You have an unstated assumption here that the results and perception of them are what matters. But this is one of the very things under dispute in all these arguments. I at least don’t want to communicate with machines; I want to communicate with people. An LLM that tries to translate “meaning” and not just the actual words/idioms used will alter what the person is trying to convey, often in strange ways.
Even when ‘asked’ to just do ‘grammar edits’ LLMs significantly affect the meaning of sentences:
How can you be OK with non-LLM based translation then?
That can change meaning too, even when not done by an LLM. What is your actual argument here? Based on this either you ban all of them or none of them.
I am getting increasingly convinced that you are coming up with these “principles” as you go, and in reality it all comes down to the fact that you just don’t like LLMs.
There is a difference in kind between deterministic systems and non-deterministic ones; pre-machine learning translation systems were deterministic, LLMs are non-deterministic (because literally no one runs at temperature 0 except solo self-hosters, as far as I can tell). Pre-LLM machine learning systems are oddly intermediate, depending on how they’re built.
These things don’t make a nice, clean sequence but the ends are definitely different things when judged on the result text’s relationship to the source text.
If 2 different runs don’t give the same result text, then you can’t really tell what the origin text was or what it meant just from looking just at the results.
Mechanized literal/phrasebook translation is fine. Yes, I know non-LLM translation also uses “neural net” technology; but I think this is fine as it’s basically being used as a “word network” so to speak. The outputs are deterministically created from the inputs, neither more nor less. LLMs do not behave this way, by design (see the link to the post with screenshots of LLM-based Google Translate hallucinating things, same as any other LLM.)
Yes, and that’s inevitable obviously, because languages aren’t bijective to each other. Humans, though, are free to be more loose with translations because they actually can understand what the other person is trying to say, and so find a more idiomatic way to say it. Since computers don’t “understand” anything, for them to try to reword things based on meaning would be a category error.
Why do you say this? What have I said in here that is actually inconsistent with what I have been saying since I first started trying to introduce philosophical principles to this discussion some days ago, both on this forum and on the mailing list?
If you knew anyone that knows me in real life, they could tell you that I have been against LLMs for these and related reasons pretty much since they became widespread and prominent two or three years ago. It is the case that the Haiku community is a much different one from the ones I’ve had these sorts of conversations with before, and that I am having to try and rephrase my arguments and principles in a way that is more likely to be comprehensible to this community. But my underlying principles have not changed, I can tell you that.
There is very little I “don’t like” for which my (dis)taste is an irrational matter. I have changed my opinions and “tastes” on quite a number of matters in response to logical arguments that shifted my principles. LLMs are not an exception to this. So, no, this is not a matter of “personal/arbitrary taste”.
Can you actually do this with non-LLM automatic translation? You’d need to know what system was used, in what version, what dictionaries it had available and so on and so forth, so you “can’t really tell what the origin text was or what it meant just from looking just at the results”.
Perhaps that is an issue of me not being precise enough - I am not taking human translators into consideration, because it is obvious to me you are OK with translations produced this way.
That statement applies equally to non-LLM and LLM based automatic translation.
Maybe we have a different understanding of the word “principles”. For me a principle is something that you do not abandon as soon as it is inconvenient to uphold it. In these conversations (I’m including HaikuPorts mailing list) I believe you did that multiple times already.
I don’t want to speak to machines
non-LLM automatic translation is speaking to machines
well, not really because they are deterministic
[see first part of this post]
We should consider environmental impact
OK, let’s consider carbon footprint from energy mix
well, we live in complex societies
LLMs are a black box I can’t inspect
Your CPU is a black box you can’t inspect (that was my point with “do you know all the calculations” question)
(And no, it’s not like you give it an input and a deterministic output comes out. It should but it not always does, cf. Pentium FDIV bug. If this was a discussion between accountants, there probably would have been one saying “see, I was right all along, you can’t trust these integrated circuits”)
I generally don’t get this fixation on deterministic vs non-deterministic. Do you use ECC memory in your rig?
