Rocky Linux Creator Built a New 'Not Linux' Kernel in 11 Days Using AI (itsfoss.com) 72
From the blog It's FOSS
Gregory Kurtzer is no stranger to the Linux world. He created CentOS, and when Red Hat messed it up, he started Rocky Linux as the community replacement. He also co-founded CIQ, the company offering enterprise support around Rocky. His latest project has nothing to do with rebuilding another Linux distro, though. He has built an entirely new operating system kernel, and he didn't write most of the code himself. An AI agent did.
Meet Ginux, Linux userspace without Linux underneath... It's an independent, Rust-first x86-64 kernel. Which simply means it is NOT a fork of the Linux kernel. It has its own memory management, scheduler, process model, filesystems, networking, and device support, all written from scratch.
The most interesting part is that even though it is not Linux, it is Linux-compatible. Ginux implements the Linux ABI (the binary interface compiled programs use to talk to the kernel), so existing Linux binaries can run on it unmodified. In the project's own terms, Linux is a "personality" that Ginux exposes, not the kernel running underneath...
Right now, Ginux runs only in a QEMU virtual machine, with no physical hardware profile, no GPU or NVMe support, and no independent security audit.
Meet Ginux, Linux userspace without Linux underneath... It's an independent, Rust-first x86-64 kernel. Which simply means it is NOT a fork of the Linux kernel. It has its own memory management, scheduler, process model, filesystems, networking, and device support, all written from scratch.
The most interesting part is that even though it is not Linux, it is Linux-compatible. Ginux implements the Linux ABI (the binary interface compiled programs use to talk to the kernel), so existing Linux binaries can run on it unmodified. In the project's own terms, Linux is a "personality" that Ginux exposes, not the kernel running underneath...
Right now, Ginux runs only in a QEMU virtual machine, with no physical hardware profile, no GPU or NVMe support, and no independent security audit.
Ginux? (Score:5, Funny)
Re: (Score:1)
So it's modeled after a large vagina?
Rumor has it OF dark money paid for the AI tokens.
I heard sponsors saying they hope this fucker, also sucks ass.
Kinda shitty. I know.
Re:Ginux? (Score:4, Funny)
Re: (Score:2)
Yes, and is compatible with all the same labiaries.
all written from scratch? (Score:4, Insightful)
all written from scratch.
I don't know if you know what that phrase actually means. Nothing from AI is "written from scratch". All of it is based off of stolen work from people who didn't give permission using a kind of social engineering no better than a petty hacker.
Re: (Score:3, Insightful)
Re: (Score:2)
Often in copyright cases for song melodies, if the artist being sued for infringement could have possibly had access to the music they’re accused of copying, even if it was something they listened to once, they can be accused of “subconsciously” infringing on the original content. One of the most notorious examples of this is Tom Petty’s claim that Sam Smith’s “Stay With Me” sounded too close to Petty’s “I Won’t Back Down.” Smith eventually h
Re: (Score:2)
It all feels so arbitrary what can be copyrighted and what's infringement. The music industry has some of the the most strict rules, while clothing designs and food preparations can't be protected at all.
Re: all written from scratch? (Score:4, Interesting)
It's not about uniqueness. If you take a photo of a pretty tree and post it online, you obviously own the copyright to that photo. If I find that same tree and take my own photo of it under very similar circumstances, and publish that, then my photo is still mine and I own the copyright to it, even though it's not very unique.
On the other hand if I take your photo and edit it with Photoshop to be subtly different, and publish it, I am in breach of your copyright. Even if those two hyopthetical photos I produced end up identical, one is in breach of copyright and one is not.
The question is whether I copy your work or not, and as you don't own any IP related to the subject (a tree), I am perfectly in my rights to take my own photo of that same tree.
Re: (Score:1)
Re: (Score:1)
On the other hand if I take your photo and edit it with Photoshop to be subtly different, and publish it, I am in breach of your copyright.
Not necessarily. My fair use overrides your copyright.
That is the question in front of the courts- the question of fair use.
And if training a model is fair use, is it fair use to use its product to compete with original copyright holders?
Ultimately, it will probably take law to clear this up for real. It may end up at the Supreme Court, but I hope they have the foresight to refuse to rule.
