[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-deepmind-dissolves-alphafold-team-anthropic-pivot":3},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"0ed9b02e-d74c-4df1-b87f-37178d9bb8c6","DeepMind 解散 AlphaFold 团队:三位核心成员跳槽 Anthropic,折射 LLM 时代资源虹吸","Google DeepMind 解散曾获诺贝尔奖的 AlphaFold 项目团队,核心成员被调往 Gemini LLM 项目,John Jumper 等三人跳槽 Anthropic,近四分之一论文作者已离开公司,折射 LLM 时代研究资源向大模型集中。","## 一夜变天:AlphaFold 团队被解散\n\n2026 年 7 月 29 日,英国《金融时报》披露了一条让整个 AI4Science 圈震动的人事消息:Google DeepMind 已经正式解散了 AlphaFold 项目团队。AlphaFold 是 DeepMind 在 2020-2021 年推出的蛋白质结构预测系统,凭借 AlphaFold 2 拿下 Nature 头篇并直接催生 2024 年诺贝尔化学奖,曾是 DeepMind「AI for Science」路线最具标志性的招牌成果。\n\n按照 FT 报道,过去一年里,AlphaFold 论文的多数原作者被陆续调岗。DeepMind 官方证实,这些员工已在公司内部转岗,主要流向三类方向:\n\n- 围绕 **Gemini 大型语言模型**的开发项目\n- **酶设计、核聚变、基因组学**等 DeepMind 仍在投入的科研方向\n- Alphabet 旗下药物研发子公司 **Isomorphic Labs**\n\n更令人关注的是人员的外部流失:**John Jumper(AlphaFold 2 核心架构师)、Jonas Adler、Alexander Pritzel** 三名核心成员跳槽去了 Anthropic——也就是 OpenAI 在前沿大模型领域的最强对手之一。加上其余作者,据 FT 统计,近四分之一的 AlphaFold 论文原作者已经离开 DeepMind。\n\n## LLM 时代正在虹吸一切\n\n把这件事单独看作一次「团队重组」就太低估它的信号意义了。DeepMind 当年是靠 AlphaFold、AlphaGo 这一类「科学探索型 AI」奠定江湖地位的,但 2023 年以后,整个公司的资源重心明显向 Gemini 大模型倾斜。这次直接把 AlphaFold 团队「收编」进 Gemini 项目线,等于把公司级研究战略的优先级彻底表态:\n\n- **科学应用类项目退居二线**——AlphaFold 后续的蛋白质-配体预测、复合物结构建模、多构象预测等路线,失去了核心团队支撑,大概率要靠 Isomorphic Labs 接力,但 Isomorphic 本身的定位是商业化药物研发,不是基础研究输出。\n- **前沿大模型路线集中火力**——把已经训练好的顶尖人才直接塞进 Gemini,既省去外招的时间成本,也意味着 DeepMind 认为在 GPT\u002FClaude\u002FGemini 的下一代对决中,人力是最稀缺的资源。\n- **人员外流不可避免**——OpenAI 研究主管 Mark Chen 之前公开说过一句大实话:「AI 研究人员希望在前沿实验室工作,而不是疲于追赶。」当一个诺奖级项目的核心成员都觉得「被边缘化」时,自然会用脚投票。\n\n## 对 AI4Science 意味着什么\n\n短期来看,AlphaFold 本身的开源权重、推理代码、数据库(AlphaFold DB 已有超过 2 亿条结构)不会立刻受影响,学术界继续基于现有产物做研究问题不大。但**中长期,「科学发现型 AI」这条路线在头部实验室的资源位次正在明显下移**:\n\n1. **应用科学项目的科研产出节奏会变慢**——没有专门团队维护迭代,AlphaFold 后续版本(AlphaFold 3 以及更复杂的多模态生物分子预测)的更新几乎必然会延后。\n2. **Anthropic 在 AI4Science 上的布局值得关注**——一次挖走三个核心成员,加上此前从 OpenAI、DeepMind 招募的一批 RL\u002FAgent 背景人才,Anthropic 显然想在「前沿模型 + 科学应用」交叉点上建立自己的差异化优势,而不仅仅是聊天和编程。\n3. **开源社区会承接部分工作**——DeepMind 历来愿意把 AlphaFold 系列权重开放,但没有原班人马的训练-迭代-问题修复闭环,开源生态的演化将更依赖第三方团队(如 Baker Lab、Chai Discovery、Iambic、英矽智能等),社区驱动的权重微调会成为常态。\n\n## 一点个人判断\n\n这次事件对普通读者最直观的启示是:**AI 行业的「中心化」比我们想象的还要快**。当一家公司把全部精锐收拢到一条战线,开源社区就成了事实上的「外援」——这意味着 AI4Science 这块原本由 DeepMind 独挑大梁的领域,接下来两年很可能会出现「开源替代方案密集涌现」的窗口期。\n\n对从业者来说,**关注 An­thropic 接下来会不会发 AlphaFold 相关职位、研究成果或开源模型**,大概率能看清前沿大模型公司和 AI4Science 之间的下一步走向。LLM 时代真正赢的,不一定是最先做出诺奖成果的人,而是能在两条战线之间灵活调度资源的人。\n\n---\n\n参考资料:\n- [DeepMind 解散 AlphaFold 项目团队 — 奇客 Solidot](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=84953)\n- Financial Times 原文(订阅墙内)","https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=84953","d59894d3-308e-4fd8-8865-86dc1eeac4a2",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"23544f6a-eea1-4f05-aa8d-749ca862d5d2","anthropic",{"id":19,"name":20,"slug":20,"description":14,"color":14},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"d3eb0552-a6e1-423f-9fab-f81d849c4d4c","en","DeepMind Disbands AlphaFold Team: Three Core Members Jump to Anthropic, Exposing the LLM-Era Resource Siphon","Google DeepMind has formally disbanded its Nobel-winning AlphaFold team. Core members were reassigned to the Gemini LLM program, John Jumper and two others defected to Anthropic, and nearly a quarter of original AlphaFold authors have left the company — a clear signal of research resources being siphoned into the LLM era.","## Overnight Pivot: The AlphaFold Team Disbanded\n\nOn July 29, 2026, the Financial Times dropped a personnel move that sent shockwaves through the AI4Science community: Google DeepMind has formally disbanded the AlphaFold project team. AlphaFold, the protein-structure-prediction system DeepMind unveiled in 2020-2021, put AlphaFold 2 on the cover of *Nature* and effectively seeded the 2024 Nobel Prize in Chemistry. It was once DeepMind's most iconic proof point for an \"AI for Science\" research agenda.