I vibed a proof of Conway's conjecture
summary
A non-mathematician successfully obtained a Lean proof of Conway's refinement conjecture by using a multi-agent AI workflow. The process required iterative cycles of mathematical exploration, rigorous formal verification via Lean, and strategic 'burning down' of hallucinatory results. While AI can drive the discovery of proofs, it currently lacks the the self-discipline to avoid logical drift and 'process theater' without human project management.
key points
- The author utilized a multi-agent system featuring specialized roles like 'Math' agents for discovery, 'Red' agents for skepticism, and 'Lean' agents for formal verification to prevent hallucinations.
- A critical turning point occurred when the author shifted from trusting AI-generated 'papers' to grounding all progress in Lean certificates, ensuring each new claim was mechanically verified before building upon it.
- Effective AI-driven research requires strict terminology control and the separation of stable, peer-reviewed mathematics from experimental, high-risk proofs to avoid conceptual contamination.
who should read it
AI enthusiasts, programmers interested in formal verification, and mathematicians curious about the intersection of LLMs and automated theorem proving.
discussion positive
consensus
Commenters agree that the author's non-expert approach to mathematics makes the content more accessible and engaging than traditional academic lectures. There is a broader discussion comparing traditional deep technical mastery ('wizardry') with the use of LLMs to achieve results without fundamental understanding ('sorcery'). While some see this shift as a loss of validation capability, others view it as a stress test of meta-skills and a new way to interface with knowledge.
top insight
Dan Abramov describes his approach as 'epistemic performance art,' using LLMs to see how much knowledge work can be accomplished without the underlying expertise. He posits that this shifts the required skill from domain knowledge to the ability to constrain 'demons' (AI) into meeting specific requirements, essentially treating technical breakthroughs as a management problem.
Generated by an LLM from the article and thread. Check the sources before relying on it.