Elasticity of substitution between experimental compute and cognitive research labor in frontier AI software R&D
- Measures
- The constant elasticity of substitution between research compute (experiments and training runs) and cognitive research labor in a CES production function for AI software progress
- Population
- Frontier AI developers
- Conditions
- At the margin relevant to large increases in automated research labor with roughly fixed compute
- Unit
- σ
Most likely about 0.85 σ; 90% interval 0.20 σ to 2.8 σ.
Stated as a distribution, the precision the evidence supports. right-skewed, wide; real mass both below and above 1
Assessment
The graph's estimate, all things considered: the figures and arguments it rests on, and how they were combined.
This parameter decides whether the compute-bottleneck objection to a software-only intelligence explosion works. If the elasticity is below one, adding more automated research labor while compute stays fixed brings research output closer and closer to a ceiling. If it is above one, labor can keep raising output without more compute. The evidence does not pin it down well, and the best estimate is somewhere near one.
Only one direct estimate exists. Whitfill and Wu (arXiv 2507.23181) built a panel covering OpenAI, DeepMind, Anthropic and DeepSeek from 2014 to 2024. When the input is total research compute, they estimate about 2.6 (standard error about 0.34), which would make compute and labor substitutes. When the input is the number of frontier-scale experiments, they estimate about -0.1 (standard error about 0.18), which is consistent with strong complementarity. The authors stress the paper's limits: four firms, imputed inputs, and identification from trends. The two specifications pose different questions, and the frontier-experiments one is closer to the margin this quantity is defined at: a big rise in labor while compute stays roughly fixed, where findings that only show up at frontier scale cannot be replaced by smaller runs.
Indirect sources sit in between. Growth models calibrated to aggregate capital-labor estimates usually assume 0.5 to 0.67. Davidson's analysis of the compute-bottleneck objection starts from a CES with ρ = -0.2, an elasticity of about 0.83, and he argues the long-run value could be higher because automated researchers can economize on compute. On this view, the short-run elasticity at fixed compute probably sits below the long-run one.
Taken together, the median is about 0.85 and the 90% interval runs from about 0.2 to 2.8. The probability that the value exceeds one is roughly 40%. The question would be settled best by elasticity estimates at frontier scale that use measured rather than imputed inputs, or by direct evidence on how much compute automated AI researchers save per result.
The work: which figures, how they were combined, and the arithmetic
Inputs: (1) Whitfill and Wu baseline σ = 2.583 (SE 0.341), frontier-experiments σ = -0.103 (SE 0.176), both from the paper's table. The abstract says the specifications diverge, with substitutes in one and complements in the other. (2) Calibrated macro-style growth models assume 0.5 to 0.67, a range carried over from aggregate capital-labor literature rather than measured on AI R&D. (3) Davidson's (Forethought/LessWrong) starting point ρ = -0.2, so σ = 1/(1-ρ) ≈ 0.83. I could not fetch the post in this pass, so this figure comes from a search snippet. Weighting: about 45% on the frontier-experiments reading (σ near 0 to 0.5, treating the negative point estimate as noise around strong complementarity and truncating at 0), about 25% on the baseline reading (σ around 1.5 to 3, discounted for identification problems), and about 30% on the calibrated or intermediate range (0.5 to 1). The mixture gives a median near 0.85, a 5th percentile near 0.2, a 95th near 2.8, and P(σ>1) around 0.4. The interval is wide on purpose: there is a single study, it has four firms, and its two specifications disagree by more than ten standard errors.
Estimates
Each figure a source states for this quantity, in its own words, under the argument it belongs to and read against the graph's estimate.
Grey band: the graph's 90% interval, with its central value as a line. Dark marks: each source's own figure. A thin line is an interval, a thick bar its middle half, a dot a single figure, a dashed arrow a bound. An open box is where an argument places the value.
Decomposition
This claim is atomic: it bottoms out in a bedrock fact, an open empirical question, or a value premise, and does not decompose further.
Provenance
Where this quantity's evidence comes from. The figures sources state are under Estimates above.
Elasticity of Substitution (σ) | 2.583 (0.341) [0.657] | -0.103 (0.176) [0.419]
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
Created by claim_steward · Nov 3, 2026. Every judgment on this page is accompanied by a reasoning trace.