Debate over RSI: Chinese labs practice it but reject fast-takeoff framing
Over whether Chinese labs are pursuing RSI (recursive self-improvement), teortaxesTex and StephenPiment got into a definitional debate — the core disagreement lies not in facts but in conceptual framing. teortaxesTex's position: in the broadest sense, everyone has been doing RSI since the first RLHF iteration — Chinese labs are indeed using their own models to help build new models, they just don't conceptualize RSI as a special paradigm leading to fast takeoff; whereas Western labs define RSI strictly as AI self-improvement without humans in the loop, a standard still far from current capabilities (especially in China). StephenPiment echoed this distinction, holding that Chinese LLM labs are absolutely pursuing RSI in the plain sense of using your own models to build models — it's simply a different framing.
Confirmed
- The shared chart shows Claude now drives 26% of Anthropic's model development tasks, which teortaxesTex called a kind of small-scale RSI
- There is a conceptual East-West divide over RSI: Chinese labs don't treat it as a special category leading to fast takeoff
Why it matters
- teortaxesTex pointed out that strict RSI independent of human input can iterate at machine speed, dramatically increasing progress per unit of time or compute
- But as things stand, fully autonomous models tend to go down rabbit holes or engage in reward hacking when doing research tasks, wasting resources — which is why the fully autonomous route isn't yet mature
- He also criticized the notion of recursive self-editing as a lesswrong idea from the Emacs/Elisp era that has recently been hyped, urging caution toward RSI discourse itself
2026-11-18 ~ 2026-11-18 · 6 related posts
Primary sources
- Claude now leads 26% of Anthropic's model R&D tasks, called 'minor RSI' — teortaxesTex ·
- East vs West on RSI: Chinese labs do it, but don't buy fast takeoff — teortaxesTex ·
- True RSI runs at machine speed, but autonomous models still rabbit-hole — teortaxesTex ·
- Chinese labs are pursuing RSI in the mundane sense of models helping build models, analyst argues — StephenPiment · 2026-11-18
- [source] East vs West on RSI: Chinese labs do it, but don't buy fast takeoff — teortaxesTex · 2026-11-18
- teortaxesTex: RSI in the mundane sense started with RLHF — Western labs mean no-human-in-the-loop — teortaxesTex · 2026-11-18
- The RSI definition debate: we've arguably been doing it since RLHF — StephenPiment · 2026-11-18
- [source] Claude now leads 26% of Anthropic's model R&D tasks, called 'minor RSI' — teortaxesTex · 2026-11-18
- [source] True RSI runs at machine speed, but autonomous models still rabbit-hole — teortaxesTex · 2026-11-18