Don't misinterpret this link as representing a final decision. It's actually three separate proposals which will be debated and then voted on.
Proposal A is "expressly forbid any contributions to Debian written with the use or assistance of large language models (LLMs) or other generative AI tools."
Proposal B is "The Debian project allows AI-assisted contributions (partially or fully generated by an LLM), provided the following conditions are met [...]"
Proposal C is "request that all contributors to Debian avoid the use of LLMs in their Debian work" without an outright ban.
Of the 3 proposals, the 2nd/B seems to be more of an “informed consent” model while the other two are seeking a comprehensive exclusion. I hope earnest dialogue and driving to a broad consensus among significant contributors is forthcoming.
Notable are the endorsements, which are balanced over all three options, perhaps indicating only 1/3 are supportive of a more permissive position. If this represents core contributors, this could be worrisome: a clear majority favoring a exclusion, but yet a sizable portion of contributors who don’t want a ban. Losing 1/3 of core contributors would be a significant loss to the project.
> Notable are the endorsements, which are balanced over all three options, perhaps indicating only 1/3 are supportive of a more permissive position.
That's not the case. First, you can't really extrapolate from endorsers to all of the Debian Developers. Second, endorsers might not have looked at the other proposals at all, and the fact that they only endorsed one option doesn't mean they don't support the others. And you can also endorse an option if you believe it is a valid, useful proposal to add to the GR, even if you don't agree with it personally.
But most importantly, when Debian votes, you don't vote for a single option, instead you rank the choices, and apart from the given options, another option "further discussion" is always added to the list of choices. It's likely that many who rank the most permissive option first will rank the lesser permissive option second, above "further discussion", or they could even give options the same rank (if they really don't think one is better than the other).
> Notable are the endorsements, which are balanced over all three options...
I don't think this is notable. The process requires five seconders, so you're just seeing the set of people who seconded asynchronously before there were obviously enough that others didn't bother. The culture is to avoid unnecessary noise and leave it for the vote.
I wonder how long software projects will even be able to ban LLM use in an age where exploits are found by LLMs. Like, if you have two forks of Debian and one uses LLMs to fix exploitable bugs and the other one doesn't, the level of security they can offer will be worlds apart. And noone in their right mind would want to use the less secure one. Similar to how noone would want to drive a car that was 100% hand built by humans when we know that machines do a much better job at precision tasks. LLMs are just another tool in the end.
I wonder how they can reconcile the stricter proposals with the LLM usage in kernel development. That seems totally untenable. I mean it all seems untenable, but with the kernel especially.
Also, what about when you inevitably get a situation where a critical vulnerability is discovered, and the only patch available is LLM generated? Do they have to wait to patch until some person who hasn't seen the LLM-generated patch does a clean room implementation?
I understand the objections to LLMs, but rejecting LLM-generated code really doesn't seem realistic.
Kernel development is excluded from this, because it's covered by "Upstream projects using LLMs for development".
On security, I worry that proposal A's "forbid any contributions to Debian written with the use or assistance of large language models", as written, excludes contributions where the LLM assisted in discovering the vulnerability. That's clearly a bad policy, and they should update their wording to clarify that.
The proposals which say no LLM generated code mean just for Debian, if upstream allows it then they will still accept it. It’s just contributions to Debian itself which would prohibit or discourage LLM contributions.
while my heart lean heavily toward A, this would serverly limit Debian contribution and would rely on contributor honesty. It might be dismissed as an attempt to gatekeep contribution. But the issue of the human cost on the receiving end of the contribution and how sophisticated attack vector it could potentially be give me some concerns about long term quality.
B being the more pragmatic, despite the (valid) code licensing concerns. but could leave some very openly LLM driven project room to contribute package.
C as a middle ground would not satisfy any party and would only push back the resolution of the status quo.
I misinterpreted it as a final decision when I first skimmed the page. The title of the linked document is "Resolution: LLM usage in Debian" and it's not obvious that it's several competing proposals until you explore the page in more detail.
The Hacker News submission title was updated to include "Three Proposals" after I posted my comment.
> A LLM (...) merely produces syntactically likely combinations of the training data
While doesn't matter much in the rest of the policy, this is a common misconception among AI skeptics. It is not the case for a long time (since RL is used heavily in the training) and a LLM may go beyond its training data.
I think the questionable term is not “syntactically likely” but “merely”. Syntactical likeness is a vast solution space that encompasses the work of a terrible developer and a genius developer. In fact this solution space is a gap wide enough to encompass all coding knowledge and expertise.
"Combinations of the training data" is also hiding a pretty common and serious misconception about modern generative AI: this seems to come from a mental model that LLMs and image generators are merely producing a collage of sorts, especially that this is some way linked to specific parts of the training data as opposed to a generalized result of building up an understanding of how to produce a sentence, piece of code, or plausible image.
The natural next argument that I see a lot is "well it's still all probabilistic", which is technically true. However, don't the atoms that make up the cells that make up a human move around and interact based on probabilities?
Saying a LLM is all just based on probabilities is pretty meaningless, even if it is true, since everything else is just based on probabilities too. What matters is the massive amount of machinery that generates those probabilities. Anything from rolling a dice to an accurate model of an entire human, or even a model of the entire universe, is all "just probabilities".
That argument is mostly used against the misconception of AI-stans that LLMs are compilers, English is a programming language, and that prompting is a programing. There are many reasons for why that is false, and LLMs being stochastic is one of the more obvious reasons.
But there are more probabilistic elements about LLMs. Inference is by default deterministic for any given input and weights, however it is made stochastic algorithmically, when the model chooses the second or third likeliest outcome with some probability. Training is stochastic, the adjustment of the weights is probabilistic, etc. etc.
You seem to make the fact that our brains are also probabilistic as some sort of a gotcha. However that has never been in dispute. Neurons fire with certain probability, this has been known since the advent of neuroscience. However this is why computers are preferred for certain tasks over human brains. When you run a program you can be certain it behaves in a certain way given a particular input and parameters.
I assume RL stands for reinforcement learning, which just means that the weights are adjusted by evaluating some loss function. If the loss function uses the training data (i.e. supervised learning) then the original assumption still holds. So I fail to see how this is a misconception.
Supervised learning and reinforcement learning are mathematically distinct. Supervised pre-training as you said maximizes the likelihood of a token sequence. RL optimizes a policy against an external reward signal or execution environment. For example, RL evaluates code against runtime execution/unit tests/proof checkers (e.g lean). Models learn new strategies and are able to produce novel code if trained in an RL environment.
Doesn't the LLM still output the next most probable token (modulus heat/desired variation) according to its model? Training shifts the probability - but doesn't change the fact that the output is a sampling based on input and the model?
It’s not sampling anything though. At each inference step the fundamental output of the process can be interpreted as a probability for each possible token how likely it is. Sampling would imply otherwise but the probabilities are the output.
Common sense testing only tells about the presence of common sense, that's it. It is clear from recent events that LLMs do definitely possess some sort of intelligence even though it might have been polarized to certain domains.
Funny because out of all the distros, Gentoo benefits from LLM usage, at least as a power user. I have fixed many-an-issues just having LLMs write patches in a way that portage automatically builds them, or rapidly put together ebuilds. I respect upstream decision though and don't share my changes (and realistically who would want them)
This tracks with what I do too (also a gentoo power user). My personal overlay is much healthier thanks to LLMs, but I don’t use it for writing code for personal use. It’s just not fun.
Yes for gentoo and their users it makes sense. Debian is much, much larger and is used pretty extensively in enterprises.
The biggest question, I think, is: does a complete LLM ban interfere with, or slow down, security patches? I think it could, as LLMs are used a lot for vulnerability discovery.
I've been running Gentoo for 20 years. What you consider "Gentoo" is simply the package manager and "ebuilds"; instructions to build upstream packages. Right now I'm using LLMs to make my own ebuilds and maintain them. The packages I am compiling are basically vanilla upstream Linux packages. It is laughable to suggest Gentoo doesn't use AI made code because upstream already has it.
Their policy doesn't cover upstream, it's only for the Gentoo project specifically. It's clearly infeasible for them (or Debian) to try to control what their upstreams do, especially with the kernel accepting LLM contributions now.
What comprises “assistance?” Getting Claude Code to find bugs, suggest a fix, but a human actually editing the source files and testing the result - is that assistance? Perhaps this policy is as it says: a guiding principle that the community is expected to follow.
