I’m going back in time to give the luddites a mecha.

I hope I’m right about how these things work because I do not understand any of the math or programming behind them. All I know is that when a person draws a fucked up hand it’s because hands are really hard to draw due to hands being in all kinds of different shapes based on what they’re doing, and having many many different planes and angles and perspective doohickeys. But a person knows a hand is a hand. Show them a human hand with a five fingers, or six fingers, or a steel hook, or a grasping robot clow, or a lobster claw, or whatever, they can identify all of those things as a category of objects called “hands”.

And the machines don’t do that. There’s no indication that the machine has a concept of an abstract object called “hand” that is the same object regardless of it’s actual appearance in a 2d image. There’s no evidence of a concept of a squarish meaty bit with four thing meat sticks and a fat shorter meat stick. It’s just blurs that resemble figures because the machine doesn’t “think” or “know” or “represent” anything when it outputs these images.

And if I’m wrong, and there is something thinking in there, and it just fucks these things up because it’s entire perceptual world consists of datasets with no attached information or context? Fuck me, idk. The whole thing freaks me the fuck out because we really are getting in to the deep water of “If we did encounter a non-human intelligence how would we even know?” The people who think these things have human-like intelligence are massively ignorant dorks but I keep having spooky dreams about symbioses between trees and mushroom networks or the way ant hives engage in complex problem solving based on the simple, rules based action of individual ants.

I’m genuinely frightened that they’re going to fix the hand problem soon.

I hate to break it to you, but that’s been varying degrees of solved for over a year. Most AI slop that just contains ludicrous basic errors is left unfixed because the people who spam AI slop are principally just incredibly stupid, lazy chuds slamming the treat button like it’s a gacha pull, but even something as basic as automatic1111 or its relatives allowed manual inpainting to reroll broken parts of the image, then added optional automatic secondary processes to detect faces and hands and run an automated inpainting process to fix them with dedicated “fix faces and hands” models.

Now currently relevant open models range from “sometimes it doesn’t fuck up” to “the only tell is that proportions are subtly off and that sometimes there are continuity issues that give away that there’s not any sort of reality behind the image nor even an internal model of what is there in places you can’t directly see, but it’s less glaring than it used to be” and it’s only getting more horrifying. Looking at “realistic” models always makes me feel sort of sick, even beyond the rote response I feel whenever I look at civitai.

I’m pretty sure the hand problem will be fixed, all right… while breaking the planet we live in a little more in the trade.

This entire Rube Goldbergian machine complex spends so much electricity and demands so much water and expels so much pollution just to try to convincingly trick us into believing it’s real, all because that’s the mandate of the owners of the damn things, and what’s expected from the slop hogs at this point.

Asking someone to paint a picture might involve paying a worker. Fuuuuuuck that, this is THE FUTURE.

Isn’t that the hand of the kid on the right? Dad’s hands seem down at his sides and under the table.