jh7370 wrote: > > I wonder if this would be a good opportunity to get GitHub CoPilot involved > > with reviewing (as a supplement to, not instead of, human reviewing)? It > > seems like this PR is mostly mechanical and both the before and after > > syntax should be well-documented, I expect, so I'd expect an AI reviewer to > > do a pretty good job if asked to specifically comment on the migration. > > So, yes, I'm definitely using LLMs to do the heavy lifting of reviewing the > code, but our [AI policy > requires](https://llvm.org/docs/AIToolPolicy.html#details) that people do > that off-github. LLMs definitely do find issues, but it's a question of > whether we want to paste those LLM tokens into github issue comments. I > personally prefer to have all that iteration done offline, and then to post > the result, rather than having a bunch of public back-and-forth with > AI-powered linters that other reviewers then have to wade through.
Thanks for that. As I get more and more used to using LLMs in my non-LLVM work, I'm definitely getting used to using them to assist with reviews. I think as more people become familiar with them, it's likely we'll want to revisit this aspect of the AI policy (especially given the recent Discourse post about architecture to support automated reviews), but in the meantime, using them locally before publicly posting the PR is generally worthwhile, if people have the skills and resources to do so. > I think this raises a question of, what do I expect from reviewers? It's not > reasonable for them to read 10k+ lines of doc changes, clearly. I'm mostly > looking for a spot check, and approval, like "I sampled 8 locations, and 8/8 > looked good, I'm confident in the methodology, this is good enough to land." Makes sense. https://github.com/llvm/llvm-project/pull/201244 _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
