Linus Torvalds Credits Claude Code on Experimental Linux Page-Cache Patches
Torvalds says Claude Code assisted the code, tests and review for an eight-patch branch; its performance numbers come from his laptop, and the work is not merged.
An AI-assisted kernel experiment
Linus Torvalds’s experimental noref-filemap-read-with-llm branch contains an eight-patch attempt to make Linux reads from the page cache less contentious when many threads access the same file. In the main commit, dated October 10, Torvalds explicitly says an LLM coding assistant, Claude Code, worked under his direction on the code, test programs, review and first draft of the commit explanation. He says he ran the measurements on his own machine. The branch drew English reporting and Hacker News attention on October 12; that attention does not change the original commit date.
The claim and its limits
The series aims to avoid repeated page-reference updates during reads by deferring reuse of removed cached pages until an RCU grace period ends. Torvalds reports that eight threads making 16-byte reads of one page rose from about 10 million to 30-33 million reads per second in his laptop test. He measured with security mitigations and hardened user-copy checks disabled and says larger machines have not been measured. His commit also identifies outstanding memory-management and testing risks. This is a proposal in a separate branch, with no stated merge target, not a performance improvement already shipped to Linux users.