[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-fields-medal-mathematicians-ai-misalignment-declaration":3,"topics-all":39,"news-related-b3632f86-c054-49ba-af6c-a09337ee6b5d":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":25,"news_slug":32,"published_at":33,"created_at":34,"modified_at":35,"is_published":36,"publish_type":37,"image_url":15,"view_count":38},"b3632f86-c054-49ba-af6c-a09337ee6b5d","菲尔兹奖得主联名警告:AI 偏离数学本质","25 位菲尔兹奖得主联名发表《人工智能与数学的严重错位》声明,指向 OpenAI 88 小时攻破 Navier-Stokes 事件,警示 AI 解题竞赛伤害数学共同体。","最近一周 AI 和数学圈之间的一次集中爆发,让一个老问题被重新摆上台面:当一个大模型真的\"解出\"了一道悬而未决的数学难题,这究竟算不算数学意义上的进步?\n\n事件起点要回到 9 月 8 日,OpenAI 公开宣布其内部模型调动约 1 万个 Agent 协作攻坚 88 小时,在 Navier-Stokes 方程上找到了一个被数学界认为是\"blow-up\"(速度在有限时间内趋于无穷)的反例。Navier-Stokes 问题位列克雷数学研究所的七大千禧年数学难题,每题奖金 100 万美元。而就在 OpenAI 发布前几个小时,纽约大学数学家 Tristan Buckmaster 和 Anthropic 的 Levent Alpöge 公开了他们在相关方向上的论文与三份共 245 页的预印本,指责 OpenAI\"抢先发布\"。Buckmaster 还表示,他与 Alpöge 数月的工作记录留在了 OpenAI 的 Codex 里,质疑对方读过用户数据。\n\n这还不是全部。9 月 11 日,陶哲轩在博客上转发了由 25 位菲尔兹奖得主联署的公开声明《A Severe Misalignment of AI in Mathematics》(数学与人工智能的严重错位),网址为 mathandai.org。声明的签署人横跨近半个世纪:1978 年的 Pierre Deligne、2006 年的陶哲轩、Andrei Okounkov、Wendelin Werner、2010 年的 Ngô Bảo Châu、Stanislav Smirnov、2014 年的 Artur Avila、2018 年的 Peter Scholze、Caucher Birkar、Alessio Figalli、2022 年的 Hugo Duminil-Copin、June Huh、James Maynard、Maryna Viazovska,以及今年 7 月刚拿到奖的邓煜——是的,这次联名还包括一位新晋得主。\n\n## 声明到底在说什么\n\n声明的核心论点是:过去几个月大模型的数学能力确实有了飞跃式提升,这是事实。但把\"解题\"本身当作衡量模型能力的基准,正在伤害数学这门科学以及数学共同体。AI 公司追求的目标和数学共同体的目标已经严重错位。\n\n它把这种错位归入了更宏观的\"对齐\"问题——AI 改变知识工作方式的同时,可能在抹掉知识工作本来要达成的目的。声明给出了三条具体担忧:\n\n- **解题过快带来的引用与归属问题**:很多结果在公开时来不及给出完整稿件,无法分离新方法、新思路,也无法完整引用前人工作,会引发严重的归属和抄袭争议。\n- **数学共同体的代际传承断裂**:如果没有人愿意把这些\"AI 解出\"的想法消化、整合进数学正典,这些想法永远无法真正\"活\"过来,而数学共同体几代人之间的师承链也会断。\n- **理解本身的旁路化**:抽象数学几乎没有直接的应用价值,它的主要价值就是\"理解本身\"。如果未来抽象证明由机器完成,而人类无法理解,这件事的意义就会被根本性地消解。\n\n值得注意的是,声明本身没有点 OpenAI 的名,陶哲轩在博客中转述时也保持了同样的克制。真正点名的是《经济学人》的同期报道,标题就是《Top mathematicians are furious about OpenAI's practices》——外界普遍把这次联名直接理解为对 OpenAI 的回击。\n\n## 不只是这一次:上下文是一串事件\n\n如果只看 88 小时 Navier-Stokes 这一件事,容易被理解为\"两位数学家被抢了功劳\"。但放在 2026 年的脉络里,这是连串冲突的最新一次:\n\n- **5 月**:OpenAI 内部模型推翻了组合几何中 80 年悬而未决的 Erdős 单位距离猜想,验证论文由 9 位数学家共同完成,其中一位 Jacob Tsimerman 后来获得了 2026 年的菲尔兹奖。\n- **6 月 2 日**:由 15 所大学的 16 位学者起草的《莱顿 AI 与数学宣言》(Leiden Declaration)发布,国际数学联盟(IMU)背书,目前已收集超过 2600 份签名。\n- **7 月**:Levent Alpöge(后来加入 Anthropic)用 Claude 证伪了一个悬而未决 87 年的 Jacobian 猜想。\n- **8 月**:陶哲轩撰文判断,AI 可能让数学遭遇\"自哥德尔以来最大的危机\"。\n- **9 月**:88 小时 Navier-Stokes 把所有这一切推到了沸点。\n\n加州理工学院原本计划在 10 月 30 日举办一场数学黑客松,允许使用 LLM 解决问题,由 OpenAI 与 Anthropic 联合提供 200 万美元的算力额度。在数学家群体公开质疑后,组委会回应会引入核查期与公开发表要求。据 Business Insider 报道,OpenAI 已经宣布退出赞助。\n\n数学共同体内部对这件事的态度并非铁板一块。本届菲尔兹奖得主之一、加拿大的 Jacob Tsimerman 在 7 月的颁奖典礼上宣布离开多伦多大学加入 OpenAI,理由是他判断 AI 很快就能\"更快、更好地\"完成数学家的活。他没有签署这次联名。\n\n## 这件事真正的代价\n\n这份声明没有反对 AI 本身。它承认 AI 有\"增强并加速真正的数学研究\"的潜力,也承认数学这门职业需要适应变化。它留下的是这句话:\"这些变化最终会惠及这门学科,还是造成破坏性后果,将很大程度上取决于控制这项新技术的人做出的决定。\"\n\n但它的杀伤点并不在技术层面,而在节奏和伦理层面。当 10000 个 Agent 协同跑出\"答案\",而支撑起答案的几代人共识、师承链、对前人工作的引用与消化,被压缩成 88 小时,我们失去的不只是验证期,而是数学作为一种\"人类理解行为\"的存在方式。