Recursive self-improvement, or RSI, is the idea of AI improving not only its performance, but its ability to make further improvements. That could involve changing its code, tools or underlying model. An AI rewriting a function is a start, not a verdict: the interesting question is whether successive versions become better at improving themselves—and whether the gains hold up outside their own tests.
What might achievement look like?
- A clear account of what changed. Which parts did the AI modify, which stayed fixed, and how much did people contribute?
- Repeated gains that travel. Successive versions improve the improvement process, with benefits on unfamiliar tasks—not merely the test used to select the winner.
- Checks outside the loop. Independent evaluation helps rule out easier tests, extra resources or manipulated scoring. Theories of an “intelligence explosion” remain theories, not an automatic consequence of a successful experiment.
AGI & RSI: this week’s developments
Research, experiments and competing claims about general intelligence and self-improving AI—what was reported, what was tested, and what remains an open question.
First roundup coming soon. No news items are included in this draft.
Please note: AGI and RSI have no universally agreed achievement test. These sections explore evidence, theories and thought experiments; they do not certify that either has been achieved. Reported results will be distinguished from independently reproduced findings.
Questions about what this means for you?
Talk to us about where AI fits in your business.
[email protected]