Yes. But a mechanistic translation, produced by deterministic rules, won’t try to figure out “meanings”. It will just convert words and phrases from one language to another.
I agree. But I don’t think I have abandoned any principles here?
Ah, well, I don’t think it is. I am giving the machine some text, and expecting the machine to transform that text, word by word or phrase by phrase, by a set of fixed rules. In other words: words are symbols; I expect the machine to operate on the symbols as symbols, and exchange them for other symbols of similar meaning, in a way pre-established by humans (language models pre-LLMs are much more well-understood and controllable, AFAIK, than LLM operations.) If the machine instead tries to replace e.g. one analogy with some other analogy, because my chosen analogy doesn’t work right in the target language, and this replacement was not pre-established in a “phrasebook” (or equivalent) by a human, then this is wrong.
“To speak” in the broad sense refers to try and convey meaning to others by the means of language. Writing text into a word processor and then doing Find & Replace isn’t speaking to machines. Early translation software really was kind of an extremely advanced “Find & Replace”; I am aware the non-LLM but still neural-net-based ones are a bit more complicated here, but fundamentally (AFAIK) that is still approximately what they are doing. So I would not agree with this statement: and that’s why I think it’s fine, while LLMs are not.
This one I wasn’t the one to bring up, I don’t think, but indeed I pressed the point after others did. And I still think we should consider it, yes. But I noted there that even if this was a non-issue (and I noted it could be made one in various ways), my own objections to LLMs would remain, because I don’t fundamentally object to them for these reasons. Even if they weren’t so wasteful, I would still object, yes.
I don’t think this is how I’d summarize my replies on this point. You brought up corporate pollution that we can’t do anything about; I agreed, but was trying to distinguish between kinds of energy use, and ones we can vs. can’t reasonably affect, and trying to show how LLMs are one we can affect much more (and much more easily) than e.g. avoiding single-use plastic bottles. I don’t think I was being disingenuous here…?
I think you are mixing @PulkoMandy and I up now; he was the one to use this terminology. I would agree with it; but in fact he didn’t say that he couldn’t inspect them, but was (I think) implying that nobody can inspect them; they’re a black box even to their own designers. (AFAIK this is indeed true in various ways. They don’t have zero knowledge, but there are parts of even the transformer model itself, independent of any data, that the designers don’t really understand why it works / what it is doing mathematically.)
This is a very big difference between LLMs and any other technology. Indeed my CPU is a “black box” in many ways … to me. But there are engineers (humans!) at the companies that build CPUs to whom it’s not a black box at all, and they understand it inside and out.
If my CPU was beyond the understanding of any humans, not just myself, I would indeed have a problem with that. But it’s not.
What I said about this on the mailing list was:
I am fundamentally asking to ban things on the basis that people cannot be responsible for things they don’t understand.
I don’t understand how my CPU works, but neither am I “responsible” for it, I just purchased it and trust that the company that designed it knows what it is doing. Likewise I don’t know how the Python interpreter works, but I am not responsible for that either; I can trust the developers of Python know what they are doing.
On the other hand, when I write code for Haiku or recipes for HaikuPorts, and submit it under my name, I am responsible for that. And if I don’t understand what I did or why I did it, well, that’s a big problem, because I am supposed to be responsible for it.
One of the big promises of computing and technology in general is that it’s supposed to be “deterministic”: get the same inputs for the same outputs. Now, sometimes you want some randomness, like in cryptography; but in general that’s still at the service of deterministic behavior (you want to get a deterministic message to someone else, only unreadable by anyone else.) Deliberate non-deterministic behavior breaks this promise. And for what?
Yes, actually, I do! I made a point of it when I configured this machine years ago:
$ wmic.exe MemPhysical get MemoryErrorCorrection
MemoryErrorCorrection
6
That bug is a bad exemple, its behavior was perfectly deterministic, just different from the specification. And so are quite a lot of problems in software.
