(Sorry to bring up a contentious topic) KDE & AI/LLM policy

Development
126

The models already exist, so the harm done during their creation cannot be undone.

Except, they’re also still being developed, retrained, tested and released. The harm is still continuing.

You do not have to pay the AI companies. There are some models that can be run locally and it will become easier/more efficient as the hardware advances.

The majority of people will be using the main LLM models that require payment either to use or to remove some limits (ChatGPT, CoPilot, Claude).
Sure, people could run a local one, but this still ignores the fact that these models still need to be trained, which is incredibly intensive and still results in increased usage of non-renewable fuels.

Plus, I wouldn’t be surprised if some of those local model makers engage in the aggressive, Denial-of-Service-like behaviour that the larger ones love to partake in.

127

Don’t get me wrong, folks, I’m not a big AI booster. I find the “de-skilling” and “increasing velocity faster than humans are capable of understanding” consequences to be particularly relevant to KDE. I believe that any usage of these tools must avoid those consequences, as they will be ruinous over the long term to careers, institutions, and products.

I think the “generative” forms of AI are especially risky on these dimensions, especially for novices, students, and junior engineers. I find all generative AI “art” to be awful and off-putting; just terrible stuff.

I’m also upset about resource usage. While yes, it has gotten more efficient over time, Jevons’ paradox means the efficiency gains will be gobbled up by increased usage. And the world does not have nearly enough carbon-free electricity generation sources coming online to render this irrelevant.

And in principle, I feel very sad about the mass theft of creators’ work, though it conjures mixed feelings because I’ve never been much of a believer in copyright and patents to begin with, and I don’t think these legal tools were ever effective in protecting creators. But clearly this level of harm is beyond the pale. And even more depressing is that it’s considered a done deal! No court on earth is going to declare it all illegal and shut down (ostensibly) the biggest drivers of their economies’ GDP growth, their companies’ productivity, and their militaries’ lethality. Depressing? Yes. But true? You know it is. Even the fig leaf of copyright protecting “the little guy” is a dead man walking.

I could go on. So yes, I hate it too!

But the thing is, it’s easy to take a performative stand against all of this online. It’s harder to actually take actions that navigate a flawed and unstable world in a way that balances your moral compass against your relevance and effectiveness.

I don’t want to be a dinosaur in my 50s, pushed out of my career path and forced to start over from scratch in a different field a decade and a half away from retirement. And I don’t want for my actions to produce this outcome for any of my employees.

Now, I don’t blame anyone who makes the choice to drop out. The tech industry is certainly a depressing place right now. But I’m not ready to; I still think KDE can be a light in the darkness, and I want to be a part of helping it continue to shine.

And on the subject of KDE, I don’t want it to fall behind. Like it or not, a single highly-skilled engineer could probably take a few months to supervise Claude or Codex to build a better version of most KDE software that took 20 years to reach its current state. And if they do? En masse? Then KDE itself falls into irrelevance not because it’s bad, but because it’s been outcompeted by newcomers — even FOSS newcomers less allergic to AI. Their output might not be sustainable over the long-term, but no users care about that, and the damage will be done. Everyone is being habituated to throwaway content these days. Perhaps even including throwaway software. “Can’t maintain it? ehh, just vibe-code a new one in a few years, when presumably the models will be even better!”

This is very depressing, but to me, the only nightmare worse than KDE accepting a bunch of unmaintainable vibe-coded junk, is that we get outcompeted by same. I think I would cry, and I think it’s a real risk.

This is why I don’t think it’s wise for us to blanket-ban all or most AI-related contributions. I would like to see KDE use some of these tools strategically and defensively, where they make sense, to prevent ourselves from losing competitiveness when we’re right on the cusp of bursting into the mainstream.

In particular, I think AI code review could make sense for KDE. We’re always critically low on code review resources; nobody enjoys doing it. AI can be quite good at pointing out obvious mistakes and small but subtle bugs. And in the cases where it makes obviously stupid suggestions, we can all just laugh and click the “Resolve” button on the thread. We can also, by reading the good-quality suggestions, become better engineers ourselves. And those skills will remain with us when the AI bubble bursts and tokens become 5x more expensive and the “free for FOSS projects” deals become “talk to sales about a competitive package”.

