This is a great question I’ve been meaning to ask myself. Thanks for asking it. Having been vehemently against, I’m a recent convert. For me it was a matter of ethics vs convenience and efficiency. There’s a limit to how fast we can type that an llm isn’t bound by. I first tried running ollama locally, but it didn’t give me the power I wanted, so now I’ve been using Lumo and ollama cloud models for coding. I’ve been impressed with the results and it’s within my ethical standards. (As someone who isn’t a programmer professionally my concerns were primarily based on capitalism. Ollama is “better” but complicated)

I’ve found two different modalities for the way I work both with different results. It comes down to how granular I can make my queries. When I’m writing something I’m more familiar with, like bash, I typically ask for things I already know how to do, but it would be slower to type. Then there’s other cases like writing elisp for emacs. Completely different. Those cases I admittedly vibe code with about the results you’d expect. It’s really given me something to think about.

i don’t trust the quality/sustainability of fully vibe coded software (ephemeral scripts/tools excluded)

but i think that AI-assisted coding (ex: ai agent wrote a patch) is good with a few conditions: as long as humans are doing most of the work, designing the overall archtecture of the project, and the AI code is reviewed

AI completions are fine too, to me, it’s literally just “smarter autocomplete” on what i was going to write anyways

this is all with the assumption that we are using local/FOSS models. chatgpt, claude, grok, gemini, and other closed LLMs can get in the bin with the other proprietary software 🗑️

Personally, I think that using LLMs is a terrible idea, especially for Free Software. It creates a reliance on proprietary software^[1]. Generated code has uncertain provenance, for the same reason that researchers can provoke LLMs to “auto-complete” famous books, undermining the license. LLMs rewrite instead of fixing, since the latter requires intelligence, so it makes less-maintainable code. And studies show that using models slows work down, rather than speeding it up.

I’ve only found two arguments in favor: The technology is (allegedly) here to stay. And sure, the studies say that it slows down work, creates bugs in most of its changes, makes everything harder to maintain, alienates the team, and erodes skill, but don’t worry, because I have a system.

I wouldn’t risk it on any project of mine, and especially not for anything released under a public license where it needs to promise people that they can use the code for any purpose.

My experience with ethically run (if unethically created) local models has been similar. Lots odd hardware to get slow mediocre answers.

But the biggest reason AI competes with plain old web search is that they’ve enshittified the web search available to us.

Corporate assholes are breaking search to force people to use their centralized harmful AI answer guessers.

We can fight back with Searx. https://wiki.debian.org/FreedomBox/Manual/Searx

The problem is that FOSS projects that are popular already have a a high quality standard because they are developed by very smart people who’s been developing for a long time.

There is inherently a problem with current AI - It doesn’t learn.

When a new developer joins a FOSS project they first spend weeks or even months learning this project’s specific coding practices, standards, and reasoning. I’ve been programming for 20 years and have a PhD. I can’t “just start contributing” to something like KDE because I’ve never looked at its codebase. I don’t know assumptions, methodology, or challenges that are baked into the codebase. I need to spend time to learn those to be a useful contributor.

AI can’t learn unless you do gradient descend. The problem that you mention about sounding elitist is that those FOSS developers spend a good amount of their time training new contributors and everyone is proud of their work. Now you have people who are not ready to be contributors tell AI to contribute and it’s doing its best but it just can’t learn and even having a 200K token context is not enough to fit all those project specific assumptions and rhetorics in.

I use AI at work, but not at home.

There was an interesting post here which reasons through when it’s ethical to use proprietary software to further FOSS. I generally agree with the argument, but AI doesn’t meet that threshold for me.

I think our goal should be for everyone to be able to use FOSS for everything. That means, if we use proprietary tools, we have to be able to replace them eventually. Otherwise, we’re just dependent.

But that doesn’t seem possible with AI yet. If there was a project making a truly open AI (open weights, open, consensually retrieved and properly licensed training data, open harnesses, open training infra etc.), and using existing AIs somehow helped with that, I’d be all for it.

Another aspect is how it shifts the way things are made. Projects are built on collaboration and shared understanding, but AI is such a different way of working - so what does it mean for some people in a project to use it and others not to? Why would I spend a week working on a very tight, clean patch when someone else can re-implement literally the whole project in an hour? And why would they contribute to this shared project at all, when they can build something tailored exactly to their needs by themselves?