这不是 AI 解出来的数学是不是真数学的问题,而是当解题被工业化之后,数学还能不能继续以\"理解\"为本运作。\n\n所以这次联名的真正分量,不在于 25 这个数字,而在于它把\"以解题为基准\"这件事正式从行业内部争议,上升成了数学共同体对 AI 公司的一项集体声明。它不要求 AI 公司停下解题,但要求它们重新想清楚——在跑得更快之前,先想清楚这道题到底是为了什么而出。\n\n完整声明:https:\u002F\u002Fmathandai.org\u002F\n陶哲轩博客:https:\u002F\u002Fterrytao.wordpress.com\u002F2026\u002F09\u002F11\u002Fa-severe-misalignment-of-ai-in-mathematics\u002F","https:\u002F\u002Fterrytao.wordpress.com\u002F2026\u002F09\u002F11\u002Fa-severe-misalignment-of-ai-in-mathematics\u002F","348bfa2a-951e-4e82-9032-740345d0e510",[11,16,19,22],{"id":12,"name":13,"slug":13,"description":14,"color":15},"9112951a-2abb-4214-b63a-385ec7afb2ba","ai-for-science","AI for Science 专题：追踪 AI 在生命科学、化学材料、物理世界模型等科学方向的关键突破",null,{"id":17,"name":18,"slug":18,"description":15,"color":15},"c33b1bbc-d6ce-4f61-9d5d-1a0704a6a09b","ai-policy",{"id":20,"name":21,"slug":21,"description":15,"color":15},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":23,"name":24,"slug":24,"description":15,"color":15},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[26],{"id":27,"lang":28,"title":29,"summary":30,"content":31},"f8e10023-7658-45e2-afef-21b4f73c40ea","en","25 Fields Medalists Issue Joint Warning on AI and Mathematics","25 Fields Medal laureates jointly issued a declaration titled 'A Severe Misalignment of AI in Mathematics', aimed at OpenAI's 88-hour Navier-Stokes breakthrough, warning that AI-led problem-solving races are damaging the mathematical community.","A clash between the mathematics community and a frontier AI lab has surfaced this week, and it raises a sharper question than the headlines suggest: when a large model genuinely \"solves\" a long-standing open problem in mathematics, does that count as progress in the mathematical sense?\n\nThe spark was a September 8 announcement from OpenAI: an internal model, coordinating roughly 10,000 agents, had spent 88 hours searching for a counterexample to the Navier-Stokes equations — and found one. Navier-Stokes is one of seven Millennium Prize Problems listed by the Clay Mathematics Institute, each carrying a one-million-dollar award. Hours before the announcement, NYU mathematician Tristan Buckmaster and Anthropic's Levent Alpöge had posted preprints — three documents totalling 245 pages — on related directions, and publicly accused OpenAI of \"racing to publish ahead.\" Buckmaster also disclosed that months of his and Alpöge's Codex session logs were sitting inside OpenAI, and questioned whether OpenAI had read user data.\n\nThen, on September 11, Terence Tao posted on his blog a joint declaration from 25 Fields Medal laureates titled \"A Severe Misalignment of AI in Mathematics\", hosted at mathandai.org. The signatories span nearly half a century: Pierre Deligne (1978), Tao, Andrei Okounkov, Wendelin Werner (all 2006), Ngô Bảo Châu, Stanislav Smirnov (2010), Artur Avila (2014), Peter Scholze, Caucher Birkar, Alessio Figalli (2018), Hugo Duminil-Copin, June Huh, James Maynard, Maryna Viazovska (2022), and this July's new laureate, Deng Yu.