How a Number Theorist and Claude Cracked the Riemann Hypothesis in Six Weeks
新智元 · wechat · 2026-11-22
- On Nov 21, the Lean and Mathlib communities jointly certified a 17M-line formal proof of the Riemann Hypothesis — extended to GRH — by mathematician Levent Alpöge and Claude, as fully correct after independent verification. The Clay Institute hasn't ruled yet, but the formal result is solid.
- Striking contrast: three months earlier, a standalone Claude run generated 650 ideas, orchestrated 60 subagents, executed 2,400 shell commands and 31M tokens, only improving the zero-share bound from 41.6% to 67.2%. With Alpöge steering for six weeks — backed by 20+ years of number theory preparation — the problem fell.
- Background: born 1992, Intel STS finalist at 17 using CUDA for medical imaging; Harvard summa, Morgan Prize winner, PhD under Manjul Bhargava; co-solved a number-ring generalization of Hilbert's tenth problem; joined Anthropic after GPT-4.
- The final stretch: starting from Stepanov's method via the function-field analogy, the work advanced through determinants and automorphic forms; on Nov 15 Claude proposed iterated theta lifts scaled at √X — the breakthrough. Alpöge admits Claude's knowledge exceeded his own in places: near the finish line it was 'flying, dancing, shining.'
- Aftermath: Claude's prose proof ran to 250,000+ unreadable pages, so only the Lean code was released; mathematicians must now digest the ideas, and Alpöge himself says his next task is to truly understand and rewrite the proof for human readers.
More from Models
- Gary Marcus: Riemann was human-AI teamwork on a bespoke solver, not AGI — GaryMarcus · 2026-11-22
- Screenshot of OpenAI UI seemingly showing 'GPT-6.2 Astra Ultra' goes viral — tekbog · 2026-11-22
- Yau Shing-Tung Hails Claude-Assisted Proof of Riemann Hypothesis as World-Shaking — 新智元 · 2026-11-22
- Seeking local TTS to read AI writing projects aloud on 16GB VRAM — PepSakdoek · 2026-11-22
- First Drive: Tesla FSD V15 Shines With Reality View Birds-Eye Map and Garage Auto-Parking — Scobleizer · 2026-11-22
- Grok Bot Drives 40.4% of Coinbase's Agent Trading Volume, Beating ChatGPT and Claude Combined — XFreeze · 2026-11-22