We now note that that we do not waive our attribution requirements for Spritely Goblins, which are present in Apache v2 (and for that matter, CC BY, CC BY-SA, and most FOSS licenses) https://codeberg.org/spritely/goblins/pulls/998/files
-
Maybe ive been reading too much into some words; or maybe Ive misunderstood.
BUT some of your conversations and comments on LLMs to me and others have had a distinct "attribution is going to be fucked in a post-LLM world, the best we can hope for is to preserve copyleft" vibe to it.
I would be hard pressed to find an example in our own comment history; as my toots self destruct after a few months. BUT it's definitely come up.
Anyway; am I wrong in that sentiment? Do you think there is a way forward with some of these LLM concerns preserving attribution?
If a statement I made seemed like I said that, I either misspoke or it was misquoted. (Also, I wouldn't have 🤬'd making a point like this in any event.)
My position is instead that enforcement of attribution requirements **by themselves** (without other copyleft provisions to provide scaffolding) has always been extremely difficult to nearly impossible in real world contexts. #LLMGenAI may make the problem more acute, but the problem has always been there.
Cc: @cwebber
-
If a statement I made seemed like I said that, I either misspoke or it was misquoted. (Also, I wouldn't have 🤬'd making a point like this in any event.)
My position is instead that enforcement of attribution requirements **by themselves** (without other copyleft provisions to provide scaffolding) has always been extremely difficult to nearly impossible in real world contexts. #LLMGenAI may make the problem more acute, but the problem has always been there.
Cc: @cwebber
-
Btw I have heard at least one prominent FOSS person (I'm choosing to not fill their notifications here by identifying them) that nobody cares about attribution requirements (and particularly not enforcing them). But to me, attribution is what I care about MORE than any other licensing requirements, more than copyleft "share-alike" style provisions, etc.
If you feel similarly, it would be good to hear on this thread. Perhaps even copying a similar notice into your project is a good idea.
@cwebber while I use #llms a lot, it actually makes me seek out relevant resources and projects and learn about them, all of those I can find, and I make sure they are cross referenced and mentioned in the docs on the project. I find that to be a pretty good solution, not only in terms of exposure and surfacing smaller efforts, but because I spend time learning about the internals and not just import a lib and struggle with the api surface. Also find a lot of the original sources (papers, textbooks, etc…) which… are not often mentioned at all anyway (!).
-
-
We now note that that we do not waive our attribution requirements for Spritely Goblins, which are present in Apache v2 (and for that matter, CC BY, CC BY-SA, and most FOSS licenses) https://codeberg.org/spritely/goblins/pulls/998/files
So if you train on our codebase, you have to retain attribution, which we consider the most important requirement to motivate development of technology in the commons.
If, somehow, you are using a statistical soup which has no way to preserve attribution, well, we don't permit or approve of dropping such a requirement, and actually, the licenses we use never have.
@cwebber that's very interesting!
I'm not sure your licenses apply to training LLMs; it's not clear to me that copyright governs the training process.
I know you think differently; that a license attaches like a spell to the bits and bytes, no matter how small the excerpt or how varied the rearrangement.
-
@cwebber that's very interesting!
I'm not sure your licenses apply to training LLMs; it's not clear to me that copyright governs the training process.
I know you think differently; that a license attaches like a spell to the bits and bytes, no matter how small the excerpt or how varied the rearrangement.
The mechanism is not the issue. I think the expression of author intent is really important!
For Web resources, we have the AI preferences from IETF:
https://datatracker.ietf.org/doc/html/draft-ietf-aipref-attach-05
But, of course, Free Software doesn't just come from one website. It's made to be spread around. It would be good to have some sort of AI prefs expressed in the package itself, no matter where it is.
-
The mechanism is not the issue. I think the expression of author intent is really important!
For Web resources, we have the AI preferences from IETF:
https://datatracker.ietf.org/doc/html/draft-ietf-aipref-attach-05
But, of course, Free Software doesn't just come from one website. It's made to be spread around. It would be good to have some sort of AI prefs expressed in the package itself, no matter where it is.
@cwebber it is for these reasons we need something better than open source. An ethical open license.
-
Btw I have heard at least one prominent FOSS person (I'm choosing to not fill their notifications here by identifying them) that nobody cares about attribution requirements (and particularly not enforcing them). But to me, attribution is what I care about MORE than any other licensing requirements, more than copyleft "share-alike" style provisions, etc.
If you feel similarly, it would be good to hear on this thread. Perhaps even copying a similar notice into your project is a good idea.
