@janellecshane first few lines: I know one letter, and how to move parts of it around
Next few: OK, here's a couple other letters, and some decorative touches
Last few: how about some kanji?

@janellecshane I thought that this was one of those joke posts, like "When you know you need to get new glasses" kind of thing

@janellecshane When you get stuck in a witch's labyrinth on the way to the optometrist

@janellecshane Pleoped

@janellecshane@wandering.shop that would make a lot of sense, from my understanding these AIs tend to generate large-scale features and then fill in the details

@janellecshane iirc at least one other generative net works by generating the image at like 1/16th the final resolution, then upscaling it using text-based neural upscaling. which would explain this kind of behavior!

@janellecshane this is likely an important finding for #GPT3 and #Copilot too, that obscure things (especially to a WIMPish web culture) will be less well defined. So cat portraits are safe...

@janellecshane amogus row 7 column 4