So Gregory Kurtzer, the guy behind CentOS and Rocky Linux, just showed off a brand-new kernel called Ginux. He says he put it together in 11 days, mostly by letting an AI agent do the coding. The odd part is that it can already boot Rocky Linux 10 with no Linux code underneath. He’s quick to say it’s a research prototype, though, not something to run your servers on.
Kurtzer announced Ginux on X last week, and the code is now on GitHub under the Apache 2.0 license. The name combines Linux, GPT, and Greg. According to the project’s README, the “G” officially stands for GPT, a joke meant to keep its author’s ego in check.
Writing a kernel from scratch usually eats up years of work from a team of seasoned engineers. Ginux didn’t go that route, and no, it isn’t a Linux fork, whatever the name says. Memory management, scheduling, filesystems, networking: all of it was written from the ground up, mostly in Rust. The only thing it borrows from Linux is the ABI, so regular Linux programs run on it without changes.
The README treats Linux as a kind of “personality” the kernel puts on for the software it runs. Right now it’s the only one that exists, but BSD gets a mention as something that might come later.
Even that single personality is far from complete. Linux 6.12’s x86-64 interface has 375 system calls, and Ginux fully implements only 82 of them. Another 245 work in part, and the remaining 48 are deliberately blocked.
On Ginux, software runs inside isolated “domains,” and each one gets its own processes, memory, and filesystem. The project says that’s a tougher boundary than a container, but it still uses ordinary OCI container images to fill each domain with files. That’s how Rocky Linux 10 makes it all the way to systemd’s multi-user target.
Kurtzer’s technical report calls his method Human-Directed Agentic Engineering. He set the architecture, milestones, and pass criteria. The agent wrote the code, booted the kernel, tested it, and fixed the failures, all in one long-running OpenAI Codex task using GPT-5.6 Sol. In 11 days, it produced about 139,000 lines of code and tests in over 1,000 commits. Real distributions like Rocky Linux exposed bugs the synthetic tests missed.
Ginux remains a research project. There has been no independent security audit, and it boots only under QEMU, with no bare-metal, GPU, or NVMe support. Kurtzer concedes that AI can now produce complex systems software, but argues that compatibility, security, and long-term maintenance “remain the real moat.”
Linux’s own developers have become more receptive to AI. Last month, Linus Torvalds used Claude Code to write a set of experimental page cache patches. Reviewers found a race condition in them, and the patches are still under review. At about the same time, kernel maintainers agreed to no longer require contributors to disclose AI-assisted code.
Maybe you would like other interesting articles?