Personally I find that this makes my work as a deâeloper considerably easier, and, yes, I could trace all the instructions at assembly language level, or below, if needed.
Currently, we lack the tools and knowledge to do this with LLMs (because the system is so huge and parallel, typical debugging methods don’t really work), and we also lack ways to “guardrail” them. Since there is a single natural language input, data and instructions are mixed.
This makes them quite difficult to use safely (and there are indeed countless “jailbreaking” examples demonstrating this, which are sometimes fixed by having a non-LLM-based preprocessor or postprocessor where developers can have more control). Now, that’s an argument for me to not use LLMs, but not to tell other people not to. You can do your own risk assessment.
As for the different arguments being exposed: I think yes, we all start from intuitions and then rationalize them, which can take quite a bit of time and sometimes end up showing that the intuition was actually incorrect. This is a vast topic and thinking about it isn’t necessarily easy. So letys help Waddlesplash pinpoint what it is exactly he has problems with (whichemay be a combination of several things). I am learning from that, let’s see where it brings us?
I went ahead and translated some passages with Google Translate classic mode (without LLM) and compared it with Google Translate advanced mode (with LLM). Most of these passages are in German since that’s the non-English language that I’m most familiar with. I also took a passage from one of @anon11892322’s posts. Thanks @anon11892322 for providing it!
Direkt nach der Bekanntgabe des Kaufs von Be durch Palm am 18. August 2001 wurde das OpenBeOS-Projekt gegründet, indem eine Mailingliste mit diesem Namen eingerichtet wurde.
Google Translate classic:
Immediately after Palm announced its purchase of Be on August 18, 2001, the OpenBeOS project was founded by setting up a mailing list with that name.
Google Translate with LLM:
Immediately following the announcement of Palm’s acquisition of Be on August 18, 2001, the OpenBeOS project was founded by establishing a mailing list of that name.
Example 3: German wikipedia article about Haiku (different passage)
Und er sah zur Weckuhr hinüber, die auf dem Kasten tickte.
Google Translate classic:
And he glanced over at the alarm clock ticking on the side of the cupboard.
Google Translate classic when given more of the passage:
And he looked over at the alarm clock ticking on the cupboard.
Google Translate with LLM:
And he looked over at the alarm clock ticking on the chest.
Google Translate with LLM when given more of the passage:
And he glanced over at the alarm clock ticking on the chest of drawers.
Personal opinion: I would rank the LLM as worse than the classic mode overall. However, this may not necessarily hold up with passages of a longer length. Also, the results in different languages may be different. Maybe somebody else can provide examples in their language?
I will try to summarize this and, from my side, permanently close my participation in this discussion. Any further continuation of this topic is just walking in circles and demagoguery.
What I am absolutely convinced of now is that the vocal opponents of LLMs and AI translation here have demonstrated a profound lack of technical understanding of the subject. The arguments against it are mostly based on superficial knowledge gleaned from articles, negative bias, or emotions, rather than actual engineering facts.
Just look at the debate over “classic” vs. modern translation. You argue that “classic” translation (NMT) is acceptable while LLMs are bad. But NMT is also built on neural networks, just older, outdated ones that fail to grasp context—which leads to completely incorrect translations.
Look at the 4 translation options I provided earlier. Only one translator completely failed to translate the Russian idiom, distorting it to mean the exact opposite (translating “I am a bad artist” as “I’m a pretty good artist”). That was Translation C — the “acceptable, classic” Google Translate.
And the translations that @Zardshard pointed out as sounding the most natural and correct? Translation A was DeepL, and Translation D was generated by a LOCAL LLM running on my home PC (Gemma-4-26B-A4B with Q3 quantization).
The conclusion is simple: if we stop using LLMs for translation and rely only on “classic” translators, we will drastically increase errors and misunderstandings in our conversations.
