PLEASE JUST LET ME EXPLAIN REDUX
AI {STILL} ISN'T AN AUTOMATIC COLLAGE MACHINE
I'm not judging anyone for thinking so. The reality is difficult to explain and requires a cursory understanding of complex mathematical concepts - but there's still no plagiarism involved. Find the original thread on twitter here; https://x.com/reachartwork/status/1809333885056217532
A longpost!
This is a reimagining of the legendary "Please Just Let Me Explain Pt 1" - much like Marvel, I can do nothing but regurgitate my own ideas.
You can read that thread, which covers slightly different ground and is much wordier, here; https://x.com/reachartwork/status/1564878372185989120
This longpost will;
Give you an approximately ELI13 level understanding of how it works
Provide mostly appropriate side reading for people who want to learn
Look like a corporate presentation
This longpost won't;
Debate the ethics of image scraping
Valorize NFTs or Cryptocurrency, which are the devil
Suck your dick
WHERE DID THIS ALL COME FROM?
The very short, very pithy version of *modern multimodal AI* (that means AI that can turn text into images - multimodal means basically "it can operate on more than one -type- of information") is that we ran an image captioner in reverse.
The process of creating a "model" (the term for the AI's ""brain"", the mathematical representation where the information lives, it's not sentient though!) is necessarily destructive - information about original pictures is not preserved through the training process.
The following is a more in-depth explanation of how exactly the training process works. The entire thing operates off of turning all the images put in it into mush! There's nothing left for it to "memorize". Even if you started with the exact same noise pattern you'd get different results.
SO IF IT'S NOT MEMORIZING, WHAT IS IT DOING?
Great question! It's constructing something called "latent space", which is an internal representation of every concept you can think of and many you can't, and how they all connect to each other both conceptually and visually.
CAN'T IT ONLY MAKE THINGS IT'S SEEN?
Actually, only being able to make things it's seen is sign of a really bad AI! The desired end-goal is a model capable of producing "novel information" (novel meaning "new").
Let's talk about monkey butts and cigarettes again.
BUT I SAW IT DUPLICATE THE MONA LISA!
This is called overfitting, and like I said in the last slide, this is a sign of a bad, poorly trained AI, or one with *too little* data. You especially don't want overfitting in a production model!
To quote myself - "basically there are so so so many versions of the mona lisa/starry night/girl with the pearl earring in the dataset that they didn't deduplicate (intentionally or not) that it goes "too far" in that direction when you try to "drive there" in the latent vector and gets stranded."
Anyway, like I said, this is not a technical overview but a primer for people who are concerned about the AI "cutting and pasting bits of other people's artworks". All the information about how it trains is public knowledge, and it definitely Doesn't Do That.
There are probably some minor inaccuracies and oversimplifications in this thread for the purpose of explaining to people with no background in math, coding, or machine learning. But, generally, I've tried to keep it digestible. I'm now going to eat lunch.
Post Script: This is not a discussion about capitalists using AI to steal your job. You won't find me disagreeing that doing so is evil and to be avoided. I think corporate HQs worldwide should spontaneously be filled with dangerous animals.
Cheers!
666 notes
·
View notes
Panda tea ceremony 🍵
Also you can check out this speedpaint ✨
This post was sponsored by my cringe uni professor👍/silly
(more about it under read more if you’re curious)
Okay so I didn’t plan on posting anything till I draw something tnmn related
But I’m finishing up my semester and one of the requirements to pass my digital illustration exam is
“upload your finished assignment to your youtube channel and generate qr code to it 🤓”
This is literally the only way to submit your work 💥💀
So well
since I’m posting the speedpaint might as well share the art 🐼🍵
127 notes
·
View notes
Prof Scarborough surveyed 55,000 people who were divided into big meat-eaters, who ate more than 100g of meat a day, which equates to a big burger, low meat-eaters, whose daily intake was 50g or less, approximately a couple of chipolata sausages, fish-eaters, vegetarians and vegans.
The analysis is the first to look at the detailed impact of diets on other environmental measures all together. These are land use, water use, water pollution and loss of species, usually caused by loss of habitat because of expansion of farming. In all cases high meat-eaters had a significantly higher adverse impact than other groups.
538 notes
·
View notes