"Basically, you introduce more randomness into the works in order to keep things from becoming overly optimized (or highly correlated, in the language of self-organized criticality)."

That sounds a lot like how I've heard AI guys talk about avoiding optimization pitfalls when training.

I guess the difference is we can run that on an accelerated timescale, whereas civilization marches to its own, slower beat.