Though LLMs are not intelligent, that’s not how they work. They are not expert engines of 80s. They are text predictors defined by billions parameters learned from the dataset.
3Blue1Brown explain how they work rather well, I think: Neural networks - YouTube.
There are any number of political questions we can indeed completely sidestep in Haiku because whatever the answer is, it doesn’t affect Haiku itself. But obviously in the course of collaborating with other people, we are going to have many interpersonal interactions. LLMs directly affect people’s behavior and the interpersonal interactions of developing software. So there is no way to avoid the “mixing” you are talking about here.
And by stating “refusing to interact with progress” you are making an assumption that the “progress” is actually an improvement over whatever it is replacing. Well, what if it isn’t? In order to know if it’s really ‘progress’ or not, we have to know: what are we ‘progressing towards’? And if this thing does not help us progress towards that, well, then it’s not really ‘progress’. Just because it “goes faster” doesn’t mean it’s ‘progress’; if it’s going somewhere we don’t want to go…
I do not agree. Human speech is the product of human intelligence. Asking a non-human non-intelligence to interpret human speech is a category error.
Code is a special type of human speech that also accords to some specification to be interpreted literally by machines. But human speech that doesn’t accord to those standards should not be interpreted by machines. If you have a “machine-readable” file format that defines how to build software, sure, let’s build a translator from this into HaikuPorts recipes. But asking a machine to interpret the products of human thought is flawed from the start.
Well, I am trying to explain the principles from which my attitude and behavior comes. Perhaps you find the principles ‘strange’, but I would hope they are at least intelligible. And if they aren’t, well, ask about whatever you find strange, and I will try and answer!
Sorry, I was perhaps unclear: when I meant “grow in patience” I didn’t mean users should just put up with hostility, I meant that forum users who are themselves hostile should learn patience (or be made to leave because they’re being hostile.) If people only interact with others who already know Haiku, then they won’t have any practice being patient with people who don’t know the basics.
I think the use of LLMs is very antisocial. You do not; but I would hope you would agree that driving away people who are “antisocial”, no matter how talented, is the right thing for a community to do to preserve its atmosphere? If so, then the question is whether we can amongst ourselves agree these things are antisocial, and not the mere fact of whether banning them would result in people being driven away. Of course it will, but that would then be the point.
X512 was trying to say that my opposition is pointless. I was trying to make the case that it’s not, even if we lose in the end. I don’t think we are actually trying to defeat the “giants” (Windows, Linux, etc.), but we are trying to carve out a space away from them, aren’t we? Well, how is that not a “romantic” proposition?
This is an interesting example, because monasteries, while populated by a select group, were historically for everyone. You could go visit one and stay for a few days at no charge, and retreat from the cares of the world for a while.
I think we should have higher ambitions than that, sure. But if that’s the niche Haiku falls into (and I think at the moment it does for many people), is that really such a bad thing?
My favorite discussions of this are from a guy named Colin Fraser :
I guess you have a bit of bias here… For example: if instead of asking in a forum, I read an entire book to find the answer, this makes me antisocial? if instead of asking in a forum, I try to figure out by myself, that makes me antisocial?
I respect and share at some point your concerns about LLM. But assuming that someone is antisocial just because the tool that uses, I guess is a bit to extreme…
Nobody said anything about reading a book or figuring things out without help. We’re talking about LLMs. And to repost a link from earlier: yes, yes they are.
Reading a book written by a human is receiving the thoughts of some other human(s). Trying to figure things out by yourself can be a good learning experience; but past a certain point it would be silly not to ask for help. And asking for help, as in “holding a conversation”, with a machine instead of a human? Yes, that’s antisocial.
If LLMs were like search engines, and only gave you exact copies of things that other humans had put time and thought into, that wouldn’t be antisocial (at least not inherently.) But they don’t, they generate new things, and that’s the whole reason they’re touted as being so much better than search engines.
The pros and cons of AI use on Haiku according to Gemini.
Pros
Alleviates Developer Burnout: The Haiku project has a very small pool of active maintainers, and some code submissions wait years for review. AI can act as a force multiplier to handle routine maintenance
Boosts Ecosystem Software: Haiku famously suffers from a lack of native applications. Proponents argue that Large Language Models (LLMs) make it easier for novice developers or non-coders to successfully port and write new software for the platform.
Accelerates Repetitive Tasks: For writing straightforward elements—such as system screensavers, simple application refactors, or boilerplate configuration code—AI can dramatically speed up development.
Maintains Operating System Viability: Some community members believe that if Haiku completely ignores modern machine learning capabilities, it risks falling behind and feeling obsolete to modern users
Cons
Legal and Licensing Risks: Haiku core developers have raised major concerns about copyright. Because AI models are trained on varied datasets, accepting AI-generated code introduces the risk of incorporating incompatible or legally ambiguous software licenses into the operating system
Code Bloat and “Slop”: Haiku is celebrated for being incredibly fast, lightweight, and cleanly written in C++. Opponents argue that AI often generates bloated, sub-optimal, or buggy code that takes humans longer to debug than writing it from scratch.
