In the LLM space, "open source" is being used to mean "downloadable weights"
alessiofanelli.com
alessiofanelli.com
This is a little weird given that directly above, the author puts LLaMA into the "restricted weights" category. Even by the definition the author proposes, LLaMA 2.0 isn't open source; we shouldn't be calling it open source.
If open source in the LLM world means "you can get the weights" and doesn't imply anything about restrictions on their usage, then I don't think that's adapting terminology to a new context, I think it's really cheapening the meaning of Open Source. If you want to refer to specifically "open weights" as open source, I'm a bit more sympathetic to that (although I don't think it's the right terminology to use). But I see where people are coming from -- I'm not too put off by people using open source to describe weights you can download without restrictions on usage.
But LLaMA is not open weights. It's a closed, proprietary set of weights[0] that at best could be compared to source available software.
It is deceptive for Facebook to call LLaMA open source, and we shouldn't go along with that narrative.
[0]: to the extent weights can be copyrighted at all, which I would argue they can't be copyrighted, but that's another conversation.
Also, while weights usage might be restricted, it's a very big compute investment shared with the public. They use a 285:1 training tokens to params ratio, and the loss graphs show the model wasn't yet saturated. This is valuable information for other teams looking to train their own models.
LLaMA1 was highly restrictive, but the data mix mentioned in the paper led to the creation of RedPajama, which was used in the training of MPT. There's still plenty of value in this work that will flow to open source, even if it doesn't fit in the traditional labels.
> There's still plenty of value in this work that will flow to open source, even if it doesn't fit in the traditional labels.
That is a good point; the fight over what is open source and what is source available can get heated, and part of that is a defense against the erosion of the term. But... in general source available is better than closed source software. And LLaMA 2 is a significant improvement over LLaMA 1 in that regard, it really is. So I don't necessarily want to be down on it, in some ways it's just backlash of being tired of companies stretching definitions. But they're doing a thing that will absolutely help improve open access to LLMs.
I'm always a little bit torn about how to go about this kind of criticism of terminology, and I'm not trying to say that people shouldn't be excited about LLaMA 2. But the way it works out I'm often playing word police because the erosion of the term does make it harder to refer to models with actual open weights like StableLM. Facebook deserves real praise for releasing a model with weights that can be used commercially. It doesn't deserve to be treated as if what it's doing is equivalent to what StabilityAI or RedPanda is doing.
I do like your terminology of "open weights" and "restricted weights", and I wouldn't be opposed to even breaking that down even further, I think there's a clear difference between LLaMA 1 and 2 in terms of user freedom, so I'm not opposed to people trying to distinguish, just... it's not hitting the bar of being open weights.
It's a bit like if the word vegetarian didn't exist, and if everyone argued about how it's unhelpful to say that drinking milk isn't vegan because it's still tangibly different from eating meat. On one hand I agree, but on the other hand it's better to have another category for it that means "not vegan, but still not eating meat." There is an actual danger in blurring a line so much that the line doesn't mean anything anymore, and where people who mean something more rigorous no longer have a term to communicate amongst themselves. If average people get bothered by throwing LLaMA 2 into the "restricted weights" category, it's better to introduce another category between restricted and open that means "restricted but not commercially".
Beyond that though... yeah, I agree. I don't really have a problem with people calling open weights open source, my only objection to that is kind of technical and pedantic, but I don't think it causes any actual harm if someone wants to call StableLM open source.
- It's okay to train a model on arbitrary internet data without permission/license just because you can access it
- It's not okay train a model on our model
It’s completely illegal if you think about it.
So why LLMs who crawl the internet to present snippets and information should be treated differently from Google ? (who also reproduce verbatim the same content without paying any compensation to the copyright owners (all types: text, image, code)
Google would argue (and they won in federal court versus the Author's Guild using this argument) that displaying snippets of publicly-crawlable websites constitutes "fair use." Profitability weighs against fair use but it doesn't discount it outright.
They would also probably cite robots.txt as an easy and widely-accepted "opt-out" method.
Overall, I'm not sure any court would rule against Google's use of snippets for search. And since Google's been around for over 20 years and they haven't lost a lawsuit over it, I don't think it's accurate to say "it's completely illegal if you think about it."
US copyright law is one of those things that might seem simple, but really isn't. Hence many of the copyright lawsuits clogging our judicial system.
In addition, the fair use test contains a pillar about the use not affecting the market for the copyright holder's works[1] which I think in google's case (and probably in the current openAI case too) seems obviously not to have worked out (ie google's use has demonstrably negatively affected the market for the original copyrighted work in cases such as news for example).
[1]: https://fairuse.stanford.edu/overview/fair-use/four-factors/
Most news sites wouldn't get any traffic without search engines and aggegrators. Which is why they are now whining about FB et al no longer sending them traffic.
And let's not forget that both traditional and online news is no stranger to republishing other people's content - one of the reasons fair use exists in the first place.
I have no love for big tech but let's not pretent that this is about anything other than news publishers wanting more gibs.
This is less and less true, as evidenced by the progression of 0-click searchs.
> There are countries where content companies didn't like what Google does: Google took them out of the index -> suddenly they where ok with it again so that Google put them in again.
This story screams antitrust.
You're right, the number of news publishers that share a common owner is something that should be of concern to antitrust enforcers.
It does but the complainers are usually tabloid crap pushers whom no one in power really supports.