AI-assisted bug triage is another potential field that might be able to benefit. Nobody enjoys doing this, and I think AI could at least take some of the edge off by fixing wrong versions, re-categorizing mis-filed tickets, asking reporters to clean up junk reports and attach backtraces, and so on. Nobody actually likes this part of bug triage (to the extent anyone likes any of it at all).

Would this entail a certain amount of moral compromise on various fronts? Yes. But let’s get real: everything in life is a moral compromise, and you’re fooling yourself if you think you didn’t already make a bunch of them.

Did you fly on any planes in the last few years? You morally compromised your environmental ideals. Do you drive a gas-powered car, or ride on any gas-powered public transit? You did it again. Do you own a home with a south-facing roof and haven’t put enough solar PV on it to make it net-zero? Ditto. Do you have any appliances powered by natural gas? Do you still eat meat? Do you ever throw away food that rotted because you misjudged how much you needed during the last grocery store run? Do you buy food in that grocery store that’s wrapped in plastic? Is there any polyester or nylon in the clothes you wear? Do you shower every day? Do you have an Amazon Prime subscription and order new products for two-day delivery? Do you use ad-blocking browser extensions? Do you listen to music on Spotify or YouTube instead of buying albums on Bandcamp? Do you watch clips of shows on YouTube instead of subscribing to their streaming services? Do you pirate content? Do you check books out of the library for free or buy them used instead of buying them new? Do you buy computers and computer components new? Do you buy junk plastic-and-particleboard Ikea furniture instead of real stuff made of wood? Do you repair everything you own that breaks instead of throwing it out and buying a new one? Do you always buy local, instead of from China? Do you really buy only things you actually need? Do you pay people to do things you could learn and do yourself? What percentage of your income to you donate to charity, anyway? Do you take every opportunity to protest injustice, or do you sometimes get tired and play video games instead? And speaking of injustice, how much of your tax revenue funds harmful things around the world? You’re culpable, you know!

We can play this game forever. I don’t find that it’s very useful to excommunicate one another over abstract moral compromises involving participation in world-spanning systems with negative externalities.

We’re all guilty in one way or another, having already made the compromises we were comfortable with. To me it feels hypocritical to say “well my system-participation-related moral compromises were minor and necessary; yours are huge and unacceptable!”

If we’re natural allies (e.g. FOSS people, KDE people), I propose that we give one another a break and treat one other as the allies we are rather than firing up the inquisition. Life is hard, and morally righteous people are already on the side of “help, don’t hurt” even if they can’t be 100% perfect. We’re all doing the best we can. I trust my fellow KDE people to at least try their best to navigate this crazy messed-up world and wrestle with how much or how little AI they want to allow into their lives and professional workflows. That’s why I don’t favor a one-size-fits-all sledgehammer policy on the topic.

128

While I understand why you would have that position, I just don’t think that policy works. (Since I can see from other projects that somebody will always see that as an invitation to heavily use AI, and then you got the whole plagiarism problem on your hands. And I don’t believe, like some others here, that AI autocomplete alone is likely safe to use.)

It would be nice if there were a happy middle for AI. But as it stands, a non-ban is effectively seen as an endorsement by some.

129

I feel like you’ve missed the point, which is that we all (including you) make moral compromises to live in a complex and unsatisfying world every day, so focusing unduly on a specific one is hypocritical.

Does this mean we should give up and abandon morality? No, I think it means we should accept messiness and imperfection, and grant others the same sympathy and leeway for moral failings related to participation in systems with negative externalities that we grant to ourselves.

A lot of us here are neurodivergent, which can often manifest as black-and-white thinking and hyper-fixation on the morality of choices. Being a moral person is good! But moralizing everything and fixating on other people’s moral failings (like using AI) while ignoring your own (like participating normally in modern society in any of the ways I mentioned in my prior post) is a failure mode.

130

Honestly, this is a pretty good and nuanced response. Sorry if my prior message came off as too hot-headed, just a topic I’m passionate on (sick of the world blowing itself up at every corner, ya know?) and I’m not very good at socialising. And I do agree with your point that we do need to at least compromise at some point, given perfectionism will lead us nowhere.