I think when we're talking about mature software, LLMs aren't the problem. It's the lack of discussion justifying the change. LLMs don't necessarily lower code quality. They put a spotlight on people "just going through the motions".
When the changes are small and properly reviewed, there's no point in caring how it was done. It should be trivial to write, but difficult to approve and merge.
Projects as significant as Debian should own the fact that they are indeed a bureaucracy and take inspiration from what has always worked long before LLMs.
I might su combining Proposal A (but only for contributions (to Debian) that actually involve the output of LLMs/generative-AI or that blindly trust their results) with Proposal C (for everything else).
However, in the case of item 4 of Proposal C, there are multiple kind of "assistance" that is possible; not all of them involve contributing the output of LLMs (e.g. the assistance might be to check for a mistake such as if a paragraph seems to be redundant or is incomplete or something like that; the result from the LLM might be incorrect but you will have to decide by yourself what (if anything) to do about it when writing your message). Even if some kind of assistance should not be banned entirely, it should still be discouraged (like the rest of Proposal C says). (Also, there are other programs that can sometimes be used for the assistance; e.g. sometimes ordinary spelling/grammar checking will do (although that won't find that an incorrect word that is spelled like a correct word in the correct grammar).)
I do not use LLM/generative-AI for any purpose, and would not accept contributions including their output into my projects, and would discourage other uses. (This seems to be similar to Proposal A and Proposal C.)
Skimming over the proposals, only option 1 even mentions copyright and licensing issues, which have been traditionally a big deal for open source software
> LLM output has very unclear legal status: it may be possible to copyright on its own merits, or not; it may be affected by all of the licenses and copyrights in the training data, or not. Debian Policy and the DFSG require absolute clarity for licensing and copyright[1][2]. Software and other contributions written conventionally by humans with unclear copyright or license status are not allowed in Debian; LLM output should not have a special exception to this.
The way the proposers portray it, this problem is a dealbreaker. Is the problem as portrayed? Do LLM vendors attempt to copyright the output?
Proposal B tries to solve one aspect of it:
> 2. Licensing and Attribution: If any pre-existing copyrighted materials (including pre-existing code licensed as free software) authored or owned by third parties are included in the AI tool’s output, prior to contributing such output to the project, the contributor should verify they have the right to submit it under the relevant open source license.
> 3. Accountability: Contributors assume full responsibility for their contributions, including vouching for the technical merit, security, license compliance, and utility of their submissions. The contributor remains solely accountable for the entirety of these contributions. Contributors should fully understand the proposed changes and be prepared to justify them.
How can a contributer know if LLM output includes "pre-existing copyrighted materials (including pre-existing code licensed as free software) authored or owned by third parties"? They can't be familiar with all pre-existing code in the world.
Maybe Debian could provide a search engine that looks for copyright violations, but do they really want to take responsibility for an issue with the risk of high liability? What if their search engine fails - maybe they could be sued for that too.
Also, that doesn't seem to address the IP of the LLM vendor.
> How can a contributer know if LLM output includes "pre-existing copyrighted materials (including pre-existing code licensed as free software) authored or owned by third parties"?
I use an LLM to write code, but 90% of that usage is code review which raises suggestions or fixes which I go and implement myself. Occasionally I also have the LLM complete a single function that I know it will do correctly because the requirements are all present and unambiguous. Because the content of the function is all based on the unique context of my codebase, it has about the same chance of reproducing someone else's code that I have, nearly zero.
Under those circumstances, I'd be comfortable taking the responsibility that the code is free of legal issues. The LLM assisted review also would make me more confident of the quality of the code and capable of explaining it and justifying it.
This is vastly different from a scenario where a developer has an agent coordinate and write most of the content. There are many different ways to use LLMs and it seems like these policies need to take them into consideration. Developer responsibility for the legality/quality of the code should be no different than non-assisted code, but they could also give examples of safe and beneficial LLM use.
>The way the proposers portray it, this problem is a dealbreaker. Is the problem as portrayed? Do LLM vendors attempt to copyright the output?
They're right that this has not been fully tested in court (though neither have some other open source licenses). There's a pretty clear direction that the wind is blowing in, though (you can see in most related cases that the judges focus on issues surrounding the core idea of using copyrighted data for training, and mostly don't give credit to the idea that all of a model's output is a derivative work of all of the input). I don't believe any LLM vendors try to claim copyright on the output of the models.
>How can a contributer know if LLM output includes "pre-existing copyrighted materials (including pre-existing code licensed as free software) authored or owned by third parties"? They can't be familiar with all pre-existing code in the world.
This is a common objection/worry. It seems relatively clear that even if most of the output of an LLM is not subject to existing copyright, some of it could well be (especially if, e.g. you feed an LLM a codebase and ask it to just output it again verbatim or with different variable and function names). What exactly the risks are if you are if this happens while you are using the LLM in good faith is pretty unclear (the vendors pretty much just say "It's at your own risk").
(This case felt a little surprising to me, BTW: https://petapixel.com/2026/07/22/dog-photographer-loses-copy... . But copyright case law is weird, arbitrary, and somewhat capricious about what aspects of a work are and aren't protected by copyright. With code this can also potentially be weird because in principle the functional parts are not protected, only the human expression components are. Untangling this can be time consuming, expensive, and ugly)
Copyright law varies across jurisdictions, and it's not obvious how training for AIs would align with https://en.wikipedia.org/wiki/Fair_dealing (and it appears a recent push to re-examine copyright law in Australia ended due to AI companies getting involved).
Exactly. This is like proposing to solve memory safety by asking programmers to simply not make any mistakes. Simply specify that "the contributor should verify" that the code has no use after free, no out of bounds access, no undefined behavior, etc., for which they "assume full responsibility".
At the same time, how can a human know if they accidentally rewrote exactly some code they read in the past from a codebase with an incompatible license?
Accidental copies happen all the time in code, music and art. Not all of it is necessarily super original, but that's beside the point and it happens.
So while we think we can hold AI tools to higher standards than humans, they are still modeled after human thinking and will make the same mistakes. Is this really the right thing to do and is it entirely practical?
This argument is almost algorithmic: Take the concern about <TECH>, substitute 'human'. Also, another popular response: don't address questions, instead try to put the person asking on the defensive.
The main point is the implicit argument that humans and machines are equivalent.
Interesting that this makes no distinction from using LLM output and using LLM assistance. An overly strict policy is better than no policy, but I wonder if using llm for conversation and analysis, but not code/config generation, was discussed.
The difference is LLMs are extremely good at analyzing code under the careful hand of an engineer. This is how most vulnerabilities are found now.
Prohibiting that might mean that security falls off.
And, if the concern is code quality, then everything before the code writing shouldn’t matter. A good engineer using an LLM for analysis and then writing a patch by-hand might create a better patch.
Debian is conservative so it makes sense that they have proposals for not using LLM's. They'll need another year or so but I am convinced it'll soon be obvious that any "no LLM" policy is untenable.
We are in the "cool new toy, let's make slop" phase, but we already see how careful use of LLM's can make high quality code faster and better than humans can do solo.
Once they are over the learning threshold, every good developer I know actually likes to use LLM's, as they allow them to realize their visions faster. If Debian has a "no LLM" policy I think they will eventually only have devs who like to type.
I should add that of course the "immoral companies burning too much energy" argument has actual merit, no matter how useful LLM's are or aren't. But I don't think Debian should die on that hill, as the only difference it will make is that we no longer have Debian.
We'll see how much people like it once the cheap subscriptions end and the actual cost of LLMs (+ profit margin) is asked of them. Although I'm suspecting that eventually, machines that are good enough to run reasonably useful models will become cheap enought that most developers will have those, but that might be some years from now.
Debian is run by volunteers, and I think Debian Developers would not like it if they would have to pay to work on it, thus there will be some aversion to (at least frontier) LLMs.
Curious what the "use" or "assistance" of LLM models means. I mostly use Gemini as a front-end to search Google without getting ads or SEO garbage, would that be forbidden?
Asking for good-faith compliance, and failing that, simply waiting for slop enthusiasts to publicly brag about using LLMs to break the rules. All evidence suggests that it is impossible for slop enthusiasts to resist outing themselves for any significant span of time.
