Gen AI Open-Source? Who cares about open-source software licensing in the AI space?
Lots of people. Any company building AI-related products or product features should care because when doing so they have 2 options: (1) pay someone so you can build your stuff on top of their proprietary AI model or (2) use a free-and-open-source model to power your AI product or AI features.
Are open source-licensed AI models comparable in quality and power to proprietary models?
Yes, for the most part, which is why open-source licensing is a hot topic in the AI space currently.
What types of licenses are we seeing amongst popular open-source AI models?
Here’s a sample from Huggingface of recently-popular models with high download counts:
As you can see, lots of permissive open-source licenses. Permissive open-source licenses grant the user wide discretion to build and sell products that use their underlying open-source-licensed AI model. If you’re a software developer, product manager, or CIO/CEO/CTO/CISO and could use a primer on copyright vs. copyleft, check out our Copyleft Guide for Software Developers article to brush up on this stuff.
In a couple of instances we have the CC-BY-NC-SA-4.0 license. These licenses categorically prohibit selling products and features that were built using this model.
Are Creative Commons licenses good for open-source AI models?
No. Creative Commons licenses are a dumpster fire for software licensing generally, including and especially in the AI space. Don’t just take some random tech lawyer’s word for it, though.
Creative Commons themselves have strongly advised against anyone applying their licenses to software.
The GNU is against using CC licenses in the software space: “Creative Commons Attribution 4.0 license (a.k.a. CC BY) – “This is a non-copyleft free license that is good for art and entertainment works, and educational works. It is compatible with all versions of the GNU GPL; however, like all CC licenses, it should not be used on software.”
I’ve read these licenses in detail. They were not designed for software, in my opinion. They were designed for text, images, video, in my opinion. The language they use is imprecise, vague, and ambiguous when read with an eye to software and data-related use cases, particularly in the context of AI.
They are perfectly fine licenses for their originally-intended use cases and I have personally used them to license my own images in the past. I am a fan of the Creative Commons licenses generally, just not for software use cases.
Is CC-BY-SA-4.0 a permissive open-source license for AI models?
No and I wouldn’t recommend licensing any AI models under it or using any models licensed under CC-BY-SA-4.0 to build a software product or feature you intend to sell in the future. See the above comments about Creative Commons licenses generally. The Share-Alike language in the license raises more questions than it answers when you’re talking about AI model licensing.
“But the people in the comment section on HackerNews said it was ok.” Good luck to you, my friend.
I get it, lots of AI models are licensed under CC-BY-SA-4.0, including very popular models. The popularity of the license does not, ipso facto, mean it’s a good license for commercial software products. There are plenty of use cases outside of commercial software products that would explain that popularity, and, in any case, ignorance is always and everywhere a human phenomenon (read, “a potential factor as well”).
CC-BY-SA is a bad license in the commercial software space because if you build anything on top of the model then the thing you’re building is, arguably, automatically free and open source. This is how companies have large portions of their codebase compromised from an intellectual property perspective. The exact language they use is if you build anything that is “derived from or based upon” the model. What, exactly, that means in real life is ambiguous. Certainly it would include a fine-tuned version of the model. Whether it would include a downstream application that is sending prompts to the model, for example, is unclear.
Are the Llama 3 and Llama 3.1 licenses open-source licenses?
Sort of. Let’s go back to the definition of “open source software” from my Copyleft Guide for Software Developers. “Open-source software is software that
- is made available to the public,
- in source code form,
- for free,
- under a license that puts certain conditions on its use in other software (potentially including SaaS/PaaS/IaaS) products.”
Answering this question requires deciding and defining what is the “source code” of a large language model (LLM). There are two schools of thought here:
- The source code of an LLM includes (i) the training data, (ii) the model weights, and (iii) the training code in source code format.
- Same as above minus the training data.
So if you fall into the first school of thought, most AI models licensed under and open-source license are not truly “open-source software.” If you fall into the second school of thought then Llama 3 is open source if you ignore what I’ll call “the spirit” of the open-source software movement.
Essentially, some believe that the word, open, in open-source serves two functions: (i) modifying “source” to form a compound adjective and (ii) connoting the purpose of open collaboration. If you believe this, the Llama 3 and Llama 3.1 licenses are not “open” in that they restrict certain use cases above and beyond what a typical open-source software license would. A similar analysis applies to their being “free” given that the licenses themselves require you to buy a license from Meta under certain circumstances (see below).
We’re about 3 layers too deep for most readers here so I will conclude this line of thinking by dropping a link to a fairly erudite conversation on the topic for those of you whose curiosity on the subject is, as yet, unsated.
I think most people asking this question about Llama 3 and Llama 3.1 being open-source are really interested in the answer to a different question, which is…
Can I build a for-profit application on top of Llama 3, legally?
The answer to that question, in most cases, is “yes but” or “yes and.” There are a couple of specific restrictions on the use of Llama 3.1-licensed (and Llama-3) models that are unusual for an open-source license:
- You have to buy a license from Meta if your company’s products have more than 700 million active users
- Specific to Llama 3 (doesn’t apply to Llama 3.1), you can’t use it to improve another LLM.
TL;DR?
Open-source LLMs are legit, lots of people are choosing stupid open-source licenses for them, you can build for-profit apps on Llama for 98% of use cases.