My first patch was to stable leading up to 7.0.

I have an old PC that I still use for everything, and with RAM prices being what they were (and still are), felt like I should put some effort into maintaining that old hardware. I started with the ethernet driver, since it seemed like it should be pretty straightforward, and I'm generally interested in telemetry.

I started by pointing Claude at the main source file and asking it to find bugs. It found maybe 6 hits, and one of them turned out to be a very real bug. I spent a few days reading code and documentation to convince myself that it was indeed a very real bug.

With that new understanding of the patterns, next I wanted to be sure I could hack it without the LLM. I did code review on the same driver looking for any resources allocated in the driver probe which could be leaked on probe failure, and found one.

The next step was automating away that tedious code review, so I learned enough about coccinelle to scan all of the PCI ethernet drivers for the same probe leak, and it found another in a different driver.

I feel like the assistance with that first step was the real value of the LLM. It's kind of like having an expert to ask questions about any part of the kernel anytime you want, with the usual caveats of interacting with LLM for any purpose.

I do use the Assisted-by tag when the LLM was involved in finding the bug/change, iterating on the implementation, or reviewing the work. I can understand why a lot of people don't. There's no context about what parts of the work were done by the LLM, unless you add some noise to the commit description. If I'm feeling proud of myself for finding a bug or a big optimization myself, adding that label to the commit kind of steals the thunder. That's one reason I set out to make a few commits without any LLM assistance. Now I'm over it, and I'm just gonna do what works and follow the guidelines.

Maybe the guidelines can evolve to work around the psychology. For example, if it were acceptable to credit the LLM with different tags, that might help. On some commits I would credit it with Suggested-by, on others Reviewed-by, for implementation Co-authored-by, or Assisted-by as a shorthand for any or all of these. The normalization of anthropomorphism is a little creepy, but there's a reason we don't just use Assisted-by when a real person helps with a commit. The reduction cuts both ways.