A Mathematician's Field Notes: AI Gives +50% Productivity but Still Can't Abstract

LessWrong 精选 · rss · 2026-11-13

A LessWrong author's first-person account of using AI for "real" math since autumn 2025:

50% actual speedup: AI finishes in minutes what once took hours, but doing math involves idea generation, exposition, reorganization — tasks AI only partially handles, so overall gains are smaller than expected (a 10× faster plane doesn't shorten door-to-door travel proportionally).
Humans aren't obsolete: AI's core weakness is that its proofs "don't make any goddamned sense" — strong technical skill but poor modular abstraction, producing novice-style "spaghetti-code math."
Abstraction is objective, not taste: after arguing with the AI to try simpler proof strategies, proofs get dramatically shorter and sometimes yield stronger results (x < 9.7 becomes x < 7.1), so human-style abstractions are "objectively" useful.
Practical tips: investing in AGENTS.md, custom skills and harnesses has poor ROI given the short half-life; the one phrase that really helped — "ask independent agents for an adversarial review."
Beware the receding green light: when AI almost gets it right is when hours vanish with nothing learned; restarting solo or "on a very tight leash" takes 30 minutes.

The author expects these weaknesses to improve but argues that as of autumn 2026, humans still have a lot to contribute to mathematics.

More from AGI Musings