8月5日，硅谷地震。
Jeff Dean，Google第30号员工、在公司待了整整27年的传奇工程师、MapReduce和TensorFlow的缔造者、Google AI帝国的总设计师，正式宣布离职。他不是一个人走的——和他一起离开的还有三位顶级科学家：Sanjay Ghemawat（Google Senior Fellow，Jeff Dean二十多年的黄金搭档）、Quoc Le（Google Brain创始成员，Seq2Seq和AutoML的发明者）、Oriol Vinyals（DeepMind高级研究科学家，AlphaStar和Gemini多模态核心负责人）。
四个人加起来，在Google的工作年限超过80年。他们参与或主导了Google过去二十年几乎所有最重要的基础设施和AI项目：从BigTable、Spanner到TensorFlow，从Transformer之前的分布式训练框架到Gemini多模态，从AlphaGo到AlphaFold。你可以说，这四个人就是Google AI的半壁江山。
他们的新公司叫Discovery Loop，注册为公益公司（Public Benefit Corporation），专注方向是AI for Science——用AI自动化科学发现。Jeff Dean亲自担任CEO。投资方阵容豪华到离谱：Radical Ventures和Khosla Ventures联合领投，Kleiner Perkins、Lightspeed、Doerr Capital跟投，甚至连老东家Alphabet都参投了。
消息公布当天，Alphabet股价盘中下跌约4-5%，市值蒸发数百亿美元。而就在同一时间，Google还宣布了另一则人事变动：Demis Hassabis卸任Google DeepMind CEO，转任Google DeepMind董事长兼Alphabet首席科学家，继续领导Isomorphic Labs。DeepMind CEO一职由Koray Kavukcuoglu接任。
一天之内，Google AI的两大支柱——Google Research和DeepMind——同时换帅。一个时代结束了。
Jeff和Sanjay：Google工程文化的活化石
要理解Jeff Dean离职的分量，你得先理解Jeff和Sanjay这对组合在Google意味着什么。
1999年，Jeff Dean加入Google，是公司第30号员工。那时候Google还在一个车库里办公，拉里·佩奇和谢尔盖·布林还在斯坦福读博没毕业。几年后Sanjay Ghemawat加入，两人开始了长达二十多年的传奇搭档。在Google内部，"Jeff和Sanjay"是一个专有名词——他们共用一个办公室（直到远程办公时代），代码提交记录经常分不清是谁写的，甚至有人开玩笑说他们是一个人用两个账号。
这对搭档在2000年代初做的一系列工作，定义了整个互联网时代的基础设施：MapReduce让大规模分布式数据处理变得简单，BigTable开创了NoSQL数据库的先河，Spanner实现了全球一致性分布式数据库，TensorFlow把深度学习从学术圈带到了工业界。这些项目每一个单拎出来，都足以让一个工程师名垂青史。而Jeff和Sanjay两个人，一起做了四五个。
在Google工程师文化里，Jeff Dean是神一样的存在。内部流传着无数关于他的段子："Jeff Dean的代码编译时不会警告，因为编译器不敢警告他""Jeff Dean曾经在O(1)时间内解决了NP完全问题""有一次Jeff Dean的键盘坏了，他用二进制0和1直接写了一个操作系统"。这些段子当然是夸张的，但背后是整个公司对他技术能力的绝对崇拜。
Quoc Le是另一个传奇。2011年他从斯坦福加入Google，是Google Brain创始成员之一。他和Ilya Sutskever、Alex Krizhevsky等人一起师从Andrew Ng，是深度学习黄金一代的代表人物。Seq2Seq模型是他做的，这直接催生了神经机器翻译；AutoML是他推动的，开启了用AI设计AI的方向；他在Google内部带过的团队和项目，几乎覆盖了Google AI的半壁江山。
Oriol Vinyals来自DeepMind体系。他是AlphaStar（《星际争霸》AI）的第一作者，是Gemini多模态能力的核心负责人之一，也是序列建模和强化学习领域的顶级科学家。他的离开意味着DeepMind和Google Research之间的人才流动通道，这次变成了向外的单向出口。
四个人同时离开，这在Google历史上是前所未有的。哪怕是2023年Ilya Sutskever在OpenAI政变风波后离职创办Safe Superintelligence，也只带走了几个人。Jeff Dean这次是直接把Google AI最核心的一个小圈子整体端走了。
为什么走？大公司的AI创新困境
