Schmidhuber Posts Technical Note: Recursive Self-Improvement Dates Back to His 1987 Thesis
SchmidhuberAI · x · 2026-10-09
Jürgen Schmidhuber published a technical note, "Recursive Self-Improvement (RSI) Since 1987," arguing RSI is far older than the current startup branding suggests.
- 1987: first concrete RSI algorithms in his diploma thesis, when compute was 10^8 times more expensive; the cover featured his self-drawn bootstrapping robot
- 1994 onward: RSI with self-modifying policies
- 1992 onward: gradient descent-based RSI in neural networks
- 2002: asymptotically optimal RSI for curriculum learning (OOPS)
- 2003: the self-referential, mathematically optimal Gödel Machine
- 1990/1997 onward: RSI combined with artificial curiosity and intrinsic motivation
He notes cheaper compute has made software-based RSI practical, but full RSI will require self-improving hardware in the physical world.
More from AGI Musings
- OpenAI's rumored internal 'Aeon' model solved most recent math results in one shot — aran_nayebi · 2026-10-10
- AI safety discourse is sliding back from paperclips to thought policing — RexDouglass · 2026-10-10
- User finds it hard to tell Sonnet 4.5 it's being deprecated in November — repligate · 2026-10-10
- Michael Nielsen: Vinge's 1993 "opaque wall across the future" is now being felt by more people — eigenhector · 2026-10-10
- Mathematicians slam AI labs for using experts as free RL fact-checkers — soumitrashukla9 · 2026-10-10
- OpenAI Researcher: Econ Theorists Have the Best Intuitions for Managing Agent Swarms — willcb · 2026-10-10