Linus Torvalds has written eight code changes for the Linux kernel using Anthropic's AI coding tool, Claude Code — a striking turn for a developer who spent years dismissing AI-generated code as simply not good enough.

The changes target a performance issue in Linux. They remain experimental and have not landed in the official kernel. But the documentation accompanying the work is unusually candid about how it was produced.

"The code, the test programs, the review that forms the basis of the list above, and the first draft of this text were produced with the help of an LLM coding assistant (Claude Code) under my supervision," Torvalds writes in the project's documentation.

Read that quote again. Not just the code. The test programs. The review. Even the first draft of the explanatory text itself. Torvalds is careful to pin down the division of labor: the assistant produced, he supervised. For the maintainer of the world's most widely deployed operating system kernel, that framing matters.

What changed Torvalds' mind?

The shift didn't happen overnight. Torvalds has been openly skeptical of AI-generated code for years, and his criticism was never ideological — it was practical. The tooling, in his view, didn't work well enough to be worth the trouble.

That position has softened in stages. Speaking at the Open Source Summit Europe in Prague, Torvalds told the audience his perspective had genuinely changed.

"I used to think the technology just wasn't very good. But now I've reached a point where I really enjoy using AI," he said at the conference.

For engineering leaders, that quote carries more weight than a vendor benchmark. Torvalds is not a marketer, and the kernel project is arguably the most heavily scrutinized codebase on earth — every patch passes through layers of human review before it ships to systems running everywhere from embedded devices to hyperscale data centers.

What does this change for enterprise engineering teams?

The practical question a CIO should ask: if the creator and chief maintainer of Linux now runs AI assistance for kernel work — including tests and review material — what does that signal for internal adoption?

A few signals stand out in how Torvalds approached it:

• Scope first. This is one focused project addressing a specific performance issue, not a mandate to generate code across the kernel tree.

• Human supervision throughout. Torvalds explicitly frames the assistant's output as produced "under my supervision" — the accountability chain still ends with a named human.

• Experimental status maintained. None of the eight changes have been merged into the official kernel. The work is public, but the bar for inclusion hasn't moved.

Full transparency. The documentation discloses exactly which tool was used and for what — vendor and tool named, in writing.

That pattern — contained scope, named supervision, unchanged merge criteria, and disclosure — is a reasonable template for any enterprise piloting AI-assisted development on production-critical code.

Why the tool choice matters

Torvalds used Claude Code, Anthropic's command-line coding assistant. The vendor specificity is notable: this isn't a generic statement about "AI" in the abstract, but a named product applied to a named problem in a named codebase.

It also lands at a moment when AI coding assistants are moving from novelty to infrastructure across enterprise engineering organizations. When the person who has reviewed more kernel code than perhaps anyone alive says the technology has crossed his personal quality bar — and that he "really enjoy[s]" using it — that's a data point worth more than most vendor claims.

The bottom line

Eight experimental patches, one named tool, one explicit supervision model — and one of software's most influential skeptics now on record as an enthusiastic user.