Recursive self-improvement describes a system that improves its own design, then uses the improved version to improve itself again. The word doing the work is recursive: the output of each round becomes the tool for the next. A compiler that compiles a better compiler is the mundane version — it happens routinely, and nothing dramatic follows. The AI version raises the stakes by proposing that the thing being improved is general intelligence itself, so that every round produces a system better at the very task of improving systems.
The claim that this produces an explosion rests on a premise that usually goes unstated: that the returns do not diminish. For the curve to run away, each round of improvement must arrive at least as easily as the one before it. Most optimisation processes behave in precisely the opposite way — the cheap gains are taken first and the later ones cost more, because the easy improvements get used up. Whether intelligence is the rare domain where this does not happen is an empirical question. It has never been settled. It is usually assumed.
This is not a question that early evidence can settle, and the reason is counterintuitive. A system with diminishing returns improves fastest at the very beginning, while the cheap gains are still available. A system with compounding returns starts slowly and only pulls ahead much later. For a long stretch, the trajectory that eventually explodes is the least impressive one in the room. Rapid progress today is consistent with both stories, which means it is evidence for neither.
Then there is the matter of what improving itself concretely requires. A modern model does not rewrite itself in place. Improvement means designing better architectures, assembling better training data, and running training that consumes enormous quantities of computation, electricity, and time — and none of those constraints dissolve because the designer happens to be a machine. Narrow versions of the loop already exist and are genuinely useful: systems that help design chips, tune their own training settings, or generate data for their successors. They are also slow, expensive, and bound by the same supply chains as everything else.
Stripped of the AI framing, recursive self-improvement is an ordinary human experience: getting better at something and discovering that the getting-better itself speeds up. Learning to read makes everything else easier to learn. A good tool helps you build a better tool. This is real, and it compounds — and anyone who has learned a difficult skill also knows that the curve eventually flattens, usually without announcing itself. The honest question is not whether machines can improve themselves. In limited ways they already do. It is how long the compounding lasts before it meets something that does not yield.