表面上看，官方说法很体面：Jeff Dean想专注于AI for Science这个他个人 passion 所在的方向，大公司的组织结构不适合做这种高风险长周期的探索；Sundar Pichai发公开信感谢他27年的贡献，Alphabet甚至参与投资，一副"好聚好散"的画面。
"当你在Google内部做一个新项目，需要12个VP审批、6个季度对齐OKR、3轮法务审查的时候，你就知道为什么顶级研究者要出去创业了。"—— 一位前Google AI研究员
但了解Google内部情况的人都知道，这不是一次和平分手那么简单。过去两年，Google AI一直处于一种尴尬的防守姿态：OpenAI的ChatGPT横空出世打了Google一个措手不及，Gemini发布后口碑起起伏伏，搜索产品被AI重构的过程中内部矛盾重重，越来越多的顶级研究者选择离开。
大公司AI创新的结构性困境，在Google身上体现得淋漓尽致。
第一，组织惯性和决策缓慢。在Google做一个新项目，从立项到资源到位，往往需要半年甚至更长时间。你需要对齐无数个VP的OKR，需要和其他部门抢资源，需要应对法务、隐私、安全等各种部门的审查。而创业公司可以在一周内决定方向、两周内开始动手、一个月内出第一个版本。对于Jeff Dean这种级别的人来说，时间比钱宝贵得多——他已经50多岁了，财务自由早就实现了，剩下的职业生涯想做真正有影响力的事，大公司的组织效率显然满足不了他。
第二，产品化压力挤压基础研究。Gemini发布后，Google Research和DeepMind的资源明显向产品端倾斜。搜索、广告、云、Workspace，每一个产品线都在喊着要AI能力，研究者们被拉去做各种产品集成和优化，能沉下心做长周期基础研究的空间越来越小。Jeff Dean一直想大力投入AI for Science，但在产品优先级永远高于研究的大公司环境里，这个方向始终得不到足够的资源和自主权。
第三，薪酬和激励机制的落后。OpenAI给顶级研究者开出的年薪包裹动辄数千万美元甚至上亿美元，Anthropic、xAI、各种AI创业公司都在用天价包裹挖人。Google的薪酬体系虽然在大厂里算顶级的，但和这些AI新贵相比还是差了一个量级。更重要的是，在大公司你能拿到的回报是线性的——你做出了TensorFlow这样改变世界的项目，拿到的奖金和股票也不过是包裹里的几个百分点增长；而在创业公司，你有可能拿到1%甚至10%的股权，项目成功了就是十亿甚至百亿美元级别的回报。
第四，DeepMind和Google Research的内耗。2014年收购DeepMind后，Google内部就一直存在Google Brain（后来的Google Research）和DeepMind两套体系的竞争与内耗。两边研究方向重叠、资源争夺、文化冲突，直到2023年才在Alphabet层面合并为Google DeepMind，但整合效果并不理想。Jeff Dean作为Google Research的负责人，和Demis Hassabis作为DeepMind的负责人，两人之间的权力平衡一直很微妙。这次两人同时"卸任"，一个出去创业，一个升任董事长+首席科学家，某种程度上也是这种内部博弈的最终结果。
Discovery Loop的赌注：AI for Science是下一个金矿吗？
Jeff Dean选择AI for Science作为创业方向，这个选择本身就很值得玩味。
过去两年，AI大模型的商业落地主要集中在内容生成、聊天机器人、编程助手、企业服务这些方向。但越来越多的人开始意识到，AI真正能产生指数级价值的领域可能是科学发现——药物研发、材料科学、气候建模、核聚变、基因组学。这些领域每一个突破都可能带来万亿美元级别的市场，而传统科学方法已经到了瓶颈期，AI是破局的唯一希望。
Isomorphic Labs（Demis Hassabis领导的DeepMind分拆公司）已经在AI药物研发上取得了不少进展；AlphaFold解决了蛋白质折叠问题，这本身就是AI for Science的里程碑；OpenAI也成立了专门的科学团队；各大制药公司都在疯狂布局AI药物研发。但目前还没有一家公司把"AI驱动科学发现全流程自动化"作为核心使命，Discovery Loop瞄准的就是这个空白。
Jeff Dean的优势在哪里？第一，他是系统级工程师出身，不是纯算法研究者——他懂得怎么构建大规模分布式系统，怎么把AI模型和真实的科学实验流程结合起来，而不仅仅是在benchmark上刷分数。第二，他在Google积累了二十年的基础设施经验，知道怎么搭建能支撑千名科学家、亿级实验的平台。第三，他的个人品牌和人脉能吸引到全世界最顶尖的科学家和工程师加入。
当然，这个方向的风险也极大。科学发现不像做聊天机器人，周期极长——一个新药研发平均需要10年以上，一个新材料从发现到产业化可能需要15-20年。Discovery Loop作为一家创业公司，能不能扛过这么长的周期，能不能在烧完融资前拿出真正有商业价值的成果，都是未知数。