Re: (Score:3)
Fair use would be if you look at the photo, understand how the tree looks and then draw a tree based on what you learned about how a tree is supposed to look. A new image, not a transformation of an existing image.
Re: (Score:3)
Educate yourself. [wikipedia.org]
There are many fair uses of copyrighted material.
Again, this is what the courts must figure out- what, exactly, about this is or is not fair use.
Re: all written from scratch? (Score:5, Informative)
Not necessarily. My fair use overrides your copyright.
You can do that for your own use as much as you want, but your fair use doesn't include publishing it. Different countries have some exemptions for educational use etc. but as a general case, I can't take a photo you published, make it black and white in Photoshop and publish it as mine.
Re: (Score:1)
but your fair use doesn't include publishing it.
Again, not necessarily.
Every country I'm aware of contains fair use provisions for transformative use. It would be strange if they didn't.
I answered with "not necessarily" because it really does have a lot more to do with the physical act of "taking your image and editing it with photoshop", not that that accurately describes how an LLM works at all.
This isn't to defend the pro-LLM argument by any means, but it is the argument, and it has succeeded in some courts, and failed in others.
Re: (Score:2)
Re: (Score:2)
Re: all written from scratch? (Score:5, Informative)
My fair use overrides your copyright.
Legally, it does not. It's a defence for use, not on overriding principle in the way your wording implies.
Re: (Score:2)
It is a statutory exception to copyright.
Where my fair use rights intersect with your copyrights, the rights granted by your copyright are null and void- and that's clearly written in the statute.
That part that is not clear, of course, is whether or not my opinion about whether something is fair use or not, is actually fair use- and from that perspective, you can say it's a defense, since I'll have to make that argument into a defense if a prosecutor doesn't buy it.
Re: (Score:2)
Or all software. Nobody starts from first principles and a formal description of the CPU's instruction set.
Re: all written from scratch? (Score:2)
Maybe, kind of.
But there are two issues here which say no.
The first and more irrelevant argument is that LLMs aren't actually thinking, their process is not like ours just because they use neural networks, which bear only a passing resemblance to how our brain might work. We don't even know all of the details there and we STILL have enough info to know they aren't the same.
But the real argument is that copyright exists to protect the work of humans to induce them to make more work, so what the software is d
Re: all written from scratch? (Score:3, Interesting)
Re: all written from scratch? (Score:5, Insightful)
The issue with this semantic argument is that the model weights already more or less "contain" several partial OS kernels, in the sense that you can squeeze them out as repeatable, token-identical structures with the right inference setup.
When humans compete in reciting longer and longer parts of the infinitely long number pi, we don't say they "derived pi from scratch up to N places" (that's also a thing humans can do, we just typically don't compete in it, because it smells like shit that should run on a computer); we say they "memorized and recited pi up to N places", and what the LLM did with the kernel here is closer to remembering and reciting, rather than "making from scratch".
Now, don't get me wrong, reciting significant parts of even vaguely functional OS kernel from memory would still be a super impressive and potentially useful skill for a human, but it's a bigger question of how useful of a skill is it for a computer. "Legacy" software can "remember" vmlinuz at cost of ~8MB of storage and mW needed for seconds to decompress it, and I'd argue that spending 11 days on building a kernel of unknown quality, unknown novelty and very limited utility is not all that useful or impressive of a function when compared with the existing ability to maintain a bit-perfect copy of the OS you already know works.
The feat is being presented as impressive, except, the impressive part of it doesn't really hold up to the specific implications of "from scratch" in the context of the question. Is the new OS likely to have existing development paradigms grandfathered in? Absolutely. Is it likely to inherit some of the bugs and vulnerabilities of existing OSes? Indubitably. Is its apparent "novelty" going to make it a compelling alternative to existing, available OSes? Unlikely in the least. Was it at least less expensive or avoided usage of some limited resource that'd a compressed image of linux core would need to work? Absolutely not. Did the developer learn enough to now repeat the feat without the help of the LLM? Probably not super likely.
If it has none of the practical advantages or even traits you'd expect of something made "from scratch", then perhaps it shouldn't be called that?
Re: (Score:2)
model weights already more or less "contain" several partial OS kernels,
No, they don't. They were trained on that data, but they are not that data. You can't extract it from weights, and I highly doubt you'd be able to reproduce it from weights, even on an exceedingly large model.