Whatever the decision is today, it's going to become utterly moot within a decade or so.
Think about how far we've come over the past three years. Programming with AI in ten years will be utterly unfathomable. It would be like showing Fable 5 to someone a decade ago. It will be unbelievable.
In twenty years, programming by hand might be somewhat akin to writing assembly today: a skill that a limited number of programmers understand, and fewer still actively practise. Writing code by hand will become the new "I want to be close to the metal."
Perhaps by 2050, the concept of a Linux "distro" will be obsolete, replaced by a GPL-licensed local AI that just directly assembles everything from scratch to meet your exact specifications. Instead of relying on a predefined package manager, it traverses a distributed network of trust to find suitable candidate parts, downloading and verifying every line of source code before compiling it all locally. Trust would be rooted in the local AI's ability to understand and evaluate it, meaning that what replaces the distro are the groups competing for the best models.
Perhaps by 2070, nobody even writes or distributes code any more. The future kernel and web browser aren't code, they're descriptions of verifiable intent. A software project is really an extremely detailed description of what the software is meant to do. Your local GPL-licensed AI would decide how to write, assemble, patch, optimise, and integrate it for your hardware and your threat model. That would also blur the distinction between kernels, operating systems, and applications. The AI doesn't construct a computing environment as we know it today with discrete applications. It builds a coherent system that satisfies all requirements, rebuilding parts whenever your priorities change.
Downloading compiled code? Implicitly trusting someone else's source code? These will become the things people only do on hardware that is sufficiently air-gapped from any outside influence.
> Think about how far we've come over the past three years.
I do, and that's why I am pretty damn certain the future of programming doesn't involve LLMs. We still haven't really improved them, even with three years of pouring resources into doing so. The models of today are still as stupid and unreliable as the models of three years ago. Given we haven't made significant progress on that in three years, I don't expect we will get significant progress in the future either.
You are saying that LLMs have not meaningfully improved since 2023? If you want me to take that claim seriously, you will need to do more than merely assert it. I am not a mind-reader.
My current working theory for the seemingly irreconcilable differences between programmers' experiences with LLMs is grounded in their ability and willingness to explain what they want in sufficient detail. AI is not a mind-reader.
A lot of people pointing out that Proposal A is untenable/ at philosophical odds with the widespread upstream LLM usage, etc. etc.
I think that's an unfortunate framing & I think A would be the best option for the opposite reasons. If they go with Proposal A, people will absolutely use AI tooling to contribute to Debian. This might be in bad faith (unlikely but possible), but it will also be done in good faith through a number of avenues:
1. people having a saturation of LLM tooling integrated into their development environment to such an extent that it can't help touching everything they code - an infrequent Debian contributor is unlikely to change their entire dev setup, stripping out, turning off & limiting various ai tooling configs, just to comply with one project's social contract.
2. some people will mislead themselves by reframing the social contract's stated intent to suit their own world view. Maybe they'll subconsciously ignore the "or other generative AI tools" & just avoid direct chatbots, categorising some other LLM assistive tooling as out of the scope of the ban. Or perhaps they'll reframe it in terms of generation: using LLM agents to investigate, understand & guide them in their work while they actively type out each line of code committed (occasionally copypasting from an LLM response but never granting it repo write access): "This doesn't count, right?".
3. Some contributors might simply not read the social contract carefully enough, miss this line, & innocently forget to mention their LLM usage when pushing patches.
And ultimately, that's all ok. That's the reality of the world we live in: the social contract is ultimately "social", it's not intended to be strictly enforcable, it's a statement of intent aimed at a community who are assumed to be (& generally are) acting in good faith.
LLM tools are great - they open up amazing new horizons of what individual devs can achieve, which is positive for open source in particular. But they also come with an incredible upscaling of open source review burdens & I think anything "less" than Proposal A actively invites a tsunami of slop.
The wording of "written with ai tools" implies generation, as do all the concerns about copyright etc. it would be beyond insane to attempt to ban asking a chat bot or an agent questions to navigate or understand a codebase, not in the least because that is completely unenforceable
> not in the least because that is completely unenforceable
You've completely missed my point. It's ALL completely unenforceable. That doesn't mean the directive is useless - it's a guideline for contributors towards a general project direction that will materially limit & reduce overt LLM usage by anyone acting in good faith (the vast majority). The enforcement doesn't need to be strict, but anything less than strict wording won't be effective in influencing people's behaviour.
This is a very difficult problem. LLM's don't learn from reasoning about trade-offs and from their personal experience, so systems made with AI assistance stop obeying the unwritten rules of why things are the way they are, that comes from the maintainers' decades of experience.
On the other hand, they can produce and test anything anyone asks in minutes, as well as find security issues that lay dormant at the seams for decades.
I find it ironic that so many OSS projects are or are considering banning LLM usage. After all, without OSS, LLMs would never have been this good at coding.
Shouldn't OSS be the first to benefit? Or should OSS be a bedrock of carefully curated code untouched by AI slop?
OSS projects have in many cases been the first to be hurt by LLMs.
As you note, LLMs are heavily trained on OSS code. This makes them very good at "license-washing" copyleft code in to a decidedly-not-cleanroom reimplementation which is different enough at first glance that it would be hard to stop without a legal battle most OSS projects could not even dream of.
Plenty of ink has also been spilled on how LLMs have led to a significant increase in administrative workload for OSS maintainers by allowing users to produce a high volume of both bug reports and pull requests that are likely to be of varying quality or adherence to project standards but unlike a new human contributor will never develop in to a better community member. It's infinite newbies forever, a computer-generated Eternal September.
On top of all that, as noted in the Debian proposal the companies training these LLMs have proven themselves to have no shame or concern for the resources of others in the process of gathering data, inefficiently crawling every single URL they can find and nowadays even using their own models to generate new potential URLs that have never existed, in ways specifically designed to make filtering, rate limiting, or blocking the traffic altogether as hard as possible while performing what are often some of the most resource-expensive requests. These tactics have imposed substantial real costs on basically everyone who hosts their own infrastructure as well as community hosting projects, without even getting in to the hardware-related cost increases every single one of us have experienced.
LLMs have gone to a lot of OSS projects, stolen their code, beat up their maintainers, and emptied their wallets. Now the LLM people are coming by wanting big projects to work with them, "for their security" like a protection racket.
Yes, it should. And not just for code. Think better, more accurate documentation, quick start guides for both developers and users, reduce yak shaving effort, etc.
> should OSS be a bedrock of carefully curated code untouched by AI slop?
Slop is slop. There's plenty of human slop in OSS.
This is nonsense, the vast majority of critical CVEs are trivial to fix once you spot them. It's very rare that you have something like spectre where the solution is non-obvious: most of them are shit like "delete a `free`" or "put the `goto fail` in braces".
I suspect the debate shouldn't be LLMs or no LLMs, but rather what level of human accountability is required.
We've accepted compilers, static analyzers, and code generators because the maintainer is still responsible for the final result. The interesting question is whether LLMs fundamentally change that responsibility, or just change the kinds of mistakes reviewers need to look for
Some of the reasons to ban LLMs are beyond what they do to code. I.e. environmental degradation, scraping without authorization, reduce human interaction etc. So I think banning them still makes sense.
All of which have been challenged though. As the stake is high, any such reason (for or against LLMs, to be fair) has to be backed with Wikipedia-level evidences at the very least.
The whole anti-LLM crusade feels like developers trying to gatekeep their own relevance. They know the tool can automate parts if not all of what gave them status, so instead of adapting, they want to declare its use illegitimate and keep themselves as mandatory middlemen. Which is already ironic enough. If Proposal A passes and is somehow enforceable, every distro that embraces AI while Debian moralizes about it will lap them within a few years. You do not stop technological displacement by banning the tools everyone else is using. You just volunteer to be displaced first.
Elementary, Redox and Gentoo already banned it. Nix and Fedora allow it with disclosure. Now Debian and Arch are debating whether to adapt or seal their fate alongside Gentoo. Nix and Fedora got it right. Debian and Arch are deciding whether they want to remain relevant or fade into irrelevance.
The idea that Debian will become irrelevant if it doesn’t accept llm output is so obviously wrong. What would that even mean? Debian has never been a bleeding edge distribution, it’s pretty much the whole appeal, that and the strict rules for packaged software. Even if true, less people using Debian is pretty much irrelevant, it’s not a company trying to grow at all cost.