Loss of Deeper Computing Knowledge: Many developers worry that relying on automation will cause newer contributors to lose their fundamental understanding of how operating systems and low-level code actually function.
Philosophical Clash: Part of Haiku’s charm is its independence from massive tech conglomerates. Integrating modern generative AI tools connects the open-source OS to resource-heavy corporate infrastructure, which directly contradicts the minimalist, self-contained ethos of the project.
If that is actually just Gemini output, then it’s specifically against forum rules: “Don’t post content generated by large language models (“LLMs”) or similar tools.”
I’m gonna have to ask you to edit that post to:
I put it into a “hide” box for the moment, but please do edit it; the humor value of having the LLMs fight this out among themselves is not lost on me. 
Congrats on your first mod action! I hope we won’t keep you too much busy in this topic 
Even as a convinced atheist, I’d like to point to something from Pope Leo XIV’s recent encyclical Magnifica Humanitas (15 May 2026).
The encyclical is not an endorsement of AI. The Pope calls for vigilance, prudence, and even a slower pace in adopting it. He also says clearly that AI is not morally neutral. I’m not pretending he agrees with me on everything; he doesn’t.
But on one point he is very clear, and it’s the point I’ve been trying to make here: decisions of this scale need open discussion with everyone, not a verdict reached in a narrow venue. If even a voice that is broadly cautious about AI argues for an open conversation, then perhaps that is also what we owe each other in this community.
That’s all. =)
I don’t follow you here.
Sure, there are differences between how LLM process and generate language, and how humans do. Sure, the LLMs don’t think, they do something different. But, clearly, the language is the same. The output of the LLM is related to the input, and we can read and understand it.
This is not very different from sourcecode. That is designed for both humans and machines to read and interpret, yet, we do so in very different ways. Why not do the same with existing human languages? That doesn’t require any supposition that LLMs are conscious or can undertand something. They run an algorithm, and that algorithm can process human languages.
This would be a great technology if it wasn’t so inefficient, and if there was a reliaple way to interface it with the other things we typically do with computers, that is, traditional fully predictible algorithms working on well-defined data. Here, we may hit a dead end, and LLMs will bring us no closer to bridging the gap between humans and their ambiguities, subtext, non-verbal things, and the cold hard world of computer algorithms.
There is also a lot of marketing and hype built on people mis-interpreting the LLMs output as intelligence, consciousness, or any other human traits. Here, there is a challenge: we are not equipped (culturally, or philosophically) to deal with this. There is now something that masters human language, but clearly isn’t and doesn’t operate like a human. Can we use it to communicate with computers more efficiently? Does “communicate” even make sense in that context? Pretty interesting questions. And maybe the risk of confusion is dangerous in some ways. Some people will lose their sanity or lose touch with reality. So we should proceed with caution.
But to go as far as saying we should not even be trying this? Isn’t it going a bit too far?
The language, the “results”, on the level of letters and words, may be the same. And there is a strong bias in the present day to ignore anything else but results, because anything but results is often much harder to make a determination about, so it’s easier to just avoid it. But that doesn’t mean those other things don’t matter either.
If I, a human, see a tree, and say “that is a tree”, I am conveying my thoughts in speech. If I write code, I am conveying my thoughts in speech. The words are signs, and what is signified is thoughts (e.g. the word “tree” signifies the concept of, well, a tree); and so when another human reads those words, and knows the same language I do (and I used the language properly), then you will have substantially same thoughts I did: “that is a tree”, or “this is a quicksort algorithm, which operates by…” etc.
So, therefore, if as you say (and I agree):
i.e., LLMs don’t have that, then it means that when an LLM generates the statement “that is a tree”, or “this is a quicksort algorithm”, it doesn’t use those words to signify thoughts, because it doesn’t have thoughts. And so what it is generating looks like human language, and is perceived as human language by humans … but the thing it came from did not intend it to have any meaning, because the thing it came from is a machine, and so the thing it is generated is “false” or “fake” language: because it uses the signs (words) that humans use to signify (and thus communicate) thoughts, without having thoughts to back up those signs.
This is what I mean by “category error”. Programming languages, same as mathematics notation, are designed such that they can both be understood by humans as thoughts, and mechanically transformed by well-defined algorithms into executable operations. Having machines process them as executable operations, and generate other operations from them, sounds fine. But processing them as human language, with abstract thought and everything else it entails, is something machines simply cannot do.
Logically, I think the conclusion from all that would be that indeed we should not even try such things, because they fundamentally break basic assumptions about what “speech” is, where it comes from, and what it is for. They fundamentally erode the meaning of the very thing we all use to communicate with.
I’m putting this here too, to make sure you see it before you post next; you do need to include the original text with any machine-translated posts.
One solution that was suggested to folks using automated translation systems was to ask them to post the original non-English version below the English; the “hide” box in the forum editor would be a good container for that.
Can you do that in the future, please?
(I will raise the issue of having this exception added to the no-LLM-output policy.)
