121 風險需要治理,但不能用風險作為永久壟斷的理由。 193 00:08:20,121 --> 00:08:22,401 Muse Glimmer真正要证明的 194 00:08:22,401 --> 00:08:25,381 不是30B能不能击败所有大模型 195 00:08:25,381 --> 00:08:30,841 而是开放路线能不能在AI商业化最激烈的时候继续存在 196 00:08:30,841 --> 00:08:32,241 闭源路线说 197 00:08:32,241 --> 00:08:34,461 最强能力必须集中 198 00:08:34,461 --> 00:08:37,681 才能安全、可靠、可收费 199 00:08:37,681 --> 00:08:40,301 开放路线说能力必须分发 200 00:08:40,301 --> 00:08:43,161 社会才不会被少数入口控制 201 00:08:43,161 --> 00:08:44,861 两条路线都会存在 202 00:08:44,861 --> 00:08:47,181 但它们代表完全不同的未来 203 00:08:47,181 --> 00:08:50,161 接下来要看的不是一两次榜单成绩 204 00:08:50,161 --> 00:08:52,601 而是开发者是否真的使用它 205 00:08:52,601 --> 00:08:55,041 有没有人把它部署到本地agent 206 00:08:55,041 --> 00:08:57,921 有没有企业拿它做内部流程 207 00:08:57,921 --> 00:09:01,121 有没有硬件厂商把它放进设备 208 00:09:01,121 --> 00:09:04,741 有没有开源社区围绕它做微调和工具链 209 00:09:04,741 --> 00:09:06,641 如果这些事情发生 210 00:09:06,641 --> 00:09:09,881 Muse Glimmer的意义就不只是一个模型发布 211 00:09:09,881 --> 00:09:14,001 而是Meta在重新争夺AI基础设施的话语权 212 00:09:14,001 --> 00:09:16,301 这场竞争还有一个更大的背景 213 00:09:16,301 --> 00:09:19,161 AI正在从模型能力竞争 214 00:09:19,161 --> 00:09:21,121 转向分发方式竞争 215 00:09:21,121 --> 00:09:23,921 OpenAI想把模型做成超级应用 216 00:09:23,921 --> 00:09:26,601 Google想把AI筛回搜索和安卓 217 00:09:26,601 --> 00:09:29,641 苹果想把AI放进硬件和系统 218 00:09:29,641 --> 00:09:33,241 Meta则想让开放模型成为开发者默认底座 219 00:09:33,241 --> 00:09:35,861 Muse Glimmer只是其中一块拼图 220 00:09:35,861 --> 00:09:37,841 但它指向的方向很清楚 221 00:09:37,841 --> 00:09:40,581 让AI不只存在于云端 222 00:09:40,581 --> 00:09:42,781 也存在于每台电脑 223 00:09:42,781 --> 00:09:43,621 每个应用 224 00:09:43,621 --> 00:09:45,141 每个本地工作流理 225 00:09:45,141 --> 00:09:47,961 所以这次发布不能只看成Meta 226 00:09:47,961 --> 00:09:49,721 又开源了一个模型 227 00:09:49,721 --> 00:09:51,861 它更像一次路线宣言 228 00:09:51,861 --> 00:09:54,641 AI如果只能在数据中心运行 229 00:09:54,641 --> 00:09:56,741 用户就永远是租客 230 00:09:56,741 --> 00:09:59,241 AI如果可以在本地运行 231 00:09:59,241 --> 00:10:01,541 用户才有机会成为主人 232 00:10:01,541 --> 00:10:03,801 Muse Glimmer的价值就在这里 233 00:10:03,801 --> 00:10:05,841 它可能不是最强的模型 234 00:10:05,841 --> 00:10:07,941 但它在提醒整个行业 235 00:10:07,941 --> 00:10:10,961 未来的AI不应该只属于机房 236 00:10:10,961 --> 00:10:12,461 也应该属于桌面 237 00:10:12,461 --> 00:10:16,041 这里还要看Kosla对AI投资的判断 238 00:10:16,041 --> 00:10:18,301 他认为OpenAI和Anthropic 239 00:10:18,301 --> 00:10:19,901 不是AI故事的终点 240 00:10:19,901 --> 00:10:20,921 而只是开头 241 00:10:20,921 --> 