Media glutton, poster, academic? Queer, fat, tired, all that good stuff. I live online. I do a podcast called Flash In The Pan also, it's great.


pervocracy
@pervocracy

I mostly agree with this list but I believe the issue with AI power consumption is that training the models uses a lot of energy, much more than just running a normal server.

My biggest concern with AI imagery is that it pollutes the Internet by subtracting information. Because it's all remixed and doesn't "know" anything, it can't add any real information to the world. If I take a picture of my pet mice, that's a new capture of reality that's being added to the corpus of human knowledge; if a computer generates the statistically predicted image of "a mouse" it's not. An AI cannot create information that is both new and true.


But it can create misinformation! I don't just mean pictures of Hillary Clinton eating babies or whatever, it's mostly not convincing enough for images like that to be a major threat to consensus reality, but every AI picture of a real object contains some misinformation about the object. It won't put the right field marks on a bird, the right buildings in a skyline, the right cliffs on a mountain. It's anti-useful for art reference and research, it might create a doom spiral where AIs trained on bad images get even worse, and... it teaches wrong things.

For some reason Facebook loves to show me images of imaginary cozy cabins. It's learned that cozy cabins are my aesthetic (okay, true) so it stuffs my feed up with these cabins that look superficially realistic but all the details are physically impossible. And each one of those has ten thousand comments that are probably mostly also AIs themselves, but some are real people asking "where is this?" They're trying to learn something about the world, maybe even hoping to stay there themselves one day. The way in which they've been misled is trivial, I guess, but it's still a net loss to what percent of the Internet is true information. And because AI can pump out this baloney in vastly higher volumes than people who actually locate and photograph real cabins, that percentage loss happens fast.

Forgetting what a cabin really looks like is kind of a silly little loss to the collective knowledge of humanity, but it's a loss, and that scares me.

(also, I think the cabin accounts exist to identify gullible people to target them for scams. so that is a concern.)


BTW since I'm on the topic, I think Cohost's new AI policy is fine. It allows for showing AI images to comment on them, it avoids getting non-slop generative art caught in the crossfire, and mainly, spares staff from having to constantly make the judgement of "is this image AI or just someone's art style?" But it also means that there won't be cabin slop accounts on here.

(yeah yeah #1 Staff Defender πŸ… but I get a little tired of how every damn Cohost Corner post gets a million comments about "wow so I guess Cohost is evil now, I'm going to subscribe to Cohost Minus and make you pay me five dollars a month")


You must log in to comment.

in reply to @pervocracy's post:

I mostly agree with this list but I believe the issue with AI power consumption is that training the models uses a lot of energy, much more than just running a normal server.

Yeah, inference (using the trained model) is fairly "cheap" in comparison.

However I think what the Tumblr OP was talking about is that like, a lot of what we already do online consumes a lot of energy. Video streaming is a big one because encoding millions of videos uploaded every day adds up really quickly. But it's a trade-off we're fine with because YouTube/Netflix/Twitch have a more obvious value to people than ChatGPT/Dall-E.

"...every AI picture of a real object contains some misinformation about the object. It won't put the right field marks on a bird, the right buildings in a skyline, the right cliffs on a mountain. It's anti-useful for art reference and research, it might create a doom spiral where AIs trained on bad images get even worse, and... it teaches wrong things."

This is exactly why AI generated images are a huge problem for online knitting and crochet communities of all things. People have been asked to make AI generated objects that cannot possibly exist, and false patterns attached to AI generated images have been sold on Etsy. When you remember that a lot of people rely on the internet for knitting and crochet info in some capacity... it's not great lol

You know, I used to feel like some loser Luddite for preferring my art reference to be on paper, and to gather up weird books with titles like "Meetings with Remarkable Trees," but at least I know these trees exist!

I feel increasingly leery about Googling reference photos for ANYTHING. Good reference is important in my profession!

EDIT: also, the sheer level of power that the companies are reporting needing is obscene. Even if they're LYING, I don't want to just be like, "Haha, it's fine because you're lying! We know you wouldn't ACTUALLY make that demand!" because if it turns out they do... uh.