Yes, you are right, every project that distributes LLaMA right now is violating Meta's agreement.

There is no “agreement” in the files being served from the llama torrent. And since model weights are not a derivative work, but machine generated from the works of others (In many cases without their consent), I’m not sure Facebook could copyright them even if they wanted to.

I've got a weird, probably untrue conspiracy theory about this.

Hugging face releases stable diffusion. It goes viral and vastly outpaces the competition in the blink of an eye. Then they get sued.

Meta sees both of these things go down. Meta needs a leg up on chat GPT, but worries about legal repercussions similar to stable diffusion.

Whoops, it leaked! Hey, we didn't say those dastardly devs could use it.

>But they may decide not to make their future models available after what is happening with the Llama weights.

I think that ship has probably sailed, in that no one is going to release weights in this way again. Either they will publish them outright (like Whisper) or they will keep them (almost) completely closed.

> I realize that some people are of the opinion that AI models (weights) cannot be copyrighted at all.

This isn't even ambiguous under US law.

Anyone in the field that might try to argue to change the law should think very carefully because there would be no credible argument for tightening copyright here that didn't also make it impossible to train on public data. So only the couple largest companies, who could demand permission to train everyone's data as part of their TOS, would be able to create state of the art AI systems.