The series changes how the kernel reads small chunks from the page cache by skipping folio reference counting, which causes contention when many threads read one file. On Torvalds's 4-core, 8-thread laptop with mitigations disabled, kernel cycles per read fall from about 2,950 to 725. Single-thread reads improve about 10%, and a contention counter drops from 30,538 to 101. In the patch message he said the code, test programs, review notes and first draft of the cover letter were produced with Claude Code under his direction, and that he ran the measurements himself. The caveats are large. The branch sits on top of 7.3-rc6, has been measured on a single machine, and is not merged. Torvalds also flags parts of the design he is unsure about. For builders, the post is a high-profile case of an agent doing performance work in a mature, review-heavy codebase, with the maintainer disclosing exactly where the model was used. It is a template for how credible AI-assisted contributions are documented: measurements, open problems and human review stated up front.