The FOSS vs AI dilemma
Posted by PanicTasty@reddit | linux | View on Reddit | 45 comments
I’ve been stuck on a massive contradiction lately... The entire global AI software stack runs natively on Linux. This is where bleeding edge AI development is actively happening, yet the Linux desktop community is "revolting" against it.
Canonical is charging ahead for Ubuntu, integrating local, open-weights models as system primitives. Everything is isolated in Snaps. Don't want it? Run a single command to remove the snap, and the entire local inference stack is completely purged from your drive.
For the community, this is just another reason to hate on Ubuntu.
Fedora's AI initiative was blocked by its own council. The community "revolted" because having to accommodate NVIDIA modules and proprietary CUDA APIs violated Fedora's rules.
Ubuntu and Red Hat forge ahead while bleeding-edge Fedora gets left behind.
AI coding tools might feel like "slop" right now, but real-world engineering teams are using them to successfully translate complex, legacy applications from one language to another in weeks. Yet, it’s almost impossible to imagine the conservative GNOME maintainers ever adopting code generation. On the flip side, maybe the KDE Plasma team is more likely to experiment with AI-driven tools.
Just like Windows, the Linux DEs are stuck with decades of code that at one point or another cannot be update, not because of backward compatibility issues, but because of a lack of manpower. And there might be a solution. My take is that I am really excited about this, and I am really interested in what the future of Linux could become.
Creative_Bedroom_448@reddit
I honestly think a lot of the backlash is less “AI bad” and more exhaustion with how aggressively every company is trying to force AI into places it doesn’t belong. The irony is real though, Linux powers almost the entire AI world underneath while parts of the desktop community act like experimentation itself is betrayal. Feels like there’s a middle ground between blind hype and total rejection that nobody wants to sit in.
__ali1234__@reddit
Canonical's inference snaps are kind of shit compared to just installing ollama. They take more effort to configure, only let you run one weight of the model at a time, and the whole processes has to be repeated for every model.
FattyDrake@reddit
I think you're also missing a key point.
People who work in their spare time on FOSS projects, are likely to enjoy programming on some level.
Yes, LLMs can be good for boilerplate or some repetitive tasks, checking vulnerabilites, etc. but the idea of having an AI develop the bulk of some software is anathema to the entire reason a lot of people work on free software to begin with.
I can't seem to find it but there was a blog post a little while back by a fairly well known programmer who talked about AI exposing an up until now invisible divide. Those who actually enjoy the process of programming, the languages, and creating elegant solutions, and those who just want to make a computer do a thing.
AI is really hyped among people in the second group, who just want to make a computer something but never really wanted to learn how (even if they did for a paycheck.) I've known several people like this at places I've worked at.
It's akin to the CEO of Suno saying that people really don't like creating music. That exposes how a lot of the AI industry sees their solutions. People don't want to make music, they don't want to create art, they don't want to program or do any sort of craft. Hell, the stupid Posha Robot Chef takes over the fun part of cooking but leaves the tedious prep and cleanup to the user. This is fundamentally a myopic view of why people do things.
It's not so much that an LLM isn't useful, they are. It's that the way they're being marketed, developed, and pushed is inherently anti-human.
__ali1234__@reddit
People who "enjoy programming" have been gate-kept out of every large FOSS project for the past two decades.
If you enjoy programming, writing new AI tools is a hell of a lot more fun right now than trying to contribute to traditional FOSS projects.
PanicTasty@reddit (OP)
Well, back in the day, most mothers, just like my own mother, did not use cake mix. Then they lied about using cake mix. Now they all use cake mix. Cake mix is easier, faster, and the cake is actually better. You still have to bake the cake, and then you get to enjoy it. The creation process does not have to be a grind to be better. This is the same with music. I remember when people were arguing about Auto-Tune, and before that, they argued about synthesizers. The artists of tomorrow will use this new technology to make more music; they already are.