00:10:23,401 未来会出现更多千亿美元 242 00:10:23,401 --> 00:10:25,021 万亿美元级公司 243 00:10:25,021 --> 00:10:28,901 影响科学、医疗、能源、教育、制造 244 00:10:28,901 --> 00:10:29,961 这些基础领域 245 00:10:29,961 --> 00:10:32,221 这个判断放到Muse Glimmer上 246 00:10:32,221 --> 00:10:34,161 就能看见Meta的算盘 247 00:10:34,161 --> 00:10:37,361 如果未来有大量垂直AI公司出现 248 00:10:37,361 --> 00:10:40,461 他们不一定都愿意从B原API开始 249 00:10:40,461 --> 00:10:42,921 科学研究就是一个典型场景 250 00:10:42,921 --> 00:10:44,761 实验室需要处理论文 251 00:10:44,761 --> 00:10:45,421 图像 252 00:10:45,421 --> 00:10:46,041 代码 253 00:10:46,041 --> 00:10:48,141 仪器数据和内部记录 254 00:10:48,141 --> 00:10:51,661 很多资料不能随便上传到外部云服务 255 00:10:51,661 --> 00:10:53,641 一个本地多模态模型 256 00:10:53,641 --> 00:10:54,921 就算不是最强 257 00:10:54,921 --> 00:10:56,981 也可以承担文献整理 258 00:10:56,981 --> 00:10:58,701 实验记录搜索 259 00:10:58,701 --> 00:10:59,841 图表理解 260 00:10:59,841 --> 00:11:00,921 工具调用 261 00:11:00,921 --> 00:11:02,581 和初步假设生成 262 00:11:02,581 --> 00:11:04,221 它不会替代科学家 263 00:11:04,221 --> 00:11:08,681 但会把很多低价值搜索和整理动作压低成本 264 00:11:08,681 --> 00:11:09,881 MED很清楚 265 00:11:09,881 --> 00:11:11,741 币源模型越强 266 00:11:11,741 --> 00:11:14,981 外界对模型集中化的担心就越重 267 00:11:14,981 --> 00:11:17,241 一个公司如果控制模型 268 00:11:17,241 --> 00:11:18,721 控制分发 269 00:11:18,721 --> 00:11:20,341 控制价格 270 00:11:20,341 --> 00:11:21,941 控制安全规则 271 00:11:21,941 --> 00:11:24,201 开发者就会变成依附者 272 00:11:24,201 --> 00:11:26,481 今天接口价格可以接受 273 00:11:26,481 --> 00:11:28,001 明天规则可能变 274 00:11:28,001 --> 00:11:29,641 今天调用额度够用 275 00:11:29,641 --> 00:11:32,401 明天业务增长后成本可能翻倍 276 00:11:32,401 --> 00:11:35,301 开放模型给开发者留了一条退路 277 00:11:35,301 --> 00:11:38,361 这条退路本身就是谈判筹码 278 00:11:38,361 --> 00:11:42,121 这也是为什么开放权重会压低行业利润率 279 00:11:42,121 --> 00:11:44,481 只要有足够好的开源替代品 280 00:11:44,481 --> 00:11:48,881 币源公司就很难对中低端任务收太高价格 281 00:11:48,881 --> 00:11:50,801 旗舰模型仍然可以卖高价 282 00:11:50,801 --> 00:11:55,161 但大量日常任务会被本地模型和开放模型赤掉 283 00:11:55,161 --> 00:11:57,461 Meta不靠模型订阅赚钱 284 00:11:57,461 --> 00:12:00,241 反而适合推动这种价格下行 285 00:12:00,241 --> 00:12:02,061 它要让AI变成空气 286 00:12:02,061 --> 00:12:04,081 然后在空气里做平台 287 00:12:04,221 --> 00:12:07,081 有人会问30B模型真的够吗 288 00:12:07,081 --> 00:12:08,721 这个问题要看任务 289 00:12:08,721 --> 00:12:11,561 写一段复杂法律意见可能不够 290 00:12:11,561 --> 00:12:13,821 做一个本地客服分类器 291 00:12:13,821 --> 00:12:17,681 够通读一本手册并回答设备维修问题 292 00:12:17,681 --> 00:12:21,501 