Discussion
I know LLMs are interesting as toys, but they are fundamentally problematic:
- Even if you self-host an LLM, its model is created unethically elsewhere so it's not really independent or ethical.
- LLM creation is extremely resource-intensive, which is burning the planet and destroying communities.
- Hardware prices are being driven so high that ordinary people cannot afford to buy computers.
- LLMs are designed to deskill people and make them dependent on corporate models.
@homegrown Imo, LLMs are okay as long as it doesn't replace your search engine or your thinking, but that's only my opinion
@homegrown i personally find LLMs laughable, useless but terribly unethical. as a creative person the blatant unethical issues are so clear that i still feel shocked when i hear people saying "its a revolution, its here to stay", as if corporate theft should just be accepted, while no serious "gain" was made in any creative field (music, illustration, animation, whtever). And yet, while "AI" is hailed as the "future", it's most recognizable output is Slop.... as a non-believer this is amusing.
I am not disagreeing with this in the context of massive LLMs like Claude et al, but there is no fundamental requirement the dataset be rife with theft. My 2005 generative music experiments used public domain scores for training and while it took a while, all ran on a lowly desktop. True, these were tiny and special purpose, like protein folding, but essentially the same. 😌
The environment impact of the commercial LLMs is due to scale and a greed for speed.
As for the chat aspect, we already knew this was a bad idea in 1956, only what was unethical then is a celebrated business sense today 😔
I think we may be talking at cross purposes? 🙂
The original post is only intended to be specifically about LLMs and their GenAI cousins. I don't mean all machine learning or procedural generation etc.
@homegrown@social.growyourown.services I feel like you're posting this to see what camp people fall into and that's kinda shitty.
Anyway, I don't support LLMs for MOST things. I do find they are remarkably good at making fictional stories for entertainment. That's... really it.
My two cents.
"I feel like you're posting this to see what camp people fall into and that's kinda shitty."
I can promise you I'm not. LLMs are just awful unethical tech.
"I do find they are remarkably good at making fictional stories for entertainment. "
LLMs aren't making anything. They're jumbling up bits of writing that have been stolen from uncredited human authors.
LLMs are laundering plagiarism. Take away the human data and LLM output turns to gibberish https://en.wikipedia.org/wiki/Model_collapse
@homegrown@social.growyourown.services Yes, I'm aware of this.
It doesn't change the fact that they are good at generating fictional stories for entertainment purposes.
I do not support LLM's, just to be clear.
They aren't generating them, they're stealing them.
They're not good at anything except repackaging work by human beings.
now do where the money for most great American museums came from
p.s. Because of the above, I do not think LLMs are compatible with the principles of self-hosting or community hosting.
Normalising LLMs is effectively handing over our brains and computers to massive corporations owned by amoral billionaires who just want power, and don't care if the world burns as a result of it.
LLMs are being openly promoted as a way to kill personal computing and human intellectual skills. Sam Altman openly says he wants to sell "intelligence as a service".
@homegrown Don't overexaggregate. LLMs are evolving rapidly, both in accuracy and in efficiency. Also, if you want to keep data private, don't expose it to internet. Companies had that data anyway and used it anyway. LLMs are just the latest thing they used it for and it's a bit more obvious.
p.p.s. Lots more problems with LLMs:
- LLMs are unreliable as they frequently produce extremely wrong answers, pizza recipes with glue, incorrect medical summaries etc.
- LLMs are entirely dependent on stolen human data. If LLMs have human data cut off, their output turns to gibberish.
- LLMs can be easily used to astroturf any online discourse.
- LLM owners can easily manipulate their users with misleading answers.
- LLMs are killing the websites they steal data from.
etc.
@homegrown One CAN run their own LLM with the AI query from end on their own hardware, with their own curated data, ethically. As it scales up it is likely to get quite slow, but may be mire useful (and FAR, FAR more ethical) than the corporate services.
@homegrown I have an 11 year old and as we tutor him in math we all had to re-learn things like long division all over again. The existence of calculators "de-skilled" me such that I no longer knew long division. are calculators unethical?
or if you prefer, twenty years ago, if you didn't speak French, and had to speak to someone who only knew French, you had to either learn French or hire a human translator. But tools like Google Translate came out that could do it using software. is Google Translate unethical?
"are calculators unethical? "
They would be if you were forced to rent calculators from massive corporations, and if maintenance of calculators used up more energy than entire countries etc.
Also, by the way, calculators are reliable and predictable in a way that LLMs are not. You cannot rely on an LLM in the way that you can with a calculator.
"is Google Translate unethical?"
It would be if it was causing massive spikes in energy use, water use, population displacement etc.
@homegrown home irrigation systems use 9 billion gallons of water per day in the US. Datacenters perhaps 500M gallons per day. Going after lawns would be a way better way to save water.
@homegrown Because as I already said, I can run a Gemma4 model on my PC with an AMD Rizen 7 5700G with Radion graphics, 32 GB RAM and 2 TB storage with no issues, and once that model was created its not continuing to use a massive amount of resources just by a few people running it on their own hardware, so by that logic its bullshit to say that people shouldn't and cannot selfhost LLMs on their own hardware.
> Even if you self-host an LLM, its model is created unethically elsewhere so it's not really independent or ethical.
based on what? i'm curious on this one.
Based on the fact that creating a practical Large Language Model takes so many resources that only corporations can afford to do so.
Self-hosting doesn't create the model, and it's large model creation that is by far the biggest problem.
@homegrown that's fair, but not entirely accurate. it's entirely possible to train a small focused model on consumer hardware. many of us have done this. but i get what you're saying.
Yup, true, it is technically possible to make small highly restricted models from scratch for niche purposes and/or as educational examples.
The problem is LLM users by definition want large models, and this makes them dependent on unethical corporate models.
Install bonfire.cafe
Get the full app experience
![jae[0]™](https://media.bsd.cafe/bsdmmedia01/accounts/avatars/115/004/535/892/913/622/original/9d78aced5693a345.png "jae[0]™")
