METR researchers outline the economics of recursive self-improvement

METR researchers outline the economics of recursive self-improvement

On July 22, 2026, METR researchers Parker and Tom describe a paper that uses simple models to examine how AI may accelerate AI research and development (R&D). They say they want to understand whether AI-driven feedback could accelerate future capabilities, an input to their assessment of risks from frontier AI development. The researchers avoid using “RSI” in a technical sense because the term has been used for different levels of feedback. These range from any link between capabilities and model improvements to feedback strong enough for super-exponential or fully autonomous growth. Instead, they examine feedback strength and whether it could produce “self-sustaining acceleration.” They identify the effect of increased model capabilities on the rate of algorithmic progress as the model’s most uncertain relationship. They describe possible bottlenecks, including data, training compute, inference compute, experiments, and algorithmic- or R&D-specific capabilities. They say the evidence for these limits is not overwhelming, so they cannot rule out extended, rapid acceleration. The authors say labs could release more data about AI’s effects on AI R&D. They describe the paper’s calibration, which suggests estimates for model parameters, as loose and a first draft. They hope to make the model more useful for operational purposes.

Publication: 22686477-4098-e574-256a-23e34b5df080

attribution — pass: The reading attributes the note and its claims to METR researchers Parker and Tom, who coauthored the paper. It also carries the source's caveat that notes may not reflect METR's views as a whole. It omits the seven other economist coauthors, but this does not misattribute anything.

We (Parker and Tom) recently coauthored a paper, “The Economics of Recursive Self-Improvement”, with 7 other economists.

summary — pass: The summary leads with the actor and the development: a paper modeling how AI may accelerate AI R&D. It explains the RSI definitional dispute in plain terms (feedback from capabilities to model improvements). It covers the key uncertainty, the possible bottlenecks, the conclusion that acceleration cannot be ruled out, and the call for more lab data. The title already expands RSI as recursive self-improvement. The question follows from the stated most uncertain relationship and the data wish list.

Instead we focus on the strength of feedback effects, and whether they are sufficiently strong for “self-sustaining acceleration.”

date — pass: The original publication is the instant 2026-07-22T07:00Z, and its UTC calendar day matches the stated July 22, 2026; the URL path also shows 2026-07-22. The source is a publisher page, not an archive copy, so no archive disclosure is needed. There is no hindsight framing. The authors' hopes for future iteration are reported as plans, not outcomes.

support — pass: Each claim is a source-stated paraphrase with an exact supporting excerpt. These cover: the simple models; algorithmic progress as the most uncertain relationship; that the evidence for bottlenecks is not overwhelming; the loose first-draft calibration; and that labs could release more data. The reading adds no unsupported benefits or costs.

There is more data relevant to RSI that the labs could be releasing.

media — pass: No media is selected, and no verified media candidates were supplied, so choosing none is appropriate.