The solution to this whole situation seems pretty simple to me. LLMs were trained on a giant mix of code, and it's impossible to disentangle it, but a not insignificant portion of their capabilities comes from GPL licenced code. Therefore, any codebase that uses LLM code is now GPL. You have a proprietary product? Not anymore.
Not saying there's a legal precedent for that right now, but it's the only thing that makes any sense to me. Either that or retain the models on only MIT/similarly licenced code or code you have explicit permission to train on.
In this emerging reality, the whole spectrum of open-source licenses effectively collapses toward just two practical choices: release under something permissive like MIT (no real restrictions), or keep your software fully proprietary and closed.
I suspect there’s a middle ground that involves either keeping tests more proprietary or a copyright license that bars using the work for AI reimplementation, or both.
I think it’s entirely reasonable to release a test suite under a license that bars using it for AI reimplementation purposes. If someone wants to reimplement your work with a more permissive license, they can certainly do so, but maybe they should put the legwork in to write their own test suite.
hopefully this continues to show how awkward the idea of "intellectual property" (IP) is until people abandon it
IP sounds good in theory but enables things like "patent trolling" by large corps and creating all kinds of goofy barriers and arbitrary questions like we're asking about if re-implementations of ideas are "really ours"
(maybe they were never anyone's in the first place, outside of legally created mentalities)
ideas seem to fundamentally not operate like physical things so asserting they can be considered "property" opens the door for all kinds of absurdities like as pondered in the OP
> I personally think all of this is exciting. I’m a strong supporter of putting things in the open with as little license enforcement as possible. I think society is better off when we share, and I consider the GPL to run against that spirit by restricting what can be done with it.
I like sharing too but could permissive only licenses not backfire? GPL emerged in an era where proprietary software ruled and companies weren't incentivized to open source. GPL helped ensure software stayed open which helped it become competitive against the monopoly proprietary giants resting on their laurels. The restriction helped innovation, not the supposedly free market.
i find his arguments on re-licensing blatantly AI-plagiarised libraries down to API compatibility confusing
they are arguments against any licence not just LGPL, I could literally plagiarise all his work, claim it's mine "clean-room" and not give him as much as a mention, by his own logic
and in his own words, he's "not interested" about the morality of it
I dunno, I’m inclined to think the WTFPL and MIT did more to help open source. And for a while during my youth there was indeed no distinction between publically accessible code and free and unencumbered code.
After cloning a test suite you're still left with ongoing maintenance and development, maintaining feature parity etc. There's a lot more than passing a test suite. If the rewrite is truly superior it deserves to become the new Ship of Theseus. But e.g. I doubt anyone's AI rewrites of SQLite will ever put a dent in its marketshare.
The legal question is a distraction. GPL was always enforced by economics: reimplementation had to cost more than compliance. At $1,100 for 94% API coverage, it doesn't. Copyleft was built for a world where clean-room rewrites were painful but they aren't anymore.
Strange this with this whole incident apart from the rewrite/LLM part is the general misundrstanding of the licences. LGPL being a pretty permissive one going as far as allowing one to incorporate it in propriety code without the linking reciprocity clause [1] and MIT is even more permissive.
Importantly these were meant to protect the USER of the code.Not the Dev , or the Company or the CLA holder - the USER is primary in the FreeSoftware world.Or at least was supposed to be , OSS muddied the waters and forgetting the old lessons learned when thing were basically bigcorp vs indie hacker trying to getthir electronic device to connect to what they want to connect to and do what they need is why were here.
Bikeshedding to eventually come full circle to understand why those decisions were made.
In a world where the large OEMs and bigcorps are increasinly locking down firmware , bootloaders , kernels and the internet. I would think a reappraisal of more enforcement that benefits the USER is paramount.
Instead we have devs looking to tear down the few user protections FLOSS provides and usher in a locked down hacker unfiendly future.
Perhaps code licensing is going to become more similar to music.
e.g. Somebody wrote a library, and then you had an LLM implement it in a new language.
You didn't come up with the idea for whatever the library does, and you didn't "perform" the new implementation. You're neither writer nor performer, just the person who requested a new performance. You're basically a club owner who hired a band to cover some tunes. There's a lot involved in running a club, just like there's a fair bit involved in operating a LLM, but none of that gives you rights over the "composition". If you want to make money off of that performance, you need to pay the writer and/or satisfy whatever terms and conditions they've made the library available under.
IANAL, so I don't even know what species of worms are inside this can I've opened up. It seems sensible, to me, that running somebody else's work through a LLM shouldn't give you something that you can then claim complete control over.
---------
Edit: For the sake of this argument, let's pretend we're somewhere with sensible music copyright laws, and not the weird piano-roll derived lunacy that currently exists in the U.S..
US courts have ruled that machine generated code cannot be copyright. Ergo, it cannot be licensed (under any license; nobody owns the copyright, thus nobody can "license" it to anyone else).
You cannot (*) use LLMs to generate code that you then license, whether that license is GPL, MIT or some proprietary mumbo-jumbo.
> But this all causes some interesting new developments we are not necessarily ready for. Vercel, for instance, happily re-implemented bash with Clankers but got visibly upset when someone re-implemented Next.js in the same way.
Note the Ship of Theseus, while a nice comparison for the title, is not - as the author eventually points out - an appropriate analogy here. A fundamental contribution to the idea of whether the identity of the entity persists or not is the continuity between intermediate states.
In the example given and discussed here the last couple of days there seems to be a process more akin to having an AI create a cast of the pre-existing work and fill it for the new one.
> Unlike the Ship of Theseus, though, this seems more clear-cut: if you throw away all code and start from scratch, even if the end result behaves the same, it’s a new ship.
