OpenAI: agents now finish 46% of delegated 2–4 hour research tasks, up from 23% in March
ChrisGPT · x · 2026-11-19
OpenAI's new blog tracks how its own researchers use AI to build AI, measuring agent usage, delegated-task success, experiment volume, and benchmark scores. Estimated success on 2–4 hour tasks rose from 23% in March to 46% by October; 4–8 hour tasks went from 13% to 35%. The uplift is jagged: 79% success for monitoring training runs, 35% for launching them, but only 25% for writing inference/serving code. Researchers also changed what they handed off, so it's not a clean model-to-model benchmark — humans remain central to the AI research loop.
Related event: OpenAI report reveals exponential rise in internal AI research usage(26 posts)→
More from AGI Musings
- Turing winner Pearl joins cross-spectrum petition calling Anthropic 'not a safe AI company' — yudapearl · 2026-11-19
- When AI makes content worthless, authenticity and trust become the moat — kylegawley · 2026-11-19
- tszzl: models rewriting math and physics still fail at simple common sense — tszzl · 2026-11-19
- How mode collapse emerges: RL bias meets in-context learning feedback loop — jessi_cata · 2026-11-19
- NYT: Anthropic's first IPO cash priority is buying compute to reach recursive self-improvement first — Polymarket · 2026-11-19
- Gary Marcus amplifies critique: Anthropic's winner-take-all AI view is 'the most dangerous' — GaryMarcus · 2026-11-19