Torvalds Posts a Kernel Patch Series Written With Claude Code, and Sources Disagree on Its Size and Speedup

Linus Torvalds has posted a Linux kernel patch series for review that he built with an A.I. coding assistant, according to LWN.net, which published its account on Tuesday, Oct. 14. The series targets contention when many threads read the same file, and review has already found a race condition.

LWN's Jonathan Corbet wrote that Torvalds mentioned at the Maintainers Summit that he had used a large language model on a long-standing problem. LWN does not name the tool. It counts ten patches.

What the series changes

Per LWN's quoted description, the series detects hot folios, the kernel's unit of cached file memory, and marks them so their release is delayed until after an RCU grace period. RCU, read-copy-update, lets readers proceed without locks while writers wait for readers to finish.

Folio data accesses are then done under RCU as well. The fast path avoids taking a reference count when the copy to user space does not fault. In the description's words: "No refcount taking needed for the fastpath."

LWN reports an "order-of-magnitude speedup" on a read()-heavy benchmark. It also says review has identified a race condition, so more work is needed.

Eight patches or ten, and three times or ten

It's FOSS describes a different shape of the same work. It reports eight patches on a branch named noref-filemap-read-with-llm in Torvalds' kernel repository, built on Linux 7.3-rc6 plus fixes, with no release target in the commit message.

It also names the tool. According to It's FOSS, Claude Code, working under Torvalds' direction, drafted the commit message, wrote the code and test programs, and did the review behind the list of known problems.

The benchmark figures differ too. It's FOSS reports that on Torvalds' laptop, with some security protections disabled, eight threads reading one file rose from about 10 million reads per second to 30 million to 33 million. That is roughly a threefold gain, not ten.

Detail	LWN, Oct. 14	It's FOSS
Patch count	10	8
Tool named	No, "a large language model"	Claude Code
Speedup	"Order-of-magnitude"	About 3x
Known problem	Race condition found in review	List of known problems in commit message

The gap may reflect a branch and a posted series that changed between versions, or different benchmarks. Neither source says. The Terminal did not retrieve the mailing-list posting itself, so it cannot settle which count is current.

1. Before, approx.10 million/s
2. After, upper figure33 million/s

Source: It's FOSS, accessed 2026-10-14; Torvalds' laptop, some security protections disabled

The problem and the remarks around it

It's FOSS says Matthew Wilcox flagged the shared page-cache counter in 2024, citing a customer workload. The fix holds evicted memory back from reuse briefly, at the cost of slightly more idle memory.

It also reports that Torvalds said in Prague that "I actually really like using AI" and "I use AI to do the things that I'm bad at." It attributes the account to ZDNet's Steven Vaughan-Nichols. In August, according to the same piece, he used A.I. to help fix a bug in the Intel Xe graphics driver.

LWN's reader comments split along familiar lines. One commenter argued that experts benefit while newcomers may copy the example without learning the kernel's rules. Another compared the tool to an excavator. LWN did not report an A.I.-disclosure tag or patch-trailer convention for such work.

The debate over how far to trust such tools is live: Rich Hickey argued A.I. is not a tool and drew a same-day reply, while a plugin claiming 98 percent less agent context shows how fast unverified claims spread. Neither the series nor the review has reached a merge. LWN says more work is needed after the race condition, and It's FOSS says the commit message gives no release target.