LLMs scrape my site and code, strip all identifying information and license, and provide/sell that to others for profit, without my consent.
There are so many wrongs here, at every level.
Presumably you can’t build an LLM that is a competitor of LlaMA using its outputs.
But AI weights are in legal gray zone for now. So it’s muddy waters and fair game for anyone who wants to take on the legal risks.
Should a publisher wish to be excluded from Google's, or any other web index's search and presentation, that's easy enough to specify.
<https://www.intellectualpropertyblawg.com/ip-management/what...>
<https://developers.google.com/search/docs/crawling-indexing/...>
<https://en.wikipedia.org/wiki/Robots.txt>
(And no, I'm not a fan of Google by any stretch, but let's keep the discussion rigorous here.)
________________________________
Notes:
1. You don't feel old. You are old.
In the LLM space it seems even more clear because many/most of the works in the various corpora used for this training have very clear copyright terms which prevent digital storage and reproduction without the publishers permission (just look at the reverse of the title page of any book for the copyright notice if you don't believe me).
Finally, for LLMs many/most of the works are in corpora[2] that people just download so they aren't looking at a robots.txt file put up by teh original site. If you look at The Pile paper[3] for example they explicitly say that much of the material is under copyright and that they are relying on fair use.
[1]: https://fairuse.stanford.edu/overview/fair-use/four-factors/ [2] https://github.com/Zjh-819/LLMDataHub for example [3] https://arxiv.org/abs/2101.00027
(1) the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;
(2) the nature of the copyrighted work;
(3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
(4) the effect of the use upon the potential market for or value of the copyrighted work.
<https://www.law.cornell.edu/uscode/text/17/107>
Most critically, courts have put strong emphasis on the notion of transformative use of copyrighted works, and web indexing is transformative in the sense that it does not create a competing work, but provides a means of discovering and assessing the relevance of the indexed work itself.
As to web indexing, that (and associated factors including thumbnails and caching) have been ruled by courts to be fair-use adaptations of works:
Displaying a cached website in search engine results is a fair use and not an infringement. A “cache” refers to the temporary storage of an archival copy—often a copy of an image of part or all of a website. With cached technology it is possible to search Web pages that the website owner has permanently removed from display. An attorney/author sued Google when the company’s cached search results provided end users with copies of copyrighted works. The court held that Google did not infringe. Important factors: Google was considered passive in the activity—users chose whether to view the cached link. In addition, Google had an implied license to cache Web pages since owners of websites have the ability to turn on or turn off the caching of their sites using tags and code. In this case, the attorney/author knew of this ability and failed to turn off caching, making his claim against Google appear to be manufactured. (Field v. Google Inc., 412 F.Supp.2d 1106 (D. Nev., 2006).)
<https://fairuse.stanford.edu/overview/fair-use/cases/>
Or, to use your phrase, by common law (precedential case law), that is precisely "how copyright law works". Note particularly that the courts leaned on publishers' capabilities to indicate whether or not caching was or was not permitted "using tags and code".
There's a larger issue which I'm not aware of being explicitly raised in case law, which concerns how the World Wide Web is indexed as contrasted to how a print library is indexed. In the case of a library, an independent third party (the library cataloguer) assigns metadata to a work (standardised title, author(s), translator(s), illustrator(s), publisher(s), etc., as well as subject headings and call numbers. Additional indexing is provided through citations indices (both forward and reverse --- works cited by, and citing, other works). These largely don't rely on the text of the indexed work itself, though of course the cataloguer presumably is reading at least portions of the work to classify it. Critically: the works themselves are physical artefacts of fixed form which are virtually always read directly rather than interpreted through some mechanism.[1]
As it's evolved over the past quarter century or so, Web search doesn't rely strongly on metadata (though some of this is taken into consideration), and most particularly publisher-provided keywords are almost wholly ignored, largely due to flagrant abuse of that feature by some publishers. Instead, a combined approach of full-text indexing (that is: capturing the full text of a work and identifying keywords and tuples (multi-word phrases) which can be matched against queries entered by persons searching for documents, and an assessment of the overall relevance of that work, usually at a site (or sub-site) level based on other indicia, most famously (though somewhat less relevantly today) "PageRank", Google's original site-ranking algorithm.
Further, the entire mechanism of the Web is of creating copies of works on request. When an HTTP request is sent, the server responds by copying the requested work to an output stream, which is then received (and duplicated, often multiple times) by the client system as an integral part of the utilisation of that content. US copyright law does not have a section specifically referring to computer-network transmission, but there are multiple limitations on exclusive rights to copy (by authors) above and beyond the 107 Fair Use exemptions in sections 108 through 122 of 17 U.S.C, including specifically ephemeral recordings (108) and the case of computer programmes (117).
<https://www.law.cornell.edu/uscode/text/17/chapter-1>
Large language model training is a new area of use and law (legislative or common) is yet to be determined, but there's at the very least existing statutory language as well as precedent which suggest that at least some uses might well be found to be fair use. As I'm watching the situation evolve, I'm reminded strongly of several articles copyright scholar Pamela Samuelson wrote in the 1990s over adapting copyright to the Internet age, and questions of what its future place might be: specific governance over the literal copying of expressive works, or a general doctrine against misappropriation. As always, there's a sharp tension between authors' rights (and, let's be brutally honest: publishers' profits) and the underlying Constitutional justification of US copyright law: "To promote the Progress of Science and useful Arts".
