Engineers improving neural networks act like breeders: they try a variant, check the result, and keep the best. They collected 935 such trials from 161 papers and measured how often a code tweak turns out beneficial, harmful, or neutral. The picture is familiar to any biologist: 68% of changes degrade performance, 19% do nothing, and only 13% yield a gain.
The shape of the outcome distribution in AI closely matches that known for mutations in the fruit fly and baker's yeast. Moreover, the emergence of new architectural tricks follows the same waves: slow at first, then explosive growth, then a plateau. Fourteen technical solutions were independently invented 3–5 times each, like wings in birds and bats.
The conclusion is simple: evolution is not about DNA or code, but about the shape of the landscape that the search wanders over. Whether it's genes or neural network parameters, the statistics of successes and failures are the same.
as written by the authors arXiv:2604.10571
We test whether artificial intelligence architectural evolution obeys the same statistical laws as biological evolution. Compiling 935 ablation experiments from 161 publications, we show that the distribution of fitness effects (DFE) of architectural modifications follows a heavy-tailed Student's t-distribution with proportions (68% deleterious, 19% neutral, 13% beneficial for major ablations, n=568) that place AI between compact viral genomes and simple eukaryotes. The DFE shape matches D. melanogaster (normalized KS=0.07) and S. cerevisiae (KS=0.09); the elevated beneficial fraction (13% vs. 1-6% in biology) quantifies the advantage of directed over blind search while preserving the distributional form. Architectural origination follows logistic dynamics (R^2=0.994) with punctuated equilibria and adaptive radiation into domain niches. Fourteen architectural traits were independently invented 3-5 times, paralleling biological convergences. These results demonstrate that the statistical structure of evolution is substrate-independent, determined by fitness landscape topology rather than the mechanism of selection.
The proportion of beneficial changes in neural networks (13%) is higher than in any studied organism, but lower than in viruses with their tiny genomes: those evolve even faster.
In Jurassic Park, chaos mathematician Ian Malcolm said that life finds a way. It turns out that artificial intelligence finds its way along the same statistical paths.