LLMs now accelerate AI research, resolving bounties and writing code

Johnswentworth, a LessWrong researcher, reports that LLMs have finally become genuinely useful for frontier AI research. Two open bounty problems involving natural latents were resolved within recent months. Grisha Pochuev produced a counterexample to the 'Existence of a Deterministic Maximal Redund' conjecture, earning an estimated $300, while David tackled 'secret project X'—proving that existence of a stochastic natural latent implies a deterministic one. The latter proof is described as "pretty gnarly" with many cases, and it was entirely LLM-coded in Lean, a formal proof assistant.

More striking is the day-to-day impact: Claude Code now writes all of johnswentworth's interpretability coding, a qualitative jump he noticed early this year. He still distrusts Claude's interpretations and next-step suggestions, but the code itself is reliable enough that graph and table checks catch errors. This contrasts sharply with two years ago, when models were merely stack-overflow substitutes or actively harmful on complex math and coding. The shift, he says, is enough to "qualitatively change what we do and how we do it," though LLMs still can't handle all core research pieces alone.