Linus Torvalds posts first LLM-assisted Linux kernel patch series, ~3x read throughput
Linus Torvalds has pushed an experimental branch named noref-filemap-read-with-llm to his Linux kernel repository, his first patch series developed with a large language model, targeting contention in the page cache when several threads make small reads from the same file. In the submission he states that the core code, test programs, review and the first draft of the text were done with Anthropic's Claude Code under his direction; per the benchmarks he posted, on a 4-core, 8-thread Core i7-8565U laptop, 8-thread 16-byte pread() throughput rose from roughly 9.9–10.4 million to 30–33 million reads per second and kernel cycles per read fell from about 2,950 to 725, while single-thread reads improved only slightly, from about 452–454 to 408–409 cycles. The approach defers memory reclamation until an RCU grace period elapses so the fast path avoids updating a shared refcount, but Torvalds calls reclamation the least-solved part of the series, reads targeting an unmapped user page add a failed exception on the new path, and the branch remains outside mainline, with LWN reporting reviewers have already flagged a race condition.
The kernel's most prominent maintainer, long a skeptic, publicly shipping LLM-assisted code signals AI coding assistants reaching conservative systems software, though the measured gains are limited to a narrow, read-heavy workload.