Linus Torvalds Pushes a Kernel Branch Co-Written With Claude Code
Linus Torvalds pushed a kernel branch named noref-filemap-read-with-llm that reworks how the page cache handles file reads by avoiding folio reference counts, reporting up to three times faster reads in his tests. The branch name's llm suffix is deliberate: Torvalds says the code, the test programs, and much of the review were drafted with Claude Code under his direction.
On this page
The branch, and the name
Linus Torvalds pushed a branch called noref-filemap-read-with-llm to his Linux kernel tree this week, and the suffix is not a joke: in the cover letter, Torvalds says most of the code and the test programs were written with Claude Code, with him doing the direction and the review. The work itself is a rework of memory management's filemap read path, replacing refcount-based page cache handling with a no-reference approach that leans on folios already mapped into userspace page tables.
The performance claim attached to it: up to three times faster reads in his tests, based on a page-table-lock design that avoids the reference-counting overhead of the existing path. The branch is separate from mainline and there is no merge commitment yet; it is an experimental proposal on the maintainer's own tree, which is exactly where the kernel's hardest reworks have traditionally been tried.
Torvalds has been saying this for a while
The AI-assisted branch is not a one-off. Torvalds has said repeatedly this year that he treats AI coding tools as useful for exactly this kind of work, and a follow-up comment quoted by Phoronix has him saying he did what he preached: using AI to make code decipherable, which he says uncovered more bugs in the surrounding code. The workflow he describes is the reviewed one: the model drafts, the human sets direction and reviews, and the tooling proves the result.
That matters because the kernel is the highest-scrutiny codebase in open source. A maintainer of Torvalds's stature attaching an llm suffix to a branch name, in public, with the attribution in the cover letter, normalizes disclosed AI assistance in infrastructure code the same way Co-authored-by trailers normalized it in application repositories.
What it means for local AI
This story sits oddly next to the frontier labs' own research claims, and usefully so. The same week OpenAI published hundreds of Lean-checked math results and withdrew three of them, the most influential maintainer in open source shipped a memory-management experiment drafted by a coding agent, with the performance claim published alongside reproducible tests and the branch available for anyone to benchmark.
The difference is not the tool; both workflows used frontier models. The difference is that the kernel's review culture runs on public diffs and measurable claims, so the AI contribution lands inside an instrument that catches errors before they become lore. For teams adopting coding agents, that is the pattern to copy from this story: disclose the assistance, attach the benchmark, and let the review do what it does.