Evolution works across a long feedback loop integrating approximately over all variance in that loop.

All actions that happen between the organism being conceived, and when it successfully reproduces.

That's an insane amount of random variation and stuff on that feedback loop.

And yet evolution is able to pick out the smallest changes to select for over time. How?

Even if it's an indirect and subtle effect, if it's consistent then with enough instances the signal pops out against the background noise.

Even long feedback loops, as long as the signal is threaded through like beads on the string, it can work.

A small but consistently different alignment of beads on the string.

In one feedback loop iteration the signal is washed out in the noise.

But in thousands of loops for thousands of individuals, that consistent asymmetry will stand out against the background noise of all of the other ones averaged together.

This is one of the reasons deep learning training works so well despite insanely long and multi-varying feedback loops as signals backpropagate.

Just jam an unreasonable amount of data through it, and all of the noise is filtered out and the signal remains.