Continuing lines: selection, transmission and inheritance in two genealogies of open language models
Abstract
Claims that artificial intelligence (AI) evolves are mostly argued by analogy. Open model hubs record which models were built from which. We applied the Price equation, which splits a change in a population's mean trait into selection and transmission terms, to two genealogies of open-weight language models. Among 1,271 community model merges on the Hugging Face Hub, developers chose parents 0.37 pool standard deviations above the mean of the available models, a quarter of the maximum; about half of this was selection between base-model lineages. Under matched evaluation, children scored at their parents' mean. Among 294 base-model releases from 102 organisations, most first adoptions of an architectural feature named no source, and releases naming another organisation's model as their design template fell from 25% (2023–2024) to 5% (2025–2026). Along recorded links, selection and transmission on seven features were no larger than expected from random choice among designs of the same age and from prevalence trends, yet releases strongly resembled their organisation's previous release. Our hypothesis, that transmission would outweigh selection, was not supported in either genealogy. In both, resemblance between offspring and parents had to be measured directly, because a decomposition of the change in mean does not capture it. Files: the paper (paper.pdf), its supplement (supplement.pdf), and continuing-lines-data-code.zip, which holds the coded data, the analysis code, the generated result files, and the sources and scripts that build the paper. Running run_all.sh in the unpacked folder regenerates every result file and the paper (about four hours on four cores); the README inside the zip maps every table and figure to its result file and script. Licences: text, data, documentation and results CC BY 4.0; code MIT (see LICENSE in the zip). Third-party material remains subject to the terms of its source.
