IT之家 · 智能时代 · 10/15/2026, 08:34:52
Torvalds Debuts First LLM-Assisted Kernel Patch: Boosts Concurrent Read Throughput 3x and Fixes Decade-Old Bugs
Linus Torvalds has submitted his first series of kernel patches developed with the assistance of an LLM (specifically Claude Code), targeting scalability bottlenecks in page reference counting. The experimental branch boosts concurrent file read throughput by approximately 3x, rising from ~10M to over 30M reads per second, while drastically reducing CPU cycles per operation. Additionally, Torvalds leveraged AI tools to identify and fix three subtle bugs that had persisted in the Linux codebase for 13 years, signaling a significant shift toward deep human-AI collaboration in core infrastructure development.
SOURCE COVERAGEFull Original Report
October 15 — Tech media outlet LWN published a blog post yesterday (October 14) reporting that Linus Torvalds, the creator of Linux, pushed a new development branch named noref-filemap-read-with-llm to his personal Git repository, officially submitting his first kernel patch series developed with assistance from a Large Language Model (LLM).
The report notes that typically, in the torvalds/linux.git repository, Torvalds maintains only the master branch and rarely develops on separate branches. However, the recent appearance of the noref-filemap-read-with-llm branch—with "llm" in its suffix—indicates kernel improvements he has been working on with the help of AI/LLMs (specifically Claude Code).
On October 13, Torvalds submitted new patches under this branch titled "Improve page reference count scalability," primarily addressing a long-standing bottleneck in Linux memory management. When multiple threads simultaneously perform high-frequency, small-scale reads of the same file, traditional mechanisms frequently trigger Page Cache reference count updates, easily causing severe race condition contention across multi-core CPUs.
Torvalds admitted that the core code, test programs, vulnerability reviews, and initial drafts of the announcement documentation for this patch set were all completed collaboratively by Anthropic’s AI coding tool, Claude Code, under his direct guidance.
After disabling certain security protections on his computer for testing purposes, Torvalds observed that under a specific test load involving 8 threads reading the same file concurrently, read throughput jumped from the original 9.9 million to 10.4 million operations per second directly to 30 million to 33 million operations per second (equivalent to rising from ~10M to over 30M), representing an approximate 3x performance improvement. Simultaneously, the number of CPU cycles consumed per single read dropped sharply from about 2,950 to 725.
IT Home Note: This patch is currently still in an experimental branch and has not yet been merged into the mainline Linux 7.4 release.
Furthermore, during the recent Maintainers Summit and another cleanup effort in the assoc_array codebase, Torvalds also utilized LLMs to refactor obscure legacy code, even uncovering three hidden bugs that had persisted in the Linux system for 13 years (dating back to the first commit in 2013).