跳到主要內容

指數時間假設評測:OpenAI、Anthropic同日推翻30年猜想 | Exponential Time Hypothesis Review: AI Disproves It

By Kit 小克 | AI Tool Observer | 2026-10-30

🇹🇼 指數時間假設評測:OpenAI、Anthropic同日推翻30年猜想

最近AI圈最熱的話題不是新模型發布,而是理論電腦科學界用了30年的指數時間假設(Exponential Time Hypothesis, ETH),竟然被兩家AI公司在同一天各自宣布推翻。OpenAI和Anthropic都說自家AI模型獨立找到了反例,而且發現日期都指向同一天:10月16日,事前互不知情。

指數時間假設是什麼?為什麼重要

指數時間假設是電腦科學裡用來證明「某些問題有多難」的基礎假設之一,主要針對3-SAT(布林可滿足性問題)這類NP問題:假設沒有演算法能在低於指數時間內解決最難版本的3-SAT。過去30年有數百篇論文引用這個假設來推導其他問題的困難度下限,算是理論電腦科學的地基之一。

OpenAI與Anthropic同日推翻,巧合到嚇到學界

OpenAI宣稱內部模型10月16日用自然語言找到反例,10月17日用Lean形式化驗證系統跑過一遍,10月26日正式公開。詭異的是,幾小時後Anthropic也發布論文,說Claude完全獨立地在同一天找到類似反例,雙方事前互不知情。兩邊的解法都收斂在大約2^(n/√log n)的時間複雜度——比指數暴增慢,但遠遠稱不上實用。

誠實講:這不代表P=NP,也沒人破解加密

  • 不影響P vs NP:OpenAI自家高層都講了,P還是很可能不等於NP,這次推翻的是比較窄的指數時間假設,不是複雜度理論的終極大哉問
  • Lean驗證只確認「邏輯一致」:不代表3-SAT或次指數時間的形式化定義真的對應學界原本的意思,目前UC San Diego與MIT的學者(包含假設原始提出者)正在審查語意翻譯是否正確
  • 沒有加密系統會因此被破解:這次的演算法速度仍遠遠不到能實際威脅RSA或比特幣的程度
  • 目前只是「聲稱」:獨立專家審查結果還沒出來,Lean檢查目前是唯一通過的外部測試

對一般AI使用者來說,這件事真正有意義的地方不是「30年假設倒了」,而是兩家頂尖AI實驗室的模型都能在無人監督下,獨立產出理論電腦科學的新結果——這代表AI輔助數學研究的能力正在跨過一個門檻。但距離「AI自主做出可信賴的原創數學研究」,還有審查這一關要過。

好不好用,試了才知道。


🇺🇸 Exponential Time Hypothesis Review: AI Disproves It

The hottest story in AI right now isn't a new model release — it's a 30-year-old theoretical computer science assumption called the Exponential Time Hypothesis (ETH) that two AI labs claim to have disproved, on the same day, without knowing about each other's work.

What Is the Exponential Time Hypothesis, and Why Does It Matter?

ETH is one of the foundational assumptions used to prove how hard certain problems are, especially 3-SAT (Boolean satisfiability). It assumes no algorithm can solve the hardest versions of 3-SAT meaningfully faster than brute force — in sub-exponential time. Hundreds of papers over three decades have relied on ETH to derive hardness results for other problems, making it one of the bedrock assumptions in complexity theory.

OpenAI and Anthropic Both Refuted It — on the Same Day

OpenAI says one of its internal models found a natural-language disproof on October 16, formalized it in the Lean proof assistant on October 17, and announced it October 26. Hours later, Anthropic published its own paper: Claude independently found a similar disproof, tracing back to the exact same date — October 16 — with neither lab aware of the other's work. Both results land around the same complexity bound, roughly 2^(n/√log n) — beating the exponential floor ETH predicted, but nowhere close to practical.

The Honest Take: This Doesn't Break Encryption or Prove P=NP

  • P vs NP is untouched — OpenAI's own leadership says P is still probably not equal to NP; this refutes the narrower Exponential Time Hypothesis, not the bigger open question in complexity theory
  • Lean verification only confirms internal logical consistency — it doesn't confirm the formal statement actually captures what complexity theorists mean by "3-SAT" or "sub-exponential time." Experts at UC San Diego and MIT, including the hypothesis's original authors, are now reviewing whether the translation is faithful
  • No cryptographic system is at risk — the new algorithm is still far too slow to threaten RSA, Bitcoin, or anything relying on hardness assumptions
  • It's a claim, not a confirmed result yet — Lean's check is currently the only independent test this has passed; peer review is still pending

What actually matters here isn't "a 30-year hypothesis fell" — it's that two frontier AI labs' models independently produced a theoretical computer science result without human guidance, on the same day. That's a real signal about where AI-assisted math research is heading. Whether it holds up to scrutiny is a separate question.

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

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code