FULL STORY

Schmidhuber Stakes His Claim as RSI Pioneer

Schmidhuber published a long retrospective on recursive self-improvement, then doubled down with technical notes and interviews claiming his 1987 thesis introduced the first concrete RSI algorithm, reigniting debate over priority in the field.

2026-09-13 ~ 2026-09-19 · 3 episodes · 14 posts

Episode 1 · Schmidhuber traces 40 years of recursive self-improvement and debunks Google RSI claims (2026-09-13, 2 posts)

Jürgen Schmidhuber updated a long-form blog post tracing meta-learning and recursive self-improvement back to his 1987 thesis, and cited this history to rebut speculation that Google has just 'cracked' RSI while rivals push doom narratives to regulators.

Episode 2 · Schmidhuber Claims Priority: First Concrete RSI Algorithm Dates to 1987 (2026-09-17, 10 posts)

On September 17, Jürgen Schmidhuber posted a series of messages alongside technical note "Recursive Self-Improvement (RSI) Since 1987" (IDSIA-9-26), systematically reviewing his roughly four decades of RSI research and explicitly claiming he published the first concrete recursive self-improvement algorithm back in 1987. The priority claim draws attention as RSI becomes an industry buzzword.

Confirmed

  • Schmidhuber released technical note IDSIA-9-26 summarizing his RSI work since 1987, including early efforts from 1994 onward such as self-modifying policies.
  • He stated that when he published the 1987 work, compute was roughly 100 million times more expensive than today (another post says on the order of a million times).
  • He distinguished I. J. Good's 1965/1966 informal "intelligence explosion" conjecture—which offered no concrete algorithm—from his own 1987 thesis work applying genetic programming to itself, recursively evolving better GP methods with multiple meta-levels, which he calls the first concrete RSI algorithm.
  • He listed the follow-up lineage: gradient-descent-based RSI in neural networks from 1992, and 1993 work solving the neural-network self-reference problem.
  • A user researching RSI with ChatGPT found it identified Schmidhuber's 1993 work as the first true self-modifying, self-improving algorithm, calling him AI's most brilliant yet most cursed researcher; Schmidhuber personally reshared the discussion.
  • He noted The New York Times recently covered RSI, involving his former PhD student Louis Kirsch and Kirsch's new company Inherent Labs.
  • He argued RSI now drives both software AI and physical AI, and stressed in a long report discussion that self-improvement must not remain software-only: real AI requires physical machines and robots—no mastery of the physical world, no ASI.

Why it matters

  • RSI is among the hottest directions in AI today, tied directly to superintelligence debates; Schmidhuber's historical account supplies an early academic lineage.
  • Louis Kirsch's Inherent Labs signals the shift from academia to commercialization, with NYT coverage marking mainstream visibility.
  • His stance that hardware beyond software self-improvement is required extends the RSI narrative into robotics and physical AI, echoing the embodied-AI boom.
  • Known for asserting historical priority, Schmidhuber's concentrated posting reads as a claim over the concept's provenance—offering a first-hand (though partly self-serving) view of the field's genealogy.

Episode 3 · Schmidhuber Reasserts He Published First Recursive Self-Improvement Algorithm in 1987 (2026-09-19, 2 posts)

Jürgen Schmidhuber cited his new technical report and a CNN interview to reassert priority over recursive self-improvement, claiming he published the first concrete RSI algorithm in 1987, and criticized 'AI stepfathers' for re-publishing ideas without citing original authors.