[jrmg]

I wonder where LLMs got this style from, since, while you did see something like it, it wasn’t nearly as widespread before the rise of ChatGPT, so not the most statistically likely form for these to be generated in.

Was it trained into them in the supervised or reinforcement phases?

Has it emerged from prompts extorting them to be friendly, somehow sneaking in from text message training as a result?

Is it now in a self-reinforcing feedback loop as training data grows to include modern Readmes?

[sdwr]

As far as I know, it was trained in because people preferred it in A/B testing

[LtWorf]

Private chats maybe?