The whole „embrace genAI or be left behind“ is toxic nonsense repeated without actually engaging with reasons people have to reject the technology. Even in the case where the full-AI future becomes true, it takes pretty much no time to catch up and get up to speed with LLM tooling.
> The whole anti-LLM crusade feels like developers trying to gatekeep their own relevance. They know the tool can automate parts if not all of what gave them status, so instead of adapting, they want to declare its use illegitimate and keep themselves as mandatory middlemen.
That's an absurd and entirely uncharitable interpretation. What it actually is, is developers recognize that LLMs suck at programming and don't want people using a tool which has proven to be so crappy.
> Other projects exploring similar decisions have elicited a common reply: "How will you enforce a ban on LLM contributions?" While enforcement could be a challenge, this is a statement of intent by the Debian community, and we trust this community to adhere to it in good faith.
No, but they are going to tell you what's acceptable if you want to contribute to their project, and kick you out if they find you broke the rules. Projects have always had their rules you had to follow if you wanted to participate.
I simply don't accept that this type of project rule has any validity. It's like telling me what editor I can use, or that I can't use Google to search for programming tips, or that I can only use an electric car to drive to the office.
Proposal A is the end of debian for non-english speakers. For those who don't speak english which is most of the world, using an LLM has become vital, because technical information is not available in their language or is extremely basic. Arch Linux is far more lenient with this.
Un what? There’s several friendly clauses among these proposals including
> 8. Any contributor who feels they cannot write in English without assistance, may write in their native language, and expect readers to use translation tools of their choice. In that case a human-written English summary would be very welcome but is not required. In any case we promise not to shame anyone for any linguistic mistakes.
this doesn't prevent you from reading documentation in whatever language you like. If you're saying this is the end of someone communicating with you through an LLM despite not speaking the language, that should never happen and I've never been in any technical organization where that was considered appropriate, if you want to work in international software projects you need to have at least a functional grasp of English
most documentation is not translated to spanish, for example. working on international projects is the exception, not the rule. Spanish speakers just use LLMs because searching on google in spanish yields no results, because the info they need is in english so google doesn't show it to spanish speakers or by searching it in spanish. Spanish speakers have their search result language set to spanish by default based on IP location and most never change it to english, where the results are far more detailed, and then they almost never search in english.
I would say it leaves communication between individuals in a gray area. And discourages translation work because it has to be done manually. I would say, LLM-translated text on documentation is better than having an LLM translate it again and again on the fly every time someone asks how to do something, because the LLM-translated text can be corrected for everyone to see, whereas conversations with LLMs stay private by default.
This set of proposals, are (sorry) just stupid. It's like saying to someone, you are not allowed to saw wood using an electric saw, you must do it by hand. What are we doing here?! LLM(s) are just a tool. Use it as such. You should own the work anyway.
Linus Torvalds said last week that LLMs were just a tool. Lorenzo Stoakes had the impeccable rebuttal that "it's a tool in the same way a nuclear bomb is just a tool" - of course it's useful but there is clearly more to it than that.
I'm happy for Debian to at least consider the other aspects of social impact, ethics, copyright, and maintainability.
It's more like: food processing tool A is known to be defective. We ask employees of this restaurant and supply chains to avoid food processing tool A. This seems eminently reasonable. The question is whether the tool is in fact defective.
That's just it though. The tool(s) aren't necessarily defective. Just because you can chop off your own hand with an chainsaw or do a botch job of sawing a plank of wood because you weren't watching what you were doing with the electric saw, does not by itself make the tools (Chainsaw, Electric Saw) defective.
Sure, kind of, if the supposed defect is only one of the reasons, and if ‘food processing tool A’ was a hot-button topic that many people have deeply philosophical and ideological opinions about. Oh, and for good measure, if the proposed ban is for ‘food processors’ in general, and a material amount of pushback is coming from a bunch of greybeard industrial cooks that used one particular food processor a couple of years ago, weren’t all that impressed, and have been railing against food processors ever since.
Yes, they are. But that’s exactly what Proposal B (the only one without moral panic) says: allow its use, but place the responsibility on whoever uploads it. If you upload broken or unlicensed code, that’s on you. Proposal B is the most pragmatic and realistic option.
Yes. this I agree with. It's teh same thing many corporations like the one I work for are doing as well. Expecting and mandating that even if you use these so-called "AI" tools. that you are a) responsible and b) accountable for the "work".
> This set of proposals, are (sorry) just stupid. It's like saying to someone, you are not allowed to saw wood using an electric saw, you must do it by hand.
Stupid? Since when has there been a lack of clarity in copyright status with electric vs. hand saws?
This has nothing to do with copyright status or copyrighted works. If it did, Debian wouldn't even be a thing. You can't be a puritan about this; How much of Debian can one say, hand on heart, is truly not inspired, or influenced by other people's work, copyrighted or not.
Probably since we all started getting our legal advice from you and every other “I think that it’s illegal so it’s illegal” developer-turned-fake-lawyer on the Internet.
Oh, wait, that never happened.
Debian Airport is debating whether to allow aircraft to land on autopilot. On the one hand it’s nice that autopilot gives more people the chance to fly, and it lets experienced pilots fly more often. On the other hand, Debian Airport has, to date, been run by a carefully curated club of “stick and rudder” pilots. It’s a club where the air safety is backed up by nothing more than the name and reputation of the pilot coming in to land.
What if Claude goes down on approach? Or you run out of OpenAI tokens just before touchdown? What’s the point of having a named pilot if the control of the aircraft is in someone else’s hands, in part or in whole?
With actual pilots you have to take an exam and pass a practical test to prove you are competent. It’s a part of the analogy that could apply well to the real world Debian project.
Don't misinterpret this link as representing a final decision. It's actually three separate proposals which will be debated and then voted on.
Proposal A is "expressly forbid any contributions to Debian written with the use or assistance of large language models (LLMs) or other generative AI tools."
Proposal B is "The Debian project allows AI-assisted contributions (partially or fully generated by an LLM), provided the following conditions are met [...]"
Proposal C is "request that all contributors to Debian avoid the use of LLMs in their Debian work" without an outright ban.
Of the 3 proposals, the 2nd/B seems to be more of an “informed consent” model while the other two are seeking a comprehensive exclusion. I hope earnest dialogue and driving to a broad consensus among significant contributors is forthcoming.
Notable are the endorsements, which are balanced over all three options, perhaps indicating only 1/3 are supportive of a more permissive position. If this represents core contributors, this could be worrisome: a clear majority favoring a exclusion, but yet a sizable portion of contributors who don’t want a ban. Losing 1/3 of core contributors would be a significant loss to the project.
> Notable are the endorsements, which are balanced over all three options, perhaps indicating only 1/3 are supportive of a more permissive position.
That's not the case. First, you can't really extrapolate from endorsers to all of the Debian Developers. Second, endorsers might not have looked at the other proposals at all, and the fact that they only endorsed one option doesn't mean they don't support the others. And you can also endorse an option if you believe it is a valid, useful proposal to add to the GR, even if you don't agree with it personally.
But most importantly, when Debian votes, you don't vote for a single option, instead you rank the choices, and apart from the given options, another option "further discussion" is always added to the list of choices. It's likely that many who rank the most permissive option first will rank the lesser permissive option second, above "further discussion", or they could even give options the same rank (if they really don't think one is better than the other).
> Notable are the endorsements, which are balanced over all three options...
I don't think this is notable. The process requires five seconders, so you're just seeing the set of people who seconded asynchronously before there were obviously enough that others didn't bother. The culture is to avoid unnecessary noise and leave it for the vote.
3 replies →
I wonder how long software projects will even be able to ban LLM use in an age where exploits are found by LLMs. Like, if you have two forks of Debian and one uses LLMs to fix exploitable bugs and the other one doesn't, the level of security they can offer will be worlds apart. And noone in their right mind would want to use the less secure one. Similar to how noone would want to drive a car that was 100% hand built by humans when we know that machines do a much better job at precision tasks. LLMs are just another tool in the end.
15 replies →
you can neither check nor enforce any conditions. So the right answer is A
[dead]
I wonder how they can reconcile the stricter proposals with the LLM usage in kernel development. That seems totally untenable. I mean it all seems untenable, but with the kernel especially.
