AI bros... it's so fricking over...
 Ape Out Shirt $21.68 |
 UFOs Are A Psyop Shirt $21.68 |
Ape Out Shirt $21.68 |
AI bros... it's so fucking over...
AI bros... it's so fricking over...
Ape Out Shirt $21.68 |
UFOs Are A Psyop Shirt $21.68 |
Ape Out Shirt $21.68 |
But bros I thought self attention is good?
So, So, So, So, So, I'm fricking saying So, So, So,
homosexual homosexual fgt gay
I have similar qualifications to this anon but I also know hashmaps I also coded with the software I have similar qualifications to this anon but I also know hashmaps I also coded the ability to work
What's wrong with the so word
it's a soi word, like "though" at the end of sentences or "y'all"
I type whatever the frick I want though
Ya'll are a bunch of Black folk
he only said "so" one time
And And And And So, fricking I'm So
So, I was looking at this twitter screencap thread and I wanted to yell Black person at the top of my lungs.
Anyone find the parallels between AI and inbreds to be hilarious?
I'd say it's more comparable to the idea of generation loss. Wait.... can inbreeding be mathematically comparable to generation loss? More studies are needed. I smell a paper.
it's more like losing your imagination as you age
source ?
I wish people BOTentlemen read more papers than solely living off of twitter/leddit threads between their gaming breaks. Shame that bot doesn't care as much for AI
https://arxiv.org/abs/2305.17493v2
>The Curse of Recursion: Training on Generated Data Makes Models Forget
>In this paper we consider what the future might hold. What will happen to GPT-{n} once LLMs contribute much of the language found online? We find that use of model-generated content in training causes irreversible defects in the resulting models, where tails of the original content distribution disappear. We refer to this effect as Model Collapse and show that it can occur in Variational Autoencoders, Gaussian Mixture Models and LLMs. We build theoretical intuition behind the phenomenon and portray its ubiquity amongst all learned generative models. We demonstrate that it has to be taken seriously if we are to sustain the benefits of training from large-scale data scraped from the web. Indeed, the value of data collected about genuine human interactions with systems will be increasingly valuable in the presence of content generated by LLMs in data crawled from the Internet.
Model collapse is inevitable for AI especially if it depended on cheap web scraping
reading papers costs money wtf this is BOT - consumer technology
if it effects our ability to coom and play vidya, we're not interested in spending
most of this board is NEET and/or robots
BASED. MY JOB IS SAVED!!! SUCK MY wiener CHAT GPEET
For now, maybe.
Didn’t read
Non sequitur.
The fact that people can do this (train on AI-generated content) doesn't mean it's what's being done right now.
duh
I've seen artcels claim this a few times now but never seen proof
you don't need proofs on twitter
just post something outrageous enough and people who want to believe it will
>he posted, on BOT
> get overdosed on digital art
> loose sensibility
oh wow, who would have guessed? he just forgot to mention that human digital art is also affected
Models mark their outputs so they know not to learn from them if they show up in future training data. That's the entire reason adversarial stego works.
So ummmmm apparently you basically already made this thread and you're ummm apparently too stupid to come up with another one?? Is this the power of human "intelligence"?
>GUYS PLEASE DONT MAKE ANYMORE AI ART OR YOUR AI ART WILL BE RUINED TRUST ME
okay
That's a pile of shit. People only post AI art that is good quality. Which means the ai art that gets posted online is already pre-filtered and is of higher average quality than the average quality of unfiltered output.
you are wrong on all accounts
HAHAHAHA AI BROS BTFO
>so
wouldnt the opposite be true?
if you think about one of these sd models as having a large number of possible outputs, if ppl start uploading all the best outputs to the internet and a model trains off of that wouldnt that just improve the model to exclude bad outputs even more?
>3.7M views
>76k likes
The more I that I actually take the time to learn how AI actually works the more and more obvious it becomes when you see people completely talking out of their ass
We trained AI with AI so the AI can emulate AI.
>People post the best outputs from SD
>All the best output begins to get incorporated into newer models
>newer models biased by the better output
>newer models produce outputs with better training
???
AI art probably needs smoothing to remove AI atrifacts before being used in training.
The artifacts likely frick up the training process and make the AI art more like adding random noise to the training set.
It's amazing how quickly AI goes to shit when you try to train it with AI pictures.
I tried to supplement a model with AI pics that had insufficient training data and just by throwing in 20% of AI pics it fricked up the model completely.
Even if the data you use looks as realistic as possible, it somehow still manages to screw up the results.
It's interesting to see how this plays out because art sites for example have gotten absolutely saturated with AI pictures and so have places like Pinterest.
If you scrape the web for that data it's not going to work.
This is going to lead to only well thought out models with hand picked training data standing out of the shit tier noise that has been introduced into the system.
It's not just the noise it's the end result that still sucks balls. You can always tell when a picture has been made with AI.
Throw those pics into the training data and you're in for a shit time and then do it few times over again and it's a disastrous result.
kys israelite
Just save it as jpeg with 50% quality before feeding it back
Problem solved
But Stable Diffusion is supposed to add the invisble watermark to generated images in order to avoid training on those images. However, there have been many Stable Diffusion models trained on Midjourney images. Sure, it's not generations deep yet, but they look fine.
Therefore?
>It's not just the noise it's the end result that still sucks balls. You can always tell when a picture has been made with AI.
I'm not talking about noise in the image. An AI generation isn't created with the same intentionality as a human drawing which means the ques that the trainer looks for won't be as relevant.
This is what I mean by noise. Even if you pick the best looking generated images, at the data level they're attributing differently to the set.
NonDB prog
>oh no the DB prog has access to that DB
Tragedy
>I was told that I would have to worry about an autonomous drone shooting a hellfire missile at me.
>Instead, the problem is shitty AI-generated pictures based on other AI-generated pictures.
This isn't actually happening, it's just bitter artist cope.
lol
why down syndrome fingers?
every day is repost day
YOU think the singularity is in 2033. I think it is in 2031!
Source: trust me bro
Generative AI? More DEGENERATIVE AI.
>the programs are now starting to pull from it
Is that how that works
yes, the age of free web scrapping and data harvesting are coming to an end.
no more free lunch from there
Anyone who has been on the internet more than a minute knows there is a lot of shitty art around.
Whether it is by curating the data set they train on or detecting in code the quality of the image, they'll find some solution to generated images poisoning the training.
Yeah, I downloaded a bunch of images from one of the boorus to train a model on and a lot of the art is so shitty I can't use it. I don't know how these "artists" really have room to criticize anything.