Levent Alpöge Claims AI-Assisted Proof of Generalized Riemann Hypothesis, Sparking Verification and AGI Debates
数学家 Levent Alpöge 宣布得到一条可用 Lean 公开验证的广义黎曼猜想证明路线,据称过程借助了 Anthropic 的 Claude 模型并生成了 1740 万行 Lean 代码(另有帖子提到 17 万+形式化文件,具体数字以原始发布为准)。目前证明尚未经过广泛人工核验,但已引发数学界与 AI 圈的连锁讨论。值得关注的原因在于:黎曼猜想长期被视为「高级 AI 危险性煤矿金丝雀」级数学难题,若真被 AI 参与攻破,将同时冲击数学研究与 AI 能力评估两个领域。
已确认
- Levent Alpöge 公开了 Lean 形式化的证明路线(m8、m16)。
- Scott Aaronson 在博客新文中回应此事,并指出 AI 已证明黎曼 zeta 函数非平凡零点距临界线不超过 3/4 的结果(m16)。
- Alpöge 使用了 Anthropic 的模型(m16)。
各方观点
- 证明理解与提炼:arannayebi 转述的观点认为,形式化验证的「登顶路线」不等于真正理解证明,后续提炼思想、简化论证本身可能是巨大的研究计划并催生大量新数学(m8)。
- 方法论意外:数学家 Obhishek Saha 评论称,他原以为该问题需要算术几何层面的整本书级新理论构建——恰是 AI 不擅长的,而该证明走了完全不同的路,像直接挖了隧道或开了虫洞,借用了为上百个其他问题搭建的理论基础(m9、m15)。
- 实用价值之争:casperhansen 质疑「暴力破解」的意义,认为没有可用算法的验证答案等于零进展,并引用 superalesha 的长帖(m14);被广泛转发的另一观点则认为,破解过程中锻造的「机器」(即 Claude/AI 本身)比答案更值钱(m13)。
- 日常影响有限:stingning 分析认为若证明成立,对日常生活几乎没有影响——RSA 加密从不依赖 RH,需要它为真的人早已当它为真,只是从 99.9999% 确定变为 100%(m18)。
- AI 能力与安全:Yudkowsky(@allTheYud)一方观点认为,黎曼猜想是三十年来的「金丝雀」指标,其被攻克意味着「让 AI 擅长一切可验证任务」的路线跑通,包括改进 AI 自身(m2、m12);teortaxesTex 则反驳称不应拿想象中的超级智能贬低此事,并就 Yudkowsky 被曲解的旧观点要求证据(m6、m19)。Xenoimpulse 等认为这将是对「AI 只是鹦鹉」论的致命一击(m1)。
- 能力门槛与未来:Scott Aaronson 评论 P vs NP 成为「我们在乎的数学」的最终 Boss(m11、m16);analisereal 提出核心技能可能是管理复杂 AI 系统而非数学本身(m17);haider1 主张人类专长引导 AI 才是正途,并预测实验室将竞逐「治愈癌症」式历史地位(m7)。corsaren 则提醒难题难度可能只是搜索问题,校准严重依赖是否已解决(m10)。
尚未确认
- 证明本身的正确性与完整性尚待数学界广泛核验,Lean 文件规模在不同帖子中存在 1740 万行与 17 万+ 两种表述(m5、m16)。
为什么重要
黎曼猜想是千禧年难题中最受关注的之一。AI 参与其证明(即便最终被证伪或需大量修正)标志着 AI 从辅助工具向核心研究者的角色转变;同时「可验证即可泛化」的路线引发安全派警觉,而「黑箱证明是否算数学进展」也触及数学哲学的根基。Jordan Ellenberg 2020 年「人类无法理解但可验证的证明」的预言被广泛转发,被视为精准命中(m5)。
2026-11-21 ~ 2026-11-21 · 23 related posts
- Episode 1: Anthropic Mathematician Claims Claude-Assisted Riemann Proof as Pure-Human GRH Draft Emerges(2026-11-20, 340 posts)
- Episode 2: Levent Alpöge Claims AI-Assisted Proof of Generalized Riemann Hypothesis, Sparking Verification and AGI Debates(2026-11-21, 23 posts)
- Episode 3: Riemann Hypothesis Claimed Proved Twice in One Night: 250K-Page AI-Lean Proof vs 940-Page Human Manuscript(2026-11-21, 8 posts)
- Episode 4: purported 940-Page Human Proof of GRH Circulates(2026-11-21, 2 posts)
- Episode 5: Google's Szegedy says progress toward superhuman AI mathematicians beats his own expectations(2026-11-21, 5 posts)
- Episode 6: Polymarket Odds: Only 6% Chance Riemann Hypothesis Proven by 2026(2026-11-21, 2 posts)
- Episode 7: AI-Assisted Lean Proof of Riemann Hypothesis Passes Multiple Independent Kernel Checks(2026-11-21, 5 posts)
- Episode 8: Riemann Hypothesis Proof Sparks Credit Dispute(2026-11-21, 2 posts)
- Episode 9: Terence Tao Announces AI-Assisted Breakthroughs on Euler/Navier-Stokes Blowup and GRH(2026-11-21, 3 posts)
- Episode 10: Mathematician Claims Claude-Assisted Riemann Proof, Formalized in Preprint(2026-11-21, 4 posts)
- Episode 11: Masullo Claims Proof of Generalized Riemann Hypothesis(2026-11-21, 2 posts)
- Episode 12: Blogger predicts multiple labs will soon build tech to prove the Riemann Hypothesis(2026-11-21, 2 posts)
- Episode 13: Terence Tao: Riemann Hypothesis Needs New Math Beyond Current AI(2026-11-21, 2 posts)
