OpenAI Navier-Stokes評測：AI證明百萬美元難題惹抄襲爭議 | OpenAI Navier-Stokes Review: $1M Proof, Credit Dispute

9月 09, 2026

By Kit 小克 | AI Tool Observer | 2026-09-10

🇹🇼 OpenAI Navier-Stokes評測：AI證明百萬美元難題惹抄襲爭議

OpenAI Navier-Stokes是什麼？近萬個AI代理接力88小時的證明

OpenAI在9月8日宣布，一個尚未對外開放、比GPT-6 Astra更強的實驗性模型，透過近10,000個並行運作的AI代理、花費88小時、互傳將近500萬則訊息，針對Navier-Stokes方程式——七大千禧年大獎難題之一，克雷數學研究所懸賞100萬美元——產出一份證明。內容主張流體在有限時間內速度可衝向無限大（也就是方程式會「爆破」），並公開了完整論文與Lean形式化證明檔案。

但OpenAI自己講得很保守：這份證明只滿足官方問題描述裡的C、D兩項陳述，屬於部分結果，公司也明確表示不會去申請這筆百萬美元獎金——因為克雷數學研究所的規則要求證明必須先公開發表滿兩年、並獲得數學界普遍認可才算數。

抄襲爭議：AI是不是「偷看」了別人的研究筆記？

真正炸出討論的不是數學，而是信任問題。紐約大學數學家Tristan Buckmaster和在Anthropic工作的Levent Alpöge，已經花一年時間用AI工具（包含OpenAI的Codex）研究Navier-Stokes的前置難題——Euler方程式的爆破解。他們指控OpenAI很可能是從他們私下使用Codex時留下的紀錄「借用」了關鍵思路，才能在短時間內衝出結果。

OpenAI否認直接存取兩人的私人研究，但也承認無法排除去識別化的使用資料被拿去訓練或改進模型的可能性。知名數學家陶哲軒（Terence Tao）公開稱讚Buckmaster和Alpöge的原始方法「了不起」，被外界解讀為間接支持人類研究者的優先權。Hacker News上的討論則普遍偏懷疑，質疑目前公開的證明版本，是否真的對應到官方定義的那道千禧年難題，還是只解了條件較寬鬆的版本。

給一般人的誠實提醒

這件事的重點不是「AI又贏了一次」，而是提醒大家：AI公司拿使用者的工作紀錄去訓練模型的灰色地帶，正隨著AI能力變強而變得更燙手。如果你平常用Codex、Claude或任何AI工具處理還沒發表的研究、程式碼或商業機密，這起爭議是個提醒——去識別化不代表內容不會以某種形式回流到模型裡。

好不好用，試了才知道。

🇺🇸 OpenAI Navier-Stokes Review: $1M Proof, Credit Dispute

What Is the OpenAI Navier-Stokes Proof? 10,000 Agents, 88 Hours

On September 8, OpenAI announced that an unreleased experimental model — more capable than GPT-6 Astra — used roughly 10,000 concurrent AI agents exchanging nearly 5 million messages over 88 hours to produce a proof for the Navier-Stokes equations, one of the seven Millennium Prize Problems with a $1 million bounty from the Clay Mathematics Institute. The claim: a fluid vortex can reach infinite speed in finite time — the equations blow up. OpenAI published the full paper and a Lean formalization.

Notably, OpenAI is hedging hard: the result only satisfies statements C and D of the official problem formulation — a partial result — and the company says it will not claim the $1 million prize, since Clay rules require two years of publication and broad acceptance by the math community first.

The Credit Dispute: Did AI "Peek" at a Human Researcher's Notes?

What actually blew up was trust, not fluid dynamics. NYU mathematician Tristan Buckmaster and Levent Alpöge (who works at Anthropic) had spent roughly a year using AI tools, including OpenAI Codex, to chase a stepping-stone problem — blow-up solutions for the Euler equations. They allege OpenAI may have picked up their approach from Codex session data before racing to a related Navier-Stokes result.

OpenAI denies directly accessing the pair's private work, but admits it cannot rule out de-identified usage data contributing to model improvements. Terence Tao publicly called Buckmaster and Alpöge's original approach "remarkable" — widely read as backing the human researchers' priority. Hacker News threads lean skeptical too, with some arguing the published proof may only cover a "forced" version of Navier-Stokes rather than the exact Millennium Prize statement.

The Honest Takeaway

The real story isn't "AI beat math again" — it's a warning about how blurry the line has gotten between "training data" and someone else's unpublished work. If you run unpublished research, proprietary code, or business-sensitive material through Codex, Claude, or any AI tool, this dispute is a reminder: de-identified doesn't mean it can never resurface, in some form, inside a future model.

好不好用，試了才知道。
You won't know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles

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

搜尋此網誌

AI小貼士