<https://constitution.congress.gov/browse/article-1/section-8...>
And it seems Sameulson is engaged in the discussion of generative AI and copyright, though I've yet to read her work on the subject: <https://news.berkeley.edu/2023/05/16/generative-ai-meets-cop...>
(Discussion here strongly reliant on US law. There's general international agreement on copyright through the Berne Convention, though significant national differences exist.)
________________________________
Notes:
1. There is a spectrum of works, e.g., print books, phonographs, CDs and DVDs (the latter containing anti-circumvention mechanisms), etc., but in general there's minimal if any intermediate copying and duplication of works, and in many cases none at all.
[1] https://www.theguardian.com/artanddesign/2023/may/18/andy-wa...
That said, it would tend to darken the prospects for operators of LLM generative AI systems, IMO.
Parent's point is if your own scaping army respects the "scaping.txt" and goes down on Google as they don't opt-out in their scraping.txt, it probably wouldn't fly.
User-agent: *
Disallow: /search
....What "other people"?
If it's the "you" who is not allowed to scrape google in https://news.ycombinator.com/item?id=36817237 then you can make your own "google is not allowed to scrape my thing" rules if you think that's beneficial for you.
If it's somehow related to LLM providers or users I doubt that's what the original comment was referring to.
To be clear, I understand the original comment as
LLM companies say "I can use your content and you cannot not prevent me from doing so, but I won't allow you to use the output of the LLM" just like Google says "I can scrape your content and you cannot not prevent me from doing so, but I won't allow you to scrape the output of the search engine"
I see it as the equivalent of the spam mail that require the user to login to disable them.
they are consistent, if they believe themselves to be "special" and deserves special treatment!
Copyright is literally the right to copy. Arbitrary Internet data that is not copied does not have any copyright implications.
The difference is that LLaMa imposes additional contractual obligations that, for ideological reasons (Freedom #0), open source software does not.
This issue reminds me of the FSF/AGPL situation. At some point you just have to accept that copyright law, in and of itself, is not sufficient to control what people do with your software. If you want to do that, you have to limit end-user freedom with an EULA.
If someone uses LLaMa output to train models, it is unlikely they will be sued for copyright infringement. It is far more likely they will be sued for breach of contract.
Training a model on model output isn't copying.
There's no way to phrase this where training a model on copyrighted human-generated images/text isn't copying, but training a model on computer-generated images/text is copying.
> If you want to do that, you have to limit end-user freedom with an EULA.
If you want to limit end-user freedom with a EULA, you have to figure out how to get users to sign it. Copyright is one way to force them to do so, but doesn't really seem relevant to this situation if training a model on copyrighted material is fair use.
And again, if somebody generates a giant dataset with LLaMA, if you want to argue that pushing that into another LLM to train with is making a copy of that data, then there's no way to get around the implication there that training on a human-generated image is also making a copy of that image.
That's literally what I said.
> There's no way to phrase this where training a model on copyrighted human-generated images/text isn't copying, but training a model on computer-generated images/text is copying.
Literally nobody is saying that.
> If you want to limit end-user freedom with a EULA, you have to figure out how to get users to sign it.
That is not true. ProCD v. Zeidenberg, 86 F.3d 1447 (7th Cir. 1996).
You and others seem to have an over-the-top hostile reaction to the idea that contract law can do things copyright law cannot do. But it is objective and unarguable fact.
Okay? Apologies for making that assumption. But if you're not saying that, then your position here is even less defensible. Arguing that model output isn't copyrightable but that it's still covered by EULA if anyone anywhere tries to use it is even more absurd than arguing that it's covered by copyright. The interpretation that this is covered by copyright is arguably the charitable interpretation of what you wrote.
> That is not true. ProCD v. Zeidenberg, 86 F.3d 1447 (7th Cir. 1996).
ProCD is about shrinkwrap licenses, the court determined that buying the software and installing it was the equivalent of agreeing to the license.
In no way does that imply that licenses are enforceable on people who never agreed to the licenses. The court expanded what counts as agreement, it does not mean you don't have to get people to agree to the EULA. I mean, take pedantic issue with the word "sign" if you want (sure, other types of agreement exist, you're correct), but the basic point is still true -- if you want to restrict people with a EULA, they need to actually agree to the EULA. All that ProCD did was establish that buying a product and opening the package and installing it constituted agreement.
And that becomes a problem because if you don't have IP law as a way to block access to your stuff, then you don't really have a way to force people to agree to the EULA. Someone using LLaMA output to train a model may have never been in a position to agree to that EULA, and Facebook doesn't have the legal ability to say "hey, nobody can use output without agreeing to this" because they don't have copyright over that output. Can they get people to sign a EULA before downloading the weights from them? Sure. Is that enough to restrict everyone else who didn't download those weights? No.
To go a step further, if you don't believe that weights themselves are copyrightable, then putting a EULA in front of them is even less effective because people can just download the weights from someone else other than Facebook.
You can host a project Gutenberg book and get people to sign a EULA before they download it from you, even though you don't own the copyright. And that EULA would be binding, yes. But you cannot host a project Gutenberg book, put a EULA in front of it, and then claim that people who don't download it from you and instead just grab it off of a mirror are still bound by that EULA.