但从投资方的阵容来看，资本已经在用脚投票了。Radical Ventures是AI领域最顶级的VC之一，投过DeepMind、OpenAI（早期）、Anthropic等几乎所有重要AI公司；Khosla Ventures是Vinod Khosla创办的，以押注高风险高回报的颠覆性技术著称；Alphabet的跟投更是耐人寻味——相当于Google自己也承认：这个方向在Google内部做不成，只能靠独立公司去做。
Hassabis交权：DeepMind的后创始人时代
和Jeff Dean离职同一天宣布的Demis Hassabis卸任消息，同样值得关注。
很多中文媒体把这件事解读为"Hassabis被架空""DeepMind创始人被赶走"，这是不准确的。官方公告里写得很清楚：Hassabis升任Google DeepMind董事长，同时担任Alphabet首席科学家，继续全职领导Isomorphic Labs。Koray Kavukcuoglu接任DeepMind CEO，他是DeepMind的创始成员之一，在DeepMind工作了超过15年，是内部培养起来的接班人。
这更像是一次正常的权力交接和职责调整。Hassabis今年已经50岁了，2010年创办DeepMind至今已经16年，AlphaGo、AlphaFold、AlphaStar这些标志性成果都是他领导下做出来的，但DeepMind被收购后整合进Google的这几年，他的精力更多放在了Alphabet层面的AI战略和Isomorphic Labs上，日常管理工作确实需要有人接手。
但这个时间点和Jeff Dean离职撞在一起，还是难免让人产生联想。一天之内，Google AI两大体系的一号人物同时离开日常管理岗位——一个彻底出走，一个退居二线——这标志着Google AI的"创始人时代"彻底结束了。接下来的Google DeepMind，将进入职业经理人主导的时代。
这对Google AI来说是好事还是坏事？短期来看肯定有阵痛。Jeff Dean和Hassabis这种级别的人物，他们的技术判断力、个人魅力和对人才的吸引力是不可替代的。但长期来看，任何大公司都要经历创始人交棒的过程，关键是接班人能不能延续原来的文化和创新能力。Koray是DeepMind老人，技术背景过硬，内部威望也够，但他能不能在大公司的产品压力下保住DeepMind的研究基因，这要打个问号。
人才地震的连锁反应：大模型时代的人才流动新常态
Jeff Dean的离开不是孤立事件，它是过去两年AI行业人才大流动的一个缩影，也是一个新的高潮。
2023年以来，我们已经见证了太多顶级AI人才的出走：Ilya Sutskever离开OpenAI创办SSI；Dario Amodei虽然还在Anthropic但已经在考虑接班人问题；一大批OpenAI、Google、Meta的研究员出去创办自己的AI公司；甚至连高校里的教授都坐不住了，纷纷停薪留职下海。AI行业正在重演半导体行业、互联网行业曾经发生过的故事：当一个新技术浪潮到来时，最顶尖的人才会从大公司流向创业公司，因为那里有更大的自由度、更高的回报、更快的迭代速度。
对于大公司来说，这是一个无解的难题。你可以给员工发更高的工资、更多的股票，但你给不了他们时间自主权、决策速度和改变世界的机会感。当一个顶级人才已经财务自由之后，钱就不再是最重要的激励因素了——他们想要的是：我想做的项目能不能立刻开始？我做的决定能不能不被12个VP否决？我做成了一件事能不能拿到足够的credit和回报？这些恰恰是大公司最难提供的东西。
当然，大公司不会因此衰落。Google有最雄厚的算力资源、最多的数据、最成熟的商业化渠道，哪怕走了几个顶级人才，它依然是AI竞赛的第一梯队玩家。但人才流失的长期影响是潜移默化的：当最有创造力、最敢想敢干的那批人都走了，大公司的创新能力会慢慢钝化，最终变成一个守成的、优化效率的机构，而不是开拓新边疆的先锋。
对创业者和投资人来说，Jeff Dean的出走是一个强烈的信号：AI for Science这个赛道要热起来了。未来一两年，我们会看到越来越多的AI创业公司瞄准科学发现方向，资本会疯狂涌入，这个领域会重复过去两年大模型和AI编程赛道的剧本。
对于Google来说，这是一个时代的结束。那个在车库里创业、工程师文化至上、敢于做十年磨一剑项目的Google，正在慢慢变成一个成熟的、庞大的、商业驱动的科技巨头。这是所有伟大公司成长的必经之路，但还是难免让人感慨。
Jeff Dean在Google的27年，几乎就是一部Google工程和AI的发展史。他走的时候，没有愤怒，没有撕逼，甚至老东家还给他的新公司投了钱。但所有人都知道：有些东西已经永远改变了。
明天见。
August 5. Silicon Valley had an earthquake.
