Hello Jan, > We do not condone the use of the current generation of > LLMs because they do not meet the definition of free > software - you can't study the weights, the training, the > alignment, nor make meaningful changes.
Yes, I agree. But my question was mainly about what we get in the output of that non-free software, not how the software was built itself. E.g., a free software written on Windows with proprietary editors. (You have answered that question in the later parts of your question.) > and have a chance of outputting snippets of copyrighted > code. Thanks for this very important correction. I was under the impression that they merely statistically generate and don’t produce verbatim copies. I read about this a little and I found some arguments that say something on the lines that this is not an LLM-specific issue. Even humans might do this unintentionally. Courts have addressed this before LLMs too, apparently. The core difficulty seems to be how one may predict the intentions of LLMs. > That is correct. I'm not aware of a legal precedent for > how much input from you is needed to an LLM for the courts > to then treat the LLM like a tool, e.g. a macro in Emacs. Very nice comparison. I hadn’t thought about it like this. > This is the problem - you don't know whether the output of > the LLM is a copy of someone else's copyrighted code. Hmm, most cases that I read before asking the question were about the copyrights being void, but not because of the point you raised here. The reason mainly was that the verdict of the court was along the lines that decided ‘if the creator is not a human, the creation cannot be protected under copyrights’. CC0 still would have been fine in such situations, because it puts you as someone who presents the work, probably owns it and probably has copyrights over it. It is recommended for works for which the status of the copyright holder(s) is vague or dubious. If a court decides that the presenter doesn’t have the copyrights, a fallback license is used that doesn’t impose any condition on the people who receive the copy of that file and if the presenter is declared to have copyrights, the same effect is achieved by explicit conditions of the main license. This would have worked fine for LLM-generated content if it had not copy pasted other people’s content, but because of your clarification, I think it is probably best not to publish it in any way. Better it stays with me only till the problems are fixed. > I'm a Savannah user, like you, and can't make the call on whether your > proposal is acceptable, but I hope the above helped. Definitely, thanks a lot! Hello Ineiev, > The FSF is working on the rules about accepting code > generated by LLMs; Great to know that! > until they are ready, GNU maintainers are asked to avoid > installing any LLM-generated code in GNU packages. Okay. Thanks. > It would be nice if you waited for the FSF as well. Will do that. > If your contributions went to GNU Savannah, you wouldn't > ask this question; therefore, savannah.nongnu.org. Sorry, I am not sure about what you mean here. Are you suggesting that I should upload the package to Savannah non-GNU and not at Savannah GNU? As of now, I have not uploaded it to either of these repositories for it ‘to go to GNU Savannah’ Let me know, Thanks. निरंजन
signature.asc
Description: OpenPGP digital signature