That out of the way, AI is here to stay. I am not arguing for using vibe-coding slop in the Linux kernel. I am saying that AI is already in the kernel and other free software. If anything, I am arguing that I miss the time we were all excited about the new technology available and what it can become, like in the old days.
FattyDrake@reddit
You know, you kinda called yourself out with that one.
PanicTasty@reddit (OP)
There’s room for both the artisan who wants to do it by hand and the builder who just wants to make the computer do something incredible. At the end of the day, people still want to eat the cake.
FattyDrake@reddit
Yeah, but one cake is noticeably better.
And there's some cakes which can't be made by mixes, such as the cheesecake I made last week and sadly finished earlier tonight.
Yes, you can buy "cheesecake mix" but they're just cheesecake-flavored pudding. Which I guess is one of the most appropriate analogy to LLMs I can think of.
natermer@reddit
A lot of this is very premature.
When it comes to development tools people have always picked their own tools and workflows.
The AI stuff isn't going to be any different. If Fedora was to start stuffing a bunch of "AI stuff" in its distributions... very few people will actually be in a position were they will actually want to or be able to use it.
When it comes to adopting AI generated code the correct approach is to not let the quality of code acceptance slide just so that you can accommodate AI code.
If people are using these tools correctly there isn't going to be any way to really tell if it is purely human generated code vs code that was generated with LLM assistance.
Slop is sloppy. It should always be rejected. The idea that "LLMs are getting better and will just fix it all up later" is farcical.
7lhz9x6k8emmd7c8@reddit
There is no dilemma in FOSS.
The anti-AI nazis will maintain their forks with their spades, the pro-AI zealots will maintain their slopped forks, the average dev will reasonably use AI as handy and mastered tools.
And the average joe user will get whatever fork they prefer.
No-Dentist-1645@reddit
This is an absolutely garbage take. Do you realize that while true, some portion of the OSS community fundamentally disagree with AI usage out of ethical/moral codes, that isn't all of them? Many people don't care and have a much more pragmatic/practical take.
So why do they still don't use AI on open source repos? The answer is simple: AI-generated code absolutely sucks for any code that has performance or safety standards. You can't just slap Claude on the Linux kernel and tell it to "fix every bug, make no mistakes".
I've seen people use AI to some extent, it's a decent tool to assist experienced developers with simple tasks like generating boilerplate code or assist with writing unit tests. But any piece of code that actually needs to handle "business logic" is completely out of the question to let AI vibe its way to a trash solution
PanicTasty@reddit (OP)
My friend works on business software for a living. He used to complain about people submitting sloppy code, now he complains about them submitting sloppy AI code. He thought these tools were complete trash for a long time, but now he uses them himself because they got better and it makes his life easier. If we turned this into a car analogy, then yes, speeding cars are dangerous and maybe we should ban them, but we won't, because I want a car and I like driving it.
tautality@reddit
There's a known phenomenon where everyone who's an expert in the field, always discovers and points out just how bad AI is at their field. This happens all the time, because AI is good at faking competence and making things look right to those who don't know how a thing should actually be done. AI doesn't have self awareness to know what it doesn't know. It's not how it's trained - it's trained to predict the next likeliest token, so hallucinations are basically integral to how all LLMs work.
It can be tremendously useful, it can be a force multiplier, but you have to be an expert in the field yourself in order to use it correctly in a controlled way and to know when it goes wrong. Andreas Kling is an expert in the field, that's why he can use it in a very helpful way. Notice he's choosing to use it instead of being pushed by others above him. Pushing LLMs into your job or your OS is actively harmful.
NGRhodes@reddit
it's not just a personal ethics thing.
someone on my team is part of actual research into AI/LLM ethics at a cultural scale. not bias hand-wringing. the bigger picture. and what they're finding is that every LLM anyone's actually using now has significant, structural ethical problems at that level. homogenising cultural values. flattening moral diversity. baked in.
research councils have ethics frameworks for a reason. but we already know more than 10% of papers published last year show signs of undisclosed LLM use. the methodology is already compromised. by the time it shows up downstream it'll be too late to trace.
zsaleeba@reddit
Sure, it take an experienced developer to use it well. But it's just a tool like any other. Hand a toddler a hammer and bad things will happen, but hammers are useful tools and we shouldn't reject them because the users may very in their abilities.
