Nice! It would totally love an auto encoder diagram that somehow captured the idea of “modelability” of the latent space (eg sander.ai/2025/04/15/l...). I have no idea how to do this, though.

I think the answer is basically that while the round trip diagram you have is a nice visualization of the highest level of what an AE does, practitioners typically use the diagram at a lower level. E.g., put a little “fire” icon next to one block to indicate training, show quantization, etc.

Would it be any fun? You could generate stuff but I don’t know how you could it hold together as a consistent puzzle over hours of play.