LLMs Are Starting To Noticeably Accelerate Our Work

A LessWrong researcher reports that large language models have meaningfully crossed into accelerating serious mathematical research. Two bounty problems involving natural latents were resolved in recent months using LLMs and the Lean proof assistant: one earning a three-hundred-dollar bounty for a counterexample, another producing a complex LLM-coded proof that a stochastic natural latent implies a deterministic one. The researcher also notes that Claude Code now handles their day-to-day interpretability coding without requiring human oversight—a qualitative leap from two years ago, when LLMs were largely unhelpful on hard problems.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton