The creator of Rocky Linux had an AI write a Linux-compatible kernel in eleven days

On October 27, Gregory Kurtzer announced an unexpected project on X: Ginux, “a new Linux-compatible operating system kernel that boots Rocky Linux 10 with systemd.” He is no newcomer. He created CentOS, then Rocky Linux, two distributions widely used on enterprise servers. This time, he hardly wrote any code himself: according to the project's documentation, everything was produced by a single Codex task, OpenAI's programming tool, using the GPT-5.6 Sol model, in eleven days.

The “G” in Ginux stands for both Greg and GPT, notes Next, which detailed the project on November 2. The kernel, written mostly in Rust, is not a copy of Linux. It has its own memory management, its own scheduler, its own file systems and its own networking, and only imitates the interface that Linux offers to programs.

A human architect, an AI that codes

Kurtzer calls his method “human-directed agentic engineering.” He set the architecture, the constraints, the milestones and the success criteria. The AI agent wrote the code, booted it, tested it, diagnosed errors and fixed them, in a loop, until it passed automated checks.

The result is impressive in size: about 139,000 lines of code and tests, according to the project's readme, and around a thousand recorded commits according to Next. Of the 375 core functions (system calls) that Linux 6.12 makes available to programs, Ginux fully implements 53 and partially implements 257, leaves 17 unsupported and deliberately rejects 48, according to the count revised on November 1 in the repository (the initial announcement claimed 82 complete and 245 partial calls).

What Ginux can't do

The limitations are numerous, and Kurtzer spells them out himself. Ginux only runs in a QEMU virtual machine, not on a real computer: it supports neither graphics cards nor NVMe drives. It has undergone no independent security audit. The version evaluated did not even pass all of its own tests, because of one failed check. The repository, published under the open-source Apache 2.0 license, describes it as a research prototype, “not for production or as a replacement for Linux.”

On October 29, responding to reactions, Kurtzer dismissed the idea that Ginux aims to compete with Linux: “That would be silly!” He says he considers this code disposable. What interests him are other questions: how long it takes to reproduce a bug, find its cause, fix it and validate the fix, and how much human intervention is really still needed.

What to take away

For an IT manager, Ginux changes nothing in the short term: nobody will run their servers on it. But the experiment gives an order of magnitude of what a programming agent can produce, with an experienced human to guide it. According to IT-Connect, Kurtzer estimates the cost of the operation at about $8,000 in API calls. It comes at a time when open-source maintainers are questioning the influx of AI-written code, and when AI is also being used to find vulnerabilities in the real Linux kernel. What happens next depends on two things: the technical report Kurtzer is publishing separately, and what developers who want to take over the project do with it.