Schmidhuber says he published the first concrete RSI algorithm back in 1987
SchmidhuberAI · x · 2026-10-04
Jürgen Schmidhuber pushes back on the current RSI hype by tracing its history: while I. J. Good speculated about an "intelligence explosion" in 1966, he offered no concrete algorithm. Schmidhuber claims his 1987 diploma thesis introduced the first one — Meta Evolution, applying genetic programming to itself across multiple meta-levels.
He lists decades of follow-up work: gradient descent-based RSI in neural networks since 1992, self-modifying policies since 1994, asymptotically optimal RSI for curriculum learning since 2002, and the self-referential Gödel Machine (2003) for mathematically optimal RSI, plus work combining RSI with artificial curiosity.
His argument: with compute now 100,000,000x cheaper than in 1987, RSI is driving the future of both software and physical AI. Details in technical note IDSIA-9-26.
More from AGI Musings
- Brain FLOP/s estimates may bound coding compute but not consciousness — JoshPurtell · 2026-10-04
- Bounding AGI timelines: Josh Purtell on Carlsmith-style Fermi estimates — JoshPurtell · 2026-10-04
- AI writes 20-30% of Microsoft's code: educators debate what schools should still teach — Stefania_druga · 2026-10-04
- Tesla exec: FSD and Optimus are making premium services a mass-market equalizer — yunta_tsai · 2026-10-04
- Cathie Wood: Youth Unemployment at 9.9%, Use AI to Start Your Own Company — rohanpaul_ai · 2026-10-04
- LLMs Aren't Markov Chains: Dev Pushes Back on "Next-Word Parrot" Claim — austinc3301 · 2026-10-04