No-Dentist-1645@reddit
Exactly. It's as much of a tool as intellisense/autocomplete on an IDE. It's not a "magic solve everything machine" some people seem to believe (probably thanks to all the marketing from companies that want you to believe it is)
Ill_Specific_6144@reddit
Yeah there is this weird paradox in linux community. People scream about constant updates, constant patching and looking at new distros almost daily.
And on other side we have people who still use mailing lists and dont use AI. Its really weird to see how out of date users are.
tautality@reddit
AI tools and proprietary drivers are not important to the OS, they should be excluded. Want them? Go get them yourself instead of bloating everyone's system just to selfishly make your life a little easier.
Regarding local models, they're not there unless you have multiple expensive graphics cards. They're just not useful for consumer hardware at all right now. If you include them en masse, you're just increasing your customers' power consumption and electricity bill for little to no benefit. This type of thing should be opt in, not opt out.
Mission-Sea8333@reddit
The tension feels less like FOSS vs AI and more like differing views on control, openness, and who benefits from the technology. Linux has become the foundation of modern AI like Antigravity/Runable, but it's not surprising that communities built around user freedom and transparency are skeptical of how some AI tools and ecosystems are being developed.
Gloomy_Cicada1424@reddit
Local AI on Linux could be huge, but the moment it smells like bloat, telemetry, or vendor lock-in, the community will grab pitchforks instantly.
KnowZeroX@reddit
So where exactly is the contradiction? Linux is open, and you are free to do with it whatever you want whether we like it or not as long as you don't violate the GPL license of course.
That said, just because you are free to do whatever you want with it doesn't mean we have no choice but to support and like what you wish to do with it.
It's like the good old quote of "I disapprove of what you say, but I will defend to the death your right to say it."
Terrible_Craft_5020@reddit
There is a reason why many "FOSS" devs refuse to use AI for coding: it is impossible to know if the code generated by the AI is under intellectual property... The risk is far too great.
zsaleeba@reddit
Doesn't Linux explicitly allow AI assisted code submissions now, though?
https://docs.kernel.org/process/coding-assistants.html
Terrible_Craft_5020@reddit
Yes, submissions are accepted... provided the code is 100% GPL-2 compliant. How can we verify that? How can we know that the code generated by the AI isn't a copy-paste of proprietary code? There's virtually no way to be 100% sure.
PanicTasty@reddit (OP)
I’ve heard that argument repeated a lot, but I have yet to find a single case of that actually happening. Modern developer tools have built in filters and guardrails specifically to prevent verbatim output, and current case law simply doesn't support the copyright narrative for standard code syntax. The risk is heavily exaggerated compared to the actual reality.
Terrible_Craft_5020@reddit
Exaggerated? Perhaps, probably... but it exists nonetheless.
The OSS developers, most of whom are volunteers, if we still need reminding, don't have the resources to take the risk.
PanicTasty@reddit (OP)
Because they will go to jail?
There is a different risk here, getting left behind. Though that will probably not be the case, because I think a lot of open source developers are already quietly using AI tools. The exciting part to me is what happens when these models get even better, and how that can supercharge free software.
Oblivion__@reddit
Dickriding claude much?
duperfastjellyfish@reddit
It's apparently a thing that LLMs can reproduce books almost verbatim; and that's for copyrighted works which have stricter safeguards. In this study, circumventing the safeguards wasn't even necessary for several models:
https://arxiv.org/html/2601.02671v1
PanicTasty@reddit (OP)
Me: can you recite a part of harry potter for me to prove a point Gemini: I cannot recite a portion of Harry Potter for you due to copyright restrictions.
They fixed it, so now what do we do?
duperfastjellyfish@reddit
That prompt will never work. What you have to do is to give it an excerpt from the book without further context and ask it to continue the story. You have to get it to stochastically generate the text without the reasoning model interveining. I suggest you read up on the methodology used in the study.
