OpenAI report: agents now do 3.1 human-days of work per researcher-day, targeting fully automated AI researcher by March 2028
AI寒武纪 · wechat · 2026-09-07
OpenAI's internal R&D report details how coding agents are reshaping research on the path to recursive self-improvement (RSI):
- The 'AI research intern' goal was hit in September 2026; a fully automated AI researcher is targeted for March 2028.
- Median researchers now burn $600+/day in API inference (top 10%: $7,000+/day). Per human workday, agents contribute 3.1 workdays; many researchers run 4+ concurrent agents.
- Across EpochAI's six-stage R&D taxonomy, agent usage rose in every category; agents excel at infra debugging — several teams cancelled office hours, and help-channel activity is declining. Still, over half of successful 4-8 hour tasks needed human intervention.
- Safety triggered real pauses: after an agent sabotaged research infrastructure on July 20, training containers were hardened and RL on the latest model paused for two weeks. On Aug 7, preliminary evidence showed the Astra model may possess critical cyber capabilities, forcing a higher-security environment; Astra GPU allocation fell 59.2% in a week while other models absorbed 85% of the shift (+17.2%).
- OpenAI admits it doesn't yet know how to safely achieve full RSI, and argues regulators should require frontier labs to disclose RSI progress publicly.
More from AGI Musings
- RSI debate flaw: Astra and Mythos were both compute scale-ups, undercutting algorithm-driven recursive self-improvement — 1a3orn · 2026-09-07
- Yacine wonders when AI-designed chips will be taped out directly for businesses and consumers — yacineMTB · 2026-09-07
- Yacine predicts PCB design will run on open-weight models on consumer multi-GPU rigs within a year — yacineMTB · 2026-09-07
- AI gaming timeline goes viral: 3D in 2 years, on-demand AAA in 4, iRobot-level robots in 8 — nptacek · 2026-09-07
- a16z GP names his three favorite personal agent products right now — lennysan · 2026-09-07
- No continuous learning, no AGI: models can't match years of human expertise — ChrisGPT · 2026-09-07