Schmidhuber: I published the first concrete RSI algorithms back in 1987
SchmidhuberAI · x · 2026-09-19
Jürgen Schmidhuber seized on CNN coverage of recursive self-improvement (RSI) to reassert priority — and to hit back at the "AI stepfathers" who republished basic AI techniques without citing their origins.
His timeline:
- 1987: first concrete RSI algorithms, when compute was 100,000,000x more expensive
- 1994: RSI with self-modifying policies
- 1992: gradient descent-based RSI in neural networks
- 2002: asymptotically optimal RSI for curriculum learning
- 2003: mathematically optimal RSI via the self-referential Gödel Machine
- 1990/1997: RSI combined with artificial curiosity and intrinsic motivation
He argues cheap compute is now making RSI drive the future of both software and physical AI, and links a new technical note (IDSIA-9-26).
More from AGI Musings
- Stephen Witt raises AI safety alarm in New York Times essay: 'It's bad. Really bad.' — MariusHobbhahn · 2026-09-19
- Anthropic warns of catastrophic AI risk while building labs where AI runs scientific equipment — VraserX · 2026-09-19
- NeurIPS desk-rejected 178 papers as AI-generated — then the detector flagged the chairs' own papers too — CSProfKGD · 2026-09-19
- AI community clashes over 'rogue agents' narrative: Redwood and METR quotes fuel the debate — Dan_Jeffries1 · 2026-09-19
- Fields Medalist Villani 'shaken' by OpenAI's claimed millennium problem solution — akbirthko · 2026-09-19
- Most Jobs Don't Involve Einstein-Level Science — and Breakthroughs May Not Boost Growth — binarybits · 2026-09-19