Comment by tikkun | Hacker News Reader

Comment by tikkun | Hacker News Reader

Here's my understanding (may be wrong!) of the license [1] and the acceptable use policy [2] in terms of what you can't do:

You are not allowed to do the following under the Falcon 180B TII License Version 1.0:

1. Use Falcon 180B to break any national, federal, state, local or international law or regulation.

2. Exploit, harm or attempt to harm minors and living beings in any way using Falcon 180B.

3. Create or share false information with the purpose of harming others.

4. Use Falcon 180B for defaming, disparaging or harassing others.

Notable: 5. Use Falcon 180B or any of its works or derivative works for hosting use, which is offering shared instances or managed services based on the work, unless you apply and are granted a separate license from TII.

Notable: 6. Distribute the work or derivative works unless you comply with several conditions such as including acceptable use restrictions, giving a copy of the license to recipients, stating changes you made, and retaining copyright and attribution notices among others.

7. Use trade names, trademarks, service marks or product names of the licensor unless required for reasonable and customary use in describing the origin of the work or reproducing the content of the NOTICE file.

[1]: https://huggingface.co/spaces/tiiuae/falcon-180b-license/blo...

[2]: https://falconllm.tii.ae/acceptable-use-policy.html

I remain skeptical that models are licensable at all. To be subject to licensing, they would have to be property. What kind of property are they?

Certainly, they are not copyrighted works. You can’t copyright mere data. You could no more copyright a model than you could a phone book, or even a matrix transformation of a list of phone numbers.

And even if they are covered by copyright, they are hopelessly tainted by the copyrighted works they are trained on without license. Without upstream licensing, licensing the model is usurping the rights of the original authors.

> You could no more copyright a model than you could a phone book

Just as an interesting side note, some jurisdiction recognize something apparently called “database right” in English (in Swedish it’s more like “catalog right”).

It’s a kind of intellectual property right for the work of compiling a database.

Perhaps applicable to the weights of a model? But the US does not recognize this as a thing

https://en.wikipedia.org/wiki/Database_right

Not really. Maps are also mere data, but they are quite successfully copyrightable. There's even a concept of trap streets [0] used to find out if someone used your data in their map without permission.

AI models don't have an established legal framework yet, but it's reasonable to assume that similar rules will apply here.

[0] https://en.wikipedia.org/wiki/Trap_street

Maps are not mere data. There is a lot of creativity involved in choosing the data to present and leave out, the style and colors, the arrangement of labels, etc. That's why maps are copyrightable. There is comparatively little creativity involved in feeding large fractions of the internet through a standard transformer model. Neither is there any significant creativity involved in the presentation of the raw weights. It's not at all clear to me that the weights are or should be copyrightable.

People will certainly try, though, and like all regulatory regimes copyright loves to expand and never voluntarily shrinks, so they may succeed. Honestly I think the most likely outcome is that model weights will be ruled as derived works of the input dataset, and courts will try to enforce that people who train models must license their entire dataset specifically for model training. Some would cheer that but I personally think it would be a disaster.

The training process is closely curated, bootstrapped, the bulk data is mixed with whatever manual data you have, and it generally requires tremendous amounts of human input and expertise for the model to be even remotely good. It's definitely not just "feeding the data to the model".

Compiling and checking and selecting and filtering and sorting the phone numbers for a phone book and printing that book and distributing it is not trivial. It involves significant work and even a few creative and editorial decisions, but on balance it's not that creative of a process, and the resulting phone book is not copyrightable.

As for mixing in your own data, if the model weights inherit the copyright of the data then every existing large language model is illegal.

Fair points, within the context that IP rights are fundamentally made up and shifting rights crafted to balance the concerns and interests of various parties.

I can't tell where this is going to land. Right now, we're seeing a number of parties trying to put metaphorical barbed wire in the newfound prairie of ML models, each struggling to influence the prevalent wisdom of how IP rights should apply to this context.

We could easily end up in a universe where LLMs are a licensing minefield where every copyright owner of any part of their training data gets rights on the model, become essentially unmanageable without relying on helpful licensing middlemen that smooth out the right for LLMs to exist, at a cost.

We could just as well end up with LLMs being recognized as not being derivative works themselves, albeit able to generate derivative works, a much less advantageous situation for creatives and their middlemen who see their creative output as being pirated, to reuse a familiar term of IP rights propaganda.

It'd be a little surprising to me if we ended up in a situation where the work needed to produce good LLMs wasn't associated with any commensurate IP rights on the results, and I expect the megacorps investing billions into this will find it in their heart to throw a few millions toward lobbying efforts to ensure that this isn't the outcome.

They don’t obviously fall under copyright, patent, or trademark—those are the only options for IP.

You can enter into general business contracts that govern how the parties make use of data. This happens all the time with all kinds of data sources: business listings, GIS, credit scores, etc.

If I copy of this kind of data, I might be breaking a contract and committing a tort, but I am not violating copyrights. If a third party gets the data without a contract in place and makes use of it, they are not violating copyright either; the liability falls on the contracted party that let the data get out.

But licenses of the kind proposed on models are inapplicable. Imagine how bizarre it would be if the phone book came with a license stating “you may only use the information for non-commercial purposes.” The phone book publisher would get laughed out of court and maybe even penalized for frivolous lawsuits.

You left out the part where they say it IS okay to use it on the backend of a hosted application or service.

The way I read this is that they are reserving their right to build an API product like OpenAI based on this model, and control that particular part of the market. So it can't just be a wrapper putting another brand on it that is open to general purpose use of the model.

But you can build hosted chat-based applications. They just need to be applying the model to some use case.

Living beings? Meaning I can't ask it how to treat mice/termite problems in my house?

It allows to use it to train other models, right?

Reply on news.ycombinator.com