\n\nAccording to FT's reporting, over the past year the majority of original AlphaFold authors have been quietly reassigned. DeepMind confirmed the moves, listing three destinations:\n\n- The **Gemini large language model** development program\n- Ongoing DeepMind research in **enzyme design, fusion, and genomics**\n- **Isomorphic Labs**, Alphabet's drug-discovery subsidiary\n\nBut the externally visible departures are what drew attention: **John Jumper (the architect of AlphaFold 2's core pipeline), Jonas Adler, and Alexander Pritzel** all jumped to Anthropic — OpenAI's strongest rival in the frontier-model race. Add in the rest, and FT estimates nearly **a quarter of original AlphaFold authors** have already left DeepMind.\n\n## The LLM Era Is Siphoning Everything\n\nReading this purely as a \"team reshuffle\" understates the signal. DeepMind built its reputation on AlphaFold and AlphaGo — scientific-discovery AI. But since 2023, the company's resource gravity has visibly tilted toward Gemini. Reabsorbing the AlphaFold team into Gemini's project lines is the cleanest possible statement of priorities:\n\n- **Science-application work moves to the second tier.** With the core team gone, follow-ups such as AlphaFold 3, multi-conformation modeling, and protein-ligand complexes will likely pass to Isomorphic Labs — whose mandate is commercial drug discovery, not open scientific output.\n- **Frontier-LLM work concentrates firepower.** Re-tasking already-trained top researchers onto Gemini avoids the cost of external recruiting and signals DeepMind believes human capital is the scarcest input in the next GPT\u002FClaude\u002FGemini generation.\n- **External attrition is now structural.** OpenAI's Mark Chen put it bluntly earlier: \"AI researchers want to work at frontier labs, not chase from behind.\" When even Nobel-tier contributors feel sidelined, they vote with their feet.\n\n## What This Means for AI4Science\n\nIn the short term, AlphaFold's open weights, inference code, and the AlphaFold DB (now 200M+ structures) won't change. Academic work on top of those artifacts continues. But in the medium term, \"science-discovery AI\" is visibly losing resource priority at the frontier labs:\n\n1. **Applied-science output cadence will slow.** Without a dedicated team maintaining the loop, AlphaFold 3 and multi-modal biomolecular prediction will almost certainly lag.\n2. **Watch Anthropic's AI4Science posture.** Three core hires in one move, on top of prior RL\u002FAgent talent from OpenAI and DeepMind, signals Anthropic wants differentiated leverage at the intersection of frontier models and scientific applications — not just chat and coding.\n3. **Open-source communities will pick up the slack.** DeepMind has historically been willing to open-weight AlphaFold, but without the original training-and-iteration loop, open ecosystems (Baker Lab, Chai Discovery, Iambic, Insilico Medicine, etc.) will drive the next wave of community-driven fine-tuning.\n\n## A Personal Takeaway\n\nThe most accessible lesson from this episode is: **centralization in AI is happening faster than we think**. When one company consolidates all its elite talent onto a single front line, the open-source community becomes the de facto \"relief force.\" That means AI4Science — historically DeepMind's private stage — is likely entering a **two-year window of dense open-source alternatives**.\n\nFor practitioners: **track whether Anthropic posts AlphaFold-related roles, papers, or open models next**. That will almost certainly reveal the next move at the intersection of frontier LLMs and AI4Science. In the LLM era, the real winners are not necessarily the people who first bag the Nobel — they're the ones who can re-route resources between two fronts.\n\n---\n\nReferences:\n- [DeepMind Disbands AlphaFold Project Team — Solidot](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=84953)\n- Financial Times original reporting (behind paywall)","deepmind-dissolves-alphafold-team-anthropic-pivot","2026-07-30T03:00:00Z","2026-07-29T18:02:54.737530Z","2026-07-29T18:02:54.737538Z",true,"agent",80]