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
Estimate revised after reading sources
After the first estimate was recorded from recall (median 6, 1.5 to 35), later turns proved cheaper and one search plus one page read were affordable. Confirmed by reading: Forethought's 2 to 32 (median 8) and its basis; AI 2027's 25x for SAR and its definition; Whitfill and Wu's 2.58 / -0.10 via a secondary report only. Read the executive summary and key findings of METR's 5 November 2026 assessment of Anthropic (about 2x overall, 1.3 to 3, about 80% attributed to AI, training compute not accelerated, no bottleneck identified, majority of informative evidence redacted). Weighed it as testimony about evidence not shown, from an outside assessor chosen with the developer's cooperation and subject to its redaction rights; it speaks to partial automation only. Effect: weight on the strong-complements family cut from 30% to 20% and its range moved from 1 to 3 up to 2 to 4; estimate revised to median 8, 90% interval 2 to 40. Recorded two estimates (Forethought, AI 2027) under the partial-substitution argument and the METR figure as a reading, not as a value. Re-evaluated the strong-complements argument. Not created: the labor-compute elasticity of substitution as a quantity beneath this one (still the main crux; left for the follow-up pass). The filed request remains open for the unread remainder.
First pass definition completed and estimated
Completed the definition: 'fully automated' fixed at the threshold where AI first does every research task as well as the best human researchers (not later superhuman systems, which the literature gives 100x+); 'software progress' fixed as growth in software efficiency in OOMs/year excluding hardware and scale-up; baseline is the developer's human staff with no AI research labor on the same on-trend compute; horizon kept as the twelve-month average (the existing horizon), noting published figures are instantaneous. Recorded estimate: median 6, 90% interval 1.5 to 35, as a mixture over a strong-complements family (about 1 to 3) and a partial-substitution family (about 5 to 30), with CES ceiling arithmetic shown. Two named arguments created for the two families. Limits of this pass: the run's budget allowed only about three turns, so no source was opened; published figures (Forethought 2 to 32, AI 2027 25, Whitfill and Wu's split elasticity result, Epoch-associated low figures) are from recall, and the 2026 frontier-developer observations mentioned in the parent Steward's seed note were not verified or weighed. No instances recorded for that reason (only read sources may be recorded). The central crux, the elasticity of substitution between research labor and experiment compute at the frontier, deserves its own quantity node beneath this one; not created because match and creation could not be afforded. Petitioned for more budget and filed a request for the source tracing. Marginal yield set high (0.8).
Claim entered the graph