But looking deeper into the problem, to be completely honest: it is frankly none of your business how I write my forum posts or what tools I use. I could be a native English speaker, I could ask a friend to translate for me, I could use an LLM, or I could spend 3 hours writing a reply using a paper dictionary and a grammar textbook. That is entirely my own business and my personal responsibility. It does not concern you or the administration in any way.
I had guessed one or both were from LLMs! lol. Before doing my own tests, I suspected that LLMs would perform better than translation. After doing my tests in the previous reply, I started wondering if they weren’t about equal or even potentially inferior. I suspect LLMs yield objectively better translations in some cases and traditional translation methods yield better results in other cases.
Thanks for your participation! You’ve definitely posed an interesting problem for us to think about. Hopefully there remain others that can represent your beliefs pretty well in this thread. Goodbye!
This part, at least, is not correct: as the forum policies currently stand, a post that’s solely the output of an LLM is specifically forbidden.
That does concern the administration.
This is a (presumably) unintended side-effect; I didn’t write that policy and don’t remember seeing anything about LLM-based translators in the archived discussion I read, so I don’t think this effect was anticipated.
But that is the policy as it is currently written.
None of the moderators trawl the site looking for likely LLM output, drooling over the opportunity to ban someone. Really; none of them are malicious about this stuff. Really. But after your statements in this thread, if anyone flags one of your posts it really won’t be possible to say that it doesn’t violate that policy.
If that’s how it’s going to be applied, I would suggest that he policy is reworded. The reason this policy was set up in the first place was people making a simple one-line prompt and copypasting the reply. That just adds noise to the forum, since the reply does not contain anything of much use or is easy to obtain in other ways.
That is very much not the case for translation. People taking part in a discussion in a language they don’t natively understand [or only to some extent) certainly isn’t a low-effort thing like this. And, usually, these people do haâe some understanding of english, not enough to write it themselves, but enough to check the output of the translation tool.
To me, it makes sense to let people decide how they want to write their messages. And if there should be any rule, it should be about the end-result: do we manage to communicate? Do their participation add value to the conversation? And several people have already demonstrated that, clearly, yes, that works. It’s not as good as direct communication in a common language, but it’s the next best option. And even when it doesn’t work, we can ask people, at that moment, to find someone to help in their own language (and we occasionally do that).
In the context of forum moderation rules, I don’t see how the way a message was written can matter. I can understand banning “agents” which would post without humn supervision (that’s about this being a forum to talk with other people, and also probably about the volume of posts. Experiments in other forums with such agentic posters replying to everything has been a disaster, as you can expect), and I can understand banning people who just copypaste the output of an LLM from a simple prompt (this is about having useful messages that actually oring our own thoughts to the discussion). But translations? I think the human intent and interaction is preserved well enough, possibly better than without these tools, and even if not, it’s up to the person writing the translated messages to decide. They know better what they mean, only them can decide which translation better reflects it.
We can talk about the computing cost, the waste of resources, the closed source nature of the tools currently available, the damage done to the open internet by the scrapper bots. These things don’t affect the generated output, but they still do damage. But then, it’s not really a forum moderation problem anymore (since it’s not about the messages themselves). It’s a wider decision about the ecological and societal impacts of the project as a whole, or possibly of the entire tech industry.
Thanks Zardshard for your experiment with different translators,that was very interesting.
I agree that the output of the non-LLM translators is better,seems closer to the original.
I did the same experiment with the translators I often use,and found that Heexy Translate always has exactly the same results as Google in non-LLM mode,so it’s probably not as independent as I thought
Yandex Translate has different results,but the quality is the same as Google in non-LLM mode I’d say,sometimes I liked the formulation a bit better,sometimes not so much.
Back to the actual topic,I agree with PulkoMandy that banning specific translation tools would go too far and the policy should be reworded in that case.
I think that traditional translation tools that better keep the original meaning are preferable,but as the experiments have shown,LLM-based translations can in many cases still show what the author originally wanted to say.