At best, it's a work inspired by the original, not even a derivative. There's no "substantial or material" replication.
Re: (Score:3)
You can't extract it from weights, and I highly doubt you'd be able to reproduce it from weights
Except that's lilerally what inference is, and you know it can't be anything but that because training (and most evidently "pre-training", but training at large) literally rely on that to function. The backprop during training is a literal "adjust the wegiths to ensure that we will able to reproduce this output giving these inputs with sufficient statistical probability".
'"substantial or material" replication' is entirely lawyer speak around how the fundamentally arbitrary legal convention will treat those
Re: all written from scratch? (Score:5, Informative)
The goal is to
Cool, and my grandma's goal could have very well been to win the Tour De France, yet it still didn't make her any more bicycle-like.
I do not need to encode 1 + 1 = 2 in order to solve for 2.
Do you have any proof of that? Because vast majority of people definitely see various forms of "1 + 1 = 2" repeat many times before they get to the level where they can derive that formula themselves, and I'd argue it's fairly easy to prove that vast majority of people actually remember that "1 + 1 = 2" and don't derive it in practice. As far as I know, we have seen LLMs generally tending towards being more reliable at repeating things that are already in the data than they are at inventing novel things, and as far as I know, there's been zero LLMs that have written operating systems "from scratch" that haven't had multiple complete operating systems in their training data.
I've been training neural nets since well before the initial Transformers paper. You vomiting a couple [overfitting mitigation methods] isn't impressive to me. We've known for a long time that some information is more useful when blurred, and often easier to memorize when less specific. Yes, of fucking course, we're aiming for generalization in AI models. But I you haven't shown anything to prove that this particular LLM isn't just sticking together pieces of kernels it has seen. Nobody calls a collage "from scratch".
Re: (Score:3, Informative)
Cool, and my grandma's goal could have very well been to win the Tour De France, yet it still didn't make her any more bicycle-like.
Your analogy is that your grandma wanting to achieve something unachievable for her is comparable to LLM pretraining specifically trying *not* to memorize token sequences?
Not very intelligent, are you?
Do you have any proof of that?
Do I have proof of that?
It's self-fucking-evident. Do you think a calculator decodes an unwrapped ALU output to answer your inputs?
But guess what- it could.
Because vast majority of people definitely see various forms of "1 + 1 = 2" repeat many times before they get to the level where they can derive that formula themselves, and I'd argue it's fairly easy to prove that vast majority of people actually remember that "1 + 1 = 2" and don't derive it in practice.
Strawman. Probably unintentional on your part given your demonstrated inability to hold a train of thought, but:
I asserted that I did not need to, you
Re: (Score:2)
Re: (Score:2)
If you anthropomorphize the term "learn", then you're correct. They learn nothing.
But then again, at that point, a heuristic spam filter learns nothing either, and yet we have called that learning for 30 years.
So, accepting that this level of flexibility is allowed for the word "learn", we can accept that the backpropagation is, in fact, learning.
Further, "learning" is the exact fucking opposite of novelty. It pains me that people as stupid as you have the boldness to share your half-b
Re: all written from scratch? (Score:2)
You cannot get the training data out of a model unless it's so overtrained on a specific piece of it that it actually impairs its usefulness for other things. So no, those works are not in there.
However, you might still be creating derivative works with an LLM, because the standard for those is "substantial recognizable elements" and the exact meaning of that can be fuzzy.
Re: (Score:2)
Re: (Score:2)
Thank you for this sensible comment.
Re: (Score:2)
Then you should have no problem identifying infringing code?
Re: (Score:1)
No, they don't. They were trained on that data, but they are not that data.
And yet the ability to produce long quotes from original data has been demonstrated a number of times.
Re: (Score:1)
The phrase "from scratch" was originally a sports analogy meaning from the starting line, like a footrace, or a facing line like in boxing. A line in chalk was scratched into the ground. So in this case it means from the beginning, which even using AI, it is. The boxers come to the ring to fight from scratch, but they've had years of training and watched countless boxing matches to learn. Both the CentOS creator and the AI agent have learned immense amounts of information from a lot of other people's open source work, but they started making a new OS from the beginning or the scratch line.
