Linus Torvalds, the man who built the Linux kernel and has spent decades rejecting sloppy code in public, is now using an AI coding assistant to help write some of his own.
He isn’t handing over the keys, though. Torvalds says AI is fine in the hands of competent developers. He has far less patience for what he calls “drive-by” patches from people who don’t understand the code they submit.
What Torvalds actually built with Claude
On October 10, 2026, Torvalds posted an experimental series of eight patches to a Git branch named ‘noref-filemap-read-with-llm’. The work was done with help from Anthropic’s Claude Code, a large language model tool built for programming.
The patches target a long-running problem in the kernel’s memory management: page-cache contention. When many threads read the same cached file, they all update shared reference counters, and that bookkeeping slows everything down. Torvalds’ patches let those threads read the cached file without touching the shared counters.
The numbers from his own laptop test are hard to ignore. Throughput climbed from approximately 10 million reads per second to 30 to 33 million. Kernel cycles per read fell from approximately 2,950 to approximately 725.
Torvalds has said openly that the AI tool helped generate code, review notes and the first drafts of commit messages. He says he handled the performance measurements himself, on his laptop.
The patches are not in the kernel yet
The work is still experimental. As of the latest reports, it has not been merged into the mainline kernel. The patches could come up for consideration in the upcoming 7.4 merge window.
The page-cache work isn’t his only AI-assisted project. Torvalds also used an LLM to help refactor the obscure ‘assoc_array’ code. He reported finding more bugs during that effort, some dating back to 2013.
Kernel maintainers drop the AI label
At the Maintainers Summit on October 8, 2026, core members of the kernel community agreed to drop the requirement for a mandatory ‘Assisted-by: LLM’ tag in commit messages. Maintainers will judge code on its quality, not on how it was produced.
Around October 9, Torvalds spoke at the Open Source Summit Europe. He described AI as a productivity tool for developers, especially for tasks outside their own areas of strength.
Two very different AI problems
Torvalds has pointed to two separate AI issues facing the kernel. The first is positive: AI tools can be useful for finding bugs, as his own assoc_array work suggests. The second is the opposite. Low-quality AI-generated patches can pile up in maintainers’ inboxes, sent by people who ran a prompt, got some code back and submitted it without understanding it.
The distinction is about who is holding the tool. An experienced kernel developer using Claude to speed up work they already understand is one thing. A stranger submitting machine-written fixes they can’t explain or defend is another.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.