Multiplier of the rate of AI software progress once AI research labor is fully automated, relative to human-only research, with compute on its prior growth trend

Framework. Research output as a CES function of compute C and labor L: Y = (a*C^rho + (1-a)*L^rho)^(1/rho), a the weight on compute. For rho < 0 the multiplier from unlimited labor is capped at a^(1/rho). - a = 0.5: rho = -0.2 gives 0.5^(-5) = 32; rho = -0.4 gives 0.5^(-2.5) = 5.7; rho = -1 gives 2; rho very negative gives 1. - a = 0.3: rho = -0.4 gives 0.3^(-2.5) = 20; rho = -1 gives 3.3. The ceiling moves from 2 to 32 across a range of rho that existing data cannot narrow, which is why the interval is wide.

Figures and their basis. - Forethought (Davidson and Houlden, 2025): 2 to 32, median 8, log-uniform. Read in this pass. Basis: surveys of AI researchers, the AI 2027 decomposition (their cross-check methods give 15x to 28x), and a simple model in which 30x more labor with exponent 0.5 gives about 5.5x. Judgment. - AI 2027 takeoff forecast (2025): 25 for a superhuman AI researcher, defined as the best human researcher run 30x faster in 30x the numbers, with training compute held fixed; 5 for a superhuman coder. Read in this pass (summary table). Basis: three methods, chiefly researcher surveys. Judgment. One line of evidence with Forethought, not two. - Strong-complements view (compute-bottleneck objection, Epoch-associated economists): about 1 to 3. Not re-read in this pass. Assumption, by analogy with economy-wide capital-labor substitution. - Whitfill and Wu (2025): elasticity 2.58 (s.e. 0.34) for all research compute, -0.10 (s.e. 0.18) for frontier experiments, as reported by a later paper; not read directly. The only measurement of the mechanism; inconclusive. - METR assessment of Anthropic (5 November 2026), executive summary and key findings read: overall acceleration about 2x (1.3 to 3), benchmark trend 1.9x (95% interval 1.5 to 2.8), about 80% attributed to AI, training compute not accelerated, no bottleneck identified, majority of the most informative evidence redacted.

Software component implied by the 2026 reading. If AI accounts for 0.8 of an extra 1.0x of capability growth, the AI-caused rate is about 1.8x the old rate. If software progress was roughly half of capability growth before and compute scaling did not speed up, the software rate rose by about 1 + 0.8/0.5 = 2.6x. With the stated uncertainty (overall 1.3 to 3; software share one third to two thirds) the range is roughly 1.3 to 5. Discounted for the benchmark index possibly overstating, non-AI factors, and evidence not shown: about 15% that the present software multiplier is below 1.5.

Combination. About 20% on a strong-complements outcome now placed at 2 to 4 (current gains are the slack before compute binds); about 65% on the partial-substitution family centred near 9 (most mass 4 to 30); about 15% on values above 30. Before weighing the 2026 reading the weights were 30/55/15 with the low family at 1 to 3 and the mixture median about 6; the reading moves the 5th percentile from 1.5 to 2 and the median from 6 to 8. Quantiles are rounded judgment: 5%: 2, 25%: 4, 50%: 8, 75%: 15, 95%: 40.

Not done: the body and appendices of the METR report (about five sixths of it), Whitfill and Wu's paper itself, the Epoch-associated arguments, and later revisions of the AI 2027 model were not read.

What would move it: a frontier-scale measurement of labor-compute substitutability; a second developer's acceleration with the automated share of research labor stated; release of the redacted evidence; evidence on whether automated researchers raise the yield per experiment.