Resume
-
LLM is great at finding Linux kernel bottlenecks, but not reliable at making clean fixes.
-
The revisions made were “ugly” and required extensive auditing, rewriting and editing of messages.
-
Build times dramatically improved: allmodconfig +36%, optional ~70%, noop ~90% faster.
It’s been a very interesting time in the Linux world, and whether that’s a good thing or not depends on your stance on using LLM. We recently saw Linus Torvalds fix a bug with AI, and Debian maintainers allowed LLM code submissions as long as the submitter took full responsibility. Now, the kernel developer reported using LLM to find the nasty bug, but the AI was a much better spotter than the fixer.
LLM helped me find the Linux kernel bug, but didn’t do much to fix it
His code required a lot of refactoring before it was ready
As noticed PhoronixLinux developer Lorenzo Stoakes of Arm sent an email details its operation in the kernel. A few single-stream obstacles made building a core more difficult than it should have been, so Stokes set out to fix them.
To help them in their quest, Stokes enlisted the help of an LLM, finding it better than solving the problem:
“The LLM was used to first identify where the barriers were and then how to improve them.
He created a lot of code, most of it terrible.
I extensively proofread and rewrote much of it, and heavily edited the commitment notices, cover letter, and comments.”
The good news is that despite the AI’s “hidden” code, the obstacles have now been fixed. Stoakes reports that “allmodconfig builds up to 36% faster, incremental builds up to ~70% faster, and noop builds up to ~90% faster,” a significant speed improvement regardless of which device is used. And while the AI did the build and maintenance, Stoaks did all the final checks themselves.
Stoaks’ findings are in line with what Linus Torvalds himself has observed about recent trends in nuclear development. LLM is getting much better at detecting bugs, and Linux kernel maintainers are full of reports of small problems that no one ever noticed. However, it seems that it will be some time before LLM itself becomes a proper contributor to the code.
