CONCEPT
Training as Artificial Selection
The structural homology between the breeder's hand and the loss function—the recognition that training a neural network is not programming but breeding, and that what survives is exactly and only what the selection criterion rewards.
Darwin opened
On the Origin of Species not with finches but with pigeons. Breeders, selecting which animals reproduce, had transformed the rock dove into pouters and fantails so unlike each other that an ornithologist seeing them wild would classify them as separate species. This was his rhetorical and logical bridge: if humans, selecting deliberately over centuries, could reshape a species so drastically, then nature, selecting relentlessly over millions of years, could do incomparably more. Artificial selection was his proof of concept for natural selection. It is also, almost literally, what happens when a
neural network is trained: a vast population of numerical weights begins essentially random, and a selection criterion—the loss function—culls them generation by generation until something competent survives. The capability is not authored by any engineer; it emerges from what the criterion rewarded. The breeder's hand is the loss function, and the people who choose it. What the model becomes is exactly what that hand selected for—including every