OpenAI report: agents succeed at 25-79% on internal research tasks, humans still central
daveholtz · x · 2026-11-19
OpenAI released a second report updating its September study on how internal AI deployments accelerate research, with new data on AI uplift across research domains.
- Jagged success rates: agents' success varies from as low as 25% (writing inference-stack code) to as high as 79% (monitoring training runs); because of this jaggedness, humans remain central to the research process.
- Surging usage: daily agent usage in October keeps rising, with the strongest growth among the heaviest users, who report large self-assessed value; agents are completing longer tasks.
- Safety focus grows fastest: researcher time on safety and alignment is growing faster than any other individual research area.
- Governance call: OpenAI argues AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI) to inform governance and pacing decisions.
Related event: OpenAI Report: AI Agents Doubling Success on Long Research Tasks(28 posts)→
More from AGI Musings
- LeCun blasts AI doomer letter; Gary Marcus fires back that non-doomers signed too — GaryMarcus · 2026-11-19
- Designer builds 200 scroll animations in 2 weekends, manifesto on design's AI-era job — soleio · 2026-11-19
- Musk on extending life beyond Earth: making sci-fi "not just fiction forever" — vitrupo · 2026-11-19
- 'Not losing your mind will be the only moat left': an AI-era take goes viral — kyliebytes · 2026-11-19
- Baidu and Xiaomi CEOs Warn of Widening Gap Between AI Hype and Returns — pstAsiatech · 2026-11-19
- PE's AI problem: hours saved don't show up anywhere a PE firm gets paid — alex_verem · 2026-11-19