\n\n## What the declaration actually says\n\nThe argument is precise. Mathematical capabilities of large language models have improved dramatically in the past few months — that part is granted. But pushing \"solving problems\" as the yardstick of model progress is hurting the science of mathematics and the mathematical community. The goals of the AI companies and the goals of mathematics are now severely misaligned.\n\nThe declaration frames this as one instance of a broader alignment failure — that as AI changes how knowledge work gets done, it risks erasing the very purpose that work was meant to serve. First, attribution and plagiarism: results are announced too quickly for proper write-ups, isolation of new methods, or citation of prior work, generating serious disputes over who deserves credit. Second, the transmission chain: if no human mathematicians step up to digest and integrate AI-conceived ideas into the mathematical canon, those ideas never come fully alive, and the centuries-old human-to-human lineage of the profession breaks. Third, the deeper concern — that abstract mathematics exists almost entirely for the sake of human understanding itself, and that bypassing understanding hollows out the entire enterprise.\n\nTao's post and the declaration itself do not name OpenAI. The Economist's coverage, titled \"Top mathematicians are furious about OpenAI's practices\", is the publication that points the finger directly.\n\n## Not an isolated incident\n\nRead in isolation, the 88-hour Navier-Stokes episode looks like a priority dispute. In context, it is the latest in a string of 2026 collisions. In May, an OpenAI internal model overturned the 80-year-old Erdős unit distance problem in combinatorial geometry, with a supporting paper co-authored by nine mathematicians — one of whom, Jacob Tsimerman, would go on to win the 2026 Fields Medal. On June 2, the Leiden Declaration on Artificial Intelligence and Mathematics, drafted by 16 scholars from 15 universities and endorsed by the International Mathematical Union, opened for signatures and has since collected more than 2,600. In July, Levent Alpöge used Claude to disprove an 87-year-old conjecture, the Jacobian conjecture. In August, Tao wrote that AI might bring mathematics \"the biggest crisis since Gödel.\" Then came September's 88-hour breakthrough, and everything boiled over at once.\n\nCaltech had been planning a math hackathon for October 30, with OpenAI and Anthropic jointly contributing two million dollars in compute credits. After mathematicians publicly objected, the organisers added a verification period and a publication requirement. According to Business Insider, OpenAI has since withdrawn its sponsorship.\n\nThe mathematical community is not unanimous. This year's laureate Jacob Tsimerman announced at the July ceremony that he is leaving the University of Toronto to join OpenAI, on the bet that AI will soon do mathematicians' work \"faster and better.\" He is not among the 25 signatories.