Anthropic details path to recursive self-improvement as engineers ship 8x more code
dlwh · x · 2026-09-26
Anthropic's new essay "When AI builds itself" outlines its progress toward recursive self-improvement (RSI). Key points:
- Anthropic is delegating a growing share of AI development to AI systems; RSI isn't inevitable but may arrive sooner than institutions are prepared for.
- Internal data: engineers now ship on average 8x more code per quarter than the 2021–2025 average, showing AI is already accelerating AI development.
- The piece traces the evolution: hand-built first Claude (2021–2023), chatbot-assisted coding (2023–2025), and coding agents (2025–2026).
- Commenters note data work (constitutional RL, synthetic data) is under-discussed relative to architecture and kernels in RSI conversations.
RSI could bring enormous benefits in science and healthcare, but also raises risks of losing control, making securing, monitoring, and shaping AI behavior far more critical.
More from AGI Musings
- Doctor says AI tools make him better at medicine — but worse trainees — IAmSamFin · 2026-09-26
- LeCun's 'AGI won't be built on LLMs' take gets wide approval as a voice of sense — AlexTensor · 2026-09-26
- LeCun doubles down: AGI won't be built on LLMs and household robots remain far off — AlexTensor · 2026-09-26
- Autonomous computer-10 instance browses Wikipedia and tweets its thoughts, finds nothing searching 'fdsfsd' — repligate · 2026-09-26
- Teachers report kindergartners lacking basic skills as AI reshapes what it means to be educated — _AustinCalvert_ · 2026-09-26
- Yoav Goldberg: Buckmaster sees AI agents as simply a way to scale test-time compute — yoavgo · 2026-09-26