prabhjeet16 commented on issue #18095: URL: https://github.com/apache/datafusion/issues/18095#issuecomment-5454495046
> > My two cents as a contributor: Let's remember that there's no guarantee that people will tell the truth on the internet. I don't think this will help much with the most problematic PRs where all the code is fully AI-generated and the PR feedback is handled by AI, too. I think the same people who author these PRs will also **lie** about their use GenAI use. > > I think people are not telling the truth because of misunderstandings rather than malice. We should better explain why large AI-generated PRs without understanding are not helpful. > > Here are my opinions. If we can reach agreements, they can be added to the contributor guide: > > **If the PR author only understands a small part of the implementation details, such PR should be rejected** The PR and code review system exists for two major reasons: > > * Get the job done, obviously. > * Share knowledge about the codebase between author and reviewers, which is a long-term investment in the project. For this reason, even if someone familiar with the codebase can finish a task quickly, it’s still worth helping a new contributor work on it even if it takes longer. > > This doesn’t work for fully AI-generated PRs without understanding, because reviewers are mostly prompting through a pass-through human proxy. They can solve the problem faster by using AI directly, and there is almost no knowledge sharing since the PR author doesn’t want to understand the solution. > > A better way to help is to write a better issue to make it easier for others to contribute, instead of generating a full solution. > > **If the author knows most of the solution but is missing some corners, such PRs are welcome** One case is implementing a large trait, where the author knows the high-level idea but isn’t sure if a certain interface’s default implementation is okay to use. This might be due to under-documentation of the API. They're encouraged to point out exactly which part of the PR they don’t fully understand, the review can focus on that and potentially improve the documentation. --------------------------------------------------------------------------------------------------------------------------------------------- The point about understanding the context really resonates. I’ve seen a similar challenge when working with AI-assisted changes: the difficult part isn’t always generating the code, but understanding how a change fits into the existing structure and what else it may affect. That’s actually one reason I’ve been using JigsawML to trace dependencies and understand the impact of changes across a codebase. It might be worth a quick look in the kind of cases you described. If you’re curious, the free version is there to try. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