Also, what about when you inevitably get a situation where a critical vulnerability is discovered, and the only patch available is LLM generated? Do they have to wait to patch until some person who hasn't seen the LLM-generated patch does a clean room implementation?
I understand the objections to LLMs, but rejecting LLM-generated code really doesn't seem realistic.
Kernel development is excluded from this, because it's covered by "Upstream projects using LLMs for development".
On security, I worry that proposal A's "forbid any contributions to Debian written with the use or assistance of large language models", as written, excludes contributions where the LLM assisted in discovering the vulnerability. That's clearly a bad policy, and they should update their wording to clarify that.
3 replies →
The proposals which say no LLM generated code mean just for Debian, if upstream allows it then they will still accept it. It’s just contributions to Debian itself which would prohibit or discourage LLM contributions.
There is also proposal D right at the bottom, which is "Accept AI contributions for Debian specific work".
Thanks, we've put the three proposals bit in the title above.
There actually seem to be 4 separate proposals right now, maybe another was just added?
Proposal D is "Accept AI contributions for Debian specific work"
Wait, can we have a vote to not link HN contributions to dynamic pages!
while my heart lean heavily toward A, this would serverly limit Debian contribution and would rely on contributor honesty. It might be dismissed as an attempt to gatekeep contribution. But the issue of the human cost on the receiving end of the contribution and how sophisticated attack vector it could potentially be give me some concerns about long term quality.
B being the more pragmatic, despite the (valid) code licensing concerns. but could leave some very openly LLM driven project room to contribute package.
C as a middle ground would not satisfy any party and would only push back the resolution of the status quo.
The title says "proposals"
Why would someone misinterpret this as a final decision?
I misinterpreted it as a final decision when I first skimmed the page. The title of the linked document is "Resolution: LLM usage in Debian" and it's not obvious that it's several competing proposals until you explore the page in more detail.
The Hacker News submission title was updated to include "Three Proposals" after I posted my comment.
Eternal September of the Spotless Mind.
It seems the title wasn't always as clear.
Note that we are still in the discussion period, and more proposals can and will be added. Some might even be withdrawn.
Mr President, a 4th proposal has hit the General Resolution.
Disappointing there's not a Proposal E "require use of LLM's by contributors"
> A LLM (...) merely produces syntactically likely combinations of the training data
While doesn't matter much in the rest of the policy, this is a common misconception among AI skeptics. It is not the case for a long time (since RL is used heavily in the training) and a LLM may go beyond its training data.
I think the questionable term is not “syntactically likely” but “merely”. Syntactical likeness is a vast solution space that encompasses the work of a terrible developer and a genius developer. In fact this solution space is a gap wide enough to encompass all coding knowledge and expertise.
"Combinations of the training data" is also hiding a pretty common and serious misconception about modern generative AI: this seems to come from a mental model that LLMs and image generators are merely producing a collage of sorts, especially that this is some way linked to specific parts of the training data as opposed to a generalized result of building up an understanding of how to produce a sentence, piece of code, or plausible image.
1 reply →
The natural next argument that I see a lot is "well it's still all probabilistic", which is technically true. However, don't the atoms that make up the cells that make up a human move around and interact based on probabilities?
Saying a LLM is all just based on probabilities is pretty meaningless, even if it is true, since everything else is just based on probabilities too. What matters is the massive amount of machinery that generates those probabilities. Anything from rolling a dice to an accurate model of an entire human, or even a model of the entire universe, is all "just probabilities".
That argument is mostly used against the misconception of AI-stans that LLMs are compilers, English is a programming language, and that prompting is a programing. There are many reasons for why that is false, and LLMs being stochastic is one of the more obvious reasons.
But there are more probabilistic elements about LLMs. Inference is by default deterministic for any given input and weights, however it is made stochastic algorithmically, when the model chooses the second or third likeliest outcome with some probability. Training is stochastic, the adjustment of the weights is probabilistic, etc. etc.
You seem to make the fact that our brains are also probabilistic as some sort of a gotcha. However that has never been in dispute. Neurons fire with certain probability, this has been known since the advent of neuroscience. However this is why computers are preferred for certain tasks over human brains. When you run a program you can be certain it behaves in a certain way given a particular input and parameters.
2 replies →
I assume RL stands for reinforcement learning, which just means that the weights are adjusted by evaluating some loss function. If the loss function uses the training data (i.e. supervised learning) then the original assumption still holds. So I fail to see how this is a misconception.
I think OP might be referring to synthetic data, which is now used to provide more data for training?
1 reply →
Supervised learning and reinforcement learning are mathematically distinct. Supervised pre-training as you said maximizes the likelihood of a token sequence. RL optimizes a policy against an external reward signal or execution environment. For example, RL evaluates code against runtime execution/unit tests/proof checkers (e.g lean). Models learn new strategies and are able to produce novel code if trained in an RL environment.
2 replies →
Doesn't the LLM still output the next most probable token (modulus heat/desired variation) according to its model? Training shifts the probability - but doesn't change the fact that the output is a sampling based on input and the model?
So do humans more or less? The only counterpoints I've heard are religious or unmeasured quantum brain something.
4 replies →
The combination of the most probable tokens doesn't necessarily have to be in the training set, thus creating something completely novel.
8 replies →
It’s not sampling anything though. At each inference step the fundamental output of the process can be interpreted as a probability for each possible token how likely it is. Sampling would imply otherwise but the probabilities are the output.
I would like to offer this comment from three days ago as counterpoint:
https://news.ycombinator.com/item?id=49013694
Common sense testing only tells about the presence of common sense, that's it. It is clear from recent events that LLMs do definitely possess some sort of intelligence even though it might have been polarized to certain domains.
What does that prove? That AIs are not RLd for counting letters in texts? Do you doubt that they could excel at it with only synthetic data?
Gentoo chose to ban LLMs two years ago. They seem to be doing well.
Funny because out of all the distros, Gentoo benefits from LLM usage, at least as a power user. I have fixed many-an-issues just having LLMs write patches in a way that portage automatically builds them, or rapidly put together ebuilds. I respect upstream decision though and don't share my changes (and realistically who would want them)
This tracks with what I do too (also a gentoo power user). My personal overlay is much healthier thanks to LLMs, but I don’t use it for writing code for personal use. It’s just not fun.
1 reply →
Gentoo is doing as well as it’s always done. It’s clearly a very niche distribution. I say this as someone that used Gentoo for years.
They are free to keep pretending that they aren’t receiving covert AI contributions, because I’d bet the farm that they are.
Debian has a little more ‘general interest’
Yes for gentoo and their users it makes sense. Debian is much, much larger and is used pretty extensively in enterprises.
The biggest question, I think, is: does a complete LLM ban interfere with, or slow down, security patches? I think it could, as LLMs are used a lot for vulnerability discovery.
I've been running Gentoo for 20 years. What you consider "Gentoo" is simply the package manager and "ebuilds"; instructions to build upstream packages. Right now I'm using LLMs to make my own ebuilds and maintain them. The packages I am compiling are basically vanilla upstream Linux packages. It is laughable to suggest Gentoo doesn't use AI made code because upstream already has it.
Their policy doesn't cover upstream, it's only for the Gentoo project specifically. It's clearly infeasible for them (or Debian) to try to control what their upstreams do, especially with the kernel accepting LLM contributions now.
You would not even know about it if they were not doing well
I wonder what fraction of Trixie already violates this requirement of the first proposal...
What comprises “assistance?” Getting Claude Code to find bugs, suggest a fix, but a human actually editing the source files and testing the result - is that assistance? Perhaps this policy is as it says: a guiding principle that the community is expected to follow.
I think when we're talking about mature software, LLMs aren't the problem. It's the lack of discussion justifying the change. LLMs don't necessarily lower code quality. They put a spotlight on people "just going through the motions".
When the changes are small and properly reviewed, there's no point in caring how it was done. It should be trivial to write, but difficult to approve and merge.
Projects as significant as Debian should own the fact that they are indeed a bureaucracy and take inspiration from what has always worked long before LLMs.
I might su combining Proposal A (but only for contributions (to Debian) that actually involve the output of LLMs/generative-AI or that blindly trust their results) with Proposal C (for everything else).