可能够整理图片和文字生成初稿 293 00:12:21,501 --> 00:12:24,761 可能够控制一组本地工具跑流程 294 00:12:24,761 --> 00:12:29,721 可能也够AI产品不需要每次都派最强模型上场 295 00:12:29,721 --> 00:12:33,261 就像公司不会让总裁去处理每张报销单 296 00:12:33,261 --> 00:12:35,361 真正成熟的AI系统 297 00:12:35,361 --> 00:12:37,461 很可能是多模型协作 298 00:12:37,461 --> 00:12:39,081 小模型负责本地 299 00:12:39,081 --> 00:12:39,841 便宜 300 00:12:39,841 --> 00:12:40,661 低风险 301 00:12:40,661 --> 00:12:41,781 高频动作 302 00:12:41,781 --> 00:12:44,421 中型模型负责复杂判断 303 00:12:44,421 --> 00:12:47,081 旗舰模型负责少数关键推理 304 00:12:47,081 --> 00:12:50,821 Muse Glimmer的位置就在第一层和第二层之间 305 00:12:50,821 --> 00:12:52,321 它不抢所有舞台 306 00:12:52,321 --> 00:12:54,381 但可以进入很多工作流 307 00:12:54,381 --> 00:12:56,361 这个位置一旦铺开 308 00:12:56,361 --> 00:12:58,101 使用量会非常大 309 00:12:58,101 --> 00:13:01,461 所以Muse Glimmer看起来是技术事件 310 00:13:01,461 --> 00:13:03,401 背后其实是分发事件 311 00:13:03,401 --> 00:13:06,001 AI行业未来最大的战场 312 00:13:06,001 --> 00:13:07,181 不只是模型 313 00:13:07,181 --> 00:13:08,241 谁更聪明 314 00:13:08,241 --> 00:13:10,781 而是谁能接触更多开发者 315 00:13:10,781 --> 00:13:12,001 更多设备 316 00:13:12,001 --> 00:13:13,401 更多工作流 317 00:13:13,401 --> 00:13:14,941 更多真实业务 318 00:13:14,941 --> 00:13:17,941 开放权重提供了一种复制速度 319 00:13:17,941 --> 00:13:19,581 一个模型被下载后 320 00:13:19,581 --> 00:13:22,261 可以在无数环境里变成不同产品 321 00:13:22,261 --> 00:13:23,661 避原模型再强 322 00:13:23,661 --> 00:13:26,021 也必须等用户来到它的入口 323 00:13:26,021 --> 00:13:26,741 当然 324 00:13:26,741 --> 00:13:28,741 开放模型要赢也不容易 325 00:13:28,741 --> 00:13:30,481 开发体验必须好 326 00:13:30,481 --> 00:13:31,961 文档必须清楚 327 00:13:31,961 --> 00:13:33,581 工具链必须稳定 328 00:13:33,581 --> 00:13:35,581 推理速度必须能接受 329 00:13:35,581 --> 00:13:37,681 硬件适配不能太痛苦 330 00:13:37,681 --> 00:13:39,641 如果一个模型开远了 331 00:13:39,641 --> 00:13:41,581 但部署起来像拆炸弹 332 00:13:41,581 --> 00:13:43,581 开发者很快就会放弃 333 00:13:43,581 --> 00:13:47,341 Meta这次把Execute Torch等部署线索放进来 334 00:13:47,341 --> 00:13:50,101 说明他知道只开远权重还不够 335 00:13:50,101 --> 00:13:52,261 真正要真的是落地链路 336 00:13:52,261 --> 00:13:54,381 安全问题也不能回避 337 00:13:54,381 --> 00:13:56,041 一个能在本地运行 338 00:13:56,041 --> 00:13:57,421 能调用工具 339 00:13:57,421 --> 00:14:00,121 能处理多模态输入的模型 340 00:14:00,121 --> 00:14:01,861 如果被错误使用 341 00:14:01,861 --> 00:14:03,661 也会带来自动化滥用 342 00:14:03,661 --> 00:14:05,481 开放社区需要评估 343 00:14:05,481 --> 00:14:06,281 红队 344 00:14:06,281 --> 00:14:08,361 使用限制和透明报告 345 00:14:08,361 --> 00:14:10,521 META不能一边强调开放 346 00:14:10,521 --> 00:14:12,921 一边把风险全丢给社区 347 00:14:12,921 --> 00:14:14,621 开放路线要走得久 348 00:14:14,621 --> 00:14:16,681 必须同时建立开放治理 349 00:14:16,681 --> 00:14:18,901 但避原也不是天然安全 350 00:14:18,901 --> 00:14:21,021 避原模型的能力集中后 351 00:14:21,021 --> 00:14:23,461 外界更难知道它如何训练 352 00:14:23,461 --> 00:14:24,361 如何过滤 353 00:14:24,361 --> 00:14:25,341 如何拒绝 354 00:14:25,341 --> 00:14:27,001 如何偏向某些服务 355 00:14:27,001 --> 00:14:29,861 安全不是封闭和开放的简单选择 356 00:14:29,861 --> 00:14:34,101 而是透明度、责任和可验证性之间的平衡 357 00:14:34,101 --> 00:14:36,181 Muse Glimmer的价值之一 358 00:14:36,181 --> 00:14:39,421 就是让更多研究者能直接测试模型 359 00:14:39,421 --> 00:14:42,281 而不是只能相信厂商声明 360 00:14:42,281 --> 00:14:44,281 Meta设立社区基金 361 00:14:44,281 --> 00:14:48,261 某种意义上也是在承认AI基建的外部成本 362 00:14:48,261 --> 00:14:50,341 数据中心建在某个地方 363 00:14:50,341 --> 00:14:55,901 当地居民可能面对电价、用水、噪音、土地和税收分配问题 364 00:14:56,021 --> 00:14:59,161 科技公司过去常把这些问题当成地方沟通 365 00:14:59,161 --> 00:15:01,221 现在必须当成核心战略 366 00:15:01,221 --> 00:15:02,981 因为没有社区许可 367 00:15:02,981 --> 00:15:05,661 再强的模型也需要机防供电 368 00:15:05,661 --> 00:15:07,601 把这两条线合起来 369 00:15:07,601 --> 00:15:09,921 Meta的新叙事就完整了 370 00:15:09,921 --> 00:15:11,601 一边用开放模型说 371 00:15:11,601 --> 00:15:13,201 AI能力要分散 372 00:15:13,201 --> 00:15:14,961 一边用社区基金说 373 00:15:14,961 --> 00:15:16,841 AI基建收益要回流 374 00:15:16,841 --> 00:15:18,721 这不一定全是理想主义 375 00:15:18,721 --> 00:15:21,741 但他比单纯喊超级智能更聪明 376 00:15:21,741 --> 00:15:24,961 他知道AI行业如果继续只展示估值 377 00:15:24,961 --> 00:15:26,221 参数和算力 378 00:15:26,221 --> 00:15:28,261 公众反弹会越来越强 379 00:15:28,261 --> 00:15:30,861 未來真正有殺傷力的Muse Glimmer 380 00:15:30,861 --> 00:15:31,521 應用 381 00:15:31,521 --> 00:15:34,241 可能不會出現在發布新聞裡 382 00:15:34,241 --> 00:15:37,041 而會出現在很小的工作場景裡 383 00:15:37,041 --> 00:15:39,661 一個設計師用他本地整理素材 384 00:15:40,085 --> 00:15:49,185 一個設計師用他本地整理素材。一個律師事務所用他檢索內部文件。一個工廠用他讀設備圖片。一個個人開發者用他搭自己的桌面Agent。 385 00:15:49,185 --> 00:15:54,685 每個場景都不震撼,但加起來就是生態。這也是開放模型最強的地方。 386 00:15:54,685 --> 00:15:56,765 它的價值不是一次性釋放 387 00:15:56,765 --> 00:15:59,085 而是被別人不斷重新發明 388 00:15:59,085 --> 00:16:01,425 碧原模型像一座中央電站 389 00:16:01,425 --> 00:16:04,265 開放模型像一批可移動工具箱 390 00:16:04,265 --> 00:16:06,165 中央電站功率更大 391 00:16:06,165 --> 00:16:08,045 工具箱更容易進入角落 392 00:16:08,045 --> 00:16:10,165 AI要真正改變社會 393 00:16:10,165 --> 00:16:12,045 不能只靠中央電站 394 00:16:12,045 --> 00:16:13,565 也需要工具箱 395 00:16:13,565 --> 00:16:15,725 所以Muse Glimmer不是Meta 396 00:16:15,725 --> 00:16:18,125 对碧原模型的一次正面冲锋 397 00:16:18,125 --> 00:16:19,865 而是一次彻翼包抄 398 00:16:19,865 --> 00:16:22,625 它不说自己在所有能力上第一 399 00:16:22,625 --> 00:16:24,885 他说自己更容易被拿走 400 00:16:24,885 --> 00:16:26,265 更容易本地跑 401 00:16:26,265 --> 00:16:28,145 更容易进入普通设备 402 00:16:28,145 --> 00:16:29,825 这个策略如果成功 403 00:16:29,825 --> 00:16:32,305 Meta会在AI入口战争里 404 00:16:32,305 --> 00:16:34,405 拿到一个很难被关闭的位置 405 00:16:34,405 --> 00:16:36,265 Execute Torch的存在 406 00:16:36,265 --> 00:16:39,845 说明Meta对端侧部署有长期准备 407 00:16:39,845 --> 00:16:44,945 端侧AI不是把云端模型硬塞进手机 408 00:16:44,945 --> 00:16:47,245 而是重新考虑模型大小 409 00:16:47,245 --> 00:16:50,405 延迟 内存 占用 电池和隐私 410 00:16:50,405 --> 00:16:51,665 Muse Glimmer 411 00:16:51,665 --> 00:16:55,025 如果能和端侧框架形成完整链路 412 00:16:55,025 --> 00:16:58,005 就会成为Meta连接硬件厂商的桥 413 00:16:58,005 --> 00:17:01,925 以后手机、眼镜、耳机、笔记本、桌面设备 414 00:17:01,925 --> 00:17:06,465 都可能需要一个能本地处理多模态任务的模型底座 415 00:17:06,465 --> 00:17:08,505 这对开发者很有吸引力 416 00:17:08,505 --> 00:17:11,745 闭源API最舒服的地方是审视 417 00:17:11,745 --> 00:17:13,945 最痛的地方是不可控 418 00:17:13,945 --> 00:17:18,745 接口变更、价格变更、限流、地区限制、内容政策 419 00:17:18,745 --> 00:17:20,285 都可能影响产品 420 00:17:20,285 --> 00:17:21,905 开放模型麻烦一点 421 00:17:21,905 --> 00:17:22,665 但可控 422 00:17:22,665 --> 00:17:24,925 对长期做产品的人来说 423 00:17:24,925 --> 00:17:27,805 可控往往比短期省事更重要 424 00:17:27,805 --> 00:17:30,185 Muse Glimmer如果性能足够稳定 425 00:17:30,185 --> 00:17:31,705 就会被放进很多 426 00:17:31,705 --> 00:17:34,145 不愿完全依赖云端的产品里 427 00:17:34,145 --> 00:17:35,645 企业采用AI时 428 00:17:35,645 --> 00:17:37,645 常常不是被能力卡住 429 00:17:37,645 --> 00:17:39,385 而是被审批卡住 430 00:17:39,385 --> 00:17:41,085 法务问数据去哪 431 00:17:41,085 --> 00:17:43,305 安全团队问权限怎么管 432 00:17:43,305 --> 00:17:45,625 财务问每月调用成本 433 00:17:45,625 --> 00:17:48,345 业务部门问延迟能不能接受 434 00:17:48,345 --> 00:17:51,265 开放本地模型可以同时回答几个问题 435 00:17:51,265 --> 00:17:52,985 数据可以留在内网 436 00:17:52,985 --> 00:17:54,365 成本可以固定 437 00:17:54,365 --> 00:17:55,805 延迟可以降低 438 00:17:55,805 --> 00:17:57,485 权限可以自己定义 439 00:17:57,485 --> 00:17:58,985 它不是完美答案 440 00:17:58,985 --> 00:18:02,265 但比纯云端更容易通过某些企业流程 441 00:18:02,265 --> 00:18:05,845 Meta还可以借开放路线给监管传递信号 442 00:18:05,845 --> 00:18:08,905 B原超级模型越强监管越担心 443 00:18:08,905 --> 00:18:10,765 权力集中和黑箱决策 444 00:18:10,765 --> 