e/acc Research Bites #1: What Sustains Technological Acceleration? A model of technological acceleration can reach a singularity while median efficiency improves by a finite factor. My research develops a dynamical model behind e/acc’s appeal to Fisher. It adds resource-dependent renewal of technological variety to selection, then establishes the conditions under which technological acceleration follows. Imagine computers performing the same task. Each design delivers a certain number of tasks per joule. Exceptionally efficient designs can pull the sector’s overall efficiency far above its typical computer. In my model, aggregate output per joule diverges in finite time, while median efficiency approaches 2,703 times its initial value in one illustrative run. Both keep improving throughout. One equation explains the widening gap: Aggregate efficiency / median efficiency = exp(V/2) V is the energy-weighted variance of log-efficiency. This relationship follows from the model’s lognormal distribution. As dispersion grows, techniques using a shrinking share of total energy come to supply most of the output. Just before the model’s singularity, median efficiency has risen about 2,126-fold, while aggregate output per joule has risen about 395,000-fold. The median continues towards a finite limit of 2,703-fold; the aggregate grows without bound. How does this connect to Fisher? Fisher’s theorem relates the contribution of natural selection to fitness improvement to additive genetic variance in fitness. My paper adds a law for how technological variety is renewed as available power and output change. With positive selection and renewal, finite-time divergence occurs when the resource-feedback exponent ζ is positive: renewal increases as aggregate efficiency rises within the resource regime considered. That exponent is a quantity the research programme needs to estimate. Acceleration means that efficiency’s percentage growth rate increases. In this model, it either starts immediately or follows an initial slowdown, with one turning point. The paper also proves explicit bounds on the singularity time. For the illustrated run numerical integration gives T ≈ 35.18 model time units, within the proven bounds of 30.39 and 67.58. These bounds depend on how strongly selection shifts energy towards more efficient technologies and how quickly homogenisation reduces efficiency differences between them. For a particular balance between these processes, the two bounds coincide, giving an exact formula for T. The divergence depends on permitting arbitrarily high individual efficiency. A fixed ceiling on every technique also bounds the aggregate. The next step is to introduce that ceiling and an explicit energy budget for experimentation, then calculate the gains that remain. Working paper forthcoming.