However, in the case of item 4 of Proposal C, there are multiple kind of "assistance" that is possible; not all of them involve contributing the output of LLMs (e.g. the assistance might be to check for a mistake such as if a paragraph seems to be redundant or is incomplete or something like that; the result from the LLM might be incorrect but you will have to decide by yourself what (if anything) to do about it when writing your message). Even if some kind of assistance should not be banned entirely, it should still be discouraged (like the rest of Proposal C says). (Also, there are other programs that can sometimes be used for the assistance; e.g. sometimes ordinary spelling/grammar checking will do (although that won't find that an incorrect word that is spelled like a correct word in the correct grammar).)
I do not use LLM/generative-AI for any purpose, and would not accept contributions including their output into my projects, and would discourage other uses. (This seems to be similar to Proposal A and Proposal C.)
i think the line is: expressing that you reputationally certify its correct and its worth the time
I certainly hope they'll choose C.
A is good, it's grounded. C is just based
The unhinged Background paragraph is disqualifying.
Skimming over the proposals, only option 1 even mentions copyright and licensing issues, which have been traditionally a big deal for open source software
From Proposal A:
> LLM output has very unclear legal status: it may be possible to copyright on its own merits, or not; it may be affected by all of the licenses and copyrights in the training data, or not. Debian Policy and the DFSG require absolute clarity for licensing and copyright[1][2]. Software and other contributions written conventionally by humans with unclear copyright or license status are not allowed in Debian; LLM output should not have a special exception to this.
The way the proposers portray it, this problem is a dealbreaker. Is the problem as portrayed? Do LLM vendors attempt to copyright the output?
Proposal B tries to solve one aspect of it:
> 2. Licensing and Attribution: If any pre-existing copyrighted materials (including pre-existing code licensed as free software) authored or owned by third parties are included in the AI tool’s output, prior to contributing such output to the project, the contributor should verify they have the right to submit it under the relevant open source license.
> 3. Accountability: Contributors assume full responsibility for their contributions, including vouching for the technical merit, security, license compliance, and utility of their submissions. The contributor remains solely accountable for the entirety of these contributions. Contributors should fully understand the proposed changes and be prepared to justify them.
How can a contributer know if LLM output includes "pre-existing copyrighted materials (including pre-existing code licensed as free software) authored or owned by third parties"? They can't be familiar with all pre-existing code in the world.
Maybe Debian could provide a search engine that looks for copyright violations, but do they really want to take responsibility for an issue with the risk of high liability? What if their search engine fails - maybe they could be sued for that too.
Also, that doesn't seem to address the IP of the LLM vendor.
> How can a contributer know if LLM output includes "pre-existing copyrighted materials (including pre-existing code licensed as free software) authored or owned by third parties"?
I use an LLM to write code, but 90% of that usage is code review which raises suggestions or fixes which I go and implement myself. Occasionally I also have the LLM complete a single function that I know it will do correctly because the requirements are all present and unambiguous. Because the content of the function is all based on the unique context of my codebase, it has about the same chance of reproducing someone else's code that I have, nearly zero.
Under those circumstances, I'd be comfortable taking the responsibility that the code is free of legal issues. The LLM assisted review also would make me more confident of the quality of the code and capable of explaining it and justifying it.
This is vastly different from a scenario where a developer has an agent coordinate and write most of the content. There are many different ways to use LLMs and it seems like these policies need to take them into consideration. Developer responsibility for the legality/quality of the code should be no different than non-assisted code, but they could also give examples of safe and beneficial LLM use.
>The way the proposers portray it, this problem is a dealbreaker. Is the problem as portrayed? Do LLM vendors attempt to copyright the output?
They're right that this has not been fully tested in court (though neither have some other open source licenses). There's a pretty clear direction that the wind is blowing in, though (you can see in most related cases that the judges focus on issues surrounding the core idea of using copyrighted data for training, and mostly don't give credit to the idea that all of a model's output is a derivative work of all of the input). I don't believe any LLM vendors try to claim copyright on the output of the models.
>How can a contributer know if LLM output includes "pre-existing copyrighted materials (including pre-existing code licensed as free software) authored or owned by third parties"? They can't be familiar with all pre-existing code in the world.
This is a common objection/worry. It seems relatively clear that even if most of the output of an LLM is not subject to existing copyright, some of it could well be (especially if, e.g. you feed an LLM a codebase and ask it to just output it again verbatim or with different variable and function names). What exactly the risks are if you are if this happens while you are using the LLM in good faith is pretty unclear (the vendors pretty much just say "It's at your own risk").
(This case felt a little surprising to me, BTW: https://petapixel.com/2026/07/22/dog-photographer-loses-copy... . But copyright case law is weird, arbitrary, and somewhat capricious about what aspects of a work are and aren't protected by copyright. With code this can also potentially be weird because in principle the functional parts are not protected, only the human expression components are. Untangling this can be time consuming, expensive, and ugly)
Copyright law varies across jurisdictions, and it's not obvious how training for AIs would align with https://en.wikipedia.org/wiki/Fair_dealing (and it appears a recent push to re-examine copyright law in Australia ended due to AI companies getting involved).
Exactly. This is like proposing to solve memory safety by asking programmers to simply not make any mistakes. Simply specify that "the contributor should verify" that the code has no use after free, no out of bounds access, no undefined behavior, etc., for which they "assume full responsibility".
At the same time, how can a human know if they accidentally rewrote exactly some code they read in the past from a codebase with an incompatible license?
Accidental copies happen all the time in code, music and art. Not all of it is necessarily super original, but that's beside the point and it happens.
So while we think we can hold AI tools to higher standards than humans, they are still modeled after human thinking and will make the same mistakes. Is this really the right thing to do and is it entirely practical?
This argument is almost algorithmic: Take the concern about <TECH>, substitute 'human'. Also, another popular response: don't address questions, instead try to put the person asking on the defensive.
The main point is the implicit argument that humans and machines are equivalent.
1 reply →
Interesting that this makes no distinction from using LLM output and using LLM assistance. An overly strict policy is better than no policy, but I wonder if using llm for conversation and analysis, but not code/config generation, was discussed.
I’m curious what makes you believe that this distinction is precise or meaningful. What exactly is the difference?
The difference is LLMs are extremely good at analyzing code under the careful hand of an engineer. This is how most vulnerabilities are found now.
Prohibiting that might mean that security falls off.
And, if the concern is code quality, then everything before the code writing shouldn’t matter. A good engineer using an LLM for analysis and then writing a patch by-hand might create a better patch.
1 reply →
Debian is conservative so it makes sense that they have proposals for not using LLM's. They'll need another year or so but I am convinced it'll soon be obvious that any "no LLM" policy is untenable.
We are in the "cool new toy, let's make slop" phase, but we already see how careful use of LLM's can make high quality code faster and better than humans can do solo.
Once they are over the learning threshold, every good developer I know actually likes to use LLM's, as they allow them to realize their visions faster. If Debian has a "no LLM" policy I think they will eventually only have devs who like to type.
I should add that of course the "immoral companies burning too much energy" argument has actual merit, no matter how useful LLM's are or aren't. But I don't think Debian should die on that hill, as the only difference it will make is that we no longer have Debian.
We'll see how much people like it once the cheap subscriptions end and the actual cost of LLMs (+ profit margin) is asked of them. Although I'm suspecting that eventually, machines that are good enough to run reasonably useful models will become cheap enought that most developers will have those, but that might be some years from now.
Debian is run by volunteers, and I think Debian Developers would not like it if they would have to pay to work on it, thus there will be some aversion to (at least frontier) LLMs.
Curious what the "use" or "assistance" of LLM models means. I mostly use Gemini as a front-end to search Google without getting ads or SEO garbage, would that be forbidden?
Even if they wanted to forbid that, how would they even detect it?
Some people want to uphold the rule even if they can't be caught.
3 replies →
Asking for good-faith compliance, and failing that, simply waiting for slop enthusiasts to publicly brag about using LLMs to break the rules. All evidence suggests that it is impossible for slop enthusiasts to resist outing themselves for any significant span of time.
Whatever the decision is today, it's going to become utterly moot within a decade or so.
Think about how far we've come over the past three years. Programming with AI in ten years will be utterly unfathomable. It would be like showing Fable 5 to someone a decade ago. It will be unbelievable.
In twenty years, programming by hand might be somewhat akin to writing assembly today: a skill that a limited number of programmers understand, and fewer still actively practise. Writing code by hand will become the new "I want to be close to the metal."
