A silly fops rants about LLMS... And Palworld for some reason.
Posted: Fri Jan 16, 2026 12:11 pm
Okay, so this is kind of a follow-on rant stemming from a conversation that happened in over in the Bark Furry Gayming chat. Some points brought up kind of bother me and it's taken me this long to process and come up with why they do. I want to put those thoughts out there to see if they make sense to anyone else.
I want to be clear that I'm not calling anyone out here; rather I'm gonna bring up some quotes and use them as a jumping-off board to talk more generally. This is gonna be another of my LLM/Generative 'AI' rants (I should really get started on making that blog so I have somewhere to put these things that isn't angy chat or here lol)
And I disagree. Using LLM weights implicitly supports its development continued use. At the same time however, I can't fully say that it is an objectively wrong position to take. If you want to do anything with computers, you basically need to have a similar style of compartmentalization. You can't have lithium batteries without lithium mines, no ethical consumption under capitalism, etc. I think it's quite important to remember that, just because you don't personally have a hand in it, doesn't mean that exploitation isn't happening. Our participation in the systems of exploitation may be unavoidable and involuntary, but I but I don't believe that excuses dismissing it as basically unrelated to us, especially when the existence of those systems depends on people like us unknowingly supporting them.
I'm typing this on a computer. That's only possible because this computer was produced by, at best, ethically dubious measures. I like computers and the internet and I think they have been and will continue to be an overall good thing; but I can't separate that fact in my mind from the truth of how it came to be, and the bad that it can be used for. So I take steps to mitigate the harm as much as possible; I don't buy into latest generation hardware. I try for refurbished or used when I can, and I don't see a point in upgrading unless I need to (The hardware I'm running is -- on average -- almost a decade old at this point.)
I think, as people striving to be good, taking steps to reduce harm as much as possible is necessary, and I view using an LLM as counter to that goal. And yeah, a lot of LLMs I've seen so far are based at least in part off of stolen work.
OpenAI's ChatGPT is fairly famous now for having stolen work in its training data. CoPilot is 'powered by OpenAI'. Even LLaMa, the makers of whom proclaim to use only publicly available works, used sources such as Common Crawl dumps and Books3 and OpenSubtitles, all of which contain copyrighted works used without permission. Claude is vague enough about it's sources that I can't tell for sure, aside from a vague 'trust me bro' vibe when it says it's webcrawler avoids password protected sites and 'robots.txt' (A claim which may technically be true so long as you can keep up with the sheer number of AI agents being released into the wild - Claude alone has gone through ANTHROPIC-AI, CLAUDE-WEB, ClaudeBot, and ClaudeUser already and they're under no obligation to announce when a 'new' agent is made and set loose.). Google's Gemini and Gemma only claim to use 'A diverse collection of web text', 'Code', 'Mathematical text', and 'Images'. What, exactly, those are, I was unable to find. I'm not really good at this sort of search, so perhaps I just missed it. Grok just claims 'publicly available' data, and that's an oft repeated claim for other LLMs as well. Is it? How could we find out? It should be noted that 'publicly available' isn't the same 'public domain'.
That's my main issue with LLMs. So far, any LLM that has been forthright enough with their training data that the data itself can be checked, has been found to have stolen a lot of stuff. Now we're seeing models trained on data that includes the output of *other models*, including those made with stolen works. In this environment, is it even possible to use a model that isn't, at best, ethically dubious if you don't make it yourself?
Hey, remember that tangent earlier, about copyright? In my opinion, an actual legal protection against others just straight taking your work and using it without permission that wasn't biased against people who don't have the money to go to court would have gone a long way to stopping this from becoming as big of an issue as it is now, especially paired with a mandatory disclosure of training data sources.
All that said, if you could get an open source model that could be verified to be trained on only public domain or freely available data -- and not the 'technically publicly available because it was on a piracy website' kind of freely available -- and run it locally on hardware you control... Then yes, I would have no issues with the technology. When I first learned of LLMs, that was the direction I had hoped it would take, in fact.
>>be like Palworld
>>I mean honestly, I feel like Palworld wasn't that bad
>>Palworld wasn't plagarism at all, they didn't steal the game from pokemon. At most they copied some of the asthetics
Okay, this one is actually unrelated, but I want to clear up something about that situation. I'm gonna ignore the nintendo lawsuit, because that's based on game mechanics and it's bs you can even trademark or copyright that. I also won't talk about the 'AI' usage claims; I think my stance on that is clear by now. Still...