And on the subject of KDE, I don’t want it to fall behind. Like it or not, a single highly-skilled engineer could probably take a few months to supervise Claude or Codex to build a better version of most KDE software that took 20 years to reach its current state.

This is probably the only part I really disagree with.

Sure, someone could use AI to program an alternative super quickly.
But KDE has a big community that is actively filing bug reports, making suggestions, fixing security holes and messing with it. KDE is (at least in my opinion) battle-tested.

An AI-made alternative won’t have this down for many years to come, given they’d have to convince the average person that it is not only worth it but to also engage in similar behaviour with their project. It will not be a stable option that the average person would be willing to consider. Enthusiasts, maybe?
I think KDE’s position is safe, even with AI looming over the internet.

I’d personally like, at the very least, for KDE to ban LLM-generated code (as I feel that a human solution will allow for more introspection), documentation (because I don’t think AI can ever really understand how to use a program or describe certain features) and media (audio, images, translations, e.t.c.). Maybe some of Herzenschein’s other suggestions too, given that I feel like their suggestions (mainly the “default” suggestions) would be beneficial.

As one can probably surmise, I’m not too fond of AI code review, bug triage and suggestions. However, for the sake of productive discussion and policy making, I think that I (and hopefully many others) will be willing to concede on this (ideally if KDE uses and/or recommends some kind of local model (maybe even an open-source one for the matter), mainly to use something with… less harm caused during its development while also being confirmed to not contain things like proprietary leaked code in their training set. I’m sure there’s at least a few models that would meet such a requirement?).

Plus, this policy can always be re-reviewed later, provided more ethical models come out or if regulations are put in place to force the big players to stop being so terrible (incredibly unlikely but a girl can hope).

131

I personally don’t think worrying about KDE’s code base running into problems and the entire movement causing issues for FOSS license adherence if plagiarism is embraced, is “unduly”.

(For clarity, let me point out people bring forward different concerns here. I don’t want to dismiss the ethics. However, it’s not just ethics.)

132

And let me add I generally agree with that.

However, that a vague AI policy will be interpreted as going full-in on AI by some contributors is a problem. A focus on pragmatism is great, but I think an LLM code ban is actually in many ways a pragmatic and not very radical approach.

(Since a ban can be easily reversed, is simpler and easier to understand than any vague guideline, and doesn’t need to extend to research and passive analysis use.)

Anyway, sorry, I don’t mean to say KDE needs to do anything here, since I’m not in the position to do so. I’m merely trying to explain why I don’t like the framing of unduliness or extreme morality.

133

Nate, how did you manage to write all these "do you"s without repeating yourself, if not using an LLM? Really amazed.

(Yes, I used an LLM to double-check. It says:

Strictly following your request: I won’t look at underlying meaning or logic at all — I’ll only check whether any sentences are the same specific question phrased differently.
After going through the whole text, none of the sentences ask the exact same specific question in different wording.

)

134

No LLM here, I refuse to use it for writing! The morality of everyday compromises is a topic I care a lot about, so I carry all that stuff in my head, ready to dump onto a page. :slight_smile:

135

As always, pick your battles. For me not using these tools is very easy battle to win, even though I am culpable in other ways. For me its about lessening the issues, because as one nerd I can’t fix everything.

Anyways, I think our current “project maintainers decide” meanwhile not the one i want is still the best.

Also i would never trust myself with these tools (especially by evil tech corps), the potential to get hooked to their gambling like nature would be very detrimental for me.

136

It may be time to open the debate within the community of contributors and then lay down some guidelines that would clarify when, if at all, using AI/LLM is acceptable in KDE.

The current state of uncertainty will just spark more controversy while it persists and allow ignorant/bad actors to take advantage of the grey area that comes with the lack of clear rules.

137

A better one of not-the-smartest things I’ve heard. Even smart and good people have feelings, so don’t take it as an insult but a sincerely friendly remark, please.

