OpenAI data: agents still fail critical research tasks, a bottleneck for recursive self-improvement

Dan_Jeffries1 · x · 2026-11-21

OpenAI released new data on how AI is changing research internally: researchers' AI use keeps rising exponentially and agents complete longer tasks, but human involvement remains central, with time spent on safety and alignment growing faster than any other research area. OpenAI argues AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI).

Dave Holtz revealed a potential bottleneck for RSI: estimates of how often OpenAI's internal agents complete tasks fully autonomously show high success rates for tasks like monitoring runs and code review, but low success rates on critical tasks such as writing inference/serving-stack code and validating experiment results. If a critical task bottlenecks everything else and AI can't do it autonomously, that's a major obstacle to further research acceleration.

Original post →

More from AGI Musings

AGI Musings channel →