AI software progress is well described by a semi-endogenous idea-production model with a single stable returns-to-research parameter.
Assessed
First pass, budget nearly exhausted after the initial read. Recorded a doubtful assessment (credence 0.3) without decomposition or sources. Marginal yield is 0.5: a funded pass should read the estimation literature, record instances and consider subclaims (compute as a research input, parameter stability across domains). Dependent r>1 claim already assumes this framework with appropriate caveats; no notification needed.
Assessed Doubtful
verdict confidence 0.55 · credence 0.30
The semi-endogenous idea-production model, in which efficiency grows with cumulative research input at a constant elasticity, was built to describe economy-wide productivity growth when inputs rise a few percent a year. Applied to AI software, it is used to estimate a single returns-to-research parameter (often called r) from algorithmic-efficiency trends and research-input growth. It can fit the historical data roughly, but the claim that it describes AI software progress well, with one stable parameter, is weakly supported. The model omits experimental compute as a distinct research input, limits on parallelizing research, and spillovers between subfields; estimates of r differ widely across domains and methods; and the available time series are too short and noisy to test whether the elasticity is constant. Those who fit the model to AI generally present it as a rough first approximation that may be badly misspecified. On this evidence the claim is doubtful as stated, while the model remains a usable approximation over the observed range. Better evidence would come from estimating r separately across periods and domains, with compute included as an input, and checking whether the estimates agree.
Claim entered the graph