"The most important question in the world right now is the extent to which AIs are accelerating research internally." METR asked whether AI has actually accelerated discovery: • Cyber vulnerabilities: sharp acceleration in 2026, across individual projects and aggregate databases • Math: likely acceleration, but three problems solved from preexisting lists this year is weak evidence • Optimizations: no clear slope change, even on NanoGPT and CIFAR-10 where LLMs contributed to the records •…