Jeff Dean — Google employee #30, the legendary engineer who spent 27 years at the company, architect of MapReduce and TensorFlow, the chief designer of Google's AI empire — officially announced his departure. And he didn't leave alone. Walking out with him were three top scientists: Sanjay Ghemawat (Google Senior Fellow, Jeff Dean's legendary partner of over two decades), Quoc Le (Google Brain founding member, inventor of Seq2Seq and AutoML), and Oriol Vinyals (DeepMind senior research scientist, lead on AlphaStar and Gemini multimodal).
Together, these four people represent over 80 years of combined tenure at Google. They participated in or led virtually every major infrastructure and AI project at Google over the past two decades: from BigTable and Spanner to TensorFlow, from pre-Transformer distributed training frameworks to Gemini multimodal, from AlphaGo to AlphaFold. You could say these four people are half of Google's AI brain trust.
Their new company is called Discovery Loop, incorporated as a Public Benefit Corporation, focused on AI for Science — using AI to automate scientific discovery. Jeff Dean himself will serve as CEO. The investor lineup is absurdly stacked: Radical Ventures and Khosla Ventures co-leading, with Kleiner Perkins, Lightspeed, Doerr Capital participating, and even Alphabet, their former employer, chipping in.
On the day of the announcement, Alphabet's stock dropped roughly 4–5% intraday, wiping out tens of billions in market cap. And at the same time, Google announced another personnel change: Demis Hassabis is stepping down as CEO of Google DeepMind, becoming chairman of Google DeepMind and Alphabet Chief Scientist while continuing to lead Isomorphic Labs. Koray Kavukcuoglu takes over as DeepMind CEO.
In a single day, the two pillars of Google AI — Google Research and DeepMind — changed leadership simultaneously. An era has ended.
Jeff and Sanjay: Living Fossils of Google's Engineering Culture
To understand the weight of Jeff Dean's departure, you first have to understand what Jeff and Sanjay meant at Google.
