Recursive self-improvement, often shortened to RSI, describes a loop in which an AI system gets good enough at AI research to make better versions of itself, which then do the same. If each round yields larger gains than the last, progress would speed up instead of following the steady curve seen in scaling laws. The idea was set out in 1965 by the mathematician I. J. Good, who called the result an “intelligence explosion.”
How it differs from ordinary AI assistance
Today’s AI tools already help researchers write code, run experiments and search for algorithms. That is assistance: people still choose goals, supply compute and decide what is released. RSI would mean the system itself takes over most of the research loop and the gains compound. Whether that is possible, and how fast it could happen, is unsettled.
Why labs mention it
Safety policies treat AI-driven acceleration as a risk because safety checks could fall behind capability. Anthropic’s Responsible Scaling Policy includes a threshold for dramatically accelerated progress attributable to AI, and the company’s amended IPO filing says it may imminently meet that definition, though it has not designated any current model as past it. See our report on Anthropic’s S-1 amendment.
What could limit it
Compute and energy supply, finite high-quality data, the difficulty of evaluating a successor on tasks nobody has measured, and human decisions to keep people in the loop can all slow or stop a runaway loop.
For the history and the evidence, read our explainer on recursive self-improvement.
