Cunningham, Trammell Paper: RSI Loops Not Yet Self-Sustaining

A new economics paper on recursive self-improvement lands on a deliberately deflationary finding: "feedback loops are not currently strong enough to generate a self-sustaining acceleration, though they appear to be strengthening."

The nine-author preprint, posted to arXiv on 14 September, is written by Tom Cunningham, Lukas Althoff, Basil Halperin, Brian Jabarian, Andrew Koh, Arjun Ramani, Phil Trammell, Parker Whitfill and Cheryl Wu. Its central move is to translate the RSI debate into economic plumbing. The authors build a sequence of models of AI progress, represent them as directed graphs, and derive a compact result: "net acceleration in AI capabilities depends on the product of elasticities across each feedback loop."

The paper draws a sharp line between two kinds of progress. "Narrow" capability is improvement on AI R&D benchmarks; "broad" capability is the kind that pays off across the economy. The distinction, they write, captures "the possibility that AI systems improve narrowly at optimizing AI R&D benchmarks without improving at broader economically valuable tasks."

The authors are equally direct about what they need from labs. They "document existing estimates of key parameters and provide a wish list of empirical objects that AI companies can measure and feasibly share publicly." The self-accelerating verdict itself is delivered as a "back-of-the-envelope calculation," not a settled result.

It arrives during a busy run of takeoff-adjacent work on our safety beat, which has logged more than two hundred alerts over the last ninety days.