Schmidhuber Maps 38 Years of Metalearning and Recursive Self-Improvement Since 1987
SchmidhuberAI · x · 2026-09-13
Jürgen Schmidhuber updated his long-form blog post tracing metalearning and recursive self-improvement (RSI) from his 1987 diploma thesis through today: gradient descent-based neural metalearning (1992-), meta-RL with self-modifying policies (1994-), artificial curiosity and intrinsic motivation (1997), asymptotically optimal metalearning for curricula (2002-), the mathematically optimal self-referential Gödel Machine (2003-), and new work since 2020. He argues in-context learning and test-time training in modern LLMs are special cases of metalearning, and notes that RSI startups are now emerging as compute gets cheaper.
More from AGI Musings
- Game theory of AI slowdown pledges: nobody slows down internally, everyone loses — haider1 · 2026-09-13
- The next SaaS shift: buying software that lets your agent work autonomously — eptwts · 2026-09-13
- AI solved a Millennium Prize Problem while he wrote about it: "We will never have normal times again" — voooooogel · 2026-09-13
- UK pilots three-week AI bootcamps in Preston to turn Neet youths into apprentices — nordicinst · 2026-09-13
- Video argues AI can't take everyone's jobs: excess demand means no mass layoffs — EGarrett28 · 2026-09-13
- Altman, Musk, and Hassabis back Amodei's call for independent AI oversight; OpenAI IPO delayed to 2027 — The Decoder · 2026-09-13