Perhaps by 2050, the concept of a Linux "distro" will be obsolete, replaced by a GPL-licensed local AI that just directly assembles everything from scratch to meet your exact specifications. Instead of relying on a predefined package manager, it traverses a distributed network of trust to find suitable candidate parts, downloading and verifying every line of source code before compiling it all locally. Trust would be rooted in the local AI's ability to understand and evaluate it, meaning that what replaces the distro are the groups competing for the best models.
Perhaps by 2070, nobody even writes or distributes code any more. The future kernel and web browser aren't code, they're descriptions of verifiable intent. A software project is really an extremely detailed description of what the software is meant to do. Your local GPL-licensed AI would decide how to write, assemble, patch, optimise, and integrate it for your hardware and your threat model. That would also blur the distinction between kernels, operating systems, and applications. The AI doesn't construct a computing environment as we know it today with discrete applications. It builds a coherent system that satisfies all requirements, rebuilding parts whenever your priorities change.
Downloading compiled code? Implicitly trusting someone else's source code? These will become the things people only do on hardware that is sufficiently air-gapped from any outside influence.
> Think about how far we've come over the past three years.
I do, and that's why I am pretty damn certain the future of programming doesn't involve LLMs. We still haven't really improved them, even with three years of pouring resources into doing so. The models of today are still as stupid and unreliable as the models of three years ago. Given we haven't made significant progress on that in three years, I don't expect we will get significant progress in the future either.
You are saying that LLMs have not meaningfully improved since 2023? If you want me to take that claim seriously, you will need to do more than merely assert it. I am not a mind-reader.
My current working theory for the seemingly irreconcilable differences between programmers' experiences with LLMs is grounded in their ability and willingness to explain what they want in sufficient detail. AI is not a mind-reader.
Be interesting to see what happens when it becomes a serious hacker target.
A lot of people pointing out that Proposal A is untenable/ at philosophical odds with the widespread upstream LLM usage, etc. etc.
I think that's an unfortunate framing & I think A would be the best option for the opposite reasons. If they go with Proposal A, people will absolutely use AI tooling to contribute to Debian. This might be in bad faith (unlikely but possible), but it will also be done in good faith through a number of avenues:
1. people having a saturation of LLM tooling integrated into their development environment to such an extent that it can't help touching everything they code - an infrequent Debian contributor is unlikely to change their entire dev setup, stripping out, turning off & limiting various ai tooling configs, just to comply with one project's social contract.
2. some people will mislead themselves by reframing the social contract's stated intent to suit their own world view. Maybe they'll subconsciously ignore the "or other generative AI tools" & just avoid direct chatbots, categorising some other LLM assistive tooling as out of the scope of the ban. Or perhaps they'll reframe it in terms of generation: using LLM agents to investigate, understand & guide them in their work while they actively type out each line of code committed (occasionally copypasting from an LLM response but never granting it repo write access): "This doesn't count, right?".
3. Some contributors might simply not read the social contract carefully enough, miss this line, & innocently forget to mention their LLM usage when pushing patches.
And ultimately, that's all ok. That's the reality of the world we live in: the social contract is ultimately "social", it's not intended to be strictly enforcable, it's a statement of intent aimed at a community who are assumed to be (& generally are) acting in good faith.
LLM tools are great - they open up amazing new horizons of what individual devs can achieve, which is positive for open source in particular. But they also come with an incredible upscaling of open source review burdens & I think anything "less" than Proposal A actively invites a tsunami of slop.
The wording of "written with ai tools" implies generation, as do all the concerns about copyright etc. it would be beyond insane to attempt to ban asking a chat bot or an agent questions to navigate or understand a codebase, not in the least because that is completely unenforceable
> not in the least because that is completely unenforceable
You've completely missed my point. It's ALL completely unenforceable. That doesn't mean the directive is useless - it's a guideline for contributors towards a general project direction that will materially limit & reduce overt LLM usage by anyone acting in good faith (the vast majority). The enforcement doesn't need to be strict, but anything less than strict wording won't be effective in influencing people's behaviour.
3 replies →
This is a very difficult problem. LLM's don't learn from reasoning about trade-offs and from their personal experience, so systems made with AI assistance stop obeying the unwritten rules of why things are the way they are, that comes from the maintainers' decades of experience.
On the other hand, they can produce and test anything anyone asks in minutes, as well as find security issues that lay dormant at the seams for decades.
There is no great answer here.
"Look at this. Look at what they make you give." - The Professor from The Bourne Identity (2002)
looks like the best proposal is the just the least bad
What do you think is the best proposal?
It doesn’t really matter for their central point.
That is how policy tends to work
I find it ironic that so many OSS projects are or are considering banning LLM usage. After all, without OSS, LLMs would never have been this good at coding.
Shouldn't OSS be the first to benefit? Or should OSS be a bedrock of carefully curated code untouched by AI slop?
> Shouldn't OSS be the first to benefit?
OSS projects have in many cases been the first to be hurt by LLMs.
As you note, LLMs are heavily trained on OSS code. This makes them very good at "license-washing" copyleft code in to a decidedly-not-cleanroom reimplementation which is different enough at first glance that it would be hard to stop without a legal battle most OSS projects could not even dream of.
Plenty of ink has also been spilled on how LLMs have led to a significant increase in administrative workload for OSS maintainers by allowing users to produce a high volume of both bug reports and pull requests that are likely to be of varying quality or adherence to project standards but unlike a new human contributor will never develop in to a better community member. It's infinite newbies forever, a computer-generated Eternal September.
On top of all that, as noted in the Debian proposal the companies training these LLMs have proven themselves to have no shame or concern for the resources of others in the process of gathering data, inefficiently crawling every single URL they can find and nowadays even using their own models to generate new potential URLs that have never existed, in ways specifically designed to make filtering, rate limiting, or blocking the traffic altogether as hard as possible while performing what are often some of the most resource-expensive requests. These tactics have imposed substantial real costs on basically everyone who hosts their own infrastructure as well as community hosting projects, without even getting in to the hardware-related cost increases every single one of us have experienced.
LLMs have gone to a lot of OSS projects, stolen their code, beat up their maintainers, and emptied their wallets. Now the LLM people are coming by wanting big projects to work with them, "for their security" like a protection racket.
> Shouldn't OSS be the first to benefit?
Yes, it should. And not just for code. Think better, more accurate documentation, quick start guides for both developers and users, reduce yak shaving effort, etc.
> should OSS be a bedrock of carefully curated code untouched by AI slop?
Slop is slop. There's plenty of human slop in OSS.
Debian should completely avoid AI slop.
Let the Salesforces and the Atlassians of the world destroy themselves with poor quality vibecode driven teams.
Using LLM to scan for security holes is another issue altogether.
Indeed it is, and it doesn't count as a "contribution" at all; the actual work is in fixing the CVE.
This is nonsense, the vast majority of critical CVEs are trivial to fix once you spot them. It's very rare that you have something like spectre where the solution is non-obvious: most of them are shit like "delete a `free`" or "put the `goto fail` in braces".
1 reply →
I suspect the debate shouldn't be LLMs or no LLMs, but rather what level of human accountability is required.
We've accepted compilers, static analyzers, and code generators because the maintainer is still responsible for the final result. The interesting question is whether LLMs fundamentally change that responsibility, or just change the kinds of mistakes reviewers need to look for
Some of the reasons to ban LLMs are beyond what they do to code. I.e. environmental degradation, scraping without authorization, reduce human interaction etc. So I think banning them still makes sense.
All of which have been challenged though. As the stake is high, any such reason (for or against LLMs, to be fair) has to be backed with Wikipedia-level evidences at the very least.
4 replies →
The whole anti-LLM crusade feels like developers trying to gatekeep their own relevance. They know the tool can automate parts if not all of what gave them status, so instead of adapting, they want to declare its use illegitimate and keep themselves as mandatory middlemen. Which is already ironic enough. If Proposal A passes and is somehow enforceable, every distro that embraces AI while Debian moralizes about it will lap them within a few years. You do not stop technological displacement by banning the tools everyone else is using. You just volunteer to be displaced first.
Elementary, Redox and Gentoo already banned it. Nix and Fedora allow it with disclosure. Now Debian and Arch are debating whether to adapt or seal their fate alongside Gentoo. Nix and Fedora got it right. Debian and Arch are deciding whether they want to remain relevant or fade into irrelevance.
