Yudkowsky doubted that a thousandfold increase in AI hardware would buy anything like a thousandfold increase in effective intelligence, while noting that such doubt is not reliable without an analytical calculation.

The test is whether a roughly 1000x increase in hardware or compute gives roughly 1000x gains in effective intelligence. True: scaling research shows strongly sublinear returns, for example 1000x compute giving modest or logarithmic gains in loss, benchmark scores or capability measures, far short of a 1000x improvement on any accepted measure. Mostly true: returns are sublinear on most measures, but a few narrow metrics show very large gains. Mixed: experts or evidence are evenly divided, or the returns depend heavily on which measure is used. Mostly false: credible measures show near-proportional capability gains from 1000x compute in important domains. False: 1000x hardware is shown to give roughly 1000x or greater gains in general effective intelligence.

Set before the research began.

1. E1 Scaling Laws for Neural Language Models arXiv (Kaplan et al.) · 23 January 2020

   “The loss scales as a power-law with model size, dataset size, and the amount of compute used for training, with some trends spanning more than seven orders of magnitude.”

2. E2 Where’s the “intelligence explosion”? Noahpinion (Substack/blog) · 27 September 2026

   “In OpenAI’s recently publicized results on unsolved math problems, success rises roughly with the log of compute over the range shown.”

3. E3 Hardware Scaling Trends and Diminishing Returns in Large-Scale Distributed Training arXiv (preprint) · 30 September 2024

   “(2) scaling the total number of accelerators for large model training quickly yields diminishing returns even when hardware and parallelization strategies are properly optimized, implying poor marginal performance per additional unit of power or GPU-hour.”

4. E4 AI Scaling Wall - Wikipedia Wikipedia

   “One 2025 analysis estimated that each tenfold rise in compute yields only one to two percentage points of MMLU-Pro accuracy.”

E1 reports that language-model loss falls as a power law in training compute, which is a sublinear relationship. E2 says success on unsolved math problems rises roughly with the log of compute. E4 cites a 2025 estimate that each tenfold rise in compute adds only one to two percentage points on MMLU-Pro. None of the evidence shows anything close to a 1000x capability gain from 1000x compute, so this meets the criteria for strongly sublinear returns.