Hey, you, who's reading this in the future: you can find out who I am today at https://rarf.zone/about/. πŸ’™


E-mail me at:
info@rarf.zone
Mastodon/Fediverse
yiff.life/@katja

I find that I can tell when art lacks the sense of purposiveness but also inconsistency that comes from a being with limited attention and focus capabilities, but also a coherent sense of "concepts" (as opposed to massive statistical maps of a given subject), making art?

Like, a human brain has very different strengths and weaknesses from a computer β€” regardless of neurotype! β€” and those really do shine through in the art we make.

Something like Stable Diffusion understands what features correlate with "large furry dad" "wolf" "in the style of Hirohiko Araki" "HDR" "eating ice cream seductively" "on a vaporetto in Venice" "sunny day", but it just isn't built to have the same working knowledge of lighting, anatomy, architecture, weather, or why people like certain art styles that guides a human.

So it produces something that's pretty convincing β€” but it does it very differently. It just throws monumental amounts of processing power over a relatively short period of time at a titanic statistical model, and by mathematically teasing details out, ends up producing that result.

Meanwhile, a human brain has nowhere near the same breadth of images analysed, let alone in the same degree of detail, and is nowhere near as fast at applying it, but we've got a shortcut: selective attention, combined with having a broader ("general", one might say!) understanding of how all these concepts fit together in a real-world context.

We know what physical features of a big, burly wolf guy are important in this picture. We know how muscle squishing against muscle or fat against fat works, instead of just how it looks in billions of pictures. We know what aspects of Araki's style are important when applied to a huge guy. If we want "high dynamic range", we have better ways of achieving that than what an AI might assume, which is imitating older HDR cameras. We know exactly how to make someone with a canine mouth eating ice cream look as fucky as possible. So on.

So, we allocate the attention we can pay to the important bits. The result might not look as skilled, but out of a need to do things economically, we put our attention where it counts.


You must log in to comment.

in reply to @kda's post:

this is also why i think ai art is interesting in the first place. im very uninterested in "ai art that is perfectly humanlike" since it tends to be extremely average looking humanlike art. i want (and actively seek out) ai art that exemplifies the ways in which the ai doesnt act or "think" like a person. i have stable diffusion locally on my computer, and consistently the most fun and interesting things i can get it to generate are giving it nonsense prompts (sometimes mixed with a few choice keywords to push towards a particular type of nonsense). even giving a totally blank prompt causes it to generate these inscrutible objects that are often so much more interesting to me than the perfect looking Trending on ArtStation lookalikes.

to me, ai is the currently closest thing we have to what talking to aliens would actually be like--beings with concepts that are so radically different from human concepts to the point where it's sort of impossible for us to fully understand them. to an image synth, the entire world begins and ends at "pixels arranged in certain ways"--human concepts of time, space, color, light, material, etc only exist in so far as they can be mapped onto a grid of pixels, which is why imagesynths so frequently create impossible to understand objects--"name one thing in this photo" unrecognizables. they're objects with almost no human preconceptions in them (or alternatively, almost all human preconception in them, including all of the unintensional, unconscious, unrecognized ones)--its no wonder their outputs look so alien, why its hard to make out the details or comprehend the "mind" inside of the neural net. i want more of that and less of "Perfectly Generic Stock Photo Generator" (or, if we are going to make the generator make stock photos, show me the ways in which it doesn't make good stock photos. the ones where the anatomy is messed up, or the geometry is impossible, or the object doesn't resolve into anything).

(i use "act" or "think" or "says" metaphorically here. stable diffusion isn't sentient but it's "simulating consciousness" in the same way a fictional character might simulate consciousness. ie: you can see hamlet how "thinks", see what he "does", hear what he "says", even though "hamlet" is not a sentient being. if you play a video game with hamlet in it, you can even "talk" to the simulated consciousness and have a "conversation" with him. in the same way, you can "talk" to a LLM or see how stable diffusion "thinks").

Yeah, that'sβ€”

A really good way of looking at it, I think? Just generally, technology being used to achieve qualitatively new things, rather than just cutting specialised labour out of a supply chain, is a much better application of it.

Because "something that thinks entirely differently from humans" is great to have for tasks that we're bad at. Not just ones we're good but slow and expensive at.

yeah! computers (and mathematics more generally) already dont think like people, they're great at certain tasks--making them "more like people" seems to be missing the point of computers.