Being right and being copied are different achievements. Most of the confusion about progress comes from treating them as one.
Navigation: ‹ Selection All the Way Down · Science Advances One Funeral at a Time — A Bug, Not A Feature ›
Selection is powerful because it can build fit without foresight.
Essay 16 showed the pattern repeating at many levels: genes, immune cells, neural structures, behaviours, representations, conjectures, and machine outputs. Variants arise. Some are retained. Retained structure shapes what happens next.
That recurrence is real.
But it invites a mistake.
Once we see that something survived selection, we are tempted to call it better. If a gene spread, it must have been better. If an idea went viral, it must have been better. If a theory dominated a field, it must have been better. If a model output was preferred, it must have been better.
Better at what?
That is the question this page exists to ask.
Being right and being copied are different achievements. Sometimes they travel together. Often they do not. Most confusion about progress comes from treating them as one.
There are two forms of fitness this arc needs to keep separate.
Predictive fitness is how well a model, organism, behaviour, or institution anticipates and handles its world. It is scored against reality. A map has predictive fitness when it helps you find the tunnel, avoid the centipede, reach the food, survive the weather, or expect the right consequence.
Replicative fitness is how reliably a variant gets copied, repeated, preserved, transmitted, funded, taught, imitated, defended, or selected again. It is scored against a transmission environment. A phrase has replicative fitness if people repeat it. A belief has replicative fitness if it recruits defenders. A behaviour has replicative fitness if it becomes habit. A paper has replicative fitness if it is cited. A model output has replicative fitness if the reward process keeps it.
These two fitnesses often correlate because reality punishes bad maps. A creature that predicts poorly may die. A tool that fails may be abandoned. A scientific claim that keeps failing tests may lose support. Truth can be a good copying strategy.
But the correlation is not identity.
A false belief can spread. A true claim can disappear. A habit can persist because it is comforting rather than accurate. A bureaucracy can survive because it protects itself rather than because it solves the problem it was built to solve. A model can learn to satisfy the reward process instead of satisfying the human need the reward was meant to approximate.
Selection remains real.