Schmidhuber Revives 2020 RSI Talk, Says His Systems Learned Self-Improvement Since 1994

SchmidhuberAI · x · 2026-09-18

Jürgen Schmidhuber resurfaced his 2020 talk on recursive self-improvement (RSI), arguing it has aged well amid today's RSI hype. His core claim: mainstream ML algorithms are designed by humans and limited by human cognitive biases, while his meta-learning systems since 1994 learn to define and redefine their own trial boundaries — something modern meta learners still rely on humans for — enabling unbounded self-improvement.

Original post →

More from AGI Musings

AGI Musings channel →