You are viewing a single comment's thread. View all
8
el_hoovy on scored.co
9 hours ago8 points(+0/-0/+8Score on mirror)3 children
nah it ain't literally JUST the prompt. it's two other things, back in the training data:
1. they are not scraping killallniggers.com for their data, they're scraping Reddit leftist subs. Reddit is a huge huge source of training data, probably Facebook and whatnot too, and that's basically pre-curated leftist training data with how aggressively jewish websites ban dissent.
2. big companies are trying a lot of unique training methods, including something that's basically inbreeding where they have one instance of an LLM judge the training output of another. image generators also do this by having LLMs caption images. at that point you are severely reinforcing what's already there and basically overfitting to what the judging LLM decides - which indeed is probably affected by a prompt to be gay and retarded, but also the training data mentioned before.
it's part of why you can often just *tell* something was written by an AI. it's one big habsburg family. image generators duck around this by having an immense gooner community that uses training data from all over the place and constantly intermix their models. lol, "genetic diversity" might be bullshit for humans, but not generative AI.
Bingo. Grok does it too. In face you can see the sources it references for its "thinking". These things are not operating on logic. They'll fall into the same exact circular reasoning as any off the street leftist or redditor. There's no logic flow or process modeling going on. If there is it takes a back seat.
6 hours ago2 points(+0/-0/+2Score on mirror)1 child
> You can't "intermix" models.
absolutely can, and at least for image generation it's downright common, civit.ai is chock full of merges for people trying to generate their five millionth anime breast. likely not how LLMs are done, though.
i don't merge models, i have no idea. [but here's a plebbit thread about it](https://www.reddit.com/r/StableDiffusion/comments/13dmeoq/checkpoint_merge/). i'm not sure why the idea is so contentious, i'm just describing tech...
Yea.. no.. it's what I thought. You're really just running them separately and then combining them as images in a separate step or steps. If you read carefully you can see where they're describing taking them into another tool where you basically manually stitch them together.
It's contentious because the underlying models mostly don't work in a way where you can _directly_ merge them. You can just train a new one with both sets of training data. The drawback is they get worse at certain things when you do this. So it's more worthwhile to have a bobs model and then a vagene model and then just sew that shit together like frakenstein.
Point being, this shit sucks, and if all it can do is accelerate basement dwelling masturbators into early onset chafed dick syndrome then I think it shows it is a dead end. All the larger youtube channels I'm watching are all taking a very vocal "we hear you, no more AI <anything> here!" People hate this shit, for good reason.
1. they are not scraping killallniggers.com for their data, they're scraping Reddit leftist subs. Reddit is a huge huge source of training data, probably Facebook and whatnot too, and that's basically pre-curated leftist training data with how aggressively jewish websites ban dissent.
2. big companies are trying a lot of unique training methods, including something that's basically inbreeding where they have one instance of an LLM judge the training output of another. image generators also do this by having LLMs caption images. at that point you are severely reinforcing what's already there and basically overfitting to what the judging LLM decides - which indeed is probably affected by a prompt to be gay and retarded, but also the training data mentioned before.
it's part of why you can often just *tell* something was written by an AI. it's one big habsburg family. image generators duck around this by having an immense gooner community that uses training data from all over the place and constantly intermix their models. lol, "genetic diversity" might be bullshit for humans, but not generative AI.
They're scraping everything they can get. They have a use for this data even if they're not presenting it to you.
> big companies are trying a lot of unique training methods
They're publishing a lot of white papers. It's not clear what they're actually doing.
> and constantly intermix their models
That's just another name multiple generation and then hand stitching the pieces together.
> lol, "genetic diversity" might be bullshit for humans, but not generative AI.
You can't "intermix" models. Even if you could the problem is overfitting. Which is why they generate them separately and then hand edit it together.
absolutely can, and at least for image generation it's downright common, civit.ai is chock full of merges for people trying to generate their five millionth anime breast. likely not how LLMs are done, though.
Ok. How?
It's contentious because the underlying models mostly don't work in a way where you can _directly_ merge them. You can just train a new one with both sets of training data. The drawback is they get worse at certain things when you do this. So it's more worthwhile to have a bobs model and then a vagene model and then just sew that shit together like frakenstein.
Point being, this shit sucks, and if all it can do is accelerate basement dwelling masturbators into early onset chafed dick syndrome then I think it shows it is a dead end. All the larger youtube channels I'm watching are all taking a very vocal "we hear you, no more AI <anything> here!" People hate this shit, for good reason.