I think you can decode a stereogram into the depth map with like 5 lines of Python, but it's been a while since I've played around with this stuff.

Meaning, maybe GPT can't (yet) see magic eye images, but it shouldn't have trouble building itself a prosthetic that allows it to see them.