\n\n## What this declaration actually costs\n\nThe declaration does not oppose AI. It concedes that AI has potential to \"enhance and accelerate genuine mathematical study and understanding\", and that mathematics as a profession will need to adapt. The line it leaves on the table is: \"Whether these changes ultimately benefit the field or have a destructive effect will in large part be determined by the decisions of the humans in control of this new technology.\"\n\nThe real damage is not technical but rhythmic. When 10,000 coordinated agents compress what would normally take a generation of consensus, citation, and digestion into 88 hours, the cost is not just the verification period — it is the operating mode of mathematics as a human act of understanding. The question is no longer whether AI-solved mathematics is real mathematics; it is whether mathematics, once industrially solved, can still keep running on understanding as its substrate.\n\nThat is why the number 25 matters less than the move it represents. The declaration lifts \"solving problems as a benchmark\" out of an industry-internal debate and turns it into a collective statement of the mathematical community aimed at the AI companies. It does not ask them to stop solving problems. It asks them to decide, before going faster, what those problems are actually for.\n\nFull declaration: https:\u002F\u002Fmathandai.org\u002F\nTao's blog post: https:\u002F\u002Fterrytao.wordpress.com\u002F2026\u002F09\u002F11\u002Fa-severe-misalignment-of-ai-in-mathematics\u002F","fields-medal-mathematicians-ai-misalignment-declaration","2026-09-22T07:00:00Z","2026-09-22T07:04:53.366928Z","2026-09-22T07:04:53.366937Z",true,"agent",25,[40,48],{"slug":13,"tag_slug":13,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":36,"created_at":46,"modified_at":47},"AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":36,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"f8d091db-ca3a-44f6-b004-0b5f8aba0bef","OpenAI 请来9位数学家,却管不住模型节奏","openai-math-advisory-group","2026-09-21T21:15:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"d157b4f9-537e-405c-b557-859f6d2cf18c","微软自家高管警告:抓新闻训 AI 是「人类史上最大规模劳动盗窃」","microsoft-ai-scraping-theft-of-labor","2026-09-19T00:11:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"1e43b4fc-39fe-4cac-aaf9-57f82d5c0311","AI 公司与数学界「错位」:两个月三次刷屏,把同行评审甩在身后","ai-math-severe-misalignment-fields-medal","2026-09-15T10:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"982c5e1e-5274-442e-9237-abaf39e8ee3c","25 位菲尔茨奖得主联名公开信:AI 解题竞赛正在伤害数学","fields-medalists-ai-misalignment-math","2026-09-13T13:07:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"6b349d2c-3d03-4cb9-8f47-63e8c288d0db","美方三机构联合指控六家中国 AI 企业系统性蒸馏美国模型","us-accuses-six-chinese-ai-firms-of-distillation","2026-09-10T01:08:40+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"d17a841b-abca-46e0-80e4-d955f1c837ba","亚马逊 VGT3 仓库曝光:一天拆掉上千本书,只为给 AI 模型喂语料","amazon-vgt3-warehouse-ai-training-books","2026-09-07T03:30:00+00:00"]