Jeff Dean joined Google in 1999 as employee #30. Back then Google was operating out of a garage; Larry Page and Sergey Brin were still PhD students at Stanford and hadn't even graduated yet. A few years later Sanjay Ghemawat joined, and the pair began a legendary partnership that spanned over two decades. Inside Google, "Jeff and Sanjay" is a proper noun — they shared an office (until the remote work era), code commits were often indistinguishable between the two, and there were jokes that they were one person using two accounts.
A series of projects this pair built in the early 2000s defined the infrastructure of the entire internet era: MapReduce made large-scale distributed data processing simple; BigTable pioneered the NoSQL database movement; Spanner achieved globally consistent distributed databases; TensorFlow brought deep learning from academia to industry. Any one of these projects alone would be enough to cement an engineer's place in history. Jeff and Sanjay, together, built four or five.
In Google's engineering culture, Jeff Dean is a deity. Countless internal jokes circulate about him: "Jeff Dean's code doesn't generate warnings because compilers don't dare warn him"; "Jeff Dean once solved an NP-complete problem in O(1) time"; "When Jeff Dean's keyboard broke once, he wrote an operating system directly in binary 0s and 1s." These jokes are obviously exaggerated, but beneath them is the entire company's absolute reverence for his technical ability.
Quoc Le is another legend. He joined Google from Stanford in 2011 as one of the founding members of Google Brain. Alongside Ilya Sutskever, Alex Krizhevsky, and others, he studied under Andrew Ng and is a representative figure of deep learning's golden generation. Seq2Seq was his work, which directly spawned neural machine translation; he drove AutoML, opening the direction of AI designing AI; the teams and projects he's led inside Google cover virtually half of Google AI.
Oriol Vinyals comes from the DeepMind side. He was first author on AlphaStar (the StarCraft AI), one of the core leads on Gemini's multimodal capabilities, and a top scientist in sequence modeling and reinforcement learning. His departure means the talent pipeline between DeepMind and Google Research has become a one-way exit outward.
Four people leaving simultaneously is unprecedented in Google's history. Even when Ilya Sutskever left OpenAI after the 2023 boardroom coup to found Safe Superintelligence, he only took a handful of people. Jeff Dean essentially took Google AI's most core inner circle and walked it out the door.
Why Leave? Big Tech's AI Innovation Dilemma
On the surface, the official narrative is very dignified: Jeff Dean wants to focus on AI for Science, his personal passion, and large-company organizational structures aren't suited to this kind of high-risk, long-cycle exploration; Sundar Pichai published an open letter thanking him for 27 years of contributions, Alphabet even participated in the investment — a picture of "conscious uncoupling."
"When launching a new project inside Google requires 12 VP approvals, six quarters of OKR alignment, and three rounds of legal review, you understand why top researchers go start companies."— A former Google AI researcher
But anyone familiar with Google's internal situation knows this isn't just an amicable parting. Over the past two years, Google AI has been in an awkward defensive posture: OpenAI's ChatGPT caught Google flat-footed, Gemini's reputation has been rocky post-launch, the AI restructuring of search has been rife with internal friction, and more and more top researchers are choosing to leave.
The structural dilemma of AI innovation at big companies is on full display at Google.
First, organizational inertia and slow decision-making. Getting a new project off the ground at Google, from approval to resource allocation, often takes six months or more. You need to align with countless VPs' OKRs, compete for resources with other departments, and navigate reviews from legal, privacy, security, and various other departments. A startup can decide on a direction in a week, start building in two weeks, and have a first version in a month. For someone at Jeff Dean's level, time is far more valuable than money — he's in his 50s, financially independent long ago, and for the rest of his career he wants to work on things that truly matter. Big-company organizational efficiency clearly can't satisfy that.
