An AI improving its own capabilities and using each improved version to improve itself again, in a compounding loop; the mechanism behind the intelligence-explosion argument toward AGI.
Recursive Self-Improvement is an AI system improving its own capabilities and then using the improved version to improve itself again, in a compounding loop. Each cycle raises the system's ability to design the next one, so progress can accelerate rather than hold steady.
It is the mechanism behind the "intelligence explosion" argument: once a system can enhance its own intelligence better than its human designers can, the gains could feed on themselves and run toward artificial general intelligence and beyond. Whether that acceleration would be fast (a hard takeoff over days) or slow (a gradual climb over years) is unsettled and debated.
Partial, bounded forms already exist. Models generate synthetic training data to train their successors, write and optimise code, help design better architectures, and curate their own fine-tuning. These are narrow improvement loops with a human in the loop, not open-ended self-rewrites.
It is a central alignment concern. A system that edits its own goals or capabilities could drift from its intended values faster than oversight can catch it, which is why self-improving reasoning models and agentic systems are watched closely.