If an author throws away a manuscript and starts again from scratch, people don’t assume there’s some double meaning there. Same goes for the baker starting over.
The end result is starting from scratch. Starting completely over. Full reset at the line. Like the race does. And perhaps the term is being misused here if no one truly starts from “scratch” in the FOSS world, because of the F in FOSS.
RIP Groklaw (Score:2)
With the publishing industry's ongoing battles against illegal copying to train AI as a background, it'll be interesting to watch how this plays out. Too bad Groklaw [wikipedia.org] isn't around anymore.
Re: (Score:3)
Is it "stolen work" to read a book and then recite it
In a public performance for profit, that's potentially a copyright infringment.
or paraphrase it, or create a story with the same themes? No. Those are new forms of work.
You might want to check some case law.
New Script Kids on the Block. (Score:1)
all written from scratch.
I don't know if you know what that phrase actually means. Nothing from AI is "written from scratch". All of it is based off of stolen work from people who didn't give permission using a kind of social engineering no better than a petty hacker.
So, you’re saying we should call them Script KiddAIs..
Re: (Score:2)
Pretty neat, like, academically. (Score:5, Insightful)
Ginux runs only in a QEMU virtual machine, with no physical hardware profile, no GPU or NVMe support, and no independent security audit.
So, as can be expected, it doesn't attempt roughly 99.9% of the real work an OS has to do to be useful as an OS someone might want to use for something.
It's perfectly good, cool, and virtuous to make academical/hobby model-OSes, but the "Built a New 'Not Linux' Kernel in 11 Days Using AI" headline is hella yellow when the thing being made is just not comparable to Linux kernel in what it's even meant to be able to do for you, but is being presented as some sort of achievement.
Re:Pretty neat, like, academically. (Score:5, Informative)
it doesn't attempt roughly 99.9% of the real work an OS has to do to be useful as an OS someone might want to use for something.
It's an extremely interesting experiment, especially for Rust as a systems language. But I think you are downplaying it a little. From the full announcement:
"What Ginux can do today
Ginux builds as a single GINUX.EFI image that boots straight through UEFI, with no Linux host kernel involved.
It runs isolated "domains", each with its own processes, address space, filesystem, and lifecycle. The project describes these as stronger boundaries than OCI containers or Linux namespaces.
It can also pull and verify OCI container images and use them to assemble a domain's root filesystem. That's how it gets Rocky Linux 10 booted all the way to systemd's multi-user.target. Selected Alpine Linux programs (like the musl and BusyBox) run too.
Ginux running Rocky Linux:
ginux domain new rocky
rocky: created (image required)
ginux domain set rocky image=rockylinux/rockylinux:10-ubi-init
ginux domain set rocky init=/sbin/init memory-mib=2048
rocky: settings accepted; configured
ginux domain start rocky --attach
ginux: OCI cache hit (2 layers)
ginux: starting Linux personality domain
systemd: Reached target Multi-User System.
Rocky Linux 10
rocky login:"
https://itsfoss.com/news/ginux... [itsfoss.com]
Re:Pretty neat, like, academically. (Score:5, Insightful)
It runs isolated "domains", each with its own processes, address space, filesystem, and lifecycle. The project describes these as stronger boundaries than OCI containers or Linux namespaces.
Again, I agree this is extremely cool and interesting but the headline puts it on the level of Linux by comparison. And unless compared including hardware support and performance, you can't really say if any of it's "stornger" qualities mean anything. World's fastest aircraft that exists exclusively as a virtual model still isn't comparable to aircraft that actually had to be built.
Re: (Score:2)
I'm just thinking there's no way all that code was actually reviewed in earnest by humans in eleven days, so... it's a cute idea but that's a "hell no" for actually running it on anything.
Re:Pretty neat, like, academically. (Score:4, Interesting)
Yeah, but it's a hell of a good demo or PoC of "domains" isn't it? I mean, we've all seen a container, we might have seen a namespace, a chroot environment, etc, but now we can actually see a domain and how it (maybe) does things better than all of them. In that sense then, you *can* judge the better or worse qualities within the confines of what's been implemented. Had he just said "I've got an idea... it's way better than containers..." - then yes, you could argue it's impossible to determine if it really is better or not.