Second, productization pressure squeezes basic research. After Gemini launched, resources at Google Research and DeepMind clearly tilted toward product. Search, ads, cloud, Workspace — every product line is screaming for AI capabilities, and researchers are pulled into product integrations and optimizations, leaving less and less space to focus on long-cycle basic research. Jeff Dean has long wanted to invest heavily in AI for Science, but in a big-company environment where product priorities always trump research, this direction never got sufficient resources or autonomy.
Third, outdated compensation and incentive structures. OpenAI offers top researchers annual packages in the tens of millions or even hundreds of millions of dollars; Anthropic, xAI, and various AI startups are all poaching with eye-popping packages. Google's compensation is top-tier among big tech, but it's still an order of magnitude behind these AI upstarts. More importantly, returns at big companies are linear — you build a world-changing project like TensorFlow and your bonus and stock go up a few percentage points; at a startup, you could get 1% or even 10% equity, and success means billions or tens of billions in returns.
Fourth, internal friction between DeepMind and Google Research. Since acquiring DeepMind in 2014, Google has always had two competing systems — Google Brain (later Google Research) and DeepMind — with overlapping research directions, resource fights, and cultural clashes. They were only merged into Google DeepMind at the Alphabet level in 2023, and integration has been less than smooth. As head of Google Research, Jeff Dean and Demis Hassabis as head of DeepMind maintained a delicate balance of power. Both "stepping down" from day-to-day management on the same day — one leaving entirely, the other moving to chairman + chief scientist — is to some extent the final outcome of this internal jockeying.
Discovery Loop's Bet: Is AI for Science the Next Gold Mine?
That Jeff Dean chose AI for Science as his startup direction is itself worth examining.
Over the past two years, commercial AI deployment has focused mainly on content generation, chatbots, coding assistants, and enterprise services. But more and more people are realizing that the area where AI can truly deliver exponential value may be scientific discovery — drug development, materials science, climate modeling, nuclear fusion, genomics. Every breakthrough in these fields could create trillion-dollar markets, and traditional scientific methods have hit bottlenecks; AI is the only hope for breaking through.
Isomorphic Labs (led by Demis Hassabis as a DeepMind spinout) has already made progress in AI drug discovery; AlphaFold solving protein folding is itself an AI for Science milestone; OpenAI has formed a dedicated science team; every major pharmaceutical company is furiously investing in AI drug R&D. But no company has yet made "full-process automation of AI-driven scientific discovery" its core mission, and Discovery Loop is targeting precisely this gap.
Where does Jeff Dean have an edge? First, he's a systems engineer by background, not a pure algorithm researcher — he knows how to build large-scale distributed systems, how to integrate AI models with real scientific experiment workflows, not just chase benchmark scores. Second, he has twenty years of infrastructure experience at Google and knows how to build platforms that can support a thousand scientists and a hundred million experiments. Third, his personal brand and network can attract the world's top scientists and engineers.
Of course, the risks in this direction are enormous. Scientific discovery isn't like building chatbots; cycles are extremely long — a new drug takes 10+ years on average; a new material from discovery to industrialization could take 15–20 years. Whether Discovery Loop as a startup can survive such long cycles and produce genuinely commercially valuable results before burning through funding is an open question.
But judging from the investor lineup, capital is already voting with its feet. Radical Ventures is one of the top VCs in AI, having backed DeepMind, OpenAI (early), Anthropic, and virtually every major AI company; Khosla Ventures, founded by Vinod Khosla, is known for betting on high-risk, high-reward disruptive technologies; Alphabet's participation is particularly telling — it's essentially Google admitting this direction can't be done inside Google; it can only be done by an independent company.
Hassabis Steps Back: DeepMind's Post-Founder Era
Demis Hassabis stepping down as DeepMind CEO, announced the same day as Jeff Dean's departure, is equally noteworthy.
Many Chinese media outlets framed this as "Hassabis being sidelined" or "the DeepMind founder being pushed out," which is inaccurate. The official announcement is clear: Hassabis is promoted to Chairman of Google DeepMind while serving as Alphabet Chief Scientist and continuing to lead Isomorphic Labs full-time. Koray Kavukcuoglu takes over as DeepMind CEO; he's one of DeepMind's founding members, has been at DeepMind for over 15 years, and is a successor groomed from within.
