Skip to content
Back to Vault

https://overreacted.io/how-i-vibed-a-proof-of-conways-conjecture/

Description: Dan Abramov’s five-week, week-by-week account of using Claude, ChatGPT and Codex to produce a Lean-checked proof of a 50-year-old math conjecture.

Saved context

https://overreacted.io/how-i-vibed-a-proof-of-conways-conjecture/ Description: Dan Abramov’s five-week, week-by-week account of using Claude, ChatGPT and Codex to produce a Lean-checked proof of a 50-year-old math conjecture. Mathematicians haven’t independently verified it yet. Why it’s useful: It’s a real record of an AI-built project going off the rails and being saved. His multi-agent setup produced about 30 “papers” he couldn’t verify, full of made-up terms and circular reasoning. He threw almost all of it away and kept the one result he could verify. The breakthrough came from narrowing the AI’s focus (“no subagents. you work alone SOLELY on reducing conway”). The lesson: when AI output piles up and you can’t verify it, rebuild around the one piece you can verify. When the AI starts inventing its own jargon, it has drifted.