@cwebber privacy is more important than attribution. https://penfount.social/@PenThoughts/117444874669122085
Also, copyleft and fair use and humans sharing with other humans in a non commercial collaborative way with and without attribution should always be protected. Not for profit, not deleting attribution and taking credit for others work, but libraries and mix tapes and punk style bricolage, and sharing ideas. -
@bkuhn @trashheap I appreciate the clarification on your position here.
I still do hold the positions that attribution are the glue that hold the commons together, and it's important that we fight to preserve it, and not give the impression that it is being voluntarily waived in the face of LLMs.
-
I like the idea behind CC Signals but it needs to ship a prototype soon.
-
@bkuhn @trashheap I appreciate the clarification on your position here.
I still do hold the positions that attribution are the glue that hold the commons together, and it's important that we fight to preserve it, and not give the impression that it is being voluntarily waived in the face of LLMs.
@cwebber @bkuhn @trashheap@tech.lgbt If bkuhn's right that permissive licenses make attribution hard to enforce, then the best way to accomplish that would be to use copyleft.
Anyway, if you keep Apache v2, it would certainly make sense to include a NOTICE file, since currently an Object derived from goblins does not need to attribute you at all. And as Apache v2 defines "Object", that seems to include AI models.
-
@cwebber @bkuhn @trashheap@tech.lgbt If bkuhn's right that permissive licenses make attribution hard to enforce, then the best way to accomplish that would be to use copyleft.
Anyway, if you keep Apache v2, it would certainly make sense to include a NOTICE file, since currently an Object derived from goblins does not need to attribute you at all. And as Apache v2 defines "Object", that seems to include AI models.
-
-
-
@bkuhn @trashheap I appreciate the clarification on your position here.
I still do hold the positions that attribution are the glue that hold the commons together, and it's important that we fight to preserve it, and not give the impression that it is being voluntarily waived in the face of LLMs.
@cwebber @bkuhn @trashheap IMO the real challenge to FOSS enforcement against LLM producers is the same one everyone else has: if the training is fair use because the resulting model is transformative, the producer doesn’t need a license (and doesn’t have to respect the terms of any existing license). It’s still early early days in terms of court decisions, but this looks like the direction we’re headed in the US (though the outcome will likely vary depending on how much memorization a model can be shown to do).
-
@cwebber @bkuhn @trashheap IMO the real challenge to FOSS enforcement against LLM producers is the same one everyone else has: if the training is fair use because the resulting model is transformative, the producer doesn’t need a license (and doesn’t have to respect the terms of any existing license). It’s still early early days in terms of court decisions, but this looks like the direction we’re headed in the US (though the outcome will likely vary depending on how much memorization a model can be shown to do).
@copiesofcopies @bkuhn @trashheap My main concern is that the resulting legislative environment is as equitable for FOSS as it is for the proprietary world, whatever the outcome is. In many ways this is why I'm happy to see the vibe-laundered versions of Adobe's suite etc, I think that will push the issue.
-
@copiesofcopies @bkuhn @trashheap My main concern is that the resulting legislative environment is as equitable for FOSS as it is for the proprietary world, whatever the outcome is. In many ways this is why I'm happy to see the vibe-laundered versions of Adobe's suite etc, I think that will push the issue.
@cwebber @bkuhn @trashheap unfortunately, we’re falling behind quickly, just in terms of the regulatory impact on who can afford to build competitive models. The CLAIM Act (just introduced) is basically designed to compel anyone who trains a model to either: (1) make it impossible for that model to accidentally cause harm, or (2) buy expensive insurance to cover the damage from inevitable harm.
-
@cwebber @bkuhn If I compile goblins and give the object file to someone, without attributing you -- as Apache v2 seems to allow sans NOTICE* -- and that person decompiles it, we've generated source code that is derived from goblins but does not attribute you. Seemingly legally.
Wouldn't the same apply to generating an LLM and using it no matter what you might ask in your README?
* ignoring moral rights
-
@cwebber @bkuhn @trashheap unfortunately, we’re falling behind quickly, just in terms of the regulatory impact on who can afford to build competitive models. The CLAIM Act (just introduced) is basically designed to compel anyone who trains a model to either: (1) make it impossible for that model to accidentally cause harm, or (2) buy expensive insurance to cover the damage from inevitable harm.
@cwebber @bkuhn @trashheap on the reverse-engineering side, we’ll see. Some interface and look & feel stuff is likely to be protected by copyright, and “clean room” efforts are probably misguided, because there’s no proving someone else’s model wasn’t trained on the target application.
-
