Do we know how GPT4V tokenizes? You’d probably need the blocks to match the token size?