Are AI feedback loops strong enough to sustain recursive self-improvement?

The paper's central claim is arithmetic. Net acceleration in AI capabilities "depends on the product of elasticities across each feedback loop," and the core loop asks how much the next generation improves for a one-unit gain in capabilities. A "back-of-the-envelope calculation suggests that feedback loops are not currently strong enough" to sustain acceleration, "though they appear to be strengthening." The excerpt's supporting material is data points rather than model output: a "roughly 4X" self-assessed productivity uplift among Anthropic researchers, which the excerpt says has "reasons to think this is likely overstated"; a "40-hour time horizon" at which Claude Mythos Preview beat human researchers; Mythos Preview beating humans "64% of the time" in April 2026, "up from 50%" for Claude models released in 2025; and OpenAI's report that internal coding inference's share of research compute "grew 100-fold in the prior six months."

The mechanism is a directed graph of feedback loops, built up from "a sequence of increasingly rich models of AI progress." Because the net figure is a product, a weak elasticity in any one loop pulls the whole result toward zero. That follows from the multiplicative form; the excerpt reports no separate result for it. The paper's second distinction separates "narrow" capabilities, improvement at "AI R&D benchmarks," from "broad" capabilities at "broader economically valuable tasks," and it keeps open the possibility that systems improve narrowly without improving broadly. The excerpt presents that split as a modeling distinction, not a finding.

Set against the nearest notes, the paper supplies economics they leave implicit. What bottlenecks define the path from AGI to superintelligence? lists recursive improvement as one of four pathways; this paper asks whether that pathway becomes self-sustaining. Can recursive self-improvement speed up the research process itself? states the AIDE2 paper's premise that automated R&D speeds the artifacts while the research process stays fixed; the elasticity model asks whether those gains feed back into the next generation. Do fixed-budget efficiency gains translate to real research progress? reads a benchmark gain as a research-efficiency gain, and the narrow-versus-broad split is the caution that such a gain may not reach economically valuable work. Are self-refinement and recursive self-improvement actually the same thing? separates open-ended recursive self-improvement from bounded refinement; the net-elasticity condition here is a quantitative test for the open-ended case that the survey distinguishes.

The excerpt does not establish the calibration. It gives no elasticity values, no detail of the "existing data" the model is calibrated with, and no output beyond the verdict. The 4X and 64% figures are self-assessments and comparisons from cited sources, and the 8–20% forecast concerns the chance that effective compute growth triples by 2029, not a measured loop strength. The excerpt does not link any of them to the model's parameters. The verdict is therefore a conclusion about current inputs at the strength of a back-of-the-envelope estimate, and "appear to be strengthening" is a direction read from cited points, not a rate. The paper's own wish list of measurable quantities is the step that would test it; until those numbers exist, "not yet self-sustaining" is a tentative reading.