> Why have the LLMs „learned“ to write PRs (and other stuff) this way?
They didn't learn how to write PRs. They "learned" how to write text.
Just like generic images coming out of OpenAI have the same style and yellow tint, so does text. It averages down to a basic tiktok/threads/whatever comment.
Plus whatever bias training sets and methodology introduced
That’s my whole point: Why does it seemingly „average down“ to a style that was not encountered „on average“ at the time that LLM training started?