1 00:00:00,000 --> 00:00:02,820 Google一天蒸發接近2000億美元 2 00:00:02,820 --> 00:00:04,860 導火索不是一張財報 3 00:00:04,860 --> 00:00:06,660 也不是一個模型翻車 4 00:00:06,660 --> 00:00:08,700 而是幾個人離開了公司 5 00:00:08,700 --> 00:00:10,240 這聽起來很誇張 6 00:00:10,240 --> 00:00:12,800 科技公司每天都有人離職 7 00:00:12,800 --> 00:00:15,100 AI圈每天都有人創業 8 00:00:15,100 --> 00:00:17,540 為什麼這次市場反應這麼大 9 00:00:17,540 --> 00:00:19,960 因為這次離開的不是普通高管 10 00:00:19,960 --> 00:00:21,500 而是Jeff Dean 11 00:00:21,500 --> 00:00:22,780 Sanjay Gamowatt 12 00:00:22,780 --> 00:00:23,560 Corkley 13 00:00:23,560 --> 00:00:24,840 Orio Vinyos 14 00:00:24,840 --> 00:00:25,860 這一組人 15 00:00:25,860 --> 00:00:27,400 這四個名字放在一起 16 00:00:27,400 --> 00:00:29,700 等於把Google過去20多年 17 00:00:29,700 --> 00:00:31,940 最值錢的兩條技術線抽出來 18 00:00:31,940 --> 00:00:35,100 一條是分布式系統和大規模計算 19 00:00:35,100 --> 00:00:37,800 一條是深度學習和前沿模型 20 00:00:37,800 --> 00:00:40,700 Jeff Dean在Google工作約27年 21 00:00:40,700 --> 00:00:42,500 是Google Chief Scientist 22 00:00:42,500 --> 00:00:44,800 也是Google Brain聯合創始人 23 00:00:44,800 --> 00:00:47,500 Sanjay Gamowatt是Google早期員工 24 00:00:47,500 --> 00:00:49,300 和Dean長期搭檔 25 00:00:49,300 --> 00:00:50,500 Google File System 26 00:00:50,500 --> 00:00:51,500 MapReduce 27 00:00:51,500 --> 00:00:52,300 Bigtable 28 00:00:52,300 --> 00:00:53,100 Spanner 29 00:00:53,100 --> 00:00:54,300 這些底層系統 30 00:00:54,300 --> 00:00:56,100 都和他們這一代人有關 31 00:00:56,100 --> 00:00:59,500 Quarker是Google Brain聯合創始人之一 32 00:00:59,500 --> 00:01:03,900 參與過Seek to Seek, AutoML,神經架構搜索等方向 33 00:01:03,900 --> 00:01:07,260 Oreal Vinyles是DeepMind研究副總裁 34 00:01:07,260 --> 00:01:09,760 也是Gemini技術負責人之一 35 00:01:09,760 --> 00:01:14,020 參與過AlphaStar,AlphaCode,Gemini這些關鍵項目 36 00:01:14,020 --> 00:01:15,700 所以市場這次恐慌 37 00:01:15,700 --> 00:01:18,160 並非在給4個人的工資定價 38 00:01:18,160 --> 00:01:21,220 而是在給Google AI的組織風險定價 39 00:01:21,220 --> 00:01:24,120 股價跌4%還是5% 40 00:01:24,120 --> 00:01:27,400 市值蒸發是1600億還是接近2000億 41 00:01:27,400 --> 00:01:29,280 數字可以有不同口徑 42 00:01:29,280 --> 00:01:31,480 但市场给出的信号很清楚 43 00:01:31,480 --> 00:01:35,440 当Google最核心的AI和基础设施人物同时离开 44 00:01:35,440 --> 00:01:37,580 投资人会立刻问一个问题 45 00:01:37,580 --> 00:01:40,320 Google的AI护城河到底还稳不稳 46 00:01:40,320 --> 00:01:42,140 这期真正要讲的 47 00:01:42,140 --> 00:01:44,260 不该停在Jeff Dean有多牛 48 00:01:44,260 --> 00:01:46,420 那样写出来只是人物履历 49 00:01:46,420 --> 00:01:48,240 真正要讲的是 50 00:01:48,240 --> 00:01:50,960 Google过去靠这些人建立了搜索 51 00:01:50,960 --> 00:01:53,660 广告 云和AI的底层系统 52 00:01:53,660 --> 00:01:55,720 现在这些人离开Google 53 00:01:55,720 --> 00:01:58,580 要做一家叫Discovery Loop的公司 54 00:01:58,580 --> 00:02:03,140 目标是用AI自动化科学和工程里的实验循环 55 00:02:03,140 --> 00:02:04,920 这不是普通创业 56 00:02:04,920 --> 00:02:07,740 这是Google技术文化的一部分 57 00:02:07,740 --> 00:02:10,380 从公司内部转移到了公司外部 58 00:02:10,380 --> 00:02:12,620 先看Jeff Dean到底代表什么 59 00:02:12,620 --> 00:02:14,400 很多人认识Jeff Dean 60 00:02:14,400 --> 00:02:16,900 是因为那些程序员段子 61 00:02:16,900 --> 00:02:20,160 比如编译器不优化Jeff Dean的代码 62 00:02:20,160 --> 00:02:22,420 是Jeff Dean优化编译器 63 00:02:22,420 --> 00:02:24,200 这些段子当然是玩笑 64 00:02:24,200 --> 00:02:26,620 但它们能流行这么多年 65 00:02:26,620 --> 00:02:28,460 说明一件事 66 00:02:28,580 --> 00:02:32,580 在工程式文化裡,Jeff Dean不是一個普通管理者 67 00:02:32,580 --> 00:02:35,280 他更像Google工程神話的象徵 68 00:02:35,280 --> 00:02:39,380 Google早期能崛起,靠的不是一個漂亮搜索框 69 00:02:39,380 --> 00:02:46,280 而是背後那套能處理全球網頁、廣告、數據、日誌和計算任務的基礎設施 70 00:02:46,280 --> 00:02:47,680 Google File System 71 00:02:47,680 --> 00:02:49,780 解決海量文件存儲 72 00:02:49,780 --> 00:02:53,780 Map Reduce 