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.
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.