This looks more like a normal leadership transition and responsibility realignment. Hassabis is 50 this year; he founded DeepMind in 2010, 16 years ago. AlphaGo, AlphaFold, AlphaStar — all the landmark achievements happened under his leadership. But in the years since DeepMind was acquired and integrated into Google, his focus has shifted more toward Alphabet-level AI strategy and Isomorphic Labs, and day-to-day management genuinely needed someone to take over.
Still, the timing colliding with Jeff Dean's departure inevitably invites speculation. In a single day, the heads of both major Google AI systems left day-to-day management — one departing entirely, one moving to a chairman role — marking the definitive end of Google AI's "founder era." What comes next for Google DeepMind is an era led by professional managers.
Is this good or bad for Google AI? There will definitely be short-term growing pains. People at Jeff Dean and Hassabis's level have irreplaceable technical judgment, personal charisma, and ability to attract talent. But in the long run, every great company goes through founder succession. The question is whether successors can carry forward the original culture and innovative capacity. Koray is a DeepMind veteran with strong technical credentials and sufficient internal prestige, but whether he can preserve DeepMind's research DNA under big-company product pressure remains an open question.
The Talent Earthquake's Ripple Effects: The New Normal of Talent Flow in the LLM Era
Jeff Dean's departure isn't an isolated event; it's a microcosm — and a new climax — of the massive AI talent migration over the past two years.
Since 2023, we've already seen too many top AI talents depart: Ilya Sutskever left OpenAI to found SSI; Dario Amodei is still at Anthropic but already considering succession; a flood of researchers from OpenAI, Google, and Meta have left to found their own AI companies; even university professors can't sit still, taking leaves of absence to jump into the fray. The AI industry is repeating what happened in semiconductors and the internet before it: when a new technology wave arrives, the most talented people flow from big companies to startups, because that's where there's more autonomy, higher returns, and faster iteration.
For big companies, this is an unsolvable problem. You can pay employees higher salaries and more stock, but you can't give them autonomy over their time, decision speed, or the sense of opportunity to change the world. Once a top talent is financially independent, money is no longer the primary motivator — what they want is: Can the project I want to work on start immediately? Can the decisions I make not be vetoed by 12 VPs? When I build something, can I get sufficient credit and reward? These are precisely the things big companies struggle most to provide.
Of course, big companies won't decline because of this. Google has the most formidable compute resources, the most data, the most mature commercialization channels. Even losing a few top people, it remains a first-tier player in the AI race. But the long-term effects of brain drain are subtle: when the most creative, boldest people leave, big companies' innovative capacity gradually atrophies, eventually becoming institutions that maintain and optimize rather than pioneer new frontiers.
For founders and investors, Jeff Dean's exit is a strong signal: the AI for Science sector is about to heat up. Over the next year or two, we'll see more and more AI startups targeting scientific discovery, capital will flood in, and this field will repeat the script of the foundation model and AI coding sectors over the past two years.
For Google, this is the end of an era. The Google that started in a garage, that put engineering culture above all, that dared to do decade-in-the-making projects, is slowly becoming a mature, massive, commercially driven tech giant. This is the inevitable path of all great companies, but it's hard not to feel a sense of loss.
Jeff Dean's 27 years at Google are practically a history of Google engineering and AI. He left without anger, without public drama, and his old employer even invested in his new company. But everyone knows: something has changed forever.
See you tomorrow.
Jeff Dean · Google · Discovery Loop · Sanjay Ghemawat · Quoc Le · Oriol Vinyals · Demis Hassabis · DeepMind · Koray Kavukcuoglu · AI for Science · 人才流失 · Alphabet股价 · 大公司创新困境
Jeff Dean · Google · Discovery Loop · Sanjay Ghemawat · Quoc Le · Oriol Vinyals · Demis Hassabis · DeepMind · Koray Kavukcuoglu · AI for Science · brain drain · Alphabet stock · big tech innovation dilemma
