Table of Contents

    Linus Torvalds Uses AI for Linux Kernel Development: What Happened

    Thursday, October 15, 2026 · 5 min read · 5 views

    Expert's summary

    A closer look at how Linus Torvalds used Claude Code and LLMs in experimental Linux kernel work, plus the rules and lessons for AI-assisted coding.

    Linus Torvalds, the creator of Linux and Git, has used AI to assist with Linux kernel development. According to the original Indonesian article, in October 2026 he directed Anthropic's Claude Code to help produce eight experimental patches addressing Linux performance. Those changes had not yet been merged into the mainline kernel at the time covered by the article.

    This drew attention because Torvalds had previously been skeptical of the AI industry. In 2024, he described much of its activity as marketing. The two kernel cases illustrate a more practical distinction: using AI as an engineering assistant while keeping technical judgment and responsibility with a human developer.

    How Linus Torvalds used AI for Linux kernel work

    The original report describes two separate efforts in October 2026: an experimental page-cache optimization and a review of older code in the kernel's associative-array implementation.

    Page-cache performance patches with Claude Code

    Torvalds created an experimental branch named noref-filemap-read-with-llm. It contained eight patches based on Linux 7.3-rc6, without a confirmed target release for merging.

    The patches addressed contention when many threads read the same file. Each thread normally updates a shared reference count, creating a bottleneck. Matthew Wilcox had raised this performance issue in 2024. Torvalds' approach avoids taking that reference in the same way while delaying memory reuse briefly to preserve safety.

    The original article reports a test on Torvalds' laptop with some security protections disabled. With eight threads reading one file, throughput rose from roughly 10 million to 30–33 million reads per second under those test conditions. This is a specific experimental result, not a general performance guarantee for Linux systems.

    Torvalds ran the measurements himself. Under his direction, Claude Code helped write code and testing programs, review known issues, and draft commit messages.

    Investigating old bugs in assoc_array

    The second effort involved assoc_array, an associative-array implementation used by the Linux kernel's keyring subsystem. David Howells of Red Hat originally wrote the library in September 2013.

    Developer Nils Ollrogge reported a bug and proposed a fix. Torvalds found the code difficult to understand, so he used an LLM to break large functions into more readable pieces. A user-space test harness then confirmed the reported bug and exposed three additional bugs attributed to the implementation's original 2013 commit.

    At the time described in the Indonesian article, the fix was expected to enter the kernel through the Linux 7.4 merge window. That was an expectation rather than a confirmed final release outcome.

    How Torvalds' position on AI evolved during 2026

    The original article outlines several developments that show a gradual shift from skepticism toward selective use:

    PeriodDevelopment
    Early 2026Torvalds used AI for a personal audio project called AudioNoise.
    March 2026Sashiko, an AI-assisted code-review system built by Google engineers, was released as open source.
    April 2026The Linux kernel introduced a policy for AI coding assistants alongside Linux 7.0.
    May 2026Torvalds raised concerns about AI-generated bug reports overwhelming the kernel security mailing list.
    July 15, 2026Torvalds stated on a kernel mailing list that Linux was not anti-AI and that those who disagreed could fork the project.
    August 2026He used AI to help fix a kernel bug through 24 debugging patches and 18 kernel boots.
    October 2026The page-cache patches and assoc_array investigation became public.

    In the August case, the original account says the AI repeatedly suggested the bug could not be fixed, but Torvalds kept investigating. The fix was included in Linux 7.3 and marked for backporting to stable kernel branches, according to that account.

    At Open Source Summit Europe in Prague, Torvalds described AI as useful when treated as a tool rather than an authority. He also acknowledged the burden that low-quality AI-generated reports can place on kernel maintainers.

    Linux kernel rules for AI-assisted contributions

    The article discusses the kernel's official AI Coding Assistants policy. Its core principle is that contributors remain accountable for their work, regardless of the tools used.

    Human review and legal sign-off

    An AI agent must not add a Signed-off-by tag on behalf of a human contributor. That sign-off is a human statement under the Developer Certificate of Origin. The person submitting the patch must review the work, check licensing requirements, and take responsibility for it.

    Attribution and licensing

    The original article states that AI-assisted contributions should be attributed with an Assisted-by tag. Any submitted code must also comply with the kernel's applicable licensing requirements, including GPL-2.0-only.

    Reproducible bug reports

    For nontrivial bugs, a reproducible test case is especially important. AI may generate plausible reports that do not describe real failures. A human maintainer needs evidence and a reliable way to reproduce the behavior before making changes.

    The article presents AI usage as optional rather than universally required or prohibited. The review process and standards for accepting a patch remain the same.

    What developers and organizations can learn

    Open-source contributors

    The assoc_array case shows a useful role for AI: explaining unfamiliar legacy code, restructuring code for readability, and creating tests that make a bug easier to verify. It does not mean a generated patch should be submitted without inspection.

    Developers contributing to projects such as the Linux kernel still need to understand the code, test it, provide the necessary attribution, and meet the project's contribution standards.

    Businesses and public-sector IT teams

    The original Indonesian article suggests that organizations—including companies, banks, startups, and public-sector teams—can borrow practical safeguards from this model.

    1. Require reproducible evidence for findings generated with AI.
    2. Record when AI assisted in making a code change.
    3. Assign a human owner to review and approve each contribution.
    4. Ensure the team has enough capacity to validate reports rather than letting AI increase the volume of unverified issues.

    These controls are particularly relevant to AI-assisted code review and security auditing, where plausible output is not the same as a verified finding.

    Computer science education

    The original article also argues for continued emphasis on operating systems, memory management, concurrency, and C programming in computer science education. Torvalds' ability to benefit from AI depended on a deep understanding of the underlying system.

    The broader lesson is to treat generated code and explanations as material to inspect, test, and challenge. Strong fundamentals make it possible to recognize mistakes, evaluate safety trade-offs, and interpret performance tests.

    The cases presented here demonstrate a human-led workflow for AI-assisted engineering: tools can help with debugging, test harnesses, code comprehension, and drafts, but humans retain responsibility for measurement, review, and decisions.

    Article FAQs

    Short answers to common questions connected to this article.

    Did Linus Torvalds use AI to write Linux kernel code?

    According to the original October 2026 report, yes. He directed Claude Code to help draft eight experimental performance patches and used an LLM to examine assoc_array code. The article reported that those experimental changes had not yet been merged at that point.

    Open FAQ URL

    Which AI tool did Linus Torvalds use?

    The original article identifies Anthropic's Claude Code in connection with the performance patch series. It also describes LLM assistance in reviewing older associative-array code.

    Open FAQ URL

    Does the Linux kernel accept AI-generated contributions?

    The original article describes AI-assisted contributions as possible when a human reviews and takes responsibility for the code, follows contribution and licensing rules, and includes the required attribution.

    Open FAQ URL

    How large is the Linux kernel codebase?

    The original article says the Linux kernel source tree exceeded 40 million lines around Linux 6.14-rc1 in early 2025, including comments and documentation. Counts depend on what is included and may change with releases.

    Open FAQ URL
    Katon Fajar

    The author

    Katon Fajar

    Hi, I am Katon Fajar Utomo. A Web Developer & SEO Specialist dedicated to building high-performance modern websites, scalable internal systems, and technical search engine optimization.

    More from Katon Fajar

    Keep reading