Schmidhuber: I Published the First Concrete Recursive Self-Improvement Algorithms Back in 1987
SchmidhuberAI · x · 2026-09-23
As Recursive Self-Improvement (RSI) dominates AI discourse, Jürgen Schmidhuber pushes back, noting he published the first concrete RSI algorithms in 1987, when compute was 100 million times more expensive. His technical note IDSIA-9-26 traces decades of related work:
- 1987: first concrete RSI algorithms
- 1990/1997: RSI combined with artificial curiosity and intrinsic motivation
- 1992: gradient descent-based RSI in neural networks
- 1994: RSI with self-modifying policies
- 2002: asymptotically optimal RSI for curriculum learning
- 2003: mathematically optimal RSI via the self-referential Gödel Machine
He argues that now compute is cheap, RSI is driving the future of both software and physical AI — a characteristic priority claim, but also a useful map of how RSI evolved from theory to frontier buzzword.
More from AGI Musings
- Extropic's beffjezos: everyone optimizes intelligence per dollar, not raw intelligence per second — beffjezos · 2026-09-23
- Researcher bets next-gen models at half the params will match first-gen SoTA — tokenbender · 2026-09-23
- Eric Weinstein: mathematicians complaining about AI in math aren't the Einsteins — tak3sh8 · 2026-09-23
- Once open-source models cover 90% of human work, costs could drop 1000x — tokenbender · 2026-09-23
- AI Twitter Debate: Stacking NVIDIA B300s Can't Beat Hackers Who Started Young, JEE Blamed for Killing Talent — itsOmSarraf_ · 2026-09-23
- Major model release gap shrinks from 73 days in 2023 to 18 days in 2026 — MickeySteamboat · 2026-09-23