讓大規模數據處理變成工程常規 73 00:02:53,780 --> 00:02:56,280 Bigtable 支撐結構化數據系統 74 00:02:56,280 --> 00:02:59,780 Spanner把全球分布式數據庫推到另一個層級 75 00:02:59,780 --> 00:03:03,260 這些東西聽起來不像今天的AI模型那麼性感 76 00:03:03,260 --> 00:03:07,160 但他們才是Google能夠擴張成超級機器的地基 77 00:03:07,160 --> 00:03:09,760 Jeff Dean和Sunjay Gamawatt 78 00:03:09,760 --> 00:03:11,860 這一代人的價值就在這裡 79 00:03:11,860 --> 00:03:14,900 他們的強項不只是寫某個產品功能 80 00:03:14,900 --> 00:03:19,000 而是能把一種新計算範式變成公司級基礎設施 81 00:03:19,000 --> 00:03:20,740 這點很關鍵 82 00:03:20,740 --> 00:03:23,640 Google真正賺錢的搜索和廣告 83 00:03:23,640 --> 00:03:25,340 表面是用戶入口 84 00:03:25,340 --> 00:03:27,340 背後是大規模計算能力 85 00:03:27,340 --> 00:03:29,000 YouTube的推薦 86 00:03:29,000 --> 00:03:30,380 Gmail的過濾 87 00:03:30,380 --> 00:03:31,800 Google Cloud的服務 88 00:03:31,800 --> 00:03:33,100 後來的TPU 89 00:03:33,100 --> 00:03:34,100 TensorFlow 90 00:03:34,100 --> 00:03:35,100 Google Brain 91 00:03:35,100 --> 00:03:35,880 Gemini 92 00:03:35,880 --> 00:03:38,300 也都離不開這種基礎設施文化 93 00:03:38,300 --> 00:03:39,600 Google最強的時候 94 00:03:39,600 --> 00:03:41,540 不只是有聰明人 95 00:03:41,540 --> 00:03:43,940 而是有一套能把聰明人寫出的系統 96 00:03:43,940 --> 00:03:46,340 放大到全球規模的工程組織 97 00:03:46,340 --> 00:03:48,580 Jeff Dean離開市場害怕的 98 00:03:48,580 --> 00:03:50,340 並非少一個職位 99 00:03:50,340 --> 00:03:52,180 而是這套組織神話 100 00:03:52,180 --> 00:03:53,440 出現了裂縫 101 00:03:53,440 --> 00:03:55,440 再看一起離開的三個人 102 00:03:55,440 --> 00:03:59,840 Sanjay Gamawad是最容易被普通觀眾低估的人 103 00:03:59,840 --> 00:04:01,440 相比Jeff Dean 104 00:04:01,440 --> 00:04:04,440 他更少出現在大眾科技新聞裡 105 00:04:04,440 --> 00:04:06,440 但在分布式系統裡 106 00:04:06,440 --> 00:04:08,440 他的份量極重 107 00:04:08,440 --> 00:04:10,940 他和Dean共同參與的系統 108 00:04:10,940 --> 00:04:13,440 塑造了Google的計算底座 109 00:04:13,440 --> 00:04:14,940 兩個人一起離開 110 00:04:14,940 --> 00:04:16,440 象徵意義很重 111 00:04:16,440 --> 00:04:18,440 這不是一個AI高管離職 112 00:04:18,440 --> 00:04:22,140 而是Google早期基礎設施時代的兩位核心人物 113 00:04:22,140 --> 00:04:24,700 同時把下一張放到外部公司 114 00:04:24,700 --> 00:04:27,240 Quacklet代表的是Google Brain時代 115 00:04:27,240 --> 00:04:31,620 Google Brain曾經是現代深度學習進入Google的關鍵入口 116 00:04:31,620 --> 00:04:34,520 CQ2Seek 神經機器翻譯 117 00:04:34,520 --> 00:04:37,380 AutoML 神經架構搜索 118 00:04:37,380 --> 00:04:40,320 這些方向都影響過整個AI行業 119 00:04:40,320 --> 00:04:43,120 Quacklet的價值不只是做過某個項目 120 00:04:43,120 --> 00:04:46,460 而是它熟悉從研究想法到模型系統 121 00:04:46,460 --> 00:04:49,500 再到自動化機器學習的整條鏈爐 122 00:04:49,500 --> 00:04:52,380 Discovery Loop要做自動化科研實驗 123 00:04:52,380 --> 00:04:54,740 Quacko這種背景非常貼合 124 00:04:54,740 --> 00:04:58,020 Oreal Vinyles代表的是Deep Mind和Gemini 125 00:04:58,020 --> 00:04:59,780 這一代前沿模型 126 00:04:59,780 --> 00:05:03,100 Alpha Star, Alpha Code, Gemini 127 00:05:03,100 --> 00:05:05,780 這些項目背後有一個共同點 128 00:05:05,780 --> 00:05:07,620 不是單純做聊天 129 00:05:07,620 --> 00:05:13,700 而是讓模型在複雜環境裡推理、規劃、生成代碼、解決任務 130 00:05:13,700 --> 00:05:16,900 Discovery Loop要自動化實驗循環 131 00:05:16,900 --> 00:05:19,180 需要的也正是這種能力 132 00:05:19,180 --> 00:05:21,480 模型不能只是回答問題 133 00:05:21,480 --> 00:05:26,980 他要提出假設、設計實驗、執行、觀察結果、再修改 134 00:05:26,980 --> 00:05:29,180 所以這四個人的組合很嚇人 135 00:05:29,180 --> 00:05:32,280 Dean和SunJay帶來大規模系統能力 136 00:05:32,280 --> 00:05:35,080 Quarko帶來機器學習自動化經驗 137 00:05:35,080 --> 00:05:38,680 Oreo帶來前沿模型和Agent任務經驗 138 00:05:38,680 --> 00:05:40,180 放在Google內部 139 00:05:40,180 --> 00:05:42,180 他們分別是基礎設施 140 00:05:42,180 --> 00:05:46,180 Google Brain、DeepMind、Gemini的關鍵節點 141 00:05:46,180 --> 00:05:47,780 放到一家新公司裡 142 00:05:47,780 --> 00:05:51,740 他們就變成了一套完整的AI科研自動化班底 143 00:05:51,740 --> 00:05:54,440 這就是Discovery Loop的真正爆點 144 00:05:54,440 --> 00:05:57,640 他不是要做一個更聰明的聊天機器人 145 00:05:57,640 --> 00:06:00,020 也不是在做一個代碼助手 146 00:06:00,020 --> 00:06:03,520 他想自動化科學和工程裡的Discovery Loop 147 00:06:03,520 --> 00:06:05,380 也就是發現循環 148 00:06:05,380 --> 00:06:07,140 傳統科研大概是這樣 149 00:06:07,140 --> 00:06:13,280 提出一個假設、設計實驗、執行實驗、觀察數據、修改假設 150 00:06:13,280 --> 00:06:14,880 再做下一輪 151 00:06:14,880 --> 00:06:16,320 這個過程非常慢 152 00:06:16,320 --> 00:06:18,720 也非常依賴人的直覺和耐心 153 00:06:18,720 --> 00:06:21,680 Discovery Loop想把這件事交給AI 154 00:06:21,680 --> 00:06:24,000 讓AI同時跑大量試驗 155 00:06:24,000 --> 00:06:25,600 自動篩選結果 156 00:06:25,600 --> 00:06:27,120 自動調整方向 157 00:06:27,120 --> 00:06:30,560 最後把科研和工程疊代速度推上去 158 00:06:30,560 --> 00:06:32,400 這件事聽起來很大 159 00:06:32,400 --> 00:06:34,080 但它有清晰落點 160 00:06:34,080 --> 00:06:35,840 Wired的報導裡提到 161 00:06:35,840 --> 00:06:41,600 Discovery Loop最開始會先把機器學習研究和工程當作第一塊場地 162 00:06:41,600 --> 00:06:42,880 原因很直接 163 00:06:42,880 --> 00:06:45,580 這幫人最懂的就是機器學習系統 164 00:06:45,580 --> 00:06:48,180 他們可以先讓AI自動優化 165 00:06:48,180 --> 00:06:50,280 AI自己的研究流程 166 00:06:50,280 --> 00:06:52,280 比如模型結構 167 00:06:52,280 --> 00:06:53,580 訓練策略 168 00:06:53,580 --> 00:06:54,780 評測方法 169 00:06:54,780 --> 00:06:56,080 代碼實現 170 00:06:56,080 --> 00:06:57,380 實驗參數 171 00:06:57,380 --> 00:06:58,980 等這套方法跑通 172 00:06:58,980 --> 00:07:00,780 再擴展到藥物發現 173 00:07:00,780 --> 00:07:01,580 材料 174 00:07:01,580 --> 00:07:02,380 芯片 175 00:07:02,380 --> 00:07:03,580 清潔能源 176 00:07:03,580 --> 00:07:05,980 健康信息等更重的科學領域 177 00:07:05,980 --> 00:07:07,580 這條路線很聰明 178 00:07:07,580 --> 00:07:10,080 很多AI for Science公司一上來就說 179 00:07:10,080 --> 00:07:11,180 要改變藥物 180 00:07:11,180 --> 00:07:11,980 材料 181 00:07:11,980 --> 00:07:12,780 能源 182 00:07:12,780 --> 00:07:14,780 故事很大但落地很難 183 00:07:14,780 --> 00:07:19,280 Discovery Loop選擇先從機器學習研究本身開導 184 00:07:19,280 --> 00:07:21,780 相當於把自己當作第一個客戶 185 00:07:21,780 --> 00:07:23,780 用AI加速AI研究 186 00:07:23,780 --> 00:07:26,780 再用更強的AI去加速其他科學 187 00:07:26,780 --> 00:07:28,780 如果這條路線跑通 188 00:07:28,780 --> 00:07:31,780 它的影響就不只是做一家創業公司 189 00:07:31,780 --> 00:07:34,780 它會改變AI研究的生產方式 190 00:07:34,939 --> 00:07:36,939 過去AI研究最貴的地方 191 00:07:36,939 --> 00:07:38,939 不只是算力 192 00:07:38,939 --> 00:07:40,939 還有實驗設計和工程迭代 193 00:07:40,939 --> 00:07:42,939 一個模型想法出來 194 00:07:42,939 --> 00:07:46,939 要寫代碼、調參數、跑實驗、看結果 195 00:07:46,939 --> 00:07:48,939 查bug、改架構、再跑 196 00:07:48,939 --> 00:07:50,739 很多失敗不會寫進論文 197 00:07:50,739 --> 00:07:52,299 但他們消耗時間 198 00:07:52,299 --> 00:07:54,419 算力和工程師精力 199 00:07:54,419 --> 00:07:56,559 誰能自動化這套循環 200 00:07:56,559 --> 00:07:59,099 誰就能把研究速度拉開 201 00:07:59,099 --> 00:08:01,399 這也是Google最該害怕的地方 202 00:08:01,399 --> 00:08:02,979 Google不缺算力 203 00:08:02,979 --> 00:08:04,039 不缺論文 204 00:08:04,039 --> 00:08:05,479 不缺模型人才 205 00:08:05,479 --> 00:08:08,299 它真正的壓力是組織速度 206 00:08:08,299 --> 00:08:10,339 大公司內部要排優先級 207 00:08:10,339 --> 00:08:11,779 要接產品線 208 00:08:11,779 --> 00:08:12,819 要顧安全 209 00:08:12,819 --> 00:08:14,559 要對齊商業計畫 210 00:08:14,559 --> 00:08:16,199 要準備發布節奏 211 00:08:16,199 --> 00:08:17,819 創業公司不一樣 212 00:08:17,819 --> 00:08:19,819 他可以围绕一个目标猛攻 213 00:08:19,819 --> 00:08:21,959 把组织摩擦降到最低 214 00:08:21,959 --> 00:08:23,659 Jeff Dean这批人 215 00:08:23,659 --> 00:08:26,319 如果在Google内部做Discovery Loop 216 00:08:26,319 --> 00:08:28,479 可能会被放进一个部门 217 00:08:28,479 --> 00:08:29,499 一套预算 218 00:08:29,499 --> 00:08:30,639 一个路线图 219 00:08:30,639 --> 00:08:32,219 一堆汇报关系里 220 00:08:32,219 --> 00:08:33,559 放到公司外部 221 00:08:33,559 --> 00:08:35,959 反而可以用创业公司速度去做 222 00:08:35,959 --> 00:08:38,399 这就是这件事最微妙的地方 223 00:08:38,399 --> 00:08:40,539 Google并没有完全切割他们 224 00:08:40,539 --> 00:08:42,679 Google官方公告里说 225 00:08:42,679 --> 00:08:45,979 Discovery Loop是独立Public Benefit Corporation 226 00:08:45,979 --> 00:08:49,979 但Google会作为Founding Investor和Cloud Partner参与 227 00:08:49,979 --> 00:08:51,439 这句话很重要 228 00:08:51,439 --> 00:08:53,959 它说明这不是试破脸式离职 229 00:08:53,959 --> 00:08:55,899 而是一种外部化安排 230 00:08:55,899 --> 00:08:58,919 台面上Google是投资人和云伙伴 231 00:08:58,919 --> 00:09:02,459 台下看Google也许是在承认一件事 232 00:09:02,459 --> 00:09:05,299 有些长期科研和高风险实验 233 00:09:05,299 --> 00:09:07,859 放在Google里面未必跑得最快 234 00:09:07,859 --> 00:09:09,099 不如放出去 235 00:09:09,099 --> 00:09:11,719 让这批人用创业公司形态做 236 00:09:11,719 --> 00:09:15,159 再让Google保留投资和算力合作关系 237 00:09:15,159 --> 00:09:17,419 Google之前也用过类似打法 238 00:09:17,419 --> 00:09:18,519 最近几年 239 00:09:18,519 --> 00:09:20,779 大科技越来越喜欢一种新动作 240 00:09:20,779 --> 00:09:23,639 不一定把所有东西都留在公司内部 241 00:09:23,639 --> 00:09:25,279 也不一定完整收购 242 00:09:25,279 --> 00:09:26,719 而是通过投资 243 00:09:26,719 --> 00:09:27,459 授权 244 00:09:27,459 --> 00:09:28,559 人才流动 245 00:09:28,559 --> 00:09:29,399 云合作 246 00:09:29,399 --> 00:09:32,259 把外部创新变成自己的生态延伸 247 00:09:32,259 --> 00:09:34,819 Google和Character AI的合作 248 00:09:34,819 --> 00:09:36,339 Google抢Windsurf 249 00:09:36,339 --> 00:09:37,939 人才和技术授权 250 00:09:37,939 --> 00:09:40,319 Microsoft和Open AI的绑定 251 00:09:40,319 --> 00:09:42,259 Amazon和Anthropic 252 00:09:42,259 --> 00:09:43,639 都有类似逻辑 253 00:09:43,639 --> 00:09:45,899 巨头不再只靠内部研发 254 00:09:45,899 --> 00:09:48,519 他们也靠资本和云资源 255 00:09:48,519 --> 00:09:50,779 把外部公司纳入势力范围 256 00:09:50,779 --> 00:09:52,819 Discovery Loop这次更特别 257 00:09:52,819 --> 00:09:53,979 因为出去的人 258 00:09:53,979 --> 00:09:56,579 来自Google自己最核心的技术权 259 00:09:56,579 --> 00:09:58,319 所以它既向人才流失 260 00:09:58,319 --> 00:09:59,739 也向战略外包 261 00:09:59,739 --> 00:10:02,819 投资人看到的就是这层矛盾 262 00:10:02,819 --> 00:10:04,519 如果这是人才流失 263 00:10:04,519 --> 00:10:06,099 那Google很危险 264 00:10:06,099 --> 00:10:07,879 AI战争打到现在 265 00:10:07,879 --> 00:10:10,159 顶级人才就是最稀缺资产 266 00:10:10,159 --> 00:10:11,139 Open AI 267 00:10:11,139 --> 00:10:12,159 Anthropic 268 00:10:12,159 --> 00:10:12,839 Meta 269 00:10:12,839 --> 00:10:13,579 Google 270 00:10:13,579 --> 00:10:15,519 XAI都在抢人 271 00:10:15,519 --> 00:10:18,199 千万美元级别签约包已经不稀奇 272 00:10:18,199 --> 00:10:20,239 Google一口气失去Jeff Dean 273 00:10:20,239 --> 00:10:21,039 Sunjay 274 00:10:21,039 --> 00:10:21,759 Quark 275 00:10:21,759 --> 00:10:23,659 Orio这种组合 276 00:10:23,659 --> 00:10:25,379 市场当然会紧张 277 00:10:25,379 --> 00:10:26,859 如果这是战略外包 278 00:10:26,859 --> 00:10:28,259 那Google也有风险 279 00:10:28,259 --> 00:10:32,199 因为最有野心的科研方向被放到公司外部 280 00:10:32,199 --> 00:10:36,739 说明Google内部可能已经不适合成在某些高风险探索 281 00:10:36,739 --> 00:10:38,099 一个科技巨头 282 00:10:38,099 --> 00:10:41,719 如果最前沿的问题要靠外部创业公司去推进 283 00:10:41,719 --> 00:10:43,059 投资人会问 284 00:10:43,059 --> 00:10:45,619 Google还是AI革命的中心吗 285 00:10:45,619 --> 00:10:48,099 这就是股价反应的真正逻辑 286 00:10:48,099 --> 00:10:50,199 市场看的不是离职人数 287 00:10:50,199 --> 00:10:52,979 而是Google AI的组织含金量 288 00:10:52,979 --> 00:10:54,999 过去Google的最大优势 289 00:10:54,999 --> 00:10:57,859 是全世界最强的AI研究人员 290 00:10:57,859 --> 00:10:59,059 愿意待在Google 291 00:10:59,059 --> 00:11:01,079 因为那里有数据 292 00:11:01,079 --> 00:11:01,919 算力 293 00:11:01,919 --> 00:11:02,599 论文 294 00:11:02,599 --> 00:11:03,439 环境 295 