Post

AI is dramatically accelerating AI research right now

Ajeya Cotra interprets a METR assessment of Anthropic as evidence that AI-assisted R&D has substantially accelerated capability progress, and argues that continued compounding could sharply shorten timelines. This is an extrapolation under strong assumptions, not an observed trajectory to superintelligence.

Much of the assessment’s most informative evidence and its attribution reasoning is redacted, limiting independent scrutiny. HN readers disputed the leap from faster software research to unrestricted compounding, pointing to compute, diminishing returns and physical-world bottlenecks; others argued that indirect effects still warrant concern.