That's not how copyright works. It doesn't require exact copies. You also can't just rephrase an existing book from scratch when the ideas expressed are essentially the same. Same with music.
At what point does the cost of reimplementation shrink below the benefits of obfuscation? Consider a new CVE in Linux. Well maybe my Linux is not the same as the public one. Maybe I just set a swarm of AI agents on making me a drop in replacement that is different but with an identical interface. Same-same but different. Right now writing your own OS to replace the entirety of Linux would be costly and error prone. Foolish. But will it always? What happens when Claude Code Infinute Opus can 1-shot a perfect reimagining in 24 hours? Or 30 minutes? Do all my servers have the same copy or are they all slightly different implementations of the same thing? I dunno.
> A court still might rule that all AI-generated code is in the public domain, because there was not enough human input in it. That’s quite possible, though probably not very likely.
Its not only likely, it is in fact the current position, at least in the US.
> For me personally, what is more interesting is that we might not even be able to copyright these creations at all. A court still might rule that all AI-generated code is in the public domain, because there was not enough human input in it. That’s quite possible, though probably not very likely.
As I understand it, the US Supreme Court has just this week ruled exactly this. LLM output cannot be copyrighted, so the only part of any piece of software that can be copyrighted is that part that was created by a human.
If you vibe-code the entire thing, it's not copyrightable. And if it can't be copyrighted that means it is in the public domain from the instant it was created and can't be licensed.
53 comments
[ 3.4 ms ] story [ 53.8 ms ] threadNot saying there's a legal precedent for that right now, but it's the only thing that makes any sense to me. Either that or retain the models on only MIT/similarly licenced code or code you have explicit permission to train on.
These are fascinating, if somewhat scary, times.
I think it’s entirely reasonable to release a test suite under a license that bars using it for AI reimplementation purposes. If someone wants to reimplement your work with a more permissive license, they can certainly do so, but maybe they should put the legwork in to write their own test suite.
https://reorchestrate.com/posts/your-binary-is-no-longer-saf... https://reorchestrate.com/posts/your-binary-is-no-longer-saf...
Even SaaSS isn't safe from that type of process:
https://news.ycombinator.com/item?id=47259485
https://github.com/albfernandez/juniversalchardet
IP sounds good in theory but enables things like "patent trolling" by large corps and creating all kinds of goofy barriers and arbitrary questions like we're asking about if re-implementations of ideas are "really ours"
(maybe they were never anyone's in the first place, outside of legally created mentalities)
ideas seem to fundamentally not operate like physical things so asserting they can be considered "property" opens the door for all kinds of absurdities like as pondered in the OP
I don't think Stallman has a real proposal to how innovation can be incentivized and compensated.
Take the example of medical innovations, sure big pharma is bad, but if they don't get to monetize their inventions, how will R&D get funded?
If you destroy IP and allow everyone to clone whatever, you will have a great result in the short term, then no one will continue R&D
I like sharing too but could permissive only licenses not backfire? GPL emerged in an era where proprietary software ruled and companies weren't incentivized to open source. GPL helped ensure software stayed open which helped it become competitive against the monopoly proprietary giants resting on their laurels. The restriction helped innovation, not the supposedly free market.
He is totally in on AI and that quote of his is self-serving. Can't we go back to flaming Unicode in Python?
they are arguments against any licence not just LGPL, I could literally plagiarise all his work, claim it's mine "clean-room" and not give him as much as a mention, by his own logic
and in his own words, he's "not interested" about the morality of it
odd
Bikeshedding to eventually come full circle to understand why those decisions were made.
In a world where the large OEMs and bigcorps are increasinly locking down firmware , bootloaders , kernels and the internet. I would think a reappraisal of more enforcement that benefits the USER is paramount.
Instead we have devs looking to tear down the few user protections FLOSS provides and usher in a locked down hacker unfiendly future.
[1] https://licensecheck.io/blog/lgpl-dynamic-linking
e.g. Somebody wrote a library, and then you had an LLM implement it in a new language.
You didn't come up with the idea for whatever the library does, and you didn't "perform" the new implementation. You're neither writer nor performer, just the person who requested a new performance. You're basically a club owner who hired a band to cover some tunes. There's a lot involved in running a club, just like there's a fair bit involved in operating a LLM, but none of that gives you rights over the "composition". If you want to make money off of that performance, you need to pay the writer and/or satisfy whatever terms and conditions they've made the library available under.
IANAL, so I don't even know what species of worms are inside this can I've opened up. It seems sensible, to me, that running somebody else's work through a LLM shouldn't give you something that you can then claim complete control over.
---------
Edit: For the sake of this argument, let's pretend we're somewhere with sensible music copyright laws, and not the weird piano-roll derived lunacy that currently exists in the U.S..
You cannot (*) use LLMs to generate code that you then license, whether that license is GPL, MIT or some proprietary mumbo-jumbo.
(*) unless you just lie about this part.
Kinda surprised nobody commented on this
In the example given and discussed here the last couple of days there seems to be a process more akin to having an AI create a cast of the pre-existing work and fill it for the new one.
That's not how copyright works. It doesn't require exact copies. You also can't just rephrase an existing book from scratch when the ideas expressed are essentially the same. Same with music.
Its not only likely, it is in fact the current position, at least in the US.
As I understand it, the US Supreme Court has just this week ruled exactly this. LLM output cannot be copyrighted, so the only part of any piece of software that can be copyrighted is that part that was created by a human.
If you vibe-code the entire thing, it's not copyrightable. And if it can't be copyrighted that means it is in the public domain from the instant it was created and can't be licensed.