00:11:03,439 --> 00:11:05,759 工程系统和产品入口 296 00:11:05,759 --> 00:11:07,859 可是AI商业化之后 297 00:11:07,859 --> 00:11:10,079 Google的内部环境变复杂了 298 00:11:10,079 --> 00:11:11,579 研究要服务Gemini 299 00:11:11,579 --> 00:11:13,279 Gemini要服务搜索 300 00:11:13,279 --> 00:11:14,939 搜索要保护广告 301 00:11:14,939 --> 00:11:16,359 广告要面对AI 302 00:11:16,359 --> 00:11:17,699 答案冲击云 303 00:11:17,699 --> 00:11:20,839 要追Microsoft和Amazon模型发布 304 00:11:20,839 --> 00:11:23,579 又要和OpenAI Anthropic对比 305 00:11:23,579 --> 00:11:25,719 这种压力会让研究人员 306 00:11:25,719 --> 00:11:28,239 很难只按科研直觉行动 307 00:11:28,239 --> 00:11:30,819 顶级人才最怕的往往不是忙 308 00:11:30,819 --> 00:11:32,999 而是被组织节奏托住 309 00:11:32,999 --> 00:11:35,259 AI时代的顶级研究者 310 00:11:35,259 --> 00:11:36,639 越来越像创业者 311 00:11:36,639 --> 00:11:38,579 他们不只是想写论文 312 00:11:38,579 --> 00:11:40,339 也想控制算力 313 00:11:40,339 --> 00:11:42,079 控制产品方向 314 00:11:42,079 --> 00:11:43,699 控制实验节奏 315 00:11:43,699 --> 00:11:45,399 控制团队文化 316 00:11:45,399 --> 00:11:47,219 大公司能给资源 317 00:11:47,219 --> 00:11:48,819 但也会给限制 318 00:11:48,819 --> 00:11:50,799 创业公司资源少一点 319 00:11:50,799 --> 00:11:52,179 但决策更快 320 00:11:52,179 --> 00:11:53,939 股权激励更直接 321 00:11:53,939 --> 00:11:55,279 方向更纯 322 00:11:55,279 --> 00:11:57,399 Jeff Dean这种级别的人 323 00:11:57,399 --> 00:11:58,759 离开Google创业 324 00:11:58,759 --> 00:12:00,879 带来的象征意义很强 325 00:12:00,879 --> 00:12:03,459 连Google的技术神话人物 326 00:12:03,459 --> 00:12:04,419 都开始相信 327 00:12:04,419 --> 00:12:06,479 有些事在Google外面 328 00:12:06,479 --> 00:12:07,359 更适合做 329 00:12:07,359 --> 00:12:10,179 这才是对Google最刺痛的地方 330 00:12:10,179 --> 00:12:12,439 当然,这不代表Google完了 331 00:12:12,439 --> 00:12:14,499 Google仍然拥有 332 00:12:14,499 --> 00:12:15,419 Gemini 333 00:12:15,419 --> 00:12:17,239 拥有DeepMind 334 00:12:17,239 --> 00:12:18,519 拥有TPU 335 00:12:18,519 --> 00:12:20,039 拥有搜索入口 336 00:12:20,039 --> 00:12:21,299 拥有YouTube 337 00:12:21,299 --> 00:12:22,599 拥有Android 338 00:12:22,599 --> 00:12:25,659 拥有全球最强的工程组织之一 339 00:12:25,659 --> 00:12:26,859 Demis Hasebis 340 00:12:26,859 --> 00:12:30,259 转任Google DeepMind Chair和Alphabet Chief Scientist 341 00:12:30,259 --> 00:12:31,339 不是退出 342 00:12:31,339 --> 00:12:35,479 而是从日常管理转向更高层的科学和战略角色 343 00:12:35,479 --> 00:12:38,879 Korekavu Korglu接管Gemini模型开发 344 00:12:38,879 --> 00:12:40,639 前沿AI研究 345 00:12:40,639 --> 00:12:43,099 Gemini App和开发者团队 346 00:12:43,099 --> 00:12:45,759 说明Google也在试图把模型 347 00:12:45,759 --> 00:12:48,499 产品和开发者生态拉得更紧 348 00:12:48,499 --> 00:12:50,259 这次领导层重组 349 00:12:50,259 --> 00:12:52,859 可以看作Google给AI战争换挡 350 00:12:52,859 --> 00:12:55,119 过去DeepMind和Google Research 351 00:12:55,119 --> 00:12:56,619 有研究传统 352 00:12:56,619 --> 00:12:58,479 Gemini有模型压力 353 00:12:58,479 --> 00:13:00,239 产品线有商业目标 354 00:13:00,239 --> 00:13:02,659 现在Google想把这些更集中的 355 00:13:02,659 --> 00:13:04,619 接到Gemini和开发者入口上 356 00:13:04,619 --> 00:13:05,539 问题是 357 00:13:05,539 --> 00:13:07,919 这种集中会提高产品执行力 358 00:13:07,919 --> 00:13:11,099 也可能让纯科研人物觉得空间变窄 359 00:13:11,099 --> 00:13:13,539 所以Jeff Dean离开和Google重组 360 00:13:13,539 --> 00:13:15,179 并非两个孤立事件 361 00:13:15,179 --> 00:13:17,099 他们像一枚硬币的两面 362 00:13:17,099 --> 00:13:19,659 一面是Google需要更产品化 363 00:13:19,659 --> 00:13:20,519 更集中 364 00:13:20,519 --> 00:13:22,139 更能和OpenAI 365 00:13:22,139 --> 00:13:23,139 Anthropic 366 00:13:23,139 --> 00:13:24,119 Microsoft 367 00:13:24,119 --> 00:13:26,139 打正面战的AI组织 368 00:13:26,139 --> 00:13:28,139 另一面是一些顶级研究者 369 00:13:28,139 --> 00:13:29,419 想去做更长期 370 00:13:29,419 --> 00:13:30,139 更开放 