What determines whether AI self-improvement actually compounds?

AI assistance to AI R&D becomes self-amplifying when a recursive reproduction number, RAI, exceeds 1. The excerpt defines it as a comparison between "the strength of recursive feedback" and "the rate at which further progress becomes more difficult." Below 1, incremental improvements weaken across development cycles; above 1, "the effects of incremental improvements compound across development cycles." The discussion restates the condition as "realized recursive gain exceeds the local hardening of the research frontier, so that RAI = χ/aσ > 1." The part that matters most is what the threshold is not. The transition "need not occur at any particular level of model capability," so a system "can therefore enter a self-amplifying regime before acceleration becomes visible, while rapid progress can also occur without self-amplification."

The model describes the rate of capability growth as depending on three things: baseline research productivity, recursive feedback, and the increasing difficulty of further progress. The introduction says recursive self-improvement is "better viewed as a property of an AI–R&D system than of an isolated, fully autonomous agent," one that can include human researchers, laboratories, compute, evaluation procedures and the organizations that deploy successor models. The framework separates "the onset, speed, and persistence of recursive amplification." RAI = 1 settles whether gains amplify, while development-cycle delay "constrains how rapidly that amplification unfolds." In the minimal model, higher baseline research productivity can accelerate progress "without changing whether the system is self-amplifying."

The neighboring note on agents that automate R&D treats the efficiency of the research process as the fixed quantity and argues that self-improvement is the way to move it against diminishing returns. This model asks a different question. It makes rising research difficulty an explicit term, and what it asks of a loop is whether its increments compound, not whether the loop raises research efficiency. The abstract's result sharpens the contrast: if a remedy works by raising productivity, this model says it changes how fast progress runs, not which side of RAI = 1 the system sits on. The survey in Are self-refinement and recursive self-improvement actually the same thing? sorts loops by what they improve and who validates the improvement. This excerpt adds that crossing RAI = 1 "is a local condition for amplification and does not by itself imply indefinite acceleration or unbounded capability growth," so onset and boundedness are separate questions. The ratio of two rates also fits the bottleneck reading in What bottlenecks define the path from AGI to superintelligence?. The excerpt says onset "need not coincide with a particular capability threshold such as AGI," which favors tracking frictions over betting on one timeline, though it never evaluates that note's four pathways.

The excerpt does not define χ, a or σ beyond their roles in the ratio, gives no estimate of them for any system, and says nothing about how they would be measured. It also does not say whether any current system sits above or below RAI = 1, or what an AGI or ASI transition compressed by feedback delay would look like in numbers. The implication is narrow but usable. The model gives a checkable criterion for any claim that AI-assisted R&D is accelerating: does the recursive gain outrun the hardening rate? It does not forecast when, or whether, that condition holds.