ZDNET’s key takeaways
- Torvalds simply launched a set of patches, which had been largely written with Anthropic’s Claude Code.
- These new patches handle an unsolved reminiscence cache slowdown drawback when utilizing a number of cores and processors.
- Linus Torvalds had been shifting nearer to favoring AI, and this was the ultimate step towards embracing it for Linux growth.
It needed to occur. Linus Torvalds introduced to the world that he used AI to jot down code for the Linux kernel. Who’d thought it? Effectively, me for one. Over the previous few years, I’ve seen Torvalds’ perspective change from skeptical to seeing AI as a useful gizmo for programmers and maintainers.
In his Git submission, he wrote, “The code, the check packages, the evaluation behind the record above and the primary draft of this textual content had been completed with an LLM coding assistant (Claude Code) below my course.”
Additionally: ‘I use AI to do the things that I’m bad at’: Linus Torvalds on why it works for him
Tackling a shared-memory bottleneck
Particularly, the repair addresses rivalry in Linux’s web page cache, which holds file knowledge in reminiscence. When a number of threads repeatedly learn small quantities of knowledge from the identical cached file, they take and launch references to the folio backing that knowledge. These reference-count updates require atomic writes to a shared cache line. This in flip forces processor cores to compete over it.
Doesn’t sound like that huge an issue, does it? Improper. In some workloads, it will probably actually sluggish issues down. Torvalds discovered that when eight threads issued 16-byte pread() calls towards the identical web page on a four-core, eight-thread Intel Core i7-8565U laptop computer, the shared cache line accounted for 99.92% of the measured cross-core load, in accordance with Hit in Modified cache line (HITMs). That slows work down significantly.
This isn’t simply Linus scratching a private itch. It’s been a identified subject since 2024. Earlier makes an attempt to repair it, although, have come to naught.
Torvalds and Claude Code’s proposed patches change when the kernel can reuse reminiscence faraway from the web page cache. Slightly than returning that reminiscence to the allocator instantly, the patches defer reuse till a read-copy-update (RCU) grace interval has elapsed. That lets a reader protected by an RCU read-side part copy knowledge immediately right into a consumer buffer with out taking a folio reference.
Additionally: Someone made a free Photoshop clone – so I had to try it on Linux
This method trades instant reminiscence reclamation for a learn path that avoids repeatedly updating a shared reference counter. The reported efficiency positive aspects are substantial inside that particular check. Eight-thread throughput rose from roughly 9.9–10.4 million reads per second to 30–33 million, whereas kernel cycles per learn fell from roughly 2,950 to 725. Single-thread reads improved extra modestly, from about 452–454 cycles to 408–409 cycles per learn. Torvalds added in his pull request (PR) that “A machine with extra CPUs, or multiple socket, ought to acquire extra; nothing greater has been measured.”
This doesn’t imply Linux as a complete runs 3 times sooner. Hardly! As a substitute, this describes experimental results from a selected small-read workload on Torvald’s private laptop computer, which he had with him in Prague at Linux Plumbers.
Quicker reads, unfinished reminiscence administration
That’s to not say Torvalds and Claude Code have fastened this subject. We’re removed from with the ability to say that. The efficiency positive aspects include engineering trade-offs. Delaying reminiscence reuse means pages can stay unavailable whereas the kernel waits for an RCU grace interval. In his Git submission, Torvalds described that reclamation subject as “the a part of this collection that’s least solved.”
Additionally: I’ve used Ubuntu for 22 years – and 26.10 might be its best release yet
There’s additionally a regression when a learn targets a consumer web page that’s not but mapped. The brand new path incurs a further failed exception earlier than falling again to the present implementation. As of Oct. thirteenth, the department remained outdoors the mainline Linux 7.3-rc6. Whether or not it can ever be merged, and on what timetable, stays up within the air.
From AI skepticism to hands-on use
Torvalds’s resolution to make use of Claude Code displays a long-term evolution in his perspective towards AI. This hasn’t been a easy reversal from opposition to help.
In December 2023, he described AI as a possible help to builders, saying, “I see AI as a software that may assist us be higher at what we do.” On the time, his curiosity in AI was round the way it may assist establish bugs, moderately than writing code. Nobody was giving critical thought to utilizing AI to jot down critical packages (laborious to consider now).
By late 2025, Torvalds was calling himself “an enormous believer in AI as a software, and saying he was extra not simply find bugs, however in sustaining code and reviewing patches. Producing new, critical code for Linux was nonetheless largely off the desk. Nonetheless, in January 2026, he started toying with AI-assisted coding for a Python audio visualizer in a pastime challenge.
Additionally: 75% of developers I surveyed prefer Claude Code – here’s why they choose it over Codex
His newest remarks, reported Oct. 9 following a dialog with Dirk Hohndel at Open Source Summit Europe in Prague, had been extra enthusiastic.
“I used to assume it was simply not superb — however now I’m on the level the place I truly actually like utilizing AI,” Torvalds stated. “I feel AI is an excellent software if you happen to use it accurately and deal with it as a software, and I discover it makes programming rather more pleasing.”
Even so, he cautioned, “I feel that you’ll want to be very cautious utilizing AI while you’re doing one thing actual and vital.” It doesn’t get a lot much less actual and vital than Linux.
The page-cache experiment places either side of his place on show. Claude Code helped produce kernel code, checks, and evaluation supplies. The ensuing department reveals promising benchmark positive aspects, but additionally the unresolved memory-reclamation questions and efficiency regression that should be addressed earlier than the experiment can grow to be manufacturing kernel code.
Nonetheless, we’ve reached a turning level. Torvalds, arguably the world’s foremost developer and maintainer, is now utilizing AI to develop new Linux kernel code. This may have been unthinkable even a yr in the past, and even now some folks nonetheless battle to think about AI including new code to Linux or every other vital program.
