Recursive Self-Improvement Hits KPIs: Claude Leads 26% of Anthropic R&D, OpenAI Runs 3.1 Agent-Days Per Human Day
机器之心 · wechat · 2026-10-10
A deep-dive recap of how Recursive Self-Improvement (RSI), an idea dormant for 60 years, became a tracked metric at frontier labs in 2026.
- Labs: Anthropic reports 80%+ of its codebase is now written by Claude, with Claude leading 26% of R&D tasks (up from 1% in March). OpenAI hit its "research intern" milestone, with research teams now running 3.1 agent-days per human workday; mid-August median daily inference spend per researcher exceeded $600.
- China: Zhipu's GLM-5.3-powered InfraAgent built a production inference service from scratch on a 100K+ domestic-chip cluster, boosting throughput 3x; MiniMax's M2.7 claims to be the first model deeply involved in its own evolution.
- Experiments: Karpathy's autoresearch ran 2 days on nanochat, delivering 11% speedup and catching a QK-Norm scaling bug he missed; Weco's AIDE² two-loop system beat a two-year hand-tuned pipeline in 8 days, though its own "ignition test" was inconclusive.
- Reality check: Princeton's "shadow evaluation" had agents independently reproduce unpublished NeurIPS submissions—both papers were rejected. Today's AI is "an extraordinary engineer" but a poor open-ended researcher; human advantage remains in research taste and judgment.
More from AGI Musings
- Humanities dissertations have collapsed, social sciences are next — with AI as one driver — gsiemens · 2026-10-23
- Greg Brockman: AI makes hacking 100x more profitable and easier, upending hacker economics — gabriel1 · 2026-10-23
- 'Nothing is real now on social media': viral demo shows AI-generated content everywhere — saibharadwaj · 2026-10-23
- Lilian Weng to Discuss Recursive Self-Improvement at SGLang Summit 2026 — BanghuaZ · 2026-10-23
- mmitchell_ai questions whether the AI safety vs security framing stems from EA — mmitchell_ai · 2026-10-23
- AI, Robotics and Abundant Energy Will Undo Millennia of Scarcity-Driven Life, Argues Kardashev-Scale Optimist — Dr_Singularity · 2026-10-23