The father of CentOS had AI write a kernel in 11 days: Ginux's 455 stars came from one news story

Ginux is a Rust kernel that Gregory Kurtzer, founder of CentOS and Rocky Linux, handed to OpenAI Codex to write. It isn't modified from Linux, yet it can boot Rocky Linux 10 all the way to systemd in a virtual machine. Released in Oct 27, by Nov 6 it had 455 stars and 57 forks (GitHub, 11-06). Almost all of those stars came from Linux media reposting one story after another, and the reasons were "who made it" and "11 days." Few people starred it because they wanted to use it. You can't borrow the author's reputation. What you can learn is how he picked a target and how he launched it, turning a technical experiment into a news story.

A "11 days" post

2026 Oct 27, Kurtzer posted a single line on X:

In 11 days, I built a new Linux-compatible operating system kernel that boots Rocky Linux 10 with systemd, all using Human-Directed Agentic Engineering (HDAE).

(11 days, I used "human-directed agentic engineering" to build a Linux-compatible new kernel that boots Rocky Linux 10 with systemd.)

It's FOSS published a report the same afternoon, titled "Rocky Linux Founder Built a 'Not Linux' Kernel in 11 Days Using AI." The article also picked up a joke from the README: the G in Ginux looks like Greg, and the author said that was "to keep my ego in check," with G officially standing for GPT. Nearly all the code was written by GPT-5.6 Sol in a single Codex task running for 11 continuous days, leaving roughly 139k lines of implementation and test code across 1,000 commits. That same day he published a technical report with a DOI on Zenodo explaining the method: humans set the architecture, constraints, milestones, and acceptance criteria. AI wrote the code, booted it, tested it, debugged it, and revised it until it passed every executable check.

The post struck the right tone. He repeatedly said this wasn't meant to replace Linux. In his words, "AI can write complex systems software, but compatibility, security, hardware support, standards, community, and long-term maintenance are the real moat." It's FOSS also stated the shortcomings up front: it only runs in QEMU, has no real-hardware support, no GPU or NVMe support, and no security audit.

How the stars climbed

(Star counts from GitHub's daily figures.)

Two things stand out. First, it never took off on Hacker News: 6 points and no comments, so these stars weren't dug up by the developer community. Instead, tech outlets copied each other: It's FOSS, Linuxiac, TechSpot, and Slashdot had nearly identical headlines. Second, Kurtzer's other repos have almost no stars, which means the stars followed the story, not his account.

The 86 comments on Slashdot also show who showed up. The highest-scored ones were making crude jokes about his name, and further down people argued about whether AI-written code counts as "from scratch" and whether model weights already contained several kernels, then drifted into copyright and fair use. Very few people discussed the kernel itself. People came for the spectacle and to take sides, which reflects where the "can AI write a kernel" debate stood at the time: in October Linus had just publicly said he used AI to write patches, and both sides were hunting for examples.

Day 18, he "killed" it

On Nov 5, Kurtzer published a long LinkedIn post titled "Built Ginux in 11 days, killed it on day 18 (well, pivoted)." He said the project was never meant to produce a usable kernel ("that would be silly"). Its purpose was to test the HDAE method: pick something you understand, that's easy to test, and that ordinary people wouldn't do alone. Having proven that, he didn't expect the "huge attention: news, articles, podcasts, videos, and drive-by code reviews" it drew. He used the same method to automatically review nearly 80 PRs and 70 issues.

The repo now contains something different: a "user-space kernel" that attaches to Linux as an ordinary process. Each isolated environment has a contract specifying which system calls are passed through, which are forwarded, and which are denied outright. The README says, "Linux must never break user space; Ginux is meant to break it on purpose." The version that could boot directly is kept on the bootable-kernel branch.

The pivot itself made news again, and from 11-03 to 11-05 the repo gained 36 to 49 stars a day. However, the new direction is close to Google's gVisor (about 20k stars, already used in production), and the author himself admits gVisor and User-Mode Linux already do part of this. Whether the new direction can hold people on its own remains to be seen over the next few weeks.

At the same time, he split the method out into its own hdae repo, packaged as a skill for Codex and Claude Code that installs with one curl command. It had 51 stars at the time. When the attention peaked, he captured it with something people could actually install.

Who's doing something similar

The last row is the most telling. moxnox and Ginux came out only five days apart and tell the same story: AI helped someone write an operating system. Its author had no name recognition and got less than half of Ginux's stars. That's roughly what the "famous name" factor is worth: the story itself earned about two hundred stars, and the author's name added over two hundred more (an inference). Most toy OSes that ordinary people vibe-coded online don't get even that many.

If it were you

C: learn how it picked its target and how it launched, but don't try to follow this project. More than half of Ginux's stars came from the "CentOS founder" identity, plus media reposting it again and again. Followers can get neither of those. It's already been covered by a whole round of Linux media, and in the user-space sandbox niche it's still overshadowed by gVisor. Writing your own "AI-written kernel," however well done, would just make you the second person telling the same story.

Best suited for: developers with years of real experience in a hardcore field (databases, compilers, browser engines, emulators, protocol stacks) who already have some standing in their community and want to use an open-source project to show what an expert can accomplish with AI. You don't need a name as big as Kurtzer's, but you do need one or two vertical outlets or communities willing to repost it for you.

Who it's not for: People new to the field who expect AI to write an OS for them. Kurtzer himself says that stuff AI-assembled by outsiders "on vibes" comes out poor quality, and media won't pick it up. Also not for people who want to build "a better gVisor" to rack up stars. That's years of engineering, not a single news hook.

How to do it: Copy the structure, swap the content. Pick something you know well that ordinary people wouldn't tackle alone, with an objective acceptance test outsiders can understand at a glance, like "boots a real distro," "passes an official test suite," or "opens a real website." Set a number of days and let the AI work nonstop, keeping the process on record. On launch day, release the post, the repo, and a method report (with a DOI or a long-form essay) together, so journalists have something to write about. Put numbers in the headline: days, lines of code, commit count, what it runs. State your weak points up front, because serious outlets will only repost what's honest. Once heat builds, immediately package the method into something people can install (a skill pack, templates, scripts) and route the traffic to your next repo.

How to get users: Don't count on Hacker News. Ginux only got 6 points there. Focus on vertical media and communities in your niche. Linux has It's FOSS, Linuxiac, and Slashdot, and databases, gaming, and security each have their own equivalents. Email editors directly, with a one-line headline and a screenshot of it running. After that, write a long-form post for every pivot and milestone. Ginux's second wave of stars came from the "killed it on day 18" article.

When to move: before: Aim for the window while the debate over "can AI write system software" is still hot. Right now someone is doing something similar every month, and "AI wrote X" stories will keep getting staler. Ideally ship before Q1 of year 2027, and pick something no one has yet claimed AI wrote.

Signs you read it wrong: ① If before year-end an unknown author's "AI wrote X in N days" project passes 2,000 stars, the story alone is enough and fame isn't required. That path is easier than C. ② If Ginux, after pivoting, still gains more than 100 stars a week in weeks with no news, the user-space contract direction holds people on its own, and following its technical direction is worth considering. ③ If "method skill pack" repos like hdae outgrow Ginux in stars, the real value is the method, and you should follow the method pack, not the project.