雅可比猜想遭AI推翻:Claude Fable 5三行反例解87年謎題 | Jacobian Conjecture Falls to Claude Fable 5's AI Proof
By Kit 小克 | AI Tool Observer | 2026-07-28
🇹🇼 雅可比猜想遭AI推翻:Claude Fable 5三行反例解87年謎題
雅可比猜想被推翻了。2026年7月20日,Anthropic數學家Levant Alpöge用Claude Fable 5只花幾行算式,就找到反例推翻這個從1939年懸而未解87年的數學難題,成為AI輔助數學研究又一里程碑事件,也讓數學界重新思考LLM能做到什麼程度。
什麼是雅可比猜想?
雅可比猜想由德國數學家Ott-Heinrich Keller於1939年提出,主張如果一個多項式映射在每個點附近都是局部可逆(雅可比行列式恒為非零常數),那麼它在整體上也必然可逆。這個看似直覺的猜想困擾數學家近九十年,無數嘗試證明或反駁都以失敗告終。
Claude Fable 5 怎麼找到反例的?
Alpöge表示,他和朋友Akhil在看世界盃比賽時隨口問了Claude Fable 5這個問題,模型給出的反例只用三行方程式,構造出一個雅可比行列式恆等於-2、卻把三個不同輸入映射到同一輸出的多項式映射,直接證明「局部可逆不代表整體可逆」。這個結果不是透過正式論文發表,而是先在社群媒體上公開,消息一出,全球數學家在幾小時內就各自驗證了正確性。
- 發現方式:非正式對話提問,而非刻意設計的證明任務
- 反例長度:僅三行方程式構造
- 驗證速度:發布後數小時內獲多位數學家獨立確認
這件事有爭議嗎?
有。目前完整的模型對話紀錄、確切提示詞與Anthropic官方論文都尚未公開,外界只能透過社群貼文拼湊經過。部分數學家提醒,在正式論文與同行審查出爐前,應謹慎看待「AI獨立解題」的敘事,人類數學家的問題設定與判斷仍是關鍵一環,Claude Fable 5更像是強力的計算與構造工具,而非取代數學家的獨立研究者。
對AI數學研究代表什麼意義?
這已是2026年以來第三起重大AI數學突破:5月OpenAI模型推翻了80年歷史的Erdős離散幾何猜想,如今Claude Fable 5又拿下雅可比猜想。伴隨這波熱潮,15所大學的16位研究者6月共同發表《萊頓宣言》,呼籲數學界為AI參與研究建立透明度、歸屬權與同行審查的規範,避免「社群媒體宣布重大突破」成為常態卻缺乏可驗證流程。
好不好用,試了才知道。
🇺🇸 Jacobian Conjecture Falls to Claude Fable 5's AI Proof
The Jacobian Conjecture has fallen. On July 20, 2026, Anthropic mathematician Levant Alpoge used Claude Fable 5 to produce a three-line counterexample disproving a problem that had stumped mathematicians for 87 years, another milestone in AI-assisted math research that has the field rethinking what large language models can actually do.
What Is the Jacobian Conjecture?
Formulated by German mathematician Ott-Heinrich Keller in 1939, the Jacobian Conjecture claimed that if a polynomial map is locally invertible everywhere (its Jacobian determinant is a nonzero constant), then it must also be globally invertible. The idea sounds intuitive but resisted proof or disproof for nearly nine decades.
How Did Claude Fable 5 Find the Counterexample?
Alpoge says he and a friend, Akhil, casually asked Claude Fable 5 about the problem during a World Cup match. The model returned a counterexample in just three lines: a polynomial map with a Jacobian determinant identically equal to -2, yet mapping three distinct inputs to the same output, directly disproving the conjecture. Alpoge posted the result on social media rather than through a peer-reviewed paper, and mathematicians worldwide independently verified its correctness within hours.
- How it was found: a casual prompt, not a purpose-built proof task
- Counterexample length: just three lines of equations
- Verification: independently confirmed by multiple mathematicians within hours
Is There Controversy?
Yes. The full model transcript, exact prompts, and an official Anthropic paper have not been publicly released; the story has been pieced together from social media posts. Some mathematicians caution against reading this as AI solving math independently: human framing and judgment still matter, and Claude Fable 5 looks more like a powerful construction tool than an autonomous researcher, at least until a formal writeup and peer review appear.
What Does This Mean for AI Math Research?
This is the third major AI math breakthrough in 2026: in May, an OpenAI model disproved the 80-year-old Erdos conjecture in discrete geometry, and now Claude Fable 5 has taken down the Jacobian Conjecture. Amid the wave, 16 researchers from 15 universities published the Leiden Declaration in June, calling for transparency, attribution, and peer-review norms as AI increasingly contributes to open mathematical problems, before announced on social media becomes the default instead of verifiable process.
好不好用,試了才知道 - worth trying before you trust the hype.
Sources / 資料來源
- The Conversation: Why a tiny social media post has mathematicians rethinking AI
- John D. Cook: Locally everywhere does not imply everywhere
- Anthropic: Claude Fable 5 and Claude Mythos 5 announcement
常見問題 FAQ
雅可比猜想是什麼?
1939年提出的數學猜想,主張多項式映射若處處局部可逆,則必定整體可逆,87年來無人證明或反駁。
Claude Fable 5是怎麼推翻雅可比猜想的?
Anthropic數學家Levant Alpoge隨口提問,模型給出僅三行方程式的反例,構造出局部可逆但非整體可逆的映射。
這個結果經過同行審查了嗎?
尚未。目前只有社群媒體貼文和數學家的獨立驗證,正式論文與完整模型對話紀錄尚未公開。
這是AI第一次解開重大數學難題嗎?
不是,這是2026年第三起,5月OpenAI模型才推翻80年歷史的Erdos離散幾何猜想。
延伸閱讀 / Related Articles
- Slopsquatting解析:AI幻覺套件名成資安地雷 | Slopsquatting: AI Coding Agents Fuel Package Supply Chain Attacks
- Qoder Security 評測:AI 編碼原生抓漏洞,免插件 | Qoder Security Review: AI Coding With Built-In Vuln Scans
- Nvidia 2500億美元挺OpenAI:俄亥俄5000億機房解析 | Nvidia's $250B OpenAI Deal: Ohio's $500B AI Mega-Campus
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