Recursive Self-Improvement (RSI)

An AI system that improves its own ability to improve itself, so each gain makes the next gain easier or faster. The output of one improvement cycle is a more capable improver, which then runs the next cycle. The concern and the interest both come from the compounding: if the loop holds, capability grows faster than linearly.

How the loop works

A system reaches a level where it can modify its own design (its code, weights, training process, or the tools it builds).

1. It applies that ability to raise its own capability.
2. The now-more-capable system repeats step 2 from a higher baseline.
3. Each pass shortens the effort needed for the next pass, which is what separates RSI from ordinary iterative optimisation where the improver stays fixed.

Where the term is used

1. AI safety and alignment research, as the mechanism behind a fast capability takeoff ("intelligence explosion", proposed by I.J. Good in 1965 as an "ultraintelligent machine").
2. Systems that automate their own engineering: models that generate training data for their successors, write and test their own code, or design better model architectures.