PhD Student Uses Multi-Agent AI to Crack a 98-Year-Old Math Problem in 48 Hours
量子位 · wechat · 2026-09-06
Ma Qianli, a Zhejiang University AI PhD student and co-founder of WujieAI, solved the odd-index case of the Colombo determinant problem (open since 1928) with a multi-agent system in 48 hours, producing 20,000 lines of Lean formalization code; the preprint and code are public.
His agent platform assigned GPT-5.6 to literature search, Fable5 to adversarial review, and DeepSeek to Lean proofs. AI generated 8 proof routes; a promising p=m-1 route burned 11 hours and 47,000+ lines without converging before a human called it off — switching to p=m-2 led to success 2h23m later. The writeup details the human-in-the-loop research paradigm: choosing problems, evaluating routes, and deciding when to abandon, plus the business opportunity of building research harnesses for scientists.
More from coding & agent
- Juggling ChatGPT, Claude and agents, users hunt for a unified AI memory setup — utkuaytac · 2026-09-06
- A 3-person team can't make multi-agent coding work: three context-sharing attempts failed — Practical_Sink401 · 2026-09-06
- Villager sim game built locally on 16GB VRAM with Qwen3.8-27B Q3 quant at 75 tok/s — Fancy-Snow7 · 2026-09-06
- mitsuhiko uses AI agent to build Java-style virtual threads for Python — mitsuhiko · 2026-09-06
- Two-person team spent 12 days and $400+ credits making a 6-min AI short film with Codex + Seedance 2.5 — APPSO · 2026-09-06
- AI Coding Will Prevent Expertise: The 'Expert Novice' Paradox in Dev Skills — bibryam · 2026-09-06