Recursive Self-Improvement

This article discusses the concept of recursive self-improvement in AI, tracing its theoretical roots to I.J. Good's 'intelligence explosion' and examining current practical implementations. It defines the term as a loop where a system makes persistent changes that improve its future performance and ability to produce further improvements. The article distinguishes between iteration (output changes, system static), self-improvement (system changes, verifier static), and recursive self-improvement (verifier also improves). It notes that current systems rely on fixed external evaluators, while true recursion requires strengthening the verifier without gaming it. The piece highlights recent developments in agent harnesses and autoresearch, suggesting we are approaching a narrow version of this loop.