Meta is actively trying to take down publicly available copies of LLaMA: https://github.com/github/dmca/blob/master/2023/03/2023-03-2...

Haha good luck with that now… it’s in the digital ether available to all on IPFS… at worst you might have to ask around for someone to help you, but its “distributed” widely enough now I don’t think even a billionaire can put this back into the bottle.

And in 6 months it will be outdated.

So long LLaMA, and thanks for all the fish. You will be remembered as the slightly-sexier version of GPT-J that was most well-renowned for... checks clipboard ...Macbook acceleration.

LLaMA is part of LLM history in a way that Bard will probably never be

That's kinda my point. LLaMA's entire history will be enlightening the great erudites who missed-out on real open models like GPT-Neo. Since Meta won't cooperate with the community to improve on it, it's a dead-end. It's anything but the genesis of local AI.

Given that free alternatives like Vicuna (from the University of California and CMU) are better than LLaMA, are freely and legally available for download, and are compatible with code like llama.cpp, even if every copy of LLaMA is taken down it will have no effect on the development of chatbots. It might even improve things as people who would otherwise go for the better known LLaMA will move towards these newer, better, models.

They are all built on top of Llama…

Yes, but that doesn't matter mow. The University of California has released Vicuna as open source. It doesn't need the Llama model to be installed at this point. Nor do you need any of Meta's code to run it either as you can use llama.cpp (not created by Meta). That's the whole point of the article. It's open source now. There's nothing Meta can do.

This is incorrect. According to the official https://github.com/lm-sys/FastChat#vicuna-weights you need the original Llama weights before applying the Vicuna diff.

It's built off of llama though - you can't get to Vicuna without having the llama model weights.

It doesn't matter if you merge the LoRA, the resulting weights are still a derived work - assuming, that is, that weights are copyrightable in the first place (which is still a big if).

If the resulting weights a derived work of LLaMA then LLaMA is a derived work of the illegally pirated Books3 dataset (a dataset of a private torrent tracker) used to train it.

There's no way ML models can be protected under copyright.

The problem is that you need to risk getting sued first to prove that point. And hoping that you have deep enough pockets to outlast Meta's lawyers.

Right. Most of the fine-tuned models we've seen so far have been by university teams.

Meta is not being very selective here. I applied for the download myself and got the links after two days (using a university email address).