Firstly, when some people said Palworld plagiarized Pokemon, they don't mean that they 'happened to look similar', or 'have similar themes' or are 'based on cultural icons or popular animals'. They mean that Palword directly ripped pokemon models from the pokemon games and used them, and that there is at least circumstantial evidence to support that.
https://icon-era.com/threads/modders-an ... emon.8960/
My opinion? At a certain point, a work becomes transformative enough to be it's own thing, and where you draw that line is going to affect how big a deal you think this is. I don't personally think there's anything wrong with Palworld copying Pokemon's style -- it's a good style and popular for a reason -- and the Palworld stye is different enough imo. I consider it at minimum transformative if not wholly original. The stuff with the model might be iffy, but as far as I know, not actually proven beyond someone on twitter mooshing the models together to show how the proportions and body shapes line up, and the quote above.
I want to be clear that I'm not calling anyone out here; rather I'm gonna bring up some quotes and use them as a jumping-off board to talk more generally. This is gonna be another of my LLM/Generative 'AI' rants (I should really get started on making that blog so I have somewhere to put these things that isn't angy chat or here lol)
Okay, as anti-LLM and anti-generative 'AI' as I can be, this is actually a good and valid point. If the environmental impact is your main concern, than locally running one and using no more resources than a typical server, than sure, that works as a valid solution. And, obviously, not stealing someone else's creative work is better than stealing it. But something still bothers me about this statement, and it took me a while to figure out why, and it's the use of downloaded LLM weights.What's unethical about that? I'm not burning electricity or stealing creativity. It lives next to me, I have the LLM weights downloaded, I write the software myself
Okay, largely I agree. Copyright is something of a mixed bag for me; actual legal protections from someone just taking your stuff and using it without any of your permission -- potentially for things you don't support -- is a good idea in theory; in practice, legal systems surrounding copyright is biased in favor of people or organizations who have resources. If you're a small creator and your work is co-opted, you're pretty much SOL unless you can somehow drum up enough money to afford going to court or someone or some group is willing to fight on your behalf. This is tangential to what I want to ask here, though; is it true that, after the harm is done and the product -- the LLM weights -- is made, no further exploitation happens if you download and use them?I don't belive in copyright. I think it's wrong, and the way the weights was obtained was wrong. But they exist, and no further effort to exploit anyone happens because I downloaded them
And I disagree. Using LLM weights implicitly supports its development continued use. At the same time however, I can't fully say that it is an objectively wrong position to take. If you want to do anything with computers, you basically need to have a similar style of compartmentalization. You can't have lithium batteries without lithium mines, no ethical consumption under capitalism, etc. I think it's quite important to remember that, just because you don't personally have a hand in it, doesn't mean that exploitation isn't happening. Our participation in the systems of exploitation may be unavoidable and involuntary, but I but I don't believe that excuses dismissing it as basically unrelated to us, especially when the existence of those systems depends on people like us unknowingly supporting them.
I'm typing this on a computer. That's only possible because this computer was produced by, at best, ethically dubious measures. I like computers and the internet and I think they have been and will continue to be an overall good thing; but I can't separate that fact in my mind from the truth of how it came to be, and the bad that it can be used for. So I take steps to mitigate the harm as much as possible; I don't buy into latest generation hardware. I try for refurbished or used when I can, and I don't see a point in upgrading unless I need to (The hardware I'm running is -- on average -- almost a decade old at this point.)
I think, as people striving to be good, taking steps to reduce harm as much as possible is necessary, and I view using an LLM as counter to that goal. And yeah, a lot of LLMs I've seen so far are based at least in part off of stolen work.
OpenAI's ChatGPT is fairly famous now for having stolen work in its training data. CoPilot is 'powered by OpenAI'. Even LLaMa, the makers of whom proclaim to use only publicly available works, used sources such as Common Crawl dumps and Books3 and OpenSubtitles, all of which contain copyrighted works used without permission. Claude is vague enough about it's sources that I can't tell for sure, aside from a vague 'trust me bro' vibe when it says it's webcrawler avoids password protected sites and 'robots.txt' (A claim which may technically be true so long as you can keep up with the sheer number of AI agents being released into the wild - Claude alone has gone through ANTHROPIC-AI, CLAUDE-WEB, ClaudeBot, and ClaudeUser already and they're under no obligation to announce when a 'new' agent is made and set loose.). Google's Gemini and Gemma only claim to use 'A diverse collection of web text', 'Code', 'Mathematical text', and 'Images'. What, exactly, those are, I was unable to find. I'm not really good at this sort of search, so perhaps I just missed it. Grok just claims 'publicly available' data, and that's an oft repeated claim for other LLMs as well. Is it? How could we find out? It should be noted that 'publicly available' isn't the same 'public domain'.