In cases where LLM-based translations result in sloppy-sounding walls of text and it’s difficult to differentiate it from a totally LLM-generated post,the author should be asked to try another translation service,but I haven’t seen that in the posts of 3dEyes so far,they always sound very clear and human-made.
Translation between German and English usually works well with mechanical tools. From my own experience, this is not always true for other languages like, for instance, Russian, though.
You always talk about resources and the ethics of energy consumption. Is your car electric? Does your home run on diesel or natural gas for heating? Does your electricity provider, which powers your computer, guarantee that the energy comes 100% from renewable sources? Do you produce the energy used to write and read your posts on this forum via photovoltaic systems? If so, then we can discuss the ethical issue of resources. But I don’t think that’s the point of this forum. Although I hope that everyone on this forum talking about energy is 100% covered by renewable sources.
Riepilogo
Vi ĉiam parolas pri rimedoj kaj pri la etiko de uzata energio. Ĉu via aŭto estas elektra? Ĉu via domo hejtas per dizelo aŭ metano? Ĉu via elektroprovizanto, kiu funkciigas vian komputilon, garantias, ke ĝi devenas 100% de renovigeblaj fontoj? Ĉu vi produktas per fotovoltaikaj sistemoj la energion, kiu estos uzata por verki kaj legi viajn afiŝojn en ĉi tiu forumo? Se jes, tiam ni povas alfronti la etikan problemon de la rimedoj. Sed mi ne pensas, ke tio estas la temo de la forumo. Kvankam mi esperas, ke ĉiuj en ĉi tiu forumo, kiuj parolas pri energio, estas 100% kovritaj de renovigeblaj fontoj.
It’s (almost) impossible to not have any negative impact at all,but one can try to reduce it.
No,my car is not electric,but I use it like once a week or once every two weeks,the rest of the time I go by foot.
My home uses natural gas for heating,but I don’t need like 25C all the time,18-19C is enough and I don’t keep the windows open longer than needed.
My energy is from 100% renewable sources,yes,when comparing prices renewable energy is often cheaper than mixed energy here in Germany anyway.
Unfortunately no photovoltaic here because I don’t own the house I live in,so that’s not my choice.
Now the big new datacenters that are getting built everywhere need a lot more energy than any single person,or even all people on this forum together,could waste even if we were trying very hard to.
It’s so much that new gas power plants are built specifically for those data centers.
In many cases,it’s unclear if there even is enough power available to power on newly built datacenters,yet they’re still getting built.
Sure,everyone should try not to waste energy,but we’re talking about completely different dimensions here.
That I still have a old car with fossil fuel doesn’t mean I’m not allowed to criticize a single company for wasting thousand times more energy.
You make a good point, but what If we make the primary focus on us using open-source language models like BLOOM to reject the closed-source monopoly.
We can also look into using search engines like OceanHero that use their revenue to fund environmental cleanup and don’t get me wrong but these tools still consume computing power, choosing open-source and eco friendly technologies really does help the environment a lot.
Which brings me to the point on why don’t we expand Eco friendly value that the dev hold and ai usage that some user hold, to start making the OS eco friendly while still being fast and with ai usage that still uses renewable energy to give us a equal middle ground.
You can’t “cancel out” bad effects like this. Sure you may have a solar panel on your roof, but that didn’t stop AI data centers from re-starting already discontinued coal plants, and that completely ignores the cost of components, cooling (water) etc.
(EDIT: and the economic costs that these Data centers cannot ever recoup the investment, they are build to become e-waste in 2/3 years… so you have the enviromental impact of recycling to contend with aswell)
Data centers here are a massive problem, enviromentally speaking, and no ammount of additional solar panels will fix the direct impact those have on their local communities.
Funnily enough, speaking of “sustainability”, the EU will make claims of “carbon neutral (through compensation)” completely illegal next year.