The idea that Debian will become irrelevant if it doesn’t accept llm output is so obviously wrong. What would that even mean? Debian has never been a bleeding edge distribution, it’s pretty much the whole appeal, that and the strict rules for packaged software. Even if true, less people using Debian is pretty much irrelevant, it’s not a company trying to grow at all cost.
The whole „embrace genAI or be left behind“ is toxic nonsense repeated without actually engaging with reasons people have to reject the technology. Even in the case where the full-AI future becomes true, it takes pretty much no time to catch up and get up to speed with LLM tooling.
> The whole anti-LLM crusade feels like developers trying to gatekeep their own relevance. They know the tool can automate parts if not all of what gave them status, so instead of adapting, they want to declare its use illegitimate and keep themselves as mandatory middlemen.
That's an absurd and entirely uncharitable interpretation. What it actually is, is developers recognize that LLMs suck at programming and don't want people using a tool which has proven to be so crappy.
[dead]
[dead]
[dead]
[dead]
tl;dr right problem, wrong solution.
[flagged]
Are you aware, that a very big chunk of Linux distributions like Ubuntu are build in top of Debian?
Why do you want Debian to die?
Anyone that has an opinion about this, already knows that, yes. Drop the condescension.
1 reply →
By Lindy effect, debian has a life expectancy of 25 more years
Just like BSD is dying. Netcraft confirms it.
The fact that some react to these like they were taken away some possession is genuinely interesting
Proposal A
Using LLM encourages data theft or unethical acts.
Proposal D, currently enacted: Use LLMs but don't tell anyone
Poor Debian. AI is killing it right now.
How can they even tell code was written by AI instead of a human?
> Other projects exploring similar decisions have elicited a common reply: "How will you enforce a ban on LLM contributions?" While enforcement could be a challenge, this is a statement of intent by the Debian community, and we trust this community to adhere to it in good faith.
Yeah, good luck with that.
emdashes
Code uses emdashes?
1 reply →
My coding TUIs all have rules to never use emdashes.
For one thing, LLMs write terrible code. So unless a human does a lot of editing, they'll be able to tell because the code is slop.
Proposals like these are arrogant and obnoxious. You aren’t going to tell me what tools I can use.
No, but they are going to tell you what's acceptable if you want to contribute to their project, and kick you out if they find you broke the rules. Projects have always had their rules you had to follow if you wanted to participate.
I simply don't accept that this type of project rule has any validity. It's like telling me what editor I can use, or that I can't use Google to search for programming tips, or that I can only use an electric car to drive to the office.
4 replies →
Were you going to contribute to Debian otherwise?
Then don’t contribute? It’s crazy how some of you feel offended that projects you’re not engaged with might decide to not use LLMs
Proposal A is the end of debian for non-english speakers. For those who don't speak english which is most of the world, using an LLM has become vital, because technical information is not available in their language or is extremely basic. Arch Linux is far more lenient with this.
Un what? There’s several friendly clauses among these proposals including
> 8. Any contributor who feels they cannot write in English without assistance, may write in their native language, and expect readers to use translation tools of their choice. In that case a human-written English summary would be very welcome but is not required. In any case we promise not to shame anyone for any linguistic mistakes.
That is Proposal C, not Proposal A. (Although I do think Proposal A as written also does not forbid translation usage to communicate.)
this doesn't prevent you from reading documentation in whatever language you like. If you're saying this is the end of someone communicating with you through an LLM despite not speaking the language, that should never happen and I've never been in any technical organization where that was considered appropriate, if you want to work in international software projects you need to have at least a functional grasp of English
most documentation is not translated to spanish, for example. working on international projects is the exception, not the rule. Spanish speakers just use LLMs because searching on google in spanish yields no results, because the info they need is in english so google doesn't show it to spanish speakers or by searching it in spanish. Spanish speakers have their search result language set to spanish by default based on IP location and most never change it to english, where the results are far more detailed, and then they almost never search in english.
2 replies →
This is a somewhat common talking point for supporting LLMs. What did non-English speakers do before LLMs?
They learned English or gave up.
2 replies →
I don't believe Proposal A forbids this usage.
I would say it leaves communication between individuals in a gray area. And discourages translation work because it has to be done manually. I would say, LLM-translated text on documentation is better than having an LLM translate it again and again on the fly every time someone asks how to do something, because the LLM-translated text can be corrected for everyone to see, whereas conversations with LLMs stay private by default.
2 replies →
This is ridiculos, think about that, will Debian project survive without a Linux kernel written with the help of LLM tools?
This set of proposals, are (sorry) just stupid. It's like saying to someone, you are not allowed to saw wood using an electric saw, you must do it by hand. What are we doing here?! LLM(s) are just a tool. Use it as such. You should own the work anyway.
Linus Torvalds said last week that LLMs were just a tool. Lorenzo Stoakes had the impeccable rebuttal that "it's a tool in the same way a nuclear bomb is just a tool" - of course it's useful but there is clearly more to it than that.
I'm happy for Debian to at least consider the other aspects of social impact, ethics, copyright, and maintainability.
An LLM is a tool.
A nuclear bomb is a tool.
And programmers are the biggest tools of them all.
/ba dum tis
You wouldn't program a computer with a nuclear bomb!!!
It's more like: food processing tool A is known to be defective. We ask employees of this restaurant and supply chains to avoid food processing tool A. This seems eminently reasonable. The question is whether the tool is in fact defective.
That's just it though. The tool(s) aren't necessarily defective. Just because you can chop off your own hand with an chainsaw or do a botch job of sawing a plank of wood because you weren't watching what you were doing with the electric saw, does not by itself make the tools (Chainsaw, Electric Saw) defective.
7 replies →
Why don't you come back to talking about software rather than using terrible analogies.
LLMs write bad code / so do people, a lot of really bad code.
LLMs write security bugs / so do people, so many god damned security bugs.
1 reply →
Sure, kind of, if the supposed defect is only one of the reasons, and if ‘food processing tool A’ was a hot-button topic that many people have deeply philosophical and ideological opinions about. Oh, and for good measure, if the proposed ban is for ‘food processors’ in general, and a material amount of pushback is coming from a bunch of greybeard industrial cooks that used one particular food processor a couple of years ago, weren’t all that impressed, and have been railing against food processors ever since.
Ironically, any SoTA LLM would be able to make a better analogy than you. Sometimes the average human is the slop generator
Yes, they are. But that’s exactly what Proposal B (the only one without moral panic) says: allow its use, but place the responsibility on whoever uploads it. If you upload broken or unlicensed code, that’s on you. Proposal B is the most pragmatic and realistic option.
Yes. this I agree with. It's teh same thing many corporations like the one I work for are doing as well. Expecting and mandating that even if you use these so-called "AI" tools. that you are a) responsible and b) accountable for the "work".
> This set of proposals, are (sorry) just stupid. It's like saying to someone, you are not allowed to saw wood using an electric saw, you must do it by hand.
Stupid? Since when has there been a lack of clarity in copyright status with electric vs. hand saws?
This has nothing to do with copyright status or copyrighted works. If it did, Debian wouldn't even be a thing. You can't be a puritan about this; How much of Debian can one say, hand on heart, is truly not inspired, or influenced by other people's work, copyrighted or not.
1 reply →
Probably since we all started getting our legal advice from you and every other “I think that it’s illegal so it’s illegal” developer-turned-fake-lawyer on the Internet. Oh, wait, that never happened.
Yeah, I agree, it's really stupid. It's like discussing whether contributors are allowed to use google search, stackoverflow, an encyclopedia, etc.
Debian Airport is debating whether to allow aircraft to land on autopilot. On the one hand it’s nice that autopilot gives more people the chance to fly, and it lets experienced pilots fly more often. On the other hand, Debian Airport has, to date, been run by a carefully curated club of “stick and rudder” pilots. It’s a club where the air safety is backed up by nothing more than the name and reputation of the pilot coming in to land.
What if Claude goes down on approach? Or you run out of OpenAI tokens just before touchdown? What’s the point of having a named pilot if the control of the aircraft is in someone else’s hands, in part or in whole?
With actual pilots you have to take an exam and pass a practical test to prove you are competent. It’s a part of the analogy that could apply well to the real world Debian project.