Comment by teruakohatu

Right now English language LLMs favour more tokens over fewer tokens and more unique tokens. For example in GPT4, “ Altogether now” is four tokens.

I don’t think visual information density is much help.