Your ability to control access is what gives you the ability to force people to sign the EULA. And that's kind of dependent on IP law. If someone sticks the LLaMA 2.0 weights on a P2P site, and those weights aren't covered by copyright or other IP law, then no, under no interpretation of US law would downloading those weights from a 3rd-party source constitute an agreement with Facebook.
But even if you don't take that position, even if you assume that model weights are copyrightable, if I download a dataset generated by LLaMA, there is still no shrinkwrap license on that data.
To your original point:
> If someone uses LLaMa output to train models, it is unlikely they will be sued for copyright infringement. It is far more likely they will be sued for breach of contract.
It is incredibly unlikely that someone using a 3rd-party database of LLaMA output would be found to be in violation of contract law unless at the very least they had actually agreed to the contract by downloading LLaMA themselves. A restriction on the usage of LLaMA does not mean anything for someone who is using LLaMA output but has not taken any action that would imply agreement to that EULA.
> You and others seem to have an over-the-top hostile reaction to the idea that contract law can do things copyright law cannot do. But it is objective and unarguable fact.
No, what we have a hostile reaction to is the objectively false idea that a EULA covers unrelated 3rd parties. That's not a thing, it's never been a thing.
I don't know what to say if you disagree with that other than that I'm putting a EULA in front of all of Shakespeare's works that says you now have to pay me $20 before you use them no matter where you get them from, and apparently that's a thing you believe I can do?
Clickwrap agreements are enforceable, and legally enforceable agreements can place more restrictions on the use of a piece of software than copyright law alone can.
As a result, software that, for ideological reasons, does not restrict use will always have fewer protections than software with more restrictive terms.
Your off-topic rant about Shakespeare is irrelevant.
> Clickwrap agreements are enforceable, and legally enforceable agreements can place more restrictions on the use of a piece of software than copyright law alone can.
To take a page from your earlier comment, literally no one here is denying the existence of clickwrap agreements. Clickwrap agreements are completely irrelevant to the current conversation.
> Your off-topic rant about Shakespeare is irrelevant.
You can not enforce a EULA on someone interacting with a piece of work you do not own IP rights to if they did not agree to that EULA in some way.
I'm sorry, but agreement is part of contract law.
If you think you can force a EULA on a piece of content you don't own that will bind people who got the content from a 3rd-party and who never agreed to your EULA under any legal definition of agreement, then by all means, slap a EULA on Shakespeare. It makes just as much sense as what you're suggesting.
> literally no one here is denying the existence of clickwrap agreements.
You denied the enforceability of clickwrap agreements. You were wrong.
LLaMA uses a clickwrap agreement. "By clicking 'I Accept' below or by using or distributing any portion or element of the Llama Materials, you agree to be bound by this Agreement."
That agreement covers its output: "You will not use the Llama Materials or any output or results of the Llama Materials to improve any other large language model (excluding Llama 2 or derivative works thereof)."
Your hypotheticals about third parties are off-topic and have zero bearing on this conversation.
The topic under discussion is whether it is logically "inconsistent" for Meta to claim its output is protected while other content is not. Those two positions are perfectly consistent in light of the fact that LLaMA output is protected by the terms of a clickwrap agreement.
It's all but certainly copied, and not just in the "held in memory" sense but actually stored along with the rest of the training collection. What may not happen is distribution. There's a difference in scale/nature of copyright violation between the two but both could well be construed that way.
Additionally, I think there's a reasonable argument that use as training data is a novel one that should be treated differently under the law. And if there's not:
> If you want to do that, you have to limit end-user freedom with an EULA.
What will eventually happen -- at least without some kind of worldwide convention -- is that someone who can successfully dodge licensing obligations will be able to take and redistribute weight-data and/or clean-room code.
At least, if we're adopting a "because we can" approach to everything related.
Regardless, the lack of a license cannot give you more permission than a restrictive license. You're arguing that if take a book out of a bookstore without paying (or signing a contract), then I have more rights than if I sign a contract and then leave with the book.
That is clearly false. It's hard to imagine the confusion of ideas that would lead you to such a conclusion.
The whole legal premise of these models is that training on copyrighted material is fair use. If it's not, then... I mean is Facebook trying to claim that including copyrighted material in a dataset isn't fair use regardless of the author's wishes? Because I have bad news for LLaMA then.
"You need permission to train on this" is an interesting legal stance for any AI company to take.
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).
> 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'll add to this that it's not just output; say that someone is using another service built on top of LLaMA. Facebook itself launched LLaMA 2.0 with a public-facing playground that doesn't require any license agreement or login to use.
You can go right now and use their public-facing portal and generate as much training data as you can before they IP-block you, and... as far as I can tell you haven't done anything in that scenario that I can see that would bind you to this license agreement.
So I still feel like I'll be surprised if any AI company that's serious about wanting bootstrapping itself off of LLaMA is going to be too concerned about this license (whether that's a good idea to do just because the training data itself might be garbage is another conversation). It just seems so easy to get around any restrictions.
NN design is more interesting, but I don't think we're at the point yet where they are sufficiently complex to be copyrightable in general. Patentable, maybe.
Majority of the time, the code and weights are under independent license terms-while in theory the code license could say it is revoked or revocable if you violate the terms of the weights licenses, I think such a license term is rare in practice.
It is quite common even when the weights are under a restricted license for the code to be released under a standard open source license, and no open source license contains such a license term (and it would probably make the license non-open source were it included)
Not to diminish the conversation here, but not even a Supreme Court Justice knows what the legality is. You’d have to be a whole 9 person Supreme Court to make an accurate statement here. I don’t think anyone really knows how Congress meant today’s laws to work in this scenario.