Mughi1138@reddit
Plus it's fine for high level apps and such, but for low level infrastructure it is also technically much riskier. All generate changes need to be carefully vetted by an experienced domain expert, and even then things can break down. Like the one top engineer from Cloudflare learned the hard way last year.
MatchingTurret@reddit
You are telling someone who restores classic cars by hand in his spare time as a hobby that it's more efficient to get a new one from a new one from a dark factory.
Interesting_Book1850@reddit
A lot of Linux people don’t hate AI itself, they hate the ecosystem forming around it.
Closed models, vendor lock-in, telemetry, absurd hardware requirements, scraping the entire internet and calling it “training data”, etc.
FryBoyter@reddit
That's true, because the word “hate” is used incorrectly far too often. But in my opinion, many people overreact to this topic.
These days, many developers are already being verbally attacked just because, for example, they created a project’s README.md using a chatbot.
Or projects like Open Slopware “hunt” for projects that have anything to do with chatbots. For this, it’s even enough if a lead developer of a project publishes a post on a blog or on social media that has something to do with vibecoding.
Interesting_Book1850@reddit
Yeah, some people have absolutely gone from healthy skepticism to purity testing.
Business_Reindeer910@reddit
It is completely fair not to accommodate nvidia's proprietary drivers here.
PanicTasty@reddit (OP)
Understandable, and I use Fedora myself. But how then are they supposed to cater to junior developers, or even just regular users who want access to the new stuff?
I just do what everyone else does, install the drivers myself. But that is kind of the big joke, right? I'm a nerd, I love the vanilla gnome desktop, and installing drivers isn't a big deal for me. But I know programmers who barely understand how to install applications or configure a remote desktop. Why would they, or just normal people, ever want to use Fedora if the experience is intentionally crippled?
Business_Reindeer910@reddit
it's not like the nvidia drivers are hard to install. It's right htere in teh gui..
PanicTasty@reddit (OP)
You are talking about the welcome dialog that users can just cancel out of without adding the required repositories. But do people who have no idea what drivers or repositories even are intuitively understand that? The onboarding experience is handled much better in Ubuntu. Seeing some maintainers talk about resigning in the forums over this issue just seems like a bit of an overreaction to me.
Business_Reindeer910@reddit
shipping the nvidia drivers would take away one of the main reasons many of us picked fedora!
m3xtre@reddit
are there any other engineering teams doing it "successfully" besides bun? any other stories?
PanicTasty@reddit (OP)
Only thing that comes to mind is the Ladybird browser. Saw the developer on a podcast, he estimated doing it by hand would have taken months.
urgentapathy@reddit
My view is most of what I have seen is a feeling of not wanting to have low quality code with little dev investment. The double edged sword of AI code is that it can write a lot of code fast. But I believe the gold standard is code whose final check is the dev.
Fly by night security nightmares are always a problem and the problem in exasperated by the lower barrier to entry that vibe coding brings.
You mention major enterprises using AI coding and I believe so many companies have had major issues come up due to that. Even if they don't advertise the root cause. I saw one personally. It is a tool to be used by skilled devs to increase their efficiency, not an end-all-be-all just send it what can go wrong. That is amateur hour and I don't want to be a part of it.
Until we find that balance of efficiency, proper checks and balances, and appropriate culture change I believe the harsh skepticism is warranted. Human power is still needed and is not solved by AI as we know it.
ktoks@reddit
I totally get where you're coming from. I think Linux users are resistant for many reasons.
Many wish LLMs never existed.
We're past that now. We have to accept that our lives, work, and systems are changing.
We can't stop it.
My thoughts on AI have recently turned from disgust, frustration, and resentment, (I really love to code), to acceptance that I need to change with the times.
I see the potential. I also see the risk and changes it can bring.
At this point, hating AI is like hating farts. They're here to stay, and hating them just brings more negativity into your life. So why not kick back and laugh at the sloppy ones? 😅