371 00:13:30,139 --> 00:13:32,079 更接近科研自动化的事情 372 00:13:32,079 --> 00:13:34,539 Google留下的是Gemini战争 373 00:13:34,539 --> 00:13:37,299 Dean他们带走的是Discovery Loop战争 374 00:13:37,299 --> 00:13:39,139 这两场战争都重要 375 00:13:39,139 --> 00:13:40,319 但节奏不同 376 00:13:40,319 --> 00:13:42,999 Gemini战争拼的是模型发布 377 00:13:42,999 --> 00:13:44,119 产品体验 378 00:13:44,119 --> 00:13:44,999 API 379 00:13:44,999 --> 00:13:46,279 开发者工具 380 00:13:46,279 --> 00:13:48,419 企业客户和搜索入口 381 00:13:48,419 --> 00:13:52,099 Discovery Loop战争拼的是科研生产方式 382 00:13:52,099 --> 00:13:53,219 自动化实验 383 00:13:53,219 --> 00:13:55,539 长期复利和跨学科突破 384 00:13:55,539 --> 00:13:56,519 短期看 385 00:13:56,519 --> 00:13:58,999 Gemini更能影响Google收入 386 00:13:58,999 --> 00:14:00,059 长期看 387 00:14:00,059 --> 00:14:02,039 Discovery Loop如果跑通 388 00:14:02,039 --> 00:14:05,359 可能会影响整个科学和工程创新速度 389 00:14:05,359 --> 00:14:08,499 这也是为什么Google还要投资Discovery Loop 390 00:14:08,499 --> 00:14:10,139 它不能完全放手 391 00:14:10,139 --> 00:14:11,979 如果Discovery Loop成功 392 00:14:11,979 --> 00:14:14,639 Google至少还能通过投资和云合作 393 00:14:14,639 --> 00:14:15,939 赤到一部分红利 394 00:14:15,939 --> 00:14:17,999 如果Discovery Loop失败 395 00:14:17,999 --> 00:14:20,679 Google失去的是几个人的外部尝试 396 00:14:20,679 --> 00:14:22,939 但仍然保留内部AI主线 397 00:14:22,939 --> 00:14:25,299 这个安排对Google来说很现实 398 00:14:25,299 --> 00:14:28,519 既放人出去追更大的科研野心 399 00:14:28,519 --> 00:14:31,559 又不让他们完全跑到竞争对手怀里 400 00:14:31,559 --> 00:14:34,799 这也是大科技处理顶级人才的新方式 401 00:14:34,799 --> 00:14:38,019 过去公司会想尽办法把人留在内部 402 00:14:38,019 --> 00:14:39,679 现在更现实 403 00:14:39,679 --> 00:14:41,859 留不住就投资 404 00:14:41,859 --> 00:14:43,899 管不住就合作 405 00:14:43,899 --> 00:14:47,399 不适合放在内部就放到外部公司 406 00:14:47,399 --> 00:14:49,499 但保留资本和云关系 407 00:14:49,499 --> 00:14:51,999 这样既减少人才彻底流失 408 00:14:51,999 --> 00:14:54,239 也给公司留一个未来选项 409 00:14:54,239 --> 00:14:55,939 可是市场不一定买账 410 00:14:55,939 --> 00:14:58,759 因为投资人不只看有没有合作 411 00:14:58,759 --> 00:14:59,939 还看控制权 412 00:14:59,939 --> 00:15:02,639 Discovery Loop是独立公司 413 00:15:02,639 --> 00:15:03,839 不是Google部门 414 00:15:03,839 --> 00:15:05,579 Jeff Dean他们的时间 415 00:15:05,579 --> 00:15:06,359 注意力 416 00:15:06,359 --> 00:15:07,219 股权激励 417 00:15:07,219 --> 00:15:08,279 团队文化 418 00:15:08,279 --> 00:15:10,039 都不再完全服务Google 419 00:15:10,039 --> 00:15:12,219 Google是partner不是老板 420 00:15:12,219 --> 00:15:14,059 这种差别很大 421 00:15:14,059 --> 00:15:16,019 一家公司最值钱的人才 422 00:15:16,019 --> 00:15:18,119 从雇员变成合作伙伴 423 00:15:18,119 --> 00:15:19,879 听起来还在生态里 424 00:15:19,879 --> 00:15:21,679 实际控制力已经下降 425 00:15:21,679 --> 00:15:23,319 这就是2000亿美元 426 00:15:23,319 --> 00:15:24,659 这个钩子该怎么用 427 00:15:24,659 --> 00:15:27,239 不能说四个人只2000亿 428 00:15:27,239 --> 00:15:27,979 那太粗 429 00:15:27,979 --> 00:15:28,899 应该说 430 00:15:28,899 --> 00:15:31,899 市场用接近2000亿美元的市值波动 431 00:15:31,899 --> 00:15:33,699 提醒Google AI时代 432 00:15:33,699 --> 00:15:34,939 最贵的资金 433 00:15:34,939 --> 00:15:36,339 最貴的資產不只是GPU 434 00:15:36,339 --> 00:15:37,939 也不是模型參數 435 00:15:37,939 --> 00:15:40,839 而是能把科學想法變成系統的人 436 00:15:40,839 --> 00:15:43,739 Jeff Dean這批人正是這種資產 437 00:15:43,739 --> 00:15:45,739 他們不是只會做模型 438 00:15:45,739 --> 00:15:47,739 也不是只會做系統 439 00:15:47,739 --> 00:15:51,159 他们知道怎么把研究想法变成基础设施 440 00:15:51,159 --> 00:15:53,899 怎么把基础设施变成公司能力 441 00:15:53,899 --> 00:15:57,099 怎么把公司能力变成长期技术优势 442 00:15:57,099 --> 00:15:58,519 这种人一旦离开 443 00:15:58,519 --> 00:16:02,219 市场会立刻重新评估公司的未来执行力 444 00:16:02,219 --> 00:16:03,479 再往深处看 445 00:16:03,479 --> 00:16:06,819 这件事还说明AI战争进入了第三阶段 446 