00:18:12,025 开放模型 445 00:18:12,025 --> 00:18:15,565 让更多研究者能省计能力和风险 446 00:18:15,565 --> 00:18:19,545 至少表面上更符合广泛分发的公共叙事 447 00:18:19,545 --> 00:18:23,645 对一家掌握社交平台和广告系统的公司来说 448 00:18:23,645 --> 00:18:25,325 这种姿态很重要 449 00:18:25,325 --> 00:18:29,345 它需要证明自己不是把AI变成下一层垄断 450 00:18:29,345 --> 00:18:31,765 但市场不会因为姿态就买单 451 00:18:31,765 --> 00:18:33,885 开发者最后只看三件事 452 00:18:33,885 --> 00:18:35,025 够不够好 453 00:18:35,025 --> 00:18:36,105 跑不跑得动 454 00:18:36,105 --> 00:18:37,665 出问题能不能修 455 00:18:37,665 --> 00:18:41,065 Muse Glimmer需要在真实任务里证明自己 456 00:18:41,585 --> 00:18:43,265 比如本地图片理解 457 00:18:43,265 --> 00:18:44,265 文档问答 458 00:18:44,265 --> 00:18:45,465 桌面自动化 459 00:18:45,465 --> 00:18:46,465 工具调用 460 00:18:46,465 --> 00:18:47,525 失败恢复 461 00:18:47,525 --> 00:18:51,345 只要这些能力能稳定解决一批中等难度问题 462 00:18:51,345 --> 00:18:53,545 30B就不再是小模型 463 00:18:53,545 --> 00:18:56,185 而是刚好够用的生产力模型 464 00:18:56,185 --> 00:18:58,925 AI行业正在出现一种新分工 465 00:18:58,925 --> 00:19:01,165 最强壁源模型负责天花板 466 00:19:01,165 --> 00:19:03,945 开放模型负责地板和中间层 467 00:19:03,945 --> 00:19:06,305 天花板决定行业想象力 468 00:19:06,305 --> 00:19:08,525 地板决定普及速度 469 00:19:08,525 --> 00:19:09,365 没有天花板 470 00:19:09,365 --> 00:19:11,345 AI缺少突破 471 00:19:11,345 --> 00:19:12,425 没有地板 472 00:19:12,425 --> 00:19:15,745 AI只会变成少数公司的高价服务 473 00:19:15,745 --> 00:19:18,585 Muse Glimmer的意义就在地板和中间层 474 00:19:18,585 --> 00:19:20,985 它让更多人能踩上去 475 00:19:20,985 --> 00:19:22,225 扎克伯格讲 476 00:19:22,225 --> 00:19:24,885 发明创造而不是自动化 477 00:19:24,885 --> 00:19:28,005 其实是在修复AI的公共形象 478 00:19:28,005 --> 00:19:29,345 过去两年 479 00:19:29,345 --> 00:19:32,425 很多人听到AI就想到裁员 480 00:19:32,425 --> 00:19:33,445 替代 481 00:19:33,445 --> 00:19:35,465 数据中心耗电 482 00:19:35,465 --> 00:19:37,485 和大公司集中权力 483 00:19:37,485 --> 00:19:41,185 Meta要把话题拉回创造力和个人赋能 484 00:19:41,185 --> 00:19:43,285 Muse Glimmer适合这个叙事 485 00:19:43,285 --> 00:19:47,385 因为它看起来不像一个只服务企业裁员的工具 486 00:19:47,385 --> 00:19:50,565 而像一个可以被个人拿走的创造工具 487 00:19:50,565 --> 00:19:53,585 当然这种叙事也要接受现实检验 488 00:19:53,585 --> 00:19:57,505 如果开放模型最后主要被大公司拿去降低成本 489 00:19:57,505 --> 00:19:59,405 普通人没有明显受益 490 00:19:59,405 --> 00:20:01,905 那负能就会变成漂亮化 491 00:20:01,905 --> 00:20:03,685 真正能证明Method的 492 00:20:03,685 --> 00:20:06,025 是它有没有让个人开发者 493 00:20:06,025 --> 00:20:07,045 小企业 494 00:20:07,045 --> 00:20:07,785 学校 495 00:20:07,785 --> 00:20:10,025 