when the model predicts the next token, it doesn't predict the next bit.

I'm saying, eventually, it will... perhaps :)