It makes me wonder if FOSS as we know it will become less important, but open standards will become more important. Since you can make software tailored exactly to your needs, clients will proliferate. But when you have to integrate with others, you still have to negotiate and come to an agreement on how to connect together, so that work will be even more valuable.

To save time, yes. To save on your skill, no.

Or in other words: only those, who are already masters, should use it.

I use massive amounts of AI for home lab dev and operations. It’s an incredible tech that unlocks so much for homelab devs. I never would have the time or patience to build out an entire system like I have now.

It has allowed me to de-google my life completely now. I run SSO with a mail server, file server, gitea, arr stack, immich for photos, and dozens more things. The whole stack monitors and acts on alerts by itself. Quite wild.

I use a couple DGX Spark boxes for inference. Currently running Deepseek v4 Flash 0731.

I have zero reliance on big tech now (aside from the annoying Google sso; that will take time to migrate away from).

So yes, I use AI just a tad. It would be foolish not to.

Some background: I’ve been a professional software dev/arch for 35 years.

LLMs can quickly produce outputs that can fool people into thinking a human wrote it, but they cannot write “good” code for the same reason they cannot write good prose, and that’s because they have no understanding of what the code means. The only thing they are rewarded with in training is “are humans able to tell that this thing wasn’t made by a human?” and if the answer is “no” it gets it’s digital training cookie. That’s all LLMs “know”.

So getting back to your question, there is no reason LLMs can’t be used to assist a programmer, or even one day quickly create bespoke software for single use/short term purposes. But a project being FOSS kind of definitionally implies collaboration to ensure reliability which is very difficult when LLMs are used, since again, they don’t have an understanding of what specific code actually means.

IMO anyone ascribing ethics or morals to LLMs (not its use but the technology itself), even in a negative light is grossly exaggerating their capabilities, often unintentionally. LLMs do not have will, they cannot develop a will. They are not curious, they cannot become curious, they cannot understand the world. They do not understand English or Javascript.

I’m not looking for an elitist lecture. I genuinely want to know the community’s thoughts.

What if the community’s thoughts are an elitist lecture? I don’t know what qualifies as that, but it makes the question a bit loaded. I’m not sure it’s the best way to get a balanced set of answers.

I don’t see anything wrong with using AI for almost anything at all (including this), as long as there is human review every time. I do not see how the free software philosophy, by itself, either endorses or condemns AI.

Personally, the local model seems to be the future for AI within FOSS, and I think it may be refined to have lower usage requirements for a specific capability to run on smaller machines, not sure how much better general purpose models will get but after trying them, I expect modest progress over the next few years.

I don’t trust the large frontier models and the push to build data centers, that’s a grift at the expense of spending that could be useful elsewhere.

I’m lucky enough to have bought a 3090 2nd hand when they weren’t insanely expensive in 2023. I’m trying “vibecoding” on my personal half-finished coding projects. My coding style is poor by most professional standards, but the edits I have had my local model (Qwen 3.8 27B UD Q4) and harness (llama.cpp and Opencode) spit out are reviewable to me, make logical sense, follows my existing conventions and does run correctly. I think it is by default on highest quality, which means writing one function takes 15m to an hour, but I’d rather it be slow and hallucinate in my code less. There are other optimizations I could make to have it run faster, but it’s not a big deal to me. I can code the functions myself in less time, but what I think it is best at is catching bugs and my mistakes.

At this point code written in part with AI is in almost everything. Morals and ethics went out the window when convenience came knocking and everything since has been justifying hypocrisy.

It is very expensive to self host your AI/LLM and have it function locally. You sure can self host it and get it to work correctly but it too slow or too inaccurate for anything useful. Maybe some new model can be made to run on comoddity hardware or that new cheap hardware can be made to run AI decently, but as of now, on all the equipment I have, I cannot get it to a usable state. You can watch videos on internet on what it takes to get it to run and still have mediocre results.

You should ask this question in !localllama@sh.itjust.works. There’s a lot more knowledgeable people there. It really isn’t as expensive as it seems to get good results locally, if you know where to look.