And a personal remark on “artificial intelligence”: It’s extremely powerful on every axis of power relevant to humans tbh, the real pace of development and the cutting edge thereof unknown to the public, and the reach thereof inescapable for civilians, sadly. It’s like highly distributed Los Alamos without imaginary corrupt people, as in Dawkins’ memes. Avoiding it may be a way to protect yourself (in the short term), but it isn’t a way to protect “the little guy”; utilizing it for that purpose, and making “the little guy” understand it better is. The world has always been full of agents, whose power levels range from “extreme” to “moderate”. The weaker ones aren’t really called agents, and the extremely powerful ones will find various uses for tautologies and use one to be extremely powerful. Some of the big corporations are so untouchable right now that they can publicly announce they would legally protect their consumers for capitalizing on the genuine sloppiness they put in the market. And, about the idea that a project wouldn’t take a net negative hit (in the medium to long term) regardless of the route taken in the face of advancements in “technology”, I’d say it’s normally better to work hard on something without putting too much trust innit, even if it is as advanced as Plasma. The faster AI agents can work, the more non-profits should invest in making them as usable by the public as possible, utilize them to debug their projects, and try to make them explain themselves adequately in the process.

Legit concern imo :money_mouth_face:

138

No large field study with good data so far suggests a long term significant productivity benefit. Seems fine to avoid it from that angle.

I personally suspect that often perceived boosts in productivity either come with long term costs, and/or may be a misinterpretation of non-AI coding skill erosion.

139

Frankly, even at its current speed of critical emergence, it doesn’t look like its long-term productivity would take long enough for any large field studies to be conducted with good data. It’s called disruption (or disturbance). Things are always fine until they are not from an angle, especially in history. One exploitable weakness of agents could be observed when they get together and try to agree on how likely an imminent conflict is, regardless of whether the disagreement would be real or not from an oberserver’s perspective. And that’s all the more reason for an observer to not be able to preceive the scale of a real quantity, such as productivity.

140

Might be kind of relevant, especially to those who think AI coding is “inevitable” or the productivity boost would be obvious.

Edit: if you get a paywall, try this link.

141

@ell1e: I want to thank you. Not just for this thread, but in general. Randomly I see posts of you on forums, github, mastodon etc and it is always a pleasure to read these. I am using different nicknames and avatars everywhere, so you will probably not recognize me. But I see you are always on the right side and do also support things for other people beside you. I really like that.

We should split up between machine learning and neuronal networks versus large language models and other huge models as well as ethical vs unethical models.

But before I start this, I want to tell something most people do not see: these “tools” are just save-files, saving a huge amount of data in a very compressed and lossy way. I call them ALA → Algorithmic Lossy Archive. You could think about a MP3 where data will cut off to saving a lot of storage by losing just a bit of information (quality). Just that the storage itself is very algorithmic (the “neuronal network”). And if you save a MP3 of a MP3 of a MP3 you lose more and more information, but it is still the same data. It is pretty similar here. Every information that was used for “training” is part of the multi-gigabyte sized file. Some become damaged, some others may completely overwritten, but we have no control about that process. The prompt is like a timestamp to read that file in a deterministic way (seeds deterministic randomizing the results). But unlike a timestamp, prompts are timestamps of infinity ways to read the archive (model). Even if we don’t find an information again, it could still be part of that archive. We just cannot decide if something was fully overwritten or or partly or is still fully available. We have to belief it is still part of the model. There is no other way to handle it without lying.

So everything that is used to save into these models should be treated as part of the model, no matter how healthy the data is. Therefor it is necessary that only data is saved that was allowed to be used. It also includes respecting licenses as CC-BY-SA (recontribute under CC-BY-SA and naming authors).

What is the difference between LLMs and many old neuronal networks? The amount of data that is required. Old models could be created by paying contributors as voice actors or just creating data by their own, because not much is required. They never required to multiply all available data-centers by 10+ times to be used as intended. They never spammed the internet with slop. In 2026 when I search for rare information, it is 100 times more hard to find any on common internet search, because the first 1000 pages will be off-topic or slop-pages. So I have to use LLM just to get an information I could find previously without. LLMs are solving problems LLMs are introducing and everything adds to the power consumption. When in previous neuronal network history was this the case? A weather-forecast neuronal network is also a huge machinery. But not every end-user has to turn on the model. Scientists also had to ask for computing time and share limited resources while the industry forces it everywhere and claiming all limited resources for themselves (RAM, power etc).

