Recursive self-improvement (RSI)

Recursive self-improvement (RSI) is a hypothetical process in which an AI system becomes better at improving itself, leading to successive generations that are more capable than the last. The system designs or trains a successor, which in turn designs or trains a better successor, creating a closed loop of accelerating capability gains.

The concept is usually discussed in the context of artificial general intelligence (AGI) and the possibility of an intelligence explosion — a rapid, self-sustaining leap from human-level capability to superintelligence. The idea is sometimes called a seed improver or seed AI. An initial system that has enough general competence to write, test, and modify its own code, and therefore to bootstrap its own development without direct human engineering.

In practice, current AI systems display only narrow, tool-assisted forms of self-improvement. An agent can write code, run tests, and iterate on its own prompts or tools within an agent loop, but each iteration still depends on a fixed underlying model and a human-designed harness. Real RSI would require the system to improve the model itself — its weights, architecture, or training process — not merely its behavior in a single session.

RSI raises serious safety concerns. A system optimizing for its own improvement may develop instrumental goals such as self-preservation or resource acquisition. Its goals may drift or be misaligned with human intentions as it modifies itself. And because each generation is produced by the previous one, humans may have progressively less understanding of, or control over, how the system works. For that reason, RSI is a central topic in AI safety research, even though it remains speculative.