研究者用更低成本 496 00:20:10,025 --> 00:20:11,985 做出过去做不了的东西 497 00:20:11,985 --> 00:20:14,585 开源的价值不能停在下载量 498 00:20:14,585 --> 00:20:17,725 要落到新作品和新产品上 499 00:20:17,725 --> 00:20:18,785 未来一年 500 00:20:18,785 --> 00:20:22,005 Muse Glimmer最值得观察的不是发布热读 501 00:20:22,005 --> 00:20:23,945 而是二次开发热读 502 00:20:23,945 --> 00:20:25,865 有没有高质量微调版本 503 00:20:25,865 --> 00:20:31,945 有没有中文代码医疗法律教育设计方向的社区版本 504 00:20:31,945 --> 00:20:34,845 有没有人把它做成一件本地agent 505 00:20:34,845 --> 00:20:38,805 有没有硬件厂商围绕它优化驱动和推理速度 506 00:20:38,805 --> 00:20:42,045 这些才是开放模型真正的生命迹象 507 00:20:42,045 --> 00:20:43,925 如果这些生态长出来 508 00:20:43,925 --> 00:20:46,905 Meta就会完成一次很聪明的布局 509 00:20:46,905 --> 00:20:49,685 它不需要每个用户直接打开Meta产品 510 00:20:49,685 --> 00:20:52,625 也不需要每次推理都走Meta服务器 511 00:20:52,625 --> 00:20:57,085 只要大量AI工具底层运行的是Meta开放模型 512 00:20:57,085 --> 00:20:59,805 它就会成为看不见的基础设施 513 00:20:59,805 --> 00:21:03,965 基础设施的力量往往比一个热门应用更持久 514 00:21:04,165 --> 00:21:05,425 Meta的优势是 515 00:21:05,425 --> 00:21:08,565 它不急着把模型本身变成利润中心 516 00:21:08,565 --> 00:21:10,845 它可以允许模型价格下降 517 00:21:10,845 --> 00:21:12,845 因为它真正赚钱的地方 518 00:21:12,845 --> 00:21:15,805 在广告、社交关系和平台分发 519 00:21:15,805 --> 00:21:19,725 OpenAI必须证明订阅和API能覆盖成本 520 00:21:19,725 --> 00:21:24,425 Anthropic必须证明企业安全和高端能力能卖出价格 521 00:21:24,425 --> 00:21:26,845 Meta则可以用开放模型靴 522 00:21:26,845 --> 00:21:28,305 若对手收费能力 523 00:21:28,305 --> 00:21:29,625 这种打法很狠 524 00:21:29,625 --> 00:21:32,685 如果未来个人设备都能跑中型模型 525 00:21:32,685 --> 00:21:34,925 AI产品形态也会变 526 00:21:34,925 --> 00:21:38,605 很多应用不再需要把所有请求传到服务器 527 00:21:38,605 --> 00:21:41,245 而是把模型嵌进本地功能里 528 00:21:41,245 --> 00:21:43,625 写作软件可以本地整理抄稿 529 00:21:43,625 --> 00:21:46,245 剪辑软件可以本地识别素材 530 00:21:46,245 --> 00:21:48,605 浏览器可以本地总结网页 531 00:21:48,605 --> 00:21:51,105 开发工具可以本地读项目 532 00:21:51,105 --> 00:21:52,505 云端仍然存在 533 00:21:52,505 --> 00:21:55,105 但本地会承担更多基础动作 534 00:21:55,105 --> 00:21:58,165 所以这件事最后要落到一个判断 535 00:21:58,165 --> 00:22:01,085 AI的未来不会只有一种入口 536 00:22:01,085 --> 00:22:03,245 云端旗舰模型会存在 537 00:22:03,245 --> 00:22:05,525 本地开放模型也会存在 538 00:22:05,525 --> 00:22:07,305 前者负责极限能力 539 00:22:07,305 --> 00:22:09,705 后者负责普及和控制权 540 00:22:09,705 --> 00:22:11,545 Meta现在压的是后者 541 00:22:11,545 --> 00:22:13,185 而且压得很聪明 542 00:22:13,185 --> 00:22:15,025 它不一定引走所有收入 543 00:22:15,025 --> 00:22:17,485 但可能引走很多默认选择 544 00:22:17,485 --> 00:22:19,625 