You see there is a fundamental difference in what we had before compared to the slop-production-industry right now. I am still totally fine with all little ethical models and would use them myself if they are labeled free software (they have to be fully transparent just as any other FOSS project, including machine readable contributor lists of the saved data). I am even fine with power wasting super large models if they are used for science (and data was not stolen), because a limited amount of super computer do not destroy our planet.

But even in past I was not fine with all neuronal networks as call-center (support) that was using customers voice to “train” their models to make support-hotlines worse. Long time before the so called “AI” hype-train arrived I was criticizing this kind of abuse. And back then I was not so deeply involved in the FOSS world as I am today.

I hope my post makes it clear where the red line should be drawn and why it makes sense to do so. It is not that any kind of neuronal network is something bad. I agree with you, a lot of models are just fine. I never would cry about Wendelstein-7X (fusion reactor) is using a model to stabilize plasma. That is the perfect use-case and I would not even consider another tool instead. But it is just not comparable to LLMs in any form.

Edit:
One last additional thing. The environmental impact is an issue of our time. Once we can produce endless amount of green energy and the calculation time of a whole mega data center can be done on a single server-rack, I would not complain any longer about environmental impacts. But we live in times where climate change is accelerating. I am dead when humanity and all other life suffers +5K, but I am also in response (as everyone else alive these days) of that future.

142

I would like to say that apart from all the countless other issues with AI that have been raised in this thread, stealing every work known to man, energy consumption, producing garbage output, de-skilling, I think its relevant to mention that as time goes on these major models will be too expensive for most people to even consider.

AI is very expensive. It’s very expensive to build the data centers, bribe the city councils to let these projects get started, buying the land, construction costs, buying all the compute hardware, the cost of water and electricity, maintenance costs, constantly re-buying the compute hardware every year or so when the next gen comes out, and so on. All of this means the actual cost to the end user are heavily subsidized by venture capital and other outside investors, who have a vested interest in this stuff taking off like wild fire so they can have any hope of making a return on their investments.

To quote Ed Zitron quoting the wall street journal,

"While Copilot might not be “…the same product it was a year ago,” very little has changed about the underlying economic mismatch: that Microslop was allowing users to burn more than their subscription costs in tokens every single month for three years. Per the Wall Street Journal in October 2023:

'Individuals pay $10 a month for the AI assistant. In the first few months of this year, the company was losing on average more than $20 a month per user, according to a person familiar with the figures, who said some users were costing the company as much as $80 a month.'"

Basically these companies are in the early stages of ensh**fication. Keep the prices artificially low to attract users, and brag about the number of users to attract business partners. Then once they’ve cornered the market, they crank the prices back up again.

So basically, once all these AI companies swap over to token based billing from a flat subscription model, the models will be come so expensive to use that most people won’t be able to use them as they are currently. And if they are never actually learned to code in the first place, then they will be out of luck when this whole thing comes crashing down.

You’re not missing out by choosing to block out slop coded nonsense, you’re not going to fall behind the curve. Having people who actually know what they are doing is our competitive advantage.

143

And to sort that “$80” in a relation: Nowadays we have tools running on smartphones that are creating profiles, using them for prompting cloud services (really huge amount of data that will be send) to get nicer answers and so on. A non English journalist was testing it and generated costs in higher 3 digit numbers per day. And since LLM-companies trying to make tools more useful, they also develop larger models requiring more energy, leading to even higher costs.

So the average of $20 these days may become the average $50 tomorrow without increasing the usage, just using more reliable tools. Service fees come on top. Companies do not want to get just a +/-0 financial result. They want to get the monopoly to milk people. The winner takes all.

144

A very interesting read from a Tor Project engineer trying to fight LLM agents DDOSing their infrastructure (although it’s mostly about LLMs enshittifying everything else): The Four Horsemen of the LLM Apocalypse - anarcat

We should not embrace LLM tool use. We should fight it.

145

@ngraham, how might one report an obvious violation of this, inadvertently or not?