That's my main issue with LLMs. So far, any LLM that has been forthright enough with their training data that the data itself can be checked, has been found to have stolen a lot of stuff. Now we're seeing models trained on data that includes the output of *other models*, including those made with stolen works. In this environment, is it even possible to use a model that isn't, at best, ethically dubious if you don't make it yourself?
Hey, remember that tangent earlier, about copyright? In my opinion, an actual legal protection against others just straight taking your work and using it without permission that wasn't biased against people who don't have the money to go to court would have gone a long way to stopping this from becoming as big of an issue as it is now, especially paired with a mandatory disclosure of training data sources.
All that said, if you could get an open source model that could be verified to be trained on only public domain or freely available data -- and not the 'technically publicly available because it was on a piracy website' kind of freely available -- and run it locally on hardware you control... Then yes, I would have no issues with the technology. When I first learned of LLMs, that was the direction I had hoped it would take, in fact.
Okay, this attitude frustrates me, because I've seen it a lot online and irl. Because we're just talking about things and not actively building a resistance or trying to organize, criticism should be dismissed as 'circle-jerk'? ~~The Stonewall Uprising was not organized when it started as a reminder.~~ This isn't exclusive to 'AI' either. I've seen sentiments both irl and online that basically amounts to 'Well, you're not actively isolating yourself from literally every aspect of capitalism, therefore you should just shut up about it.' How exactly do you think collective organizing happens, if people aren't supposed to talk about it?Are we organizing collectively right now? Because if we're not, it's just a circle-jerk about AI.
>>be like Palworld
>>I mean honestly, I feel like Palworld wasn't that bad
>>Palworld wasn't plagarism at all, they didn't steal the game from pokemon. At most they copied some of the asthetics
Okay, this one is actually unrelated, but I want to clear up something about that situation. I'm gonna ignore the nintendo lawsuit, because that's based on game mechanics and it's bs you can even trademark or copyright that. I also won't talk about the 'AI' usage claims; I think my stance on that is clear by now. Still...
Firstly, when some people said Palworld plagiarized Pokemon, they don't mean that they 'happened to look similar', or 'have similar themes' or are 'based on cultural icons or popular animals'. They mean that Palword directly ripped pokemon models from the pokemon games and used them, and that there is at least circumstantial evidence to support that.
https://icon-era.com/threads/modders-an ... emon.8960/
Secondly, Pokemon has a distinct visual style. Palworld isn't even the first non-nintendo game to have cute critters as a feature, but it is the first (to my knowledge) to be quite blatantly trying to copy pokemon's style. Some of those early designs were basically just recolors and mashups of different pokemon. Seriously, take a look at a Palworld Cremis and tell me that's not just a big Eevee Gigantomax recolor with a slightly different fluff. Now, mind you, I don't think there's anything particularly wrong with that, but it's hard not to see the similarities.“You cannot, in any way, accidentally get the same proportions on multiple models from another game without ripping the models. Or at the very least, tracing them meticulously first,” one senior character artist told VGC anonymously, adding: “I would stand in court to testify as an expert on this.” They explained: “To give you an idea of how impossible this is, sometimes we have to copy one mesh to another when we make sequels to games, for example, redrawing an NPC from one game to another, and even when we rework those old models, they only SOMETIMES match this closely due to rigging changes that might need to happen. “There have been times when dozens of artists are given the same concept art to create a 3D model, for example, during art tests for jobs. I’ve seen 30 artists try to make the same horse using the exact schematics. “None were as close to each other as these Palworld models are to the Pokémon models. None. The silhouettes and proportions here are near-perfect matches.”
My opinion? At a certain point, a work becomes transformative enough to be it's own thing, and where you draw that line is going to affect how big a deal you think this is. I don't personally think there's anything wrong with Palworld copying Pokemon's style -- it's a good style and popular for a reason -- and the Palworld stye is different enough imo. I consider it at minimum transformative if not wholly original. The stuff with the model might be iffy, but as far as I know, not actually proven beyond someone on twitter mooshing the models together to show how the proportions and body shapes line up, and the quote above.