Firstly, llama is not just the weights, but also the code alongside it. The weights may or may not be copyrightable, but the code is (and possibly also the network structure itself? that would be important if true but I don't know if it would qualify).

Secondly, you can write what you want in a copyright license: you could write that the license becomes null and void if the licensee eats too much blue cheese if you want.

Following from that, if you were to train on the outputs of the AI, you may not be guilty of copyright infringement in terms of doing the training (both because AI output is not copyrightable in the first place, something which seems pretty set in precedent already, and possibly also because even if it was, it gets established that it is fair use like any other data), but if it means your license to the original code is revoked then you will at the very least need to find another implementation that can use the weights, or (if the weights can be copyrighted, which I would argue is probably not the case, if you follow the argument that the training is fair use, especially if the reasoning is that the weights are simply a collection of facts about the training data, but it's very plausible that courts will rule differently here).

This could wind up with some strange situations where someone generating output with the intent of using it for training could be prosecuted (or at least forced to cease and desist) but anyone actually using that output for training would be in the clear.

I agree it is extremely "have your cake and eat it" on the part of the AI companies: They wish to both bypass copyright and also benefit from the restrictions of it (or, in the case of OpenAI, build a moat by lobbying for restrictions on the creation and use of the models themselves, by playing to fears of AI danger).