Artistic depiction of AI models training on AI generated data.
this post was submitted on 12 Jul 2024
394 points (97.4% liked)
Animemes
2803 readers
856 users here now
Memes related to anime. Animemes.
Rules
- Don't be a shithead.
- Posts must be a meme with anime or related to anime or weeb culture.
- Use NSFW tag for lewd/ecchi. No explicit hentai.
- Nothing illegal, copyrighted, etc
- Repost only if the last post is 6 months old.
founded 2 years ago
MODERATORS
That's a hard-ass album cover
Yeah that djents to the max
Least disturbing 40k lore
Some have called it the Habsburg Singularity.
I'd love for this to be true, but I'd need to see some proof. It feels like wish-fulfillment.
https://arxiv.org/abs/2306.07899 here's a paper that found that one of the biggest sources for LLM training data is corrupted by people using AI to complete the tasks. There are plenty of papers out there that show the effects of this, which they call "model collapse".
While I don't have definitive proof I've seen both AI fanarts that are so good it's hard to tell if it's AI or not and AI abominations that are so bad and artificial it want to make you puke so I guess it really depends on the model.
Funny enough adobe now offers AI generated stock photos that are closer to those abominations rather than anything good. Though if you think about it AI stock art is so pointless... There are already so many stock photos you can choose from. Why would you go out of your way to choose a photo that looks almost the same as regular stock photos but people have 4 arms in it...
Why would you go out of your way to choose a photo that looks almost the same as regular stock photos but people have 4 arms in it…
You never know what smut some people are writing. Maybe that 4 armed horror was exactly what they needed for their cover.
I've heard training on synthetic data is fine now. Most datasets are augmented with synthetic data.
AI "human-centipeding" itself with art and word salads Is going to make confirmed human generated content an extremely valuable resource. That's partly why literally everybody is foaming at the mouth trying to scrape all of our shit. The signal (human) to noise is only going to get worse, The dead Internet is almost here if it isn't already
Age of AI: Asset-flip Regenerated
“Death internet theory evolved”
I mean we knew about this a couple of years ago no?
Everybody be worried about Skynet, and I'm all, we need to trudge through Multiplicity first.
yeah I've seen this before
OP's source is that they want it to be true
More training has been more important than more input, once they hit critical mass for enough input.
Hello, I coined the term "hallucinations" way back in 2019.
I heard "inbreeding" from another author and I've been using that in formal circles and I think it's hilarious.
Yes, we're inbreeding the visual models ;)))))
Well shit, look at that, TIL. Nothing is new under the sun.
Still crazy though!
I used the term "inbreeding" even before that, when finetuning ESRGAN on its own output.
That being said, it really isn't as much of a problem for diffusion models as it is for others.
Anyone who played with a photocopier as a kid could have told you this was going to happen.