Surely the ideal choice for the

> learned token embeddings

for the pixel intensities in M3 would be... a single scalar value between 0 and 1? (That is, a "vector" of length 1, consisting of just the original pixel intensity.)

Am I misunderstanding something? Or is this maybe a case of "Yes, you could just do that much simpler thing in this simplified example, but in general you need the complicated approach that we're trying to demonstrate"?

Really nice article! This clarified a lot to me about how positional embeddings work and how other useful context might be embedded in other models.