Seriously? [LWN.net]

Seriously?

Posted May 14, 2026 15:20 UTC (Thu) by jzb (editor, #7867)
In reply to: Seriously? by gordonmessmer
Parent article: Friction in Fedora over AI developer desktop initiative

FWIW, when writing the article, I thought it was explicit that you'd included in your non-goals things like AI services and LLM-powered applications by default. Reading it again, it still seems to me that that was your intent, but it specifically talks about privacy and remote services, but does not rule out a local LLM-driven application as I read it.

In other words, one could find wiggle room for working LLM-driven features into Fedora if they run from on-device models rather than (say) calling out to Claude or whatnot the way that the current proposal is written. If that's not the intent (and I'm guessing it is not), then the proposal could be modified to say so explicitly.

As I understand it and read the proposal, your responses, and Jef's responses, the drivers here are:

Create something Fedora-based that makes it easier to develop CUDA applications, etc.
To do so specifically as a Fedora initiative under the "Fedora Project" name with Fedora community blessing.

It is the second one that seems to be the real sticking point for many on both sides. I mean, Universal Blue is already doing some or all of the first. There's plenty of room to do that as a remix without any need for council blessing.

Particularly from Jef's responses, and earlier comments from McGrath and others, I get the sense that it is not sufficient from Red Hat/IBM's POV to have a thing that works to do AI development based on Fedora bits: it must be a thing that Fedora proper is working toward as a project as a Fedora AI Initiative™ with an eye to having an edition later on.

In other words, the real crux of the biscuit seems to be whether Fedora puts its blessing on something AI with all that entails; endorsing LLM-driven development/projects, rethinking its stance around out-of-tree kernel modules, and specifically enabling proprietary stacks like CUDA. Fedora, as a community-driven project, says "no thanks, we don't want anything with Fedora AI™ as its name on it" doesn't seem to be an acceptable result here. Rejecting, say, "Vigor-LLM enabled editing" as a feature might be allowed.

But, perhaps I am misunderstanding.

Seriously?

Posted May 14, 2026 15:39 UTC (Thu) by farnz (subscriber, #17727) [Link] (8 responses)

Related to that is that the only concrete bit of the proposal I can find so far boils down to "make it easier to run proprietary software stacks on your NVidia hardware, via a kernel that supports NVidia proprietary userspace, and some better tools for running containers with that proprietary userspace atop that kernel".

If there were things like easier ways to run models from Hugging Face on your open source ROCm/OneAPI/Vulkan/OpenCL stack, or even packaging up ZLUDA so people on non-NVidia hardware can use CUDA projects, it might feel less controversial. But there's a distinct lack of concrete things other than "open source userspace for AI is bad, so we should lean into proprietary userspace support, because the Fedora Freedom foundational principle is less important than AI", which is concerning from a distance.

Now, I've not otherwise joined in the conversation, so I may be missing something here, but it really does feel like this is "AI as motivation to abandon Free Software", not "push forward Free Software as the basis for your AI world".

Seriously?

Posted May 14, 2026 16:47 UTC (Thu) by pizza (subscriber, #46) [Link]

> But there's a distinct lack of concrete things other than "open source userspace for AI is bad, so we should lean into proprietary userspace support, because the Fedora Freedom foundational principle is less important than AI", which is concerning from a distance.

It may be "concerning" from a distance, but up close it feels more like a rug pull after getting slapped in the face.

>Now, I've not otherwise joined in the conversation, so I may be missing something here, but it really does feel like this is "AI as motivation to abandon Free Software", not "push forward Free Software as the basis for your AI world".

In theory, it's the latter. But in practice.... it will be the former.

Seriously?

Posted May 14, 2026 17:12 UTC (Thu) by rgmoore (✭ supporter ✭, #75) [Link] (2 responses)

But there's a distinct lack of concrete things other than "open source userspace for AI is bad, so we should lean into proprietary userspace support, because the Fedora Freedom foundational principle is less important than AI", which is concerning from a distance.

It should come as no surprise that a Free Software operating system is a casualty when the major goal is to play nicely with proprietary LLMs. Yes, there might be some value to a system that could be used for local LLM development, but we all know the big goal is to work well when using the big players like OpenAI and Anthropic, who aren't even close to Free Software. The only reason to care about a FOSS OS is because it's cheap and customizable, not because there's any philosophical attachment to Free Software.

it really does feel like this is "AI as motivation to abandon Free Software", not "push forward Free Software as the basis for your AI world".

