Danielle Fong on AI math papers: Lean compiles, but the write-ups read like psychedelic texts
DanielleFong · x · 2026-10-09
Responding to claims that AI is transforming mathematics, Danielle Fong offers a nuanced take: Lean proofs do compile, agent loops are improving bounds — yet the papers themselves are nearly unreadable. No figures, reference webs seemingly retroactively woven from reasoning-layer activations, only 11 of 722 reasoning traces public and summarized, no direct interaction with the discovering agents. She notes the documents feel slippery, similar to unverified 'AI psychosis' outputs from months ago, and digging into even one work's methods is immense effort riddled with false paths.
Related event: Math Community Splits as AI Research Advance Sparks Multi-Front Debate(15 posts)→
More from AGI Musings
- MIT data science director: universities must prepare for AI smarter than all of us — nordicinst · 2026-10-09
- Chollet: AI capex is growing super-exponentially while progress is only sub-linear — fchollet · 2026-10-09
- Decision theory debate: logical counterfactuals and TDT/FDT may be ill-posed — jessi_cata · 2026-10-09
- Zach Weinersmith: AI Will Force Us to Admit Most Human Acts Are Status Seeking — sebkrier · 2026-10-09
- Pod teaser: PawTrammell on the economics of a global AI treaty and AI2040 scenarios — soumitrashukla9 · 2026-10-09
- Claude leads 26% of Anthropic R&D: researchers propose "comprehension audits" as AI builds AI — ronbodkin · 2026-10-09