Back to aeroplanes... if you made the world fastest aeroplane, but it only had one seat, that still makes it useful to work out how you might one day make a faster passenger aeroplane, even if the demo model couldn't fit so much as a cabin bag, let alone any passengers.
Re:Pretty neat, like, academically. (Score:4, Interesting)
World's fastest aircraft that exists exclusively as a virtual model still isn't comparable to aircraft that actually had to be built.
Yet that virtual wind tunnel and other virtual tests are today the early steps in aircraft design.
...
I think the original announcement frames things well:
"For Ginux, that meant one persistent OpenAI Codex task running GPT-5.6 Sol for eleven days. It produced approximately 139,000 retained lines of implementation and test code, across more than 1,000 commits. Architecture documents kept the agent aligned, and real userspaces like Rocky Linux caught bugs that synthetic tests missed.
The "human-directed" part matters. This isn't as simple as typing "write me a kernel, don't make mistakes" into a chatbot. It's a person with decades of systems experience steering an agent that does the heavy lifting of coding.
Linux isn't going anywhere because a prototype boots systemd in a VM. Decades of hardware enablement, code review, and community are not something an AI agent produces in eleven days.
But let's take a step back to realize what happened here. A new kernel created from scratch, that runs an enterprise distro. And it took just 11 days with GPT 5.6 (not even 6.1, the latest model).
AI won't replace expertise. An expert who already knows how kernels work uses AI to squeeze years of grunt work into days.
When people with deep domain knowledge pair it with frontier AI agents, projects that once needed a large team and several years start to look like weekend side projects.
Ginux won't replace Linux, but it does show how far (and how fast) an skilled person can now go. "
Re: (Score:2)
It's perfectly good, cool, and virtuous to make academical/hobby model-OS
Dr. Tanenbaum, is that you?
Re: (Score:2)
Somewhere, someone must have built a server farm of all different computers with different hardware plugged in, for continuous integration testing operating systems. I've never heard of this publicly though.
Re: (Score:2)
Writing your own OS-Kernel... (Score:3)
... is a popular past time. Many people do this, often for fun.
It's so popular, there's even a Wiki for it
https://wiki.osdev.org/Expande... [osdev.org]
Re: (Score:2)
Year of the Hurd! (Score:2)
So you are saying there is a chance for Hurd?
Re: (Score:2)
Re: (Score:2)
I wonder how much Redox the AI author knew? (Score:2)
Redox OS already exists in exactly this space, albeit without the AI coding.
Just wondering how much Redox has seeped into Ginux.
Re: (Score:2)
Given the way that these work, the more likely source of the concepts that were then turned into Rust came from Linux itself.
Maybe the Gnu guys can finish Hurd now (Score:2)
Well (Score:2)
How does it perform ? (Score:2)
How fast is it, how large is the binary ? Even if the numbers are not great this is amazing for a first release.
Just as importantly: how maintainable is it, how easy to find out what happens when things go wrong ?
Still linux just stolen... (Score:1)
Stop pretending this is creation, it is theft from known code.
Re: (Score:2)
If you use a known assembler instruction, you are stealing known code.
Real time, parallel code/feature generation? (Score:2)
This brings up an interesting point. What would keep someone from making a kernel like this, except updated in real time in a clean room fashion to match the Linux kernel and features?
This is a huge can of worms. On one hand, it presents bugs which are not in the original kernel. Most likely show-stoppers, perhaps CVE scale 9 bugs. On the other hand, the attack surface is completely different, and security through obscurity may work.
Where this can be useful would be moving apps to Rust, Ada, or SPARK, s
I did this too (Score:3)
I did this too, in less than 11 days, and mine runs on bare metal on a raspberry pi. You can download the image here. It also runs it's own busybox, and sbase (suckless.org rewrite of busybox) and kubectl and quake, doom, helm, duck db, ffmpeg etc etc I did a write up about it here back in October: Bluetail OS you can download a copy of the 0.0.1 release here [osdev.org]https://bluetailos.s3.us-west-... [amazonaws.com] runs in qemu of course, on a raspberry pi 3 A+ profile on 512mb memory
Lots of progress since then, refactored a bunch of stuff runs a lot better on bare hardware and has better usb mouse/keyboard support. Oh and it has a usb network adapter as well.