mchoo7 wrote: > > Assisted-by: GPT > > I don't know how LLVM handles this type of metadata, but in FreeBSD, one of > the general consensuses is that the model versions should be called out in > commits: the metadata helps document whether or not there are issues, e.g., > functional, security, etc, which could have been introduced by a given model. > Moreover, it helps document whether or not there could be issues with > licensing for LLM generated code--just in case a future lawsuit brings any of > the generated outputs into question as off-the-shelf/hosted LLMs trained via > a number of unknown inputs (Reddit, Repology, Stack Overflow) which have > different licenses with different advertising/disclosure requirements. The > only LLMs that one can reliably claim are "license poisoning free" require > the end-user to do their own training on specific materiel (many > closed-source models and open-source models currently fail this requirement).
In LLDB, I've been using `Assisted-by: Claude` and `Assisted-by: GPT` for last few months and no lldb reviewers pointed it out so far. I read the AI policy document, but they don't seem to have consensus what to document (e.g. cli tool used, full model name, effort (light/medium/high/extra high/etc), subscription vs API credit). https://github.com/llvm/llvm-project/pull/229176 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
