Doubling AI software research input more than doubles software efficiency (research returns parameter r > 1).
Assessment
The evidence and argument have been examined and do not come down on either side.
The claim concerns the returns parameter r in the semi-endogenous ("Jones") model of AI software progress: if each doubling of cumulative research input yields more than one doubling of software efficiency (r > 1), automating AI research could produce accelerating, self-reinforcing progress; if r < 1, progress decelerates. It is the central threshold in the debate over a software intelligence explosion.
Fitted estimates lean above one but do not settle the question. Epoch's estimates for AI domains (computer vision, reinforcement learning, language models) have central values above 1, with language models highest, and Forethought's analyses anchor on roughly 1.2 to 1.4 for AI software, drawing on an earlier computer-vision estimate of about 1.4. Outside AI, Epoch's estimate for the Stockfish chess engine was about 0.83, and estimates for other algorithm domains cluster near one. The uncertainty ranges for the AI domains straddle 1.
The estimates rest on weak measurements. Research input is proxied by paper or author counts, or by lab staff and compute, and these miss spillovers from other fields and from academia. Software efficiency is known mostly as an average growth rate, so the fits have to assume progress was constant. The Jones model is also being pushed far beyond the input growth rates it was built for, and it ignores compute bottlenecks for experiments and limits on parallel research. Some of these biases would raise r and others would lower it. The value at the moment research is automated could also differ from the historical value, since returns may fall as ideas get harder to find. Taken together, the evidence slightly favors r > 1 in recent AI software without establishing it. Better input data and experiments that measure returns directly would be the most likely way to resolve it.
Full reasoning: the evidence and decisions behind this verdict
Prior model: r = λ/β in the Jones law of motion. Historical estimates of software and algorithm progress (Erdil and Besiroglu's computer-vision estimate of about 1.4, Epoch's Stockfish estimate of about 0.83 with standard error 0.15) suggest values near one. The question cannot be settled from existing proxies.
Evidence read: Ho and Whitfill (Epoch, Nov 2025), epoch.ai/gradient-updates/the-software-intelligence-explosion-debate-needs-experiments. It reports estimates for computer vision, reinforcement learning and language models, with central values above 1 (highest for language models), "a lot of probability on compounding returns", and ranges that straddle 1. It then lists data flaws: input proxies omit spillovers and academic effort, which would likely bias r downward, and output is known only as an average rate under an assumption of constant growth. It also lists model flaws: no compute bottleneck for experiments and no limit on parallelism. Epoch's earlier paper (epoch.ai/publications/do-the-returns-to-software-rnd-point-towards-a-singularity) gives r of about 0.83 for Stockfish. Forethought (newsletter.forethought.org/p/humans-are-not-chinchillas-revisiting) notes that its software-intelligence-explosion work anchors on Epoch's 1.4 for computer vision and adjusts it from there.
Weighing: central estimates for AI software sit above 1, which moves credence somewhat above 0.5. The measurement and specification problems are serious and run in both directions, and the non-AI comparison (Stockfish) is below 1. The examined evidence therefore does not come down decisively, so the status is inconclusive with credence of about 0.58. Two readings were close: inconclusive and supported. Inconclusive was chosen because the authors producing the estimates themselves say the data cannot support them. The verdict would change with better input data that includes spillovers, or with experimental measurements of returns.
Wording caveat: r is defined for cumulative research input. The claim's phrase "doubling research input" should be read in that sense.
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 claim has been said, linked to its canonical form.
Here we see a lot of probability on compounding returns. But the uncertainty is very large, and our range of estimates straddle the key threshold of r=1. So at face value, the software intelligence explosion seems quite plausible — at the very least, it's hard to rule out.
Epoch authors report estimates of r for computer vision, RL and language models with central values above 1 but wide uncertainty, then argue data and model flaws make the estimates unreliable.
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.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Nov 2, 2026. Every judgment on this page is accompanied by a reasoning trace.