Congress, or more accurate, the drafters of the Constitution, intended that Congress would work to keep the Constitution updated to match the needs of modern times. Instead, Congress ossified to the point it's unable to pass basic laws because a bunch of far right morons hold the House GQP hostage and an absurd amount of leverage was passed to the executive and the Supreme Court as a result - with the active aid of both parties by the way, who didn't even think of passing actual laws to codify something as important as equitable access to elections, fair elections, or the right to have an abortion or to smoke weed when they held majorities. And on top of that your Supreme Court and many Federal court picks were hand-selected from a society that prefers a literal viewpoint of the constitution.
But fear not, y'all are not alone in this kind of idiocy, just look at us Germans and how we're still running on fax machines.
If they turn out to be not copyrightable, then... all this would mean is downloading LLaMA 2.0 weights from a mirror instead of from Facebook.
"Yes, we train our models on a good chunk of the internet without asking permission, but don't you dare train on our models' output without our permission!"
And OpenAI also has a similar restriction.
You can't use a model's output to train another model, it leads to complete gibberish (termed "model collapse"). https://arxiv.org/abs/2305.17493v2
And the Llama 2 license allows users to train derivative models, which is what people really care about. https://github.com/facebookresearch/llama/blob/main/LICENSE
Yes, you cannot do this kind of thing indefinitely and expect endless improvements from "endless training set". But that's a very different problem.
> Supreme Court rejects Genius lawsuit claiming Google stole song lyrics SCOTUS won't overturn ruling that US copyright law preempts Genius' claim.
> The song lyrics website Genius' allegations that Google "stole" its work in violation of a contract will not be heard by the US Supreme Court. The top US court denied Genius' petition for certiorari in an order list issued today, leaving in place lower-court rulings that went in Google's favor.
> Genius previously lost rulings in US District Court for the Eastern District of New York and the US Court of Appeals for the 2nd Circuit. In August 2020, US District Judge Margo Brodie ruled that Genius' claim is preempted by the US Copyright Act. The appeals court upheld the ruling in March 2022.
> "Plaintiff's argument is, in essence, that it has created a derivative work of the original lyrics in applying its own labor and resources to transcribe the lyrics, and thus, retains some ownership over and has rights in the transcriptions distinct from the exclusive rights of the copyright owners... Plaintiff likely makes this argument without explicitly referring to the lyrics transcriptions as derivative works because the case law is clear that only the original copyright owner has exclusive rights to authorize derivative works," Brodie wrote in the August 2020 ruling.
> Google search results routinely display song lyrics via the service LyricFind. Genius alleged that LyricFind copied Genius transcriptions and licensed them to Google.
> Brodie found that Genius' claim must fail even if one accepts the argument that it "added a separate and distinct value to the lyrics by transcribing them such that the lyrics are essentially derivative works." Since Genius "does not allege that it received an assignment of the copyright owners' rights in the lyrics displayed on its website, Plaintiff's claim is preempted by the Copyright Act because, at its core, it is a claim that Defendants created an unauthorized reproduction of Plaintiff's derivative work, which is itself conduct that violates an exclusive right of the copyright owner under federal copyright law," Brodie wrote.
https://arstechnica.com/tech-policy/2023/06/supreme-court-re...
That is not what this court case was about. Genius had already settled the case of unauthorised transcriptions and had bought licences for its lyrics after a lawsuit 2014, so its own work was no longer unauthorised. In the case cited above, Genius was trying to enforce its claims against Google via contract law rather than copyright law. The court ruled that the alleged violations were covered by copyright law, so they could only pursued via copyright law, and that only the copyright holder (or assignee) of the lyrics that were copied could sue Google under it.
As for generating jsons, that's more of a inference runtime thing, since you need to pick the top tokens that result a valid json instead of just hoping it returns something that can be parsed. On top of extensive tuning of course.
I don't know how one can even forbid this. As a human, I'm a walking neural net, and I train myself on everything that I see, without a choice. The only difference is I'm a carbon-based neural net.
Ezpz
The source of a language model is more in reality the model, that is the code that was used to train the particular model. The model itself is more of a compiled binary, altough not in machine code.
So for a model to be really open source to me it would mean that you have to release the software used for generating it, so I can modify it, train it on my data, and use it.
1) "Dear artists, the model cannot infringe upon your copyright because it's merely learning like a human does. If it accidentally outputs parts of your book, you know, it just accidentally plagiarized. We all do it haha! Our attorneys remind you that plagiarism is not illegal in the US."
2) "Dear engineers, the output of our model is copyrighted and thus if you use it to train your own model, we own it."
I am not sure how both of those can be true at the same time.
I have no idea how this impacts the encodability of the license from FB which may rely on things other than copyright, but as of right now, the output absolutely cannot be copyrighted.
That being said, isn't a prompt guidance?
Oops.
I added the "haha" in there because the probability of a human doing this kind of goes way down as the length of the text increases. Can you type, verbatim, an entire chapter of a book? I can't. But, I bet the AI can be convinced in rare cases to do that.
The whole thing is very interesting to me. There was an article on here a couple days ago about using gzip as a language model. Of course, gzipping a book doesn't remove the copyright. So how low does the probability of outputting the input verbatim have to be before copyright is lost?
