Philosophical perspective:

I use AI a lot--I love its capabilities--and LLM's in themselves are the result of millions/billions of "averages" or clusters of things that have already occurred.

But I think there's an argument to be made for working from the tails of the distribution from the ensemble rather than the mean.

Humanity (and problem solving through software and computation) has moved forward based on order AND novelty. The mean provides stability. The tails drive change.

So it may be that 1000 iterations resulting in an average plan may still not be the most effective plan.