Evolution by natural selection

Darwin's argument in the Origin needs three facts. Individuals in a population vary. Some of that variation passes from parent to offspring. And many more individuals are born than can survive, so there is, in his words, a frequently recurring struggle for existence. Put those together and any heritable variation that gives even a slight edge will be found more often in the next generation, and the one after. Darwin called this natural selection. It is not a force and it has no goal; it is bookkeeping. The traits that reproduce best become common, which can mean faster, or better hidden, or smaller, or more cooperative, depending entirely on what the environment punishes at the time.

Peppered moths are the best-measured case. The dark carbonaria form became common near industrial British cities and has declined in post-industrial Britain. Between 2001 and 2007 Michael Majerus released 4,864 moths in a rural Cambridgeshire garden and recorded which survived bird predation; Cook, Grant, Saccheri and Mallet published the analysis in 2012. Dark moths survived each day at 91 percent of the rate of pale ones, a selection coefficient of about 0.1 per day, and over the same years the local carbonaria frequency fell from 0.12 to 0.01. A 9 percent daily disadvantage sounds small; compounded over a season and several generations it removes a trait. The transfer is to anything with variation, copying and differential survival: products in a market, ideas in a culture, variants in an A/B test, prompts in an evolutionary search.

Two misreadings. 'Survival of the fittest' makes it sound like a contest of strength; fitness in the theory means only reproductive success in a given environment, and the fittest thing can be a parasite or a dull-coloured moth. Social Darwinism took a description of what happens to populations and turned it into a claim about what ought to happen to people; nothing in the mechanism licenses that step. There is also a practical caution for the transfer. Markets and A/B tests select for whatever the measurement rewards, which is often not what you want, and they need real variation to work on: a test between two near-identical variants selects for noise.