Contributions in the age AI

[erimatnor]

We welcome contributions from the community, and with the advent of AI, it has become easier than ever for people to contribute. However, while AI makes contributing easier, AI-generated code also presents significant challenges.

There's a real risk of being flooded with AI slop or simply becoming overwhelmed by the sheer volume of contributions. Therefore, we're likely going to work on a contribution process for the project. This is some of my current thinking around this:

1. The general guideline: In most cases, you need to do more than simply generate code using an LLM. It might be better to start with an issue and provide real value, like benchmarks, reproduction cases for bugs, etc.
2. Simple bug fixes involving just a few lines of code will probably be readily accepted, provided they clearly demonstrate the bug and include proper tests.
3. Performance patches need to demonstrate actual performance improvements, rather than simply submitting hundreds of lines of AI-generated code claiming a certain improvement. Even if the claimed improvements are real, reviewers have no easy way to validate them without investing a significant amount of time. Therefore, such PRs must include benchmarks that are easy to run and verify. They must also demonstrate that there are no regressions in other areas.
4. Fixing existing issues: Simply pointing your agent at the issue tracker and having it generate code for each issue is probably a good way to get your PRs rejected and yourself blocked. The core development team could do the same themselves, but there's probably a reason they haven't. One reason may be that there's a roadmap and schedule to maintain. Another may be that we're still weighing the pros and cons of different approaches to a fix or improvement. Obviously, if an issue boils down to a simple bug fix, point (2) above applies.
5. Opening multiple PRs at the same time or in a quick succession without having coordinated with a core contributor that this is OK will most likely lead to dismissal of those PRs and, potentially, blocking the author.
6. Most AI models are notoriously bad at writing code comments. One way to make life easier for human reviewers is to read the comments yourself and try to improve them.

To conclude, this is an evolving topic, and I suspect most people are still trying to figure out how to deal with the new reality of AI-assisted coding. So are we, and we ask for your patience and understanding. Feel free to share your opinions and feedback on this in the thread.