Navier-Stokes評測：OpenAI稱破千禧年難題,遭25位菲爾茲得主抗議 | Navier-Stokes Review: OpenAI's Proof Ignites a Credit War
By Kit 小克 | AI Tool Observer | 2026-09-21
🇹🇼 Navier-Stokes評測：OpenAI稱破千禧年難題,遭25位菲爾茲得主抗議
Navier-Stokes方程是流體力學最古老的未解問題之一,也是克雷數學研究所懸賞100萬美元的七大千禧年難題之一。2026年9月,OpenAI宣布一款尚未公開的下一代模型,動用約1萬個協同運作的AI agent,產出一份通過Lean形式驗證的Navier-Stokes存在性與光滑性證明。消息一出立刻登上各大科技媒體頭條,但緊接著爆發的不是數學界的掌聲,而是一場「誰的功勞」的公開對撞。
OpenAI的Navier-Stokes證明是怎麼跑出來的?
OpenAI表示,近100個agent先在約50小時內解決較簡單的Euler方程「有限時間爆破」問題,接著把資源全數轉向Navier-Stokes本身,約1萬個agent並行運作88小時後產出完整證明。這是agent群體規模化解數學題的一次實測,不是單一模型單次推理,而是龐大agent陣列協作驗證出來的結果。
為什麼數學家指控OpenAI搶功?
紐約大學數學家Tristan Buckmaster公開指控,他與Anthropic研究員Levent Alpöge早在8月15日就私下用AI模型證出Euler方程會爆破,這是尚未發表的研究。他表示OpenAI數學家Sébastien Bubeck提議讓他掛名共同作者,條件是排除Alpöge,理由是Alpöge任職於競爭對手Anthropic;對話中甚至被問「你為什麼要毀掉自己的職涯」。更讓他不安的是,他在私下研究中大量使用OpenAI的Codex,不排除自己的解法因此間接流入了OpenAI自家的證明工程。
菲爾茲獎得主怎麼回應?
25位菲爾茲獎得主(數學界最高榮譽)聯署公開信,警告AI實驗室把解出知名難題當成秀模型能力的公關素材,正在侵蝕數學界的歸屬與信任文化。加上今年稍早由國際數學聯盟背書、超過4000人連署的《萊頓宣言》,學界對LLM證明的態度已從好奇轉為警戒。OpenAI隨後撤回對Caltech一場學生自辦AI數學松的100萬美元贊助。陶哲軒也在部落格撰文稱此事是「AI在數學領域一次嚴重的失準」,但同時肯定Buckmaster與Alpöge的原始研究「確實了不起」。
Kit小克怎麼看
技術成就跟這場風波其實是兩件事。agent群體確實把Navier-Stokes逼出了形式化證明,Lean驗證不是能隨便造假的東西。但把它包裝成「AI獨力解開千禧年難題」,卻不提背後有人類數學家先做出關鍵突破、雙方在credit分配上明顯沒談攏,這樣的敘事並不誠實。如果你是研究者,這件事最實用的教訓是:用AI coding工具做未發表的原創研究前,先想清楚自己的輸入會不會變成別人的素材。好不好用,試了才知道。
🇺🇸 Navier-Stokes Review: OpenAI's Proof Ignites a Credit War
OpenAI says roughly 10,000 coordinated AI agents produced a Lean-verified proof of the Navier-Stokes Millennium Prize Problem in September 2026 — one of seven $1 million Clay Institute challenges. The announcement went viral within hours, but what followed wasn't mathematical celebration. It was a public credit fight that's still escalating, and it says more about AI labs than about fluid dynamics.
How Did OpenAI's Navier-Stokes Proof Actually Happen?
Nearly 100 agents first cracked the simpler "Euler blowup" question in about 50 hours, OpenAI says, then roughly 10,000 agents ran for 88 hours to complete the full Navier-Stokes existence-and-smoothness proof, verified in the Lean proof assistant. That's a real demonstration of agent swarms grinding through formal math — not one model producing a lucky one-shot answer.
Why Do Mathematicians Say OpenAI Took Credit It Didn't Earn?
NYU mathematician Tristan Buckmaster says he and Anthropic researcher Levent Alpöge privately proved the related Euler blowup result on August 15 — weeks before OpenAI's announcement. He alleges OpenAI mathematician Sébastien Bubeck offered him co-author credit only if Alpöge, an Anthropic employee, was excluded, and that the conversation turned pointed, with Bubeck allegedly asking "why would you ruin your career?" Buckmaster also flagged a subtler risk: he'd used OpenAI's Codex heavily during his private research, raising the possibility his own approach leaked into OpenAI's effort.
What Did Fields Medalists Do in Response?
Twenty-five Fields Medalists — math's highest honor — signed an open letter warning that AI labs racing to solve famous problems as capability demos, rather than genuine scientific contributions, is eroding math's attribution culture. It follows the Leiden Declaration, backed by over 4,000 signatories and endorsed by the International Mathematical Union earlier this year. OpenAI pulled its $1 million sponsorship of a Caltech student math hackathon after mathematicians objected. Fields Medalist Terence Tao called it "a severe misalignment," while still praising Buckmaster and Alpöge's underlying math as "remarkable."
Kit's Take
The agent swarm genuinely produced a formally verified proof — Lean doesn't fake-check. But branding it as "AI solves a Millennium Prize problem," without crediting the human mathematicians whose private work triggered the race, isn't honest framing. The practical lesson: if you're doing unpublished original research and feeding it through a lab's coding assistant, assume your approach isn't fully private. 好不好用，試了才知道 — try it yourself before you believe the headline.
Sources / 資料來源
- TechCrunch: OpenAI fought dirty on career-making math problem, says NYU mathematician
- TechCrunch: OpenAI's feud with mathematicians is only escalating
- Terence Tao: A Severe Misalignment of AI in Mathematics
常見問題 FAQ
什麼是Navier-Stokes千禧年難題?
它是描述流體運動的方程組,克雷數學研究所自2000年懸賞100萬美元徵求證明其解的存在性與光滑性,是七大千禧年難題中唯一橫跨物理與數學的題目。
OpenAI真的解開Navier-Stokes了嗎?
OpenAI公布的證明通過Lean形式驗證系統檢查,技術上具備可信度,但外部數學社群尚未完成獨立同行審查,結論仍待確認。
這場爭議對用AI做研究的人有什麼提醒?
使用AI coding工具處理未發表的原創研究時,要留意自己的解法或資料是否可能被納入該公司其他產品或模型的訓練與推理過程。
What is the Navier-Stokes Millennium Prize Problem?
It's a set of equations describing fluid motion; the Clay Institute has offered $1 million since 2000 for proving whether smooth solutions always exist.
Did OpenAI actually solve Navier-Stokes?
OpenAI's proof passed Lean formal verification, which is technically credible, but independent peer review by the math community hasn't concluded yet.
What's the takeaway for researchers using AI tools?
If you're doing unpublished original research with a lab's coding assistant, assume your approach could inform that company's own models or products.
延伸閱讀 / Related Articles
- Jalapeño評測：OpenAI用AI設計晶片,宣稱甩輝達3.6倍 | Jalapeño Review: OpenAI's AI-Designed Chip Beats Nvidia
- PhantomRaven評測：抓臭蟲賞金獵人用AI寫木馬,被AI抓包 | PhantomRaven Review: AI Wrote the Malware, AI Caught It
- Plugin4Shell評測：AI編碼工具爆零點擊漏洞,2家沒修 | Plugin4Shell Review: 4 AI Coding Agents, 2 Still Unpatched
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言