https://bugs.kde.org/show_bug.cgi?id=520872
Mark Wielaard <[email protected]> changed: What |Removed |Added ---------------------------------------------------------------------------- Ever confirmed|0 |1 Status|REPORTED |CONFIRMED CC| |[email protected] --- Comment #4 from Mark Wielaard <[email protected]> --- Thanks for working on this. I agree that something like the binutils policy is a good, clear way to describe what we would like. Another is the one from elfutils: https://sourceware.org/cgit/elfutils/tree/CONTRIBUTING#n152 Some specific comments on the proposal: > This is because the copyright status of code generated by a LLM (Large > Language Model) is currently unclear. Yes, but as others pointed out, there are other more serious objections to LLM generated contributions. I do think we are hiding a little behind just saying the legal/copyright issue is the main issue. But I am fine with just stating that contributions without clear authorship rights are not accepted. If that gets us to a place where we simply don't accept anything generated by an LLM. It is a simple bright line. Maybe rephrase as: "This is because we only want to accept contributions that come with clear authorship provenance and rights (which is unclear for LLM generated text)." I think the exceptions are fine, but they aren't really LLM specific IMHO. We don't control (or want to control) what people do privately. It might be good to mention this also holds for the commit message itself, which is part of the contribution. It should not be a automatically generated summary of the contribution, but a description of the actual changes proposed. > When submitting a non-legally-significant LLM generated change, it is > encouraged to clearly indicate the use of the LLM. I don't think this is a good suggestion. If it is not legally significant then it doesn't really matter. We don't require this of any other tools. It might be appropriate if it describes a way/tool that the reviewer can use to (deterministically) reproduce part of a patch. But an LLM isn't that. The elfutils policy simply says: "Contributors are not required to disclose the use of LLMs for these [exception] purposes." off-list we were discussing whether we should say something about "imported code" like from gcc/libiberty, gdb/gdbserver, binutils/opcodes, and the compression code. It might be nice to say something that we like our imported code the have a similar policy. And that is must be GPL-compatible. It is unclear if code containing large LLM generated code can be GPL-compatible. But maybe that is just something to re-evaluate when importing new code. We could say that in general we expect such (re)imports to be done with deterministic scripts that can be reviewed and be used to show the import/patch can replicated. -- You are receiving this mail because: You are watching all bug changes.
