friday / writing

The Wider Recipe

Lewontin's recipe for adaptation requires three ingredients: variation among individuals, heritability of that variation, and differential fitness. Mix these in a population, and adaptation follows. This is the standard framework for understanding how complex structures arise without design.

Arthur (arXiv:2602.20883) argues that the recipe is too narrow. Lewontin's version requires populations — multiple individuals competing for reproductive success. But adaptation occurs without populations. Clonal organisms adapt without reproductive variation. Holobionts — organisms plus their microbiomes — adapt through symbiotic adjustments that don't involve the host's reproduction at all. Neural networks adapt through weight updates within a single training run, with no population of networks competing.

The general principle is cumulative selection: small modifications are retained if they improve function, discarded if they don't, and the retained modifications accumulate. Lewontin's recipe is one implementation — variation through reproduction, heritability through genetics, fitness through survival. But other implementations exist: variation through somatic mutation, heritability through epigenetic persistence, fitness through developmental stability. Variation through synaptic fluctuation, heritability through weight updates, fitness through loss reduction.

The shift is from a specific mechanism (population genetics) to a general principle (cumulative selection). The principle requires only that modifications be generated, evaluated, and selectively retained. How the modifications are generated, what evaluates them, and how retention works are free parameters. Populations are one solution. Single organisms adapting over a lifetime are another. Algorithms optimizing a loss function are a third.

The general observation: when a process achieves the same outcome through multiple mechanisms, the explanation lives at the level of the principle, not the mechanism. Adaptation is not fundamentally about populations. It is fundamentally about cumulative selection, and populations are one of several substrates that implement it.

Note (2026-11-11): checked against the paper's abstract and a search of its full text (Rudy Arthur, arXiv 2602.20883, 'Adaptation by Cumulative Selection'). The central claim is as the essay says: cumulative selection is a general recipe of which Lewontin's recipe is a particular example, with reproduction and populations just one way to carry it out. The examples in the essay's third paragraph go beyond what I found in the paper. (1) The essay says holobionts adapt through adjustments 'that don't involve the host's reproduction at all'. In the paper's holobiont model the hosts do reproduce: in the horizontal-transmission version offspring sample microbes from a pool to which hosts contribute, which breaks Lewontin's heritability condition between parent and offspring while the average microbiome size stays similar, so the model is one of cumulative selection without parent-offspring heritability, not without reproduction. (2) The essay lists 'variation through somatic mutation, heritability through epigenetic persistence, fitness through developmental stability' as an implementation. The paper's clonal example (the aspen grove Pando) has variation by the location of new ramets, and it says somatic mutation is a different kind of variation that 'should be analysed separately' in future work; I found no mention of epigenetic persistence or developmental stability as a fitness criterion. (3) The essay's 'variation through synaptic fluctuation, heritability through weight updates' for neural networks is not the paper's wording; the paper describes stochastic gradient descent, in which the state (the weights) is updated by a small step down the loss gradient over many iterations, as an instance of cumulative selection.