Schmidhuber: I published the first concrete RSI algorithms back in 1987
SchmidhuberAI · x · 2026-09-18
Responding to the current buzz around Recursive Self-Improvement, Jürgen Schmidhuber argues the idea is far older than most assume, citing his own timeline: the first concrete RSI algorithms in 1987 (when compute was 100 million times more expensive), gradient-descent-based RSI in neural networks since 1992, self-modifying policies since 1994, asymptotically optimal RSI for curriculum learning since 2002, the mathematically optimal self-referential Gödel Machine since 2003, and RSI combined with artificial curiosity since 1990/1997. He says cheap compute now makes RSI the driving force behind both software and physical AI.
More from AGI Musings
- Drugs can program subjective experience, so the brain is computational — ctjlewis · 2026-09-18
- Is Consciousness Computation? A Debate From Roundworm Neurons to Drug-Altered Minds — ctjlewis · 2026-09-18
- Are 'Applications' Just a Historical Compromise? Dynamic AI Logic Could Kill SaaS — YogeshMalik · 2026-09-18
- mark_k says he's fighting two overlapping ideologies: anti-AI and wokeness — mark_k · 2026-09-18
- Critics blast LW/EA p(doom) culture: thousands reportedly consumed by AI doomerism — banteg · 2026-09-18
- UK most AI-exposed economy but lacks compute and leverage, new policy paper warns — dc_lawrence · 2026-09-18