Yep it can work for "next-pixel prediction" too where the token is a pixel. If I have enough samples of arrangements of pixels, it's not fundamentally different

Yep it can work for "next-pixel prediction" too where the token is a pixel. If I have enough samples of arrangements of pixels, it's not fundamentally different than samples of arrangements of words. I'll check out that book - thanks!