METR's hidden Notes page: unpolished research on agent safety and recursive self-improvement

dejavucoder · x · 2026-09-29

METR maintains a little-known Notes page of rough research updates: a per-action blocking monitor for safer evals, whether LLMs have accelerated discovery rates in cyber/math/algorithms, metrics of agent ability, the economics of recursive self-improvement, an argument that Anthropic's 8x code-merge stat implies >2x researcher uplift from coding agents, NanoGPT speedrun evidence on AI R&D acceleration, and fine-tuning experiments on CoT controllability.

Original post →

More from Research

Research channel →