@us3r1d Ha, thank you, fair and practical suggestion, and a sneaky way to get you to pick up some Italian along the way. 
Happy to follow it from now on, and thanks for raising the policy point on my behalf, that means a lot.
Let’s test the convention straight away:
@us3r1d Ahah, grazie, suggerimento giusto e pratico, e un modo furbo per farti imparare un po’ di italiano cammin facendo. Felice di seguirlo d’ora in poi, e grazie per aver sollevato la questione di policy per mio conto, significa molto.
Testiamo subito la convenzione.
Is anyone opposed to adding a “contrib” repo for software outside the scope of the non-LLM policy? Similar to what Debian based systems have.
So we all can peacfully co-exist?
I think, at the moment, the haikuports mailing list has shelved the discussion on whether to have a policy about LLM-assisted code in ported software.
Frankly, I would think everyone would be ecstatic to see a new repo organization formed that welcomed LLM-assisted code.
(I suggested this already; I even offered to help with the setup and registered org names for such a thing at github and codeburg and I will hand them over to anyone who wants to give it a shot.)
The only real problem is that haikuports does a lot of work, since they distribute fully-built packages. So it takes a fleet of automated build systems as well, rather than just a git repo.
From my experience, LLMs greatly changed how I approached problems, but it did not change how often I collaborate.
I can ask all the “noob” questions to the LLMs and prevent wasting other people’s time. I can ask Gemini what AVX-512 is, then ask Copilot where in the Haiku codebase is AVX and other x86 initialization done.
But I cannot just ask the LLMs what the best way to integrate AVX-512 into Haiku is. It is an art that we work together on.
I need to address two specific points that have crossed a personal and ethical line for me.
It is incredibly easy for a native English speaker, on an English-speaking forum, to philosophize about the “purity of human speech” and criticize translation tools. I have a large English vocabulary and can read everything here freely. But reading and formulating complex, nuanced thoughts in a non-native language are two very different things.
What is the alternative you are proposing? Should I pull a giant paper dictionary out of my closet and spend half a day manually translating my thoughts, just to satisfy someone’s philosophical aversion to AI? I am not going to do that. My manual translation, due to lack of conversational practice, would be 100% worse than even the worst machine translation. The reality is simple: without modern translation tools (which are practically all neural-network-based today), I simply won’t write anything on this forum. If the goal is to filter out non-native speakers and create a native-English-only club, then you are on the right track.
Regarding this policy of forcing users to label LLM-assisted text and hide it under spoilers: do you realize what this practically means? Since I use machine translation (which has AI under the hood) to communicate clearly, it means I would have to label and hide absolutely every single post I make on this forum.
Do you know what this reminds me of? In my country, the government forces journalists, activists, and dissidents to prepend a mandatory “Foreign Agent” disclaimer to every single post they make, in order to stigmatize them and mark them as “undesirables” or “not entirely ours.”
Seeing an open-source community adopt a similar tactic of forced stigmatization - asking people to wear a digital warning label or sit in a “quarantine box” just because they use modern tools to communicate on equal footing - is a disgrace.
I will not participate in this practice out of principle. I prefer to be judged by my actions, my ports, and my code, not by the tools I use to bridge the language barrier.
You are making that sounding like there is a magical process behind the “thoughts”.
It is likely that during your childhood, your parents, your teachers, or anyone around you taught you that the object with this specific appearance is a “tree”. The whole process formed neural pathways from your vision receptors to the part of the brain that handles the abstract concept of “trees” and maybe another pathway between that concept to the English word “tree”.
I then see the word “tree” and link it back to my own abstract thought of a “tree” (based on what I learned) and understand what you are referring to.
The same goes to all well-defined concepts like a “quicksort algorithm”. But for more abstract concepts, maybe we link the English words to entirely different concepts, depending on the experience that shaped our neural pathways.
LLMs kinda work similarly - the input will trigger certain nodes in the graph, resulting in the output. The only difference is that our own neural pathways are much much more complex than a few billion parameters on a data center.
I therefore don’t think it is a good idea to reject LLM output entirely. I personally treat LLM output as any “untrusted” source of information - allow it to make a claim, force it to cite trusted sources, and then form my own thought.
First, thank you for taking the time to translate your posts into English; I have no idea what your native language is, but I’m quite certain that I can’t read it.
But this? This is a bit over the top.
Asking people who use automated translation systems to post the source text is not stigmatizing, and I have no idea where the “quarantine box” bit comes from.
It’s asking for a courtesy to the reader of letting them refer to what you actually meant to say when there’s ambiguity or when something appears to be insulting or violate a forum policy
Having the source text is necessary for moderation; we need the source text to actually understand the poster’s intent when there’s an issue.
“Hey; the machine translated it that way. I certainly didn’t accuse anyone’s mother of that.”
(paraphrased, not anyone’s actual response here)
That makes it impossible to moderate fairly.
This wouldn’t be necessary if Discourse had a translate button like Mastodon has; then everyone could just post in their native language and expect the reader to deal with translation.
I will promise to look into whether this is available and we can add it here, but it’s not here right now.