Which, given what it was trained on, is utterly ridiculous.

Yup, but that being said, Llama is GPLv3 weather Meta likes it or not. Same as ChatGPT and all the others. ALL of them can perfectly reproduce GPLv3 licensed works and data, making them derivative work, and the license is quite clear on that matter. In fact up until recently you could get chatGPT to info dump all sorts of things with that argument, but now when you try you will hit a network error, and afterwards it seems something breaks and it goes back to parroting a script on how it's under a proprietary license.

This is interesting but it has not been proven in court, right?

I don't see how this follows at all. Github isn't GPL3 just because it stores and gives you back gpl3 code

Is that easier to enforce than having AI only trained in a legal way (=obeying licenses / copyright law)?

Yes. Having training obey copyright is a big coordination problem that requires copyright holders to group together to sue meta (and prove they broke copyright, which is not something proven before for LLM).

Whereas meta suing you into radioactive rubble is straightforward.