00:16:06,819 --> 00:16:09,139 第一阶段是模型能力战 447 00:16:09,139 --> 00:16:10,819 谁的模型更聪明 448 00:16:10,819 --> 00:16:12,539 谁的benchmark更高 449 00:16:12,539 --> 00:16:14,279 谁的发布会更震撼 450 00:16:14,279 --> 00:16:16,539 第二阶段是产品入口战 451 00:16:16,539 --> 00:16:17,839 Chad GPT 452 00:16:17,839 --> 00:16:18,819 Gemini 453 00:16:18,819 --> 00:16:19,539 Claude 454 00:16:19,539 --> 00:16:20,479 Copilot 455 00:16:20,479 --> 00:16:21,419 Cursor 456 00:16:21,419 --> 00:16:22,499 Perplexity 457 00:16:22,499 --> 00:16:24,979 大家抢用户每天打开哪个AI 458 00:16:24,979 --> 00:16:27,659 第三阶段是科研自动化展 459 00:16:27,659 --> 00:16:30,079 谁能让AI自己加速 460 00:16:30,079 --> 00:16:31,179 AI研究 461 00:16:31,179 --> 00:16:34,779 谁就能把下一轮技术进步的速度拉开 462 00:16:34,779 --> 00:16:37,519 Discovery Loop打的正是第三阶段 463 00:16:37,519 --> 00:16:39,819 它不是在做一个更好的聊天框 464 00:16:39,819 --> 00:16:40,879 而是在问 465 00:16:40,879 --> 00:16:43,339 能不能让AI自动提出假设 466 00:16:43,339 --> 00:16:44,179 写代码 467 00:16:44,179 --> 00:16:44,999 跑实验 468 00:16:44,999 --> 00:16:46,119 分析结果 469 00:16:46,119 --> 00:16:47,959 再生成下一轮实验 470 00:16:47,959 --> 00:16:50,259 如果这个循环越来越自动化 471 00:16:50,259 --> 00:16:52,679 人类研究者的角色就会变 472 00:16:52,679 --> 00:16:55,659 人类不再亲手试每一个实验 473 00:16:55,659 --> 00:16:57,119 而是设计目标 474 00:16:57,119 --> 00:16:58,279 约束方向 475 00:16:58,279 --> 00:17:00,379 判断结果和控制风险 476 00:17:00,379 --> 00:17:03,019 这会带来一个很大的商业结果 477 00:17:03,019 --> 00:17:05,839 科研速度本身会变成竞争壁垒 478 00:17:05,839 --> 00:17:08,979 过去公司拼的是谁有更多研究员 479 00:17:08,979 --> 00:17:12,019 未来可能拼谁有更快的实验系统 480 00:17:12,019 --> 00:17:15,639 一个团队如果能同时跑一万条研究路径 481 00:17:15,639 --> 00:17:17,619 自动淘汰失败方案 482 00:17:17,619 --> 00:17:19,599 快速堆出有效结果 483 00:17:19,599 --> 00:17:21,999 它的创新速度会非常恐怖 484 00:17:21,999 --> 00:17:24,099 生物医药材料 485 00:17:24,099 --> 00:17:24,979 芯片 486 00:17:24,979 --> 00:17:25,679 AI 487 00:17:25,679 --> 00:17:26,599 模型 488 00:17:26,599 --> 00:17:27,359 能源 489 00:17:27,359 --> 00:17:29,659 都可能被这种方式重写 490 00:17:29,659 --> 00:17:31,699 这也是为什么Discovery Loop 491 00:17:31,699 --> 00:17:34,459 选择Public Benefit Corporation这个形式 492 00:17:34,459 --> 00:17:37,259 它想讲的不是普通Says故事 493 00:17:37,259 --> 00:17:40,339 而是科学和工程创新基础设施 494 00:17:40,339 --> 00:17:42,039 它需要商业资本 495 00:17:42,039 --> 00:17:43,999 也需要公共利益叙事 496 00:17:43,999 --> 00:17:50,459 因为它涉及的可能是药物、能源、健康、科学、发现这些大问题 497 00:17:50,459 --> 00:17:52,159 但这条路也有风险 498 00:17:52,159 --> 00:17:55,519 自动化科研不是按一下按钮就能出现突破 499 00:17:55,519 --> 00:17:57,459 科学实验有噪音 500 00:17:57,459 --> 00:17:59,239 现实世界有成本 501 00:17:59,239 --> 00:18:00,319 数据会偏 502 00:18:00,319 --> 00:18:01,579 模型会误判 503 00:18:01,579 --> 00:18:04,239 自动化系统会找到看起来有效 504 00:18:04,239 --> 00:18:05,739 但没有意义的捷径 505 00:18:05,739 --> 00:18:07,839 机器学习研究本身还好 506 00:18:07,839 --> 00:18:10,279 因为实验环境更容易数字化 507 00:18:10,279 --> 00:18:12,839 一旦进入药物、材料 508 00:18:12,839 --> 00:18:14,339 實驗室自動化 509 00:18:14,339 --> 00:18:17,119 就會遇到物理世界的速度 510 00:18:17,119 --> 00:18:17,659 設備 511 00:18:17,659 --> 00:18:18,519 樣本 512 00:18:18,519 --> 00:18:20,479 合規和失敗成本 513 00:18:20,479 --> 00:18:23,039 所以Discovery Loop還遠不到穩穩贏 514 00:18:23,039 --> 00:18:24,999 它的優勢是人太強 515 00:18:24,999 --> 00:18:26,119 方向太大 516 00:18:26,119 --> 00:18:29,199 Google還在背後提供投資和雲資源 517 00:18:29,199 --> 00:18:31,579 它的風險是目標太難 518 00:18:31,579 --> 00:18:32,739 周期太長 519 00:18:32,739 --> 00:18:36,659 科學自動化不一定像軟件自動化那樣快速驗證 520 00:18:36,659 --> 00:18:39,019 這就讓這期的判斷更有層次 521 00:18:39,019 --> 00:18:39,979 短期看 522 00:18:39,979 --> 00:18:42,239 這是Google AI領導層政黨 523 00:18:42,239 --> 00:18:47,799 