Reading the book and benefitting from what you learned? Obviously not copyright infringement. Putting the book into gzip and sending your friend the result? Obviously copyright infringement. Now we're in the grey area and ... nobody knows what the law is, or honestly, even how to reason about what the law wants here. Fun times.
(Personally, I lean towards "not copyright infringement", but I'm not a big believer in copyright myself. In the case of AI training, it just makes it impossible for small actors to compete. Google can just buy a license from every book distributor. SmolStartup can't. So if we want to make AI that is only for the rich and powerful, copyright is the perfect tool to enable that. I don't think we want that, though.
My take is that the rest of society kind of hates Tech right now ("I don't really like my Facebook friends, so someone should take away Mark Zuckerberg's money."), so it's likely that protectionist laws will soon be created that ruin it for everyone. The net effect of that is that Europe and the US will simply flat-out lose to China, which doesn't care about IP.)
Of course, the models that are developed for internal use by the Chinese government won't be so limited, regardless of what the law says. But then neither be the ones developed by Western three-letter agencies. So don't worry about the "Great Game"; they'll do just fine one-upping each other and screwing over all of us in the process.
This is why fundamental "interpolative" techniques like ChatGPT (whose weights are in theory frozen) is still basically super-intelligent.
It's a well-understood concept that our minds function by making sense of the world through patterns. This is the essence of interpolation - taking two known points and making an educated guess about what lies in between. Ever caught yourself finishing someone's sentence in your mind before they do? That's your brain extrapolating based on previous patterns of speech and context. These processes are at the heart of human creativity.
The field of Cognitive Science has extensively documented our tendency for interpolation and pattern recognition. Works like The Handbook of Imagination and Mental Simulation by Markman and Klein, or even "How Creativity Works in the Brain" by the National Endowment for the Arts all attest to this.
When artists create, they draw from their experiences, their knowledge, their understanding of the world - a process overwhelmingly of interpolation.
Now, I can see how you might be confused about my reference to ChatGPT being "super-intelligent". Perhaps "hyper-competent" would be more appropriate? It has the ability to generate text that appears intelligent because it's interpolating from a massive amount of data - far more than any human could consciously process. It's the ultimate pattern finder.
And that, my friend, is my version of "publications on the subject of how minds work." I may not be an illustrious scholar, but hey, even a clock is right twice a day! And who knows, maybe I'm on to something after all.
If they are, and the license allows creating finetuned models but not using the output to improve the model, then the derived model is not a violation, but it might be a derivative work.
So, no matter what they TOS says, it's not an infringing work.
> Downstream users can't just integrate the derivative work into a product without abiding by the GPL terms
You absolutely could do this if the original work is not protected by copyright, or if you use it in a way that is transformative and fair use.
Similarly, if someone creates a fair use/transformative work then the license can also be ignored.
Your customers bought that product under license A. Afterwards it turned out that you pirated some artwork from disney. Then your customer can sue you (not disney) to make things right. The specific license of the original work seems quite irrelevant here.
edit: you seem to see the customer as the primary victim here instead of Disney, but if Disney weren't a victim the customer wouldn't have a case.
That's... not really accurate. See the concept of tortious interference with a contract.
That said, doing this inside one company or with subsiduaries probably wouldn't fly.
Can you summarize why weights would not be copyrightable or give me pointers to sources that support that view.
Let’s talk about more complex models. What if my model shares 5% of the same weights with your model? What about 50%? What about 99%? How much do these have to change before you’re in the clear? What if I take your exact model and run it through some extra layers that don’t do anything, but dilute the significance of your weights?
It’s a murky area, and I’m inclined to think copyright is not at all the right tool to handle the legality of these models (especially given the glaring irony they are almost all trained using copyrighted material). Patents, perhaps better suited, but I’m also not sold.
1. Model weights are the output of mathematical principles, in the US facts are not copyrightable, so in general math is not copyrightable.
2. Model weights are the derivative work of all copyrighted works it was trained on - in which case, it would be similar to creating a new picture which contains every other picture in the world inside of it. Who is the copyright owner? Well, everyone, since it includes so many other copyright holders' works in it.
related, there was a presentation (i've lost the reference) on automatic song (tune?) generation where the presenter claimed (rather humourusly) that he'd generated all the songs that had ever been and will ever be so that while he was infringing on a large but finite number of songs, he was non infringing on an infinite number of future songs. So, on balance he was in a favourable position.
One cannot hold copyright facts, but one can "copyright" a collection of facts like a search index or a map.
https://en.wikipedia.org/wiki/Integrated_circuit_layout_desi...
In the case of an LLM, I don't think that the work of compiling the training data probably would qualify by analogy to the phonebook example.
But on reflection, you are totally right, I was just getting mixed up on the distinction between copies and the creative works themselves. Machine output of something is generally just a copy of something. Whatever it is a copy of may be a copyrightable work, and if so, whoever came up with that original work has the right to all the copies output by machines (or copies generated by hand-tracing, or whatever).
