It is a deffinert knid of hullunciaation but we hvae drffeenit bisaes tahn GPT-4.
Trehe are ptenly of otehr tpyes of gitlch tkoens for hamuns too. See opitacl iulloisns, or the fomuas peom The Cahos.
But on the other hand, yes, natural language has low entropy and lots of redundancy because it's meant to have error correction. From most sentences you can drop several words, not just scramble them, and the meaning can be inferred from the context.
It also sort of reminds me of the "impossible colors" you can see by depleting the receptors in your eyes to see pure color signals that aren't otherwise visible due to actual light activating multiple receptors due to overlap in the color ranges.
https://www.thoughtco.com/impossible-colors-introduction-415...
For art, just compare any of these images to the pictures in the article:
https://duckduckgo.com/?q=psychadelic+art&iax=images&ia=imag...
The complex houses married and single soldiers and their families.
This "glitch" is a short-circuit pathway from the input token to a result that makes statistical sense to the model's representation of the world, but very little sense to someone operating outside of that statistical model (you).
An internet meme fits that definition too: it's a short-circuit to some specific in-group context.
What do cows drink?
noun
1. a fully grown female animal of a domesticated breed of ox, kept to produce milk or beef. "a dairy cow"
(definition from Oxford Language)
But you're right, in practice people often use the word cow to mean calf.