Linus Torvalds Linux AI Stance: From 90% Marketing to Claude Code

The Linus Torvalds Linux AI stance has flipped in 2026, and October made it official. The creator of Linux now builds kernel patches with Anthropic’s Claude Code. He also called rules against AI code “UTTERLY STUPID” as the kernel dropped its AI disclosure tag. Here is what he did, what he said and where he still draws the line.

On October 9, Torvalds sat down with Dirk Hohndel at Open Source Summit Europe in Prague. He admitted he used to think AI was “just not very good,” TNW reported.

However, that view is gone. “Now I’ve reached a point where I really enjoy using AI,” he said, as Computerworld quoted. He mostly uses it for tasks he is weak at.

In fact, he even praised vibe coding as “a wonderful way to find joy in programming” and a “gateway drug” for new programmers.

Two days later, the words turned into code. As Phoronix first reported, a branch called noref-filemap-read-with-llm appeared in his own kernel tree.

Notably, the patch text is open about the tool. The code, tests and first draft were “done with an LLM coding assistant (Claude Code) under my direction,” Torvalds wrote. He ran the measurements himself.

Claude Code is one of the agents we compare in our list of the best AI coding tools. Here, though, it went after a problem kernel developers had fought for two years.

Specifically, the target was the page cache. When many threads make small reads of the same file, each read bumps a shared reference count. Oracle engineer Matthew Wilcox flagged that bottleneck in 2024, The New Stack noted.

Torvalds’ first eight patches nearly tripled throughput on his laptop. Eight threads went from 10 million to as many as 33 million reads per second.

A revised 10-patch series did far more on his 64-thread Threadripper. There, throughput jumped from 13.3 million to 304 million reads per second. Still, Torvalds warned that real workloads would gain much less than this synthetic test.

However, review did its job. Kernel developer Kiryl Shutsemau found a race in filemap_read_rcu_batch(). In rare cases, it could expose data that had never been written, according to The New Stack.

Torvalds suggested a fix, and Shutsemau folded it into a revised patch. Wilcox also questioned the new PG_rcu_free flag, but Torvalds rejected his alternatives.

In the end, Torvalds handed the work to Shutsemau, who had studied the same bottleneck before. As of mid October, nothing had been merged into the mainline kernel.

Since January, kernel docs asked contributors to add an Assisted-by tag when they used AI tools. On October 8, the Maintainers Summit in Prague agreed to deprecate it, XDA reported.

Still, not everyone liked that. When a developer pushed back on the mailing list, Torvalds replied: “Do you not see how UTTERLY STUPID those kinds of rules would be?”

He then mocked the idea with a new rule: “don’t use compilers for any of your kernel submissions.” To him, what matters is whether a developer understands the code, not which tool typed it.

The Claude Code patches were not his first test. In August, he fixed an Intel GPU driver bug after what he called a “debug session from hell, enormously helped by an AI doing much of the grunt-work,” Phoronix reported.

Overall, that session took 24 debug patches and 18 kernel boots. In the end, the fix was a one-liner, swapping a round_up() for a round_down().

Then, in October, he used LLM tools to untangle hard-to-read code in the kernel’s assoc_array library. “I did what I preach,” he wrote. The tests confirmed the reported bug and found three more dating back to 2013.

In short, the shift took two years. The table below tracks each step, with the outlet that reported it.

Being pro AI does not mean anything goes. In May, he complained that AI-assisted bug hunting had padded a release with trivial late fixes, and he promised to be “a bit more hardnosed,” It’s FOSS reported.

In addition, maintainers feel the load too. About three quarters of Maintainers Summit talks focused on making AI tools less stressful, according to TNW.

Meanwhile, other projects went the other way. NetBSD has banned AI code since 2024, and System76 banned it from COSMIC. Game engines face the same fight, as our report on the Godot engine AI code ban shows.

If Torvalds’ praise for vibe coding made you curious, start with our picks of the best vibe coding tools. They are ranked for beginners and for developers who already ship code.

Sources

Phoronix (October 11, 2026), Phoronix (October 12, 2026), Phoronix (August 21, 2026), The New Stack (October 14, 2026), TNW (October 16, 2026), XDA (October 16, 2026), Computerworld (October 14, 2026), It’s FOSS (July 17, 2026). *Photos: LinuxCon Europe Linus Torvalds 05 by Krd, CC BY-SA 4.0, resized and placed on a blurred background; LinuxCon Europe Linus Torvalds 01 by Krd, CC BY-SA 4.0, cropped and resized.*