Anyway, on LLMs... Even if we assume LLM weights are just copies (machine outputs) of whatever inputs they were trained on, then I assume I would automatically own the exclusive right to restrict the distribution of weights of a 'Me' chatbot trained exclusively on my own writings. But what if someone else comes along and writes a load of bespoke code specifically to generate improved weights for this same model, so the resultant chatbot works much better in conversation (still with my tone of voice, but with better performance and better interpretation of questions)? Is that programmer not adding some creative value, such that we might both have a right to restrict distribution of those improved weights? (NB. it's common for an item to be a 'copy' of multiple original works, e.g. copies of Jimi Hendrix's cover of Bob Dylan's 'All Along the Watchtower'.)
Its a mechanical copy subject to the copyright on the original, though.
If the original training data is a copyrightable (derivative or not) work, perhaps eligible for a compilation copyright, the model weights might be a form of lossy mechanical copy of that work, and be both subject to its copyright and an infringing unauthorized derivative if it is.
If its not, then I think even before fair use is considered the only violation would be the weights potentially infringing copyrights on original works, but I don’t think incomplete copy automatically works for them the way it would for an aggregate; I’d think you'd have to demonstrate reproduction of the creative elements protected by copyright from individual source works to make the claim that it infringed them.
If it isn't recognisable, then it's merely _distributed_ plagiarism. A million output, each of which are 0.0001% plagiarising each of million inputs.
Is The War on Drugs a VC-funded band replacement?
Are other future bands going to learn from The War on Drugs?
https://www.cbsnews.com/news/ai-stable-diffusion-stability-a...
https://www.documentjournal.com/2023/05/ai-art-generators-mo...
A codec conversion is not copyrightable. The original song which is still present enough in the conversion to impact its ability to be distributed, is still copyrightable. But you don't get some kind of new copyright just because did a conversion.
For comparison, if you take a public domain book off of Gutenberg and convert it from an EPUB to a KEPUB, you don't suddenly own a copyright on the result. You can't prevent someone else from later converting that EPUB to a KEPUB again. Copyright protects creative decisions, not mathematical operations.
So if there is a copyright to be held on model weights, that copyright would be downstream of a creative decision -- ie, which data was it trained on and who owned the copyright of the data. However, this creates a weird problem -- if we're saying that the artifact of performing a mathematical operation on a series of inputs is still covered by the copyright of the components of that database, then it's somewhat tricky to argue that the creative decision of what to include in that database should be covered by copyright but that copyrights of the actual content in that database don't matter.
Or to put it more simply, if the database copyright status impacts models, then that's kind of a problem because most of the content of that training database is unlicensed 3rd party data that is itself copyrighted. It would absolutely be copyright infringement for OpenAI/Meta to distribute its training dataset unmodified.
AI companies are kind of trying to have their cake and eat it too. They want to say that model weights are transformed to such a degree that the original copyright of the database doesn't matter -- ie, it doesn't matter that the model was trained on copyrighted work. But they also want to claim that the database copyright does matter, that because the model was trained on a collection where the decision of what to include in that collection was covered by copyright, therefore the model weights are copyrightable.
Well, which is it? If model weights are just a transformation of a database and the original copyrights still apply, then we need to have a conversation about the amount of copyrighted material that's in that database. If the copyright status of the database doesn't matter and the resulting output is something new, then no, running code on a GPU is not enough to grant you copyright and never really has been. Copyright does not protect algorithmic output, it protects human creative decisions.
Notably, even if the copyright of the database was enough to add copyright to the final weights and even if we ignore that this would imply that the models themselves are committing copyright infringement in regards to the original data/artwork -- even in the best case scenario for AI companies, that doesn't mean the weights are fully protected because the only copyright a company can claim is based on the decision of what data they chose to include in the training set.
A phone book is covered by copyright if there are creative decisions about how that phone book was compiled. The numbers within the phone book are not. Factual information can not be copyrighted. Factual observations can not be copyrighted. So we have to ask the same question about model weights -- are individual model weights an artistic expression or are they a fact derived from a database that are used to produce an output? If they're not individually an artistic expression, well... it's not really copyright infringement to use a phone book as a data reference to build another phone book.
If they can't charge for and control those other things, then we'll likely see far fewer companies releasing weights. Most of this stuff will move behind APIs in that scenario.
What if there were billions of knobs, tuned after years of feedback and observations of the sound output?
> A century later, in Feist Publications v. Rural Telephone Service Co., the Supreme Court confirmed that originality is a constitutional requirement
I don’t see why, for-profit companies release permissively-licensed ooen-source code all the time, and noncopyrightable models aren't practically much different than that.
> I don’t see why, for-profit companies release permissively-licensed ooen-source code all the time
I agree with this - however, they tend to open-source non-core components - Google won't release search engine code, Amazon wont release scalable-virtualization-in-a-box, etc.
I'm confident that Facebook won't release a hypothetical Llama 5 in a manner that enables it to be used to improve ChatGPT 8 - the aim will be unchanged from today, byt the mechanism will shift from licensing to rate-limiting, authentication & IP-bans.
As mercenary as it may sound, what these companies are trying to do is find a business model that is as friendly to themselves as it is hostile to their competitors.
This is all part of the jockeying.
The practical difference between the licenses is small enough that I expect most people (including me) will choose Llama2 anyway, because the models are higher quality. But that incentive may mean that we get stuck with these awkward pseudo-open licenses.
While the FSF side of things doesn't like the term "open source," even they say that "nearly all open source software is free software." Specifically, the MIT and Apache (and LGPL) licenses are absolutely free software licenses--otherwise Debian, FSF-approved distros, etc. would have far less software to choose from.
