Welcome to the main community in programming.dev! Feel free to post anything relating to programming here!
Cross posting is strongly encouraged in the instance. If you feel your post or another person's post makes sense in another community cross post into it.
Hope you enjoy the instance!
Rules
Follow the wormhole through a path of communities !webdev@programming.dev
This is a great question, but I feel that the better question is: If a model was trained on open source code, should all code created by that model then be subject to the licensing from the code it was trained on?
I'm really hoping we can fix the current issues with the Supreme Court before this question comes before it because it's going to be hilarious when every company that has used AI assisted coding is forced to publish their source code.
when every company that has used AI assisted coding is forced to publish their source code
I guarantee you that will never happen. In the US? Do you live here? You think tech billionaires would ever suffer that kind of consequence?
This is a great question, but I feel that the better question is: If a model was trained on open source code, should all code created by that model then be subject to the licensing from the code it was trained on?
I don't think it works like that, because if you use GPL code, then all derivative works of the code must be Open Source too. But you no longer know the original project and maintainers anymore and the original source code is "lost". So just using GPL code to train wouldn't solve the issue I think. The source code and license MUST give permission to do so, I feel like. And then all original source codes contributors that has been used to train have to be included in every project that makes use of LLMs output. I have no idea in detail, but its more complicated then just declaring the output to be Open Source automatically.
I would have looked into generating AI art if it required the same license as the training art.
Training data is frequently copied and used regardless of license. It appears the outputs of machine generated art not "human creativity", thus not eligible for copyright.
From what I understand from the article, generating code does not grant copyright, although of it is proven that most of the work is human made then you can have a copyright (to what percentage has to be manual work to have a copyright? Afaik, nothing is in the article). So with this in mind, I see that you cannot apply a license what essentially does not have a copyright (if most effort is not manual) but maybe someone can prove me wrong.
Never knew photographs were once argued to be uncopyrightable due to being a product of a machine. Neat history.
Would it be possible (legally) to make a new version of GPL stating that code generated by AI trained on the GPL copy lefted code would have to respect the GPL as well?
There are some unanswered legal questions around this. In the US there is fair use, so US companies argue that training falls under that. If they’re right it’s difficult to prohibit because fair use literally means they don’t need a license in the first place. If they’re wrong and they’re found to be making derivative works, then there is likely no need as the terms of the GPL would then apply to their models too and they would not survive that. Other countries have different situations.
Any license prohibiting use for AI training would not be an OSI approvable open source license as OSI prohibits discrimination by field of endeavor.
Prohibiting AI training would not work, in my opinion… but trying to prevent the privatization of AI production could be interesting, isn't it?
Considering as of now generated stuff without much change does not have a copyright it is also not possible to apply a license. I am not a lawyer but I am quite sure about this one.
Training requires scraping the data and using it. But there is a license to each code attached, and copyright. If the LLM was trained only on code that gives the permission to do so, then I think the generated code could be licensed under the terms of the code it was licensed too. Because its not the LLM that is responsible, but the user of the program. But that is far from reality and not how LLMs are trained and used on.
It's a good analysis, but not really specific to code, the exact same ideas apply to AI-generated text, images, music, etc.
Could you just, use the original code that was fed to the LLM, as a dependency? Or is <package manager name> add dependency libdoohickey now too lame, due to hey clanker, generate me code that does things, make no mistakes, you're a great programmer, make sure it has no serious security issues, make sure it is well optimized rolled out and is only "marginally" worse?
We don't need to find ways to license something that shouldn't exist.
But it does exist.
Exactly, and that's where the problem lies and where something needs to be done.
Like licensing?
Right but why would we want to license vibecoded slop?
Irrelevant question. For licensing and copyright questions, quality of code is not important.
See Microsoft and Google.
Why would we want to license ~~vibecoded slop~~ code built with a system that is (in most cases) probably and admittedly actively using data from sources with incompatible licences or no licences at all.
probably and admittedly actively using data from sources with incompatible licences or no licences at all
You answered your own question. I don't know what you try to say, but the article and our discussion is explaining if and why licensing code that is vibe coded is possible or makes sense at all. And you point the problem out that this is problematic and probably not possible in good faith. So its not about what we want, but about if and what is possible.
Indeed, though I was just rephrasing the question from the person you were replying to.
Responded to the wrong comment, apologies.
No, like banning LLMs.
You do live in reality, do you?
Ah yes, the reality that consists in giving up on all proper solutions out of being "down to earth", effectively solving nothing.
The proper not possible solution... Sure...
People like you are the reason I gave up on the things getting better. What you people propose has nothing to do with how the world works we live in. It's all based on how the world should work.
Good luck with banning llms, seems realistic.
Yeah, I forgot how easy it is to hide datacenters and massive data collection. Really, really hard to enforce, huh.
You can, as of this very moment, download a 16 GB LLM and run it on your local computer. Or any bigger version, if you happen to have a beefy computer. LLMs are software. Good luck banning software. This is the internet.
ALSO, even if you do happen to be able to ban AI training facilities and such in the U.S., and those models become outdated, you can even go download some Chinese models too! They are not going away.
What the fuck are you talking about? 🙄
I don’t think the existence of something bad and needing to deal with that are mutually exclusive. Of course one should tackle the root of a problem. But you still need to bail water while the boat has a hole in it.
But licensing is trying to remove the water without doing anything about the hole. Or, best case situation, saying "we will fix the hole when the water is unmanageable, but for now we'll just try to empty it faster than it fills".
If you ban LLMs, the problem will disappear. It's not like you can really hide a datacenter so easily, or the size of the data required. Anything less than that will do nothing