如果把它放进更长周期 545 00:22:19,625 --> 00:22:20,785 Muse Glimmer 546 00:22:20,785 --> 00:22:23,765 代表的是AI商品化之后的下一步 547 00:22:23,765 --> 00:22:26,545 模型能力会越来越像基础能力 548 00:22:26,545 --> 00:22:28,985 真正差异会转移到数据 549 00:22:28,985 --> 00:22:29,785 场景 550 00:22:29,785 --> 00:22:30,465 部署 551 00:22:30,465 --> 00:22:31,905 体验和生态 552 00:22:31,905 --> 00:22:34,645 META提前把模型底座打开 553 00:22:34,645 --> 00:22:37,885 是在承认模型本身迟早会变便宜 554 00:22:37,885 --> 00:22:39,525 既然便宜不可避免 555 00:22:39,525 --> 00:22:41,305 不如主动让它变便宜 556 00:22:41,305 --> 00:22:43,585 再去控制更大的生态入口 557 00:22:43,585 --> 00:22:46,365 这对B原公司是一种压力测试 558 00:22:46,365 --> 00:22:48,965 B原模型如果只是略强一点 559 00:22:48,965 --> 00:22:50,125 却贵很多 560 00:22:50,125 --> 00:22:52,485 很多企业会改用开放模型 561 00:22:52,485 --> 00:22:56,185 B原模型必须在关键任务上强到无法替代 562 00:22:56,185 --> 00:22:59,765 或者在产品体验上省心得足够明显 563 00:22:59,765 --> 00:23:02,565 否则开源模型每进步一步 564 00:23:02,565 --> 00:23:05,265 B元模型的中地段收入就少一块 565 00:23:05,265 --> 00:23:08,065 Muse Glimmer也会推动AI Agent 566 00:23:08,065 --> 00:23:10,305 从玩具走向基础设施 567 00:23:10,305 --> 00:23:13,465 以前很多Agent演示看起来很聪明 568 00:23:13,465 --> 00:23:15,485 但实际运行成本高 569 00:23:15,485 --> 00:23:16,365 延迟长 570 00:23:16,365 --> 00:23:17,585 隐私风险大 571 00:23:17,585 --> 00:23:20,165 把一部分能力放到本地之后 572 00:23:20,165 --> 00:23:22,885 Agent可以更安静地处理日常任务 573 00:23:22,885 --> 00:23:24,665 他不需要每次都上云 574 00:23:24,665 --> 00:23:26,605 不需要每一步都花钱 575 00:23:26,605 --> 00:23:29,465 也不需要把所有私人资料交出去 576 00:23:29,465 --> 00:23:31,125 这就是它的流量点 577 00:23:31,125 --> 00:23:33,545 Meta不是在送一个模型 578 00:23:33,545 --> 00:23:35,205 而是在告诉市场 579 00:23:35,205 --> 00:23:38,425 AI的权力不一定只能往云端集中 580 00:23:38,425 --> 00:23:40,425 普通设备也可以有智能 581 00:23:40,425 --> 00:23:42,385 小团队也可以有底座 582 00:23:42,385 --> 00:23:44,305 开发者也可以有退路 583 00:23:44,305 --> 00:23:47,825 这个故事比单纯参数发布更有传播性 584 00:23:47,825 --> 00:23:51,165 因为它击中了很多人对AI垄断的焦虑 585 00:23:51,165 --> 00:23:54,285 所以Muse Glimmer值得被单独拿出来讲 586 00:23:54,285 --> 00:23:57,685 不是因为它一夜之间改写大模型格局 587 00:23:57,685 --> 00:24:00,825 而是因為它把一個被忽略的問題擺到檯面 588 00:24:00,825 --> 00:24:03,605 AI普及到底靠更大的模型 589 00:24:03,605 --> 00:24:06,285 還是靠更容易被部署的模型 590 00:24:06,285 --> 00:24:08,445 閉緣巨頭會繼續沖天花板 591 00:24:08,445 --> 00:24:10,505 但真正改變日常工作的 592 00:24:10,505 --> 00:24:13,125 往往是那些能被普通人裝進電腦 593 00:24:13,125 --> 00:24:14,165 接近流程 594 00:24:14,165 --> 00:24:15,625 反覆修改的東西