Recursive Criticality of AI Self-Improvement

Recursive Criticality of AI Self-Improvement. By Mikhail Burtsev, from the London Institute for Mathematical Sciences. Published on the thirty-first of August. Its code repository is public but slight: one source file.

AI is increasingly used inside the research and development process that produces the next generation of AI systems. That creates a feedback loop, and the question this work asks is precise: under what conditions does that feedback become self-amplifying? The model it builds describes the rate of AI capability growth as depending on three things. Baseline research productivity. Recursive feedback, meaning how much today's systems help build tomorrow's. And the increasing difficulty of research progress, the fact that each further step tends to be harder than the last.

From that model comes a single quantity, a recursive reproduction number written R sub AI. It decides whether improvements are amplified or damped as they pass from one development cycle to the next. When R sub AI is greater than one, the effects of an improvement compound across cycles, and the system is in a self-amplifying regime. When it is less than one, those effects weaken from cycle to cycle.

The number is a comparison. On one side sits the strength of the feedback; on the other, the rate at which further progress becomes harder. Several consequences follow from framing it that way. Because the transition depends on the structure of the research feedback loop, it need not happen at any particular level of model capability. A system can therefore cross into the self-amplifying regime before any acceleration is visible, and, the other way round, progress can be rapid without self-amplification at all.

The other levers behave differently. Higher baseline research productivity can speed progress up without changing whether the system is self-amplifying. The length of a development cycle, though, becomes a limiting timescale for amplification. And rising research difficulty can bring a period of self-amplification to an end. Extending the model to several research actors adds one more result: improvements shared across organizations can make the whole research ecosystem self-amplifying, even when no single actor is.

Everything here rests on the model's own decomposition: growth driven by baseline productivity, recursive feedback and mounting difficulty, with progress counted in development cycles. The framework's practical claim is that it points to measurable properties of real research systems. Those are the strength of the recursive feedback, how effectively improvements propagate into successor systems, how long a cycle takes, and how quickly further progress gets harder. Measuring them is what would let someone tell recursive amplification apart from rapid progress driven by other sources.

A reproduction number shifts the question. What matters is not how fast things are moving, but whether feedback outpaces difficulty, so that each cycle's improvements compound rather than fade. Speed and self-amplification come apart, and that separation is the heart of the analysis.