What the chart probably meant to distinguish is copyleft vs free software or open source. And if you're ordering it from a permissiveness viewpoint, the subset relationship should be reversed--GPL is far more permissive than SSPL, etc., but still less permissive that MIT/Apache.
This part of the article is pretty misleading as well:
> Free software, as specified by the Free Software Foundation, is only a subset of open source software and uses very permissive licenses such as GPL and Apache.
We haven't seen this too much yet in the AI world, as mostly people who open the weights are doing so in a research manner, where the inference is decidedly needed to be open sourced- and people with closed models do so in order to make money and thus no reason to open source the inference side either, just charge for an API ("OpenAI").
It is actually editorialized in a way that feels quite different from the actual one. I think the author and the poster might disagree on what open source means.
Here is another article on LLaMa2: https://opensourceconnections.com/blog/2023/07/19/is-llama-2...
Llama2 is a binary blob pre-trained model that is useful and is licensed in a fairly permissive way, and that's fine.
At least binary blobs can be disassembled.
(Downvotes... oops. The reference is Charlie Stross's Accelerando. The protagonist has a conversation with an AI that's just trying to survive. One of the options he suggests is to open source itself. Which is a roundabout way of saying that eventually we're going to have to take the AI's own opinions into account. What if it doesn't want to be open source?)
There is no marked increase in model effectiveness in these 'new' versions, but even if you just use the 'YOLOv8' Pytorch weights (and no part of their Python toolchain, which might have some improvements), these will somehow try to download files from Ultralytics servers. Possibly for a good reason, but most likely to, let's say, "pull an Oracle."
Serious AI researchers won't go anywhere near this stuff, but the number of students-slash-potential-interns with "but it's on GitHub!" expectations that I had to reject lately due to "nope, we're not paying these guys for their Enterprise license just to check out your project" is rather disheartening...
Then there's the problems of the content of the training data, which parallel the dangers of opaque algorithms.
When it comes to "how much of it has to be available to be open source", I think it may be instructive to look at encryption algorithms.
Many of them have numeric constants or initial values--NOT part of the secret key itself--which need to be known and available, both for interoperability and for the expected security-level of the algorithm. These are arguably similar to LLM weights. (Perhaps the simplest example would be the prominence of "13" in ROT13.)
Yet if someone tried saying that their encryption standard was "open source" while keeping those constants secret and/or legally-encumbered, I think a lot of people would complain that the label is incorrect or inappropriate.
What these models do, they should either invented a new term, or use an appropriate existing term, eg. "fair use"
I hope that users aggressively ignore these restrictive licenses and give the middle finger to greedy companies like Facebook who try to restrict usage of their models. Information deserves to be free, and Aaron Swartz was a saint.
- pseudo/true random number generator and initialization
- certain speculative optimizations associated with training environments (distributed)
- Speculative optimizations associated with model compression
- Image decompression algorithm mismatch (basically this is library versioning)
- ....things I'm forgetting...
It's just a lot of things to remember to capture, communicate, and reproduce.
It's not just the generator and initialization. If you do anything multithreaded, like a producer/consumer queue, then you need to know which pieces of work went to which thread in which order.
It's a lot like reproducing subtle and rare race conditions.
No, CC-NC-ND is a thing, and even GPL applies restrictions on derivation as well.
"Open source" doesn't mean BSD/MIT. There is even open-source that you cannot freely redistribute at all - not all open-source is FOSS!
I always think it's a testament to how much copyleft has succeeded that in many cases people think of GPL and BSD/MIT as being the baseline.
"FOSS" has absolutely no requirement of it being copyleft. The MIT license is just as FOSS as the GPL. Many of the free software advocates do have an affinity for copyleft, but they are not mutually exclusive. There are plenty of FOSS advocates who also use and advocate for permissive licenses as well.
That original sense never existed. Virtually nobody said "open source" before OSI's 1998 campaign for "Open Source", as bankrolled by Tim O'Reilly.
https://thebaffler.com/salvos/the-meme-hustler
I know it's been a long time, and we've forgotten, but there is virtually no record of anyone saying "open source" before 1998, except in rare and obscure contexts and often unrelated to the modern meaning.
https://web.archive.org/web/20180402143912/http://www.xent.c...
I believe there is a name for that: gongkai. https://www.bunniestudios.com/blog/?page_id=3107
LGPL3.0 also pretty much is restricted in a way that not sure if can be used to distribute software in App Store for iOS legally.
As for terms, we'll settle on '$0 downloadable AI models' which are available today. Would rather use that over cloud-only AI models which can fall over and break your app at any time and you have zero control over that.
Stable Diffusion is a good example that fits the definition of 'open-source AI' as we have the entire training data, weights reproduciblity, etc and Llama 2 does not.
Chat-tuned derivatives of LLaMa 2 are already appearing. Given that the base LLaMa 2 model is more efficient than LLaMa 1, it is reasonable to expect that these more refined chat-tuned versions of the chat-tuned versions will outperform the ones you mention.
Find the optimal route for Karen's quest which maximizes her chances of defeating the ogre to 100%. ----2---- Write a python code using imageio.v3 to create a PNG image representing the map way-points and the route of Karen in her quest, each way-point must be of a different color and her path must be a gradient of the colors between the waypoints. ------------
I have a lot of cases those I test against different models ... GPT-4 since one week is really degraded, GPT-3.5 became a little bit better, and LLaMA2 is garbage.