中期看这是顶级AI人才从大厂内部流向高风险创业的信号 524 00:18:47,799 --> 00:18:51,939 长期看这是AI科研自动化开始进入主战场 525 00:18:51,939 --> 00:18:53,159 对Google来说 526 00:18:53,159 --> 00:18:55,919 最坏的情况不是定他们离开后失败 527 00:18:55,919 --> 00:19:00,499 失败了Google只是损失几位人才的一段外部尝试 528 00:19:00,499 --> 00:19:03,759 真正最尴尬的情况是Discovery Loop成功 529 00:19:03,759 --> 00:19:06,359 如果它成功就会证明一件事 530 00:19:06,359 --> 00:19:08,939 Google最有想象力的科研项目 531 00:19:08,939 --> 00:19:11,379 放在Google外面反而跑得更快 532 00:19:11,379 --> 00:19:13,579 这会给所有大公司敲警钟 533 00:19:13,579 --> 00:19:14,659 Meta会问 534 00:19:14,659 --> 00:19:17,499 最顶级AI人才是不是更适合小团队 535 00:19:17,499 --> 00:19:19,179 OpenAI会问 536 00:19:19,179 --> 00:19:21,659 研究自由和产品压力怎么平衡 537 00:19:21,659 --> 00:19:23,179 Anthropic会问 538 00:19:23,179 --> 00:19:26,179 安全文化和高速商业化如何共存 539 00:19:26,179 --> 00:19:27,859 Google自己更要问 540 00:19:27,859 --> 00:19:31,619 为什么一个曾经最吸引科研人才的地方 541 00:19:31,619 --> 00:19:35,139 需要把未来实验循环放到外部公司 542 00:19:35,139 --> 00:19:38,259 这才是Jeff Dean离职创业的真正分量 543 00:19:38,259 --> 00:19:40,699 他不只是一个人的职业选择 544 00:19:40,699 --> 00:19:42,999 也不只是一次普通创业 545 00:19:42,999 --> 00:19:45,959 它是一种技术组织方式的迁移 546 00:19:45,959 --> 00:19:48,979 Google曾经用分布式系统改变互联网 547 00:19:48,979 --> 00:19:51,579 用深度学习改变AI研究 548 00:19:51,579 --> 00:19:55,019 现在这批参与建造系统的人离开Google 549 00:19:55,019 --> 00:19:57,619 要去做自动化科学发现系统 550 00:19:57,619 --> 00:20:00,459 这个故事本身就有强烈的历史反差 551 00:20:00,459 --> 00:20:04,699 过去他们帮Google建了一台全球最强的信息机器 552 00:20:04,699 --> 00:20:06,579 现在他们要在Google外面 553 00:20:06,579 --> 00:20:09,619 尝试建一台自动发现新知识的机器 554 00:20:09,619 --> 00:20:12,439 市场的恐慌正来自这个反差 555 00:20:12,439 --> 00:20:13,939 Google当然还很强 556 00:20:13,939 --> 00:20:17,099 但强公司最怕的不是今天输了 557 00:20:17,099 --> 00:20:19,679 而是最早看见下一代饭事的人 558 00:20:19,679 --> 00:20:22,379 不再把下一代放在公司内部做 559 00:20:22,379 --> 00:20:25,159 这期最后可以给一个明确判断 560 00:20:25,159 --> 00:20:26,999 Jeff Dean离职创业 561 00:20:26,999 --> 00:20:29,439 不会立刻改变Google的收入 562 00:20:29,439 --> 00:20:31,979 也不会马上决定Gemini输赢 563 00:20:31,979 --> 00:20:35,139 但他会改变市场看Google AI的方式 564 00:20:35,139 --> 00:20:39,699 以前投资人默认Google有最深的AI人才词 565 00:20:39,699 --> 00:20:41,539 最强的科研传统 566 00:20:41,539 --> 00:20:42,959 最好的基础设施 567 00:20:42,959 --> 00:20:45,739 现在这个默认值被打了一个折扣 568 00:20:45,739 --> 00:20:47,239 Google仍然有资源 569 00:20:47,239 --> 00:20:48,839 但资源不等于速度 570 00:20:48,839 --> 00:20:50,599 Google仍然有天才 571 00:20:50,599 --> 00:20:52,579 但天才开始有外部选项 572 00:20:52,579 --> 00:20:54,719 Google仍然能投资未来 573 00:20:54,719 --> 00:20:57,079 但未来不一定完全长在Google里面 574 00:20:57,079 --> 00:20:59,319 这就是AI时代最残酷的地方 575 00:20:59,319 --> 00:21:02,319 过去科技巨头靠平台所用户 576 00:21:02,319 --> 00:21:04,059 靠薪酬所人才 577 00:21:04,059 --> 00:21:05,799 靠算力所研究 578 00:21:05,799 --> 00:21:08,999 现在最顶级的人才可以带着声望 579 00:21:08,999 --> 00:21:12,259 资本 云合作和一整个方向走出去 580 00:21:12,259 --> 00:21:14,879 巨头不一定拦得住 只能跟头 581 00:21:14,879 --> 00:21:17,099 所以这件事不是Google的终局 582 00:21:17,099 --> 00:21:18,719 更像Google的警报 583 00:21:18,719 --> 00:21:21,719 AI战争已经不只是模型参数战 584 00:21:21,719 --> 00:21:23,639 也不只是产品入口战 585 00:21:23,639 --> 00:21:28,379 它正在变成人才 组织速度和科研自动化的战争 586 00:21:28,379 --> 00:21:30,659 Jeff Dean这批人离开Google 587 00:21:30,659 --> 00:21:33,139 给市场看的不是一次离职 588 00:21:33,139 --> 00:21:35,579 而是一张新战场的地图 589 00:21:35,579 --> 00:21:37,499 谁能把实验循环自动化 590 00:21:37,499 --> 00:21:40,939 谁就可能掌握下一代AI进步的加速度 591 00:21:40,939 --> 00:21:43,539 Google这次没有完全失去它们 592 00:21:43,539 --> 00:21:46,319 但它也不再完全拥有它们 593 00:21:46,319 --> 00:21:47,519 这中间的差别 594 00:21:47,519 --> 00:21:50,279 就是那一天市场最害怕的东西