I might be a tiny bit more generous and classify this as "AI is all that matters" rather than "AI as motivation to abandon Free Software", but it certainly comes from a place of not caring about Free Software values.

Seriously?

Posted May 15, 2026 0:31 UTC (Fri) by AdamW (subscriber, #48457) [Link] (1 responses)

"Yes, there might be some value to a system that could be used for local LLM development, but we all know the big goal is to work well when using the big players like OpenAI and Anthropic, who aren't even close to Free Software."

Well, no, it really isn't.

You don't even have to buy our bona fides for this. Just ask yourself that: how does that help RH or IBM? RH / IBM don't have a good way to make money off people directly using frontier models from other companies.

RH *can* make money off people running models on hardware they control, so it makes sense that this is what RH would want to enable.

Aside from that, just look at the proposal; most of the concrete bits are ultimately about making CUDA work. Which you don't need at all if you're just going to have OpenAI / Anthropic run the model for you.

Finally, there's no need for an "objective" to make it possible to use model providers like OpenAI / Anthropic on Fedora. You already can. I mean, they're really just web services. There isn't really anything for operating system developers to *do* to enable the use of remote model providers. We already give you a network connection and a bunch of text editors and programming languages and shells and web browsers. What else would you need?

Seriously?

Posted May 15, 2026 7:56 UTC (Fri) by rwmj (subscriber, #5474) [Link]

This is a good argument for including the proprietary bits in RHEL. It doesn't seem to have much to do with Fedora however.

Seriously?

Posted May 14, 2026 18:27 UTC (Thu) by Cyberax (✭ supporter ✭, #52523) [Link] (1 responses)

> If there were things like easier ways to run models from Hugging Face on your open source ROCm/OneAPI/Vulkan/OpenCL stack

Uhm... I'm using Ollama Web-UI running in a podman container on a machine with dual AMD R9700 on stock Fedora. It just works. I can just select a model from the dropdown menu, load it, and it will run as long as it fits in RAM/VRAM.

NVidia is still ahead when you want to squeeze every last drop of performance with custom kernels, but for casual use it's not at all needed.

Seriously?

Posted May 14, 2026 19:06 UTC (Thu) by farnz (subscriber, #17727) [Link]

That was just an example of the sorts of things they could be offering if there was more to the proposal than "make it easier to run proprietary userspace on a Fedora system", not something that was definitely hard today.

Seriously?

Posted May 14, 2026 19:27 UTC (Thu) by AdamW (subscriber, #48457) [Link] (1 responses)

I think the irony here is that it's *already* rather easy to "run models from Hugging Face on your open source ROCm/OneAPI/Vulkan/OpenCL stack", because it's all open and there already. Pretty much just do 'dnf install ollama' or 'dnf install ramalama' and have at it. We don't need an objective for that because it's just...software. We know how to package software! We've been doing it for a while!

This was the point I was sort of trying to make with my "outside of the kernel module / CUDA stuff, what *is* this?" post - that it feels weird because we're calling it an "AI" objective but in practice it winds up being a lot about "NVIDIA hardware enablement" because that's the only part of it that's actually at all difficult or different from "just packaging software", which we've been doing forever.

Proprietary software enablement, not hardware

Posted May 15, 2026 10:57 UTC (Fri) by farnz (subscriber, #17727) [Link]

Yeah - that's kinda my point. When you strip this proposal down to the concrete actions that will be taken, it's just "we should enable people to use NVidia's proprietary userspace on Fedora", and nothing else - calling it "hardware enablement" isn't really accurate, either, because it's not just about enabling the use of NVidia hardware, but about enabling the use of NVidia's proprietary APIs, too.

I would feel very differently about it if it was (say) "we're going to make ZLUDA work on all the big Mesa Vulkan drivers so that you could use CUDA code on AMD, Intel or NVidia hardware", or even "we're going to support running ML models on NVidia proprietary Vulkan to the same extent they're supported on RADV and ANV". But instead, it's "well, NVidia's proprietary CUDA userspace is popular, so we should support that".