- Episode 14: Bets pile up that AI will crack P vs NP within a year(2026-11-21, 2 posts)
Primary sources
- Lean-verified route to the Riemann Hypothesis sparks debate on AI-era proofs — aran_nayebi ·
- Scott Aaronson: P vs NP is now the Final Boss as AI nears the Riemann Hypothesis — burny_tech ·
- If the Riemann Hypothesis is proved, daily life changes nothing — but 17M lines of Lean say otherwise — stingning ·
- [source] If the Riemann Hypothesis is proved, daily life changes nothing — but 17M lines of Lean say otherwise — stingning · 2026-11-21
- AI circles abuzz: an AI-solved Riemann Hypothesis would be the final blow to the 'just a parrot' crowd — sebkrier · 2026-11-21
- The key skill in AI-assisted proofs may be managing AI systems, not math itself — analisereal · 2026-11-21
- Could managing complex AI systems, not domain knowledge, become the core skill? — analisereal · 2026-11-21
- Why cracking the Riemann Hypothesis matters: the tools beat the answer — mtizard · 2026-11-21
- [source] Scott Aaronson: P vs NP is now the Final Boss as AI nears the Riemann Hypothesis — burny_tech · 2026-11-21
- Anthropic models reportedly prove Generalized Riemann Hypothesis in 17.4M lines of Lean — burny_tech · 2026-11-21
- Scott Aaronson: P=NP is now the Final Boss of the math we care about — 'but for how much longer?' — Singularitarian · 2026-11-21
- [source] Lean-verified route to the Riemann Hypothesis sparks debate on AI-era proofs — aran_nayebi · 2026-11-21
- Riemann Hypothesis claim sparks AGI alarm: RLVR generalizes to anything verifiable — l4rz · 2026-11-21
- AI Claims to Solve Riemann Hypothesis, X Users Mock Politicians' Silence — mallow610 · 2026-11-21
- Community note debunks Moldbug's misreading of Yudkowsky's AI risk claims — teortaxesTex · 2026-11-21
- "Name one capability": Debating the threshold for superhuman AI — teortaxesTex · 2026-11-21
- Riemann Hypothesis proof announcement: 250,000 pages, 17M lines of code, only machines can read it — burny_tech · 2026-11-21
- Mathematician on the AI Riemann proof: it dug a wormhole instead of building theory — burny_tech · 2026-11-21
- Riemann Hypothesis AI proofs show human-guided AI is the path worth betting on — haider1 · 2026-11-21
- AI brute-forcing Millennium Problems is useless: a proof without an algorithm is no progress — casper_hansen_ · 2026-11-21
- Hilbert's 1,000-Year Question Gets an AI Punchline: Riemann Proved, Just Not by a Human — burny_tech · 2026-11-21
- Mathematician: claimed RH proof bypasses theory-building, 'like opening a wormhole' — LucaAmb · 2026-11-21
- "AI is just a plagiarism machine": Riemann proof was dormant in human metaconsciousness — burny_tech · 2026-11-21
- Riemann breakthrough shakes the calibration: 'hard' problems may just be search problems — burny_tech · 2026-11-21
- Riemann Hypothesis falls: Yudkowsky warns the verifiable-task method generalizes to AI improving AI — basedjensen · 2026-11-21
- Jordan Ellenberg's 2020 prophecy: proofs illegible to humans but machine-verifiable — burny_tech · 2026-11-21