Breaking the Llama Community License | Hacker News Reader

Breaking the Llama Community License

notes.victor.earth

144 pointsmkl81 comments

This is why I've always considered the weights-vs-source debate to be an enormous red herring that skips the far more important question: are the weights actually "Open" in the first place?

If Llama released everything that the most zealous opponents of weights=source demand they release under the same license that they're currently offering the weights under, we'd still be left with something that falls cleanly into the category of Source Available. It's a generous Source Available, but removes many of the freedoms that are part of both the Open Source and Free Software Definitions.

Fighting over weights vs source implicitly cedes the far more important ground in the battle over the soul of FOSS, and that will have ripple effects across the industry in ways that ceding weights=source never would.

I don't think most people in the weights-vs-source debate misunderstands this, it's just the the current "open-source" models for the most part do not even meet the bar of source-available, so talking about if the license is actually Open is not the current discussion.

In today's world, if Meta did release full source they used to create Llama, there are only about a dozen institutions that have the capacity to actually do anything with that, and no one has that kind of spare capacity just lying around. So the question of having the source for now in this case is less about being able to do something with it, and more about behind able to examine what's going into it. Aside from making it so it won't tell me how to make cocaine or bombs, what other directives has it been programmed with on top of the intial training run. That's what's important here, so I disagree that is a red herring. Both aspects are important here, but the most important one is to not let Mark Zuckerberg co-opt the term Open Source when it's only model available, and definitely not even actually Open at that.

It gets even weirder with Llama 4: https://www.llama.com/llama4/use-policy/ [Update: Apparently this has been the case since 3.2!]

> With respect to any multimodal models included in Llama 4, the rights granted under Section 1(a) of the Llama 4 Community License Agreement are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union. This restriction does not apply to end users of a product or service that incorporates any such multimodal models.

This is especially strange considering that Llama 3.2 also was multimodal, yet to my knowledge there was no such restriction.

In any case, at least Huggingface seems to be collecting these details now – see for example https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Inst...

Curious to see what Ollama will do.

> Curious to see what Ollama will do.

I don't think they care. I'm pretty sure Llama itself trained on a bunch of copyrighted data. Have licence agreements actually mattered?

> I'm pretty sure Llama itself trained on a bunch of copyrighted data.

Every good, "SotA" model is trained on copyrighted data. This fact becomes aparent when models are released with everything public (i.e. training data) and they score significantly behind in every benchmark.

Research team from orielly found out openai trained on copyirghted books

prob got a sub...

https://ssrc-static.s3.us-east-1.amazonaws.com/OpenAI-Traini...

Technically, those were there since 3.2's Use Policy. I've summarized the changes of the license here: https://notes.victor.earth/how-llamas-licenses-have-evolved-...

Oh, I must have compared the license of 3.2 to the usage policy of 4 then or made some other error; I was convinced this was a new restriction!

Thank you, also for that article – the tabular summary of changes across the two is great!

AFAIK it’s still an open question whether there is any copyright in model weights, for various reasons including the lack of human authorship. Which would mean that if you didn’t separately and explicitly agree to the license by clicking through something, there is no basis for the “by using this model” agreement.

Of course you probably don’t have enough money to get a ruling on this question, just wanted to point out that (afaik) it is up for debate. Maybe you should just avoid clicking on license agreement buttons, if you can.

I'm in the "model weights aren't copyrightable" camp myself. I think the license exists largely to shield Meta from liability or the actions of third parties.

Humans make many choices that effect the trained weights: curation and choice of datasets, training schedules and hyperparameters. If these choices are made with an eye towards the generated results, rather than mechanically based upon test scores, why doesn't that rise up to the minimal level required to get a copyright on the weights?

It might. As I say, it’s up for debate. A judge might look at the 1kB of hyperparameters versus the 1TB of training data, and the 10 person-years of human effort versus 100,000 GPU-years of computer effort, and conclude differently.

Does Google have copyright of their search index? Never tested, as far as I know.

"sweat of the brow" was rejected in Feist, the effort involved makes no difference. The minimum creativity bar for obtaining copyright is very low but above an alphabetical ordered listing of all of your customers.

> AFAIK it’s still an open question whether there is any copyright in model weights

There's definitely copyright when you ask the model to spit out Chapter 3 of a Harry Potter book and it literally gives it to you verbatim (Which I've gotten it to do with the right prompts). There's no world where the legal system gives Meta a right to license out content that never belonged to them in the first place.

A lot of posts here saying this is irrelevant or whatever - unix was mostly developed out in the open and got into all sorts of trouble when big companies started cracking down on copyright ownership. This blog post isn't entreating you to necessarily behave differently w/r/t your Llama usage, just to be aware the license is restrictive, doesn't really line up with Meta calling it "open source", and eventually this could create consequences down the road. The post doesn't even mention it but there are safer models to use right now (deepseek) that have permissive licensing.

@dang the title rename seems worse: "you're probably ..." is rather meaningfully different than "breaking the ..." as the latter sounds like it's instructions on how to "break the ..."

This was posted by the author a couple of weeks ago, but didn't get any traction: https://news.ycombinator.com/item?id=43504429

If a rogue laptop violates the Llama license and no one is around to enforce it, did it really break the Llama license?

Seriously, I genuinely wonder what the purpose of adding random unenforceable licenses to code/binaries. Meta knows people don’t read license agreements, so if they’re not interested in enforcing a weird naming convention, why stipulate it at all?

The guy who wrote this article is likely the only person to self-report to the police that he might have broken a dumb rule published by a huge company that manipulates and deceives millions of people every minute... haha

Also see: Open Source AI Definition Erodes the Meaning of “Open Source”

https://sfconservancy.org/blog/2024/oct/31/open-source-ai-de...

> The TLDR here, IMO is simply stated: the OSAID fails to require reproducibility by the public of the scientific process of building these systems, because the OSAID fails to place sufficient requirements on the licensing and public disclosure of training sets for so-called “Open Source” systems. The OSI refused to add this requirement because of a fundamental flaw in their process; they decided that “there was no point in publishing a definition that no existing AI system could currently meet”. This fundamental compromise undermined the community process, and amplified the role of stakeholders who would financially benefit from OSI's retroactive declaration that their systems are “open source”. The OSI should have refrained from publishing a definition yet, and instead labeled this document as ”recommendations” for now.

Does a thief have the right to demand nobody re-steals what was stolen in the first place?

If we accept the existence of intellectual property in the first place, all AI is blatant and unmitigated theft.

If we do not accept it, Llama has no right to enforce such terms.

That's a pretty reductionist view. Even maximally pro-IP laws usually have significant carve-outs for derivative works or other scenarios; see for example the US's extensive "fair use" doctrine.

Reply on news.ycombinator.com