I know this might be hard news but it needs to be said... if you want to put your time into working on open source LLMs, you need to get behind something you have a real (and yes, open source) license for.
I'm also curious what the copyright status of these models even is, given the "algorithmic output isn't copyrightable" thing and that the models themselves are essentially the algorithmic output of a machine learning algorithm on third party data. What right does Meta have to impose restrictions on the use of that data against people who downloaded it from The Pirate Bay? Wouldn't it be the same model if someone just ran the same algorithm on the same public data?
(Not that that isn't an impediment to people who don't want to risk the legal expenses of setting a precedent, which models explicitly in the public domain would resolve.)
That's my question as well. The models are clearly derivative works based on other people's copyrighted texts.
Only a twisted court system would allow Google/OpenAI/Facebook to build models on other people's work and then forbid other people to build new models based on GOF's models.
That's not that clear either. (Sometimes it's more clear. If you ask the model to write fan fiction, and it does, and you want to claim that isn't a derivative work, good luck with that.)
But the model itself is essentially a collection of data. "In Harry Potter and the Philosopher's Stone, Harry Potter is a wizard" is a fact about a work of fiction, not a work of fiction in itself. Facts generally aren't copyrightable. If you collect enough facts about something you could in principle reconstruct it, but that's not really something we've seen before and it's not obvious how to deal with it.
That's going to create a practical problem if the models get good enough to e.g. emit the full text of the book on request, but the alternative is that it's illegal to make a model that knows everything there is to know about popular culture. Interesting times.
But LLMs aren't trained to learn facts like "Harry is a wizard", they're trained to reproduce specific expressions like "You're a wizard, Harry".
That is, they're trained by prompting them with a selection from a (probably copyrighted) text and weights are adjusted to make it more likely they'll output the next word of the text.
They're not a collection of general facts, they're a collection of estimates about which word follows which other words, and the order of words is the essence of copyright in text.
Pedants have been complaining about this kind of thing for years. If you generate random data, no one has a copyright on that. But if you XOR it with a copyrighted work, the result is indistinguishable from random data. No one could tell you which was generated randomly and which was derived from the copyrighted work. But XOR them back together again and you get the copyrighted work.
Things like that get solved pragmatically, not mathematically. There is no basis for saying that one set of random bits is infringing and the other isn't, but if you're distributing them for the sole purpose of allowing people to reconstitute the copyrighted work, you're going to be in trouble.
Now we have something with different practicalities. The purpose of training the model on existing works is so that it can e.g. answer questions about Harry Potter, which the majority wants to be possible and is the same class of thing that search engines need to be able to do. But the same model can then produce fan fiction as an emergent property, so what now?
Basically, completely open source with a non-commercial license. I'm not sure why so many people keep saying it 'leaked'. It's just using open source weights not directly from the provider in a way that violates the software license.
There is no reason to believe they are, which means any restriction placed on the weights themselves is bullshit.
From the US Copyright Office[1]: "A mere listing of ingredients is not protected under copyright law", and from their linked circular on that page [2]: "forms typically contain empty fields or lined spaces as well as words or short phrases that identify the content that should be recorded in each field or space".
A list of identifiers and their weights seems pretty explicitly not protected under one or the other of these.
[1] https://www.copyrightlaws.com/copyright-protection-recipes/ [2] https://www.copyright.gov/circs/circ33.pdf
It does, if you choose to keep them internally as a trade secret.
It does, if you share it only with people you contract with not to disclose it.
But, for copyright specifically, rather than “some legal protection” more generally, the “it takes millions of dollars” argument is a financial recasting of the “sweat of the brow” concept which has been definitively rejected by the courts.
The point here is that copyright requires a human to have created something using their own labor/creativity.
The result of an algorithm run by a machine isn't a creative work under these definitions.
Why would they be? I mean, what is the specific argument that they fall within the scope of the definition of what is copyrightable under US law?
'copyright' seems to refer to a specific string of symbols. literary or musical styles can't be copyrighted, but individual sequences of letters or notes can. Basically, if a derivative work copypasted without adding anything, it's considered plagiarism and/or violation of copyright.
Weights in a model would similarly be a long string of symbols that someone went to considerable trouble to collate and therefore fall under copyright.
Meta’s leaked model has been a factor in spurring open source development, whether or not it is open source; the article also discusses the practical market of effect of leaked-but-not-open things like Meta’s model combined with the impracticality of prosecuting the vast hordes of individuals using it, and particularly notes that the open-except-for-the-base-model work on top of it is a major benefit for Meta (who can use the work directly) that locks out everyone else in the commercial market (who cannot), and leaning into open source base models is a counter to that.
https://github.com/facebookresearch/llama/pull/73/files
Two months later, Facebook hasn't merged the change, but they also haven't deleted it or tried to censor it in any way. I find that hard to explain unless the leak really was intentional; with pretty much any large company, this kind of thing would normally get killed on sight.