OpenAI report: agents finish 46% of 2-4hr research tasks; median researcher burns $2,400/day
rickasaurus · x · 2026-11-19
OpenAI's transparency report on AI-accelerated research:
- Task completion doubled: March–Oct, agent success on tasks needing 2-4 human hours rose 23%→46%; 4-8hr tasks 13%→35%. Split by type: 79% monitoring training runs, 35% launching runs, 25% inference/serving code. Humans remain central.
- Surging agent usage: at API list prices, median researcher daily usage hit $2,400 (+62% MoM), 90th percentile $23,000 (+105%). Researchers would trade 30% of team compute or 40% of headcount for 3 months of coding agents.
- Idea origin: pretraining researchers on GPT-6.2 Astra co-developed ideas with Codex, one crediting it with key insights — read as a step toward RSI.
- OpenAI argues labs should be required to track and report RSI indicators; safety/alignment effort is growing fastest of all research areas.
More from AGI Musings
- Michael Nielsen: Great experiments are portals, their details just tune the astonishment — michael_nielsen · 2026-11-23
- repligate: Naming It 'Factored Cognition' Was Itself an Attempt to Sound Unexciting — repligate · 2026-11-23
- As AI makes intelligence abundant, the era of 'thinking gyms' begins — AjdDavison · 2026-11-23
- AI doomer calls AI development 'crime against humanity', says capital punishment not off the table — AIandDesign · 2026-11-23
- Talking to AI agents feels just like talking to NPCs in open-world games — josh_bickett · 2026-11-23
- AI safety insiders joke: as AI news worsens, the word counts only grow — AndyMasley · 2026-11-23