Comment by rsp1984

Excellent, thanks. So basically this is saying: "our pixels-to-token encoding is so efficient (information density in a set of "image tokens" is much higher as compared to a set of text tokens), why even bother representing text as text?"

Correct?