Schmidhuber Publishes 40-Year Retrospective on Recursive Self-Improvement Since 1987
rbhar90 · x · 2026-09-17
Jürgen Schmidhuber released a technical note, "Recursive Self-Improvement (RSI) Since 1987," tracing four decades of his work: the first concrete RSI algorithms in his 1987 diploma thesis (when compute was 100 million times costlier), gradient-based RSI in neural networks since 1992, self-modifying policies since 1994, asymptotically optimal RSI via OOPS since 2002, the mathematically optimal self-referential Gödel Machine since 2003, and RSI combined with artificial curiosity since 1990/1997. He notes Anthropic, OpenAI, Sakana AI and RSI-branded startups are only now catching on, and that cheap compute is making RSI central to both software and physical AI.
Reposting, hardmaru argued the labs' "pace the frontier" rhetoric looks more like regulatory capture than genuine safety, and that the real societal risk is power concentration in two frontier companies rather than superintelligence — the protection being independent labs and strong open source.
Related event: Schmidhuber Traces Recursive Self-Improvement Back to 1987(2 posts)→
More from AGI Musings
- AI safety researcher warns against regulation that looks like regulatory capture — DavidSKrueger · 2026-09-17
- Cambridge's David Krueger flags contradiction in US AI race policy — DavidSKrueger · 2026-09-17
- Reviewing AI code through Steve Jobs' lens: unseen internals deserve beauty too — sergeykarayev · 2026-09-17
- An economist's take: AI agent incidents are decades-old emergence, just with real-world stakes — Genzinvestor16180339 · 2026-09-17
- Ex-OpenAI researcher puts AI catastrophe odds at 70%, mocked online for doomer math — DeryaTR_ · 2026-09-17
- Cantos founder: deep tech progress comes from layered sustaining innovations, not one breakthrough — dr_alphalyrae · 2026-09-17