> their style is based on the mean

A base model that has only been pre-trained has output that is based on the mean, but most advances in capability the last three years have come from post-training. The influence of RL dominates every observable characteristic of coding agents today and steers the model towards strategies and practices that are objectively successful for the LLM.

I think the best code for an LLM is different than the best code for a human. Lots of redundancy, lots of unnecessary and overly defensive code is being reinforced by their training because it is successful there.

They still take it to far and do all kinds of stupid shit - if you look at the code you'll see it all the time but it actually doesn't cause many problems. Its much more important to focus steering effort on feature scope and architecture. I still ask it to refactor code from time to time and its pretty good at doing that.