OpenAI says recursive self-improvement has begun: top researchers burn $23K/day on agents
新智元 · wechat · 2026-11-19
OpenAI's new research-acceleration report argues the recursive self-improvement (RSI) loop is forming—AI is now materially helping build the next generation of AI.
Key data: median researcher spend on agents hit $2,400/day in October (up 62% MoM), with top-decile users burning $23K/day (+105%). Average 'agent work-days' per researcher jumped 11x to 5.7 (8 for pre/post-training teams); over 1 in 10 researchers run 16+ concurrent agents. In an internal survey, the median researcher said they'd trade away 40% of teammates or 30% of compute to keep coding agents better than GPT-5.2.
While developing GPT-6.2 Astra, core research insights emerged from dialogues with Codex—researchers admit it spotted key ideas and unblocked dead ends ('research taste'). Yet it's still a 'jagged frontier': zero-intervention success on 2-4 hour R&D tasks doubled to 46%, but writing low-level inference code sits at 25%. Humans are shifting from creators to reviewers. OpenAI calls for mandatory public RSI metrics, noting its internal ECI capability index is rising 16.4 points/year on just 2% of total R&D compute.
More from AGI Musings
- Prediction: in 20 years no one will use non-personalized UIs — StewartalsopIII · 2026-11-19
- "China should stop trying to build frontier models": hot take sparks debate — rosstaylor90 · 2026-11-19
- Erik Hoel: philosopher Nagel's denial of LLM consciousness isn't an obvious mistake — erikphoel · 2026-11-19
- Philosopher Thomas Nagel singles out Anthropic's approach to AI as 'abominable' — GaryMarcus · 2026-11-19
- Zvi's AI #195: Anthropic IPO day, Claude Fable 5.2 vs GPT-6.2, Google's new models — TheZvi · 2026-11-19
- Bob McGrew: Superintelligence Needs the Physical World — the Hard Takeoff Is Wrong — No Priors · 2026-11-19