Comment by the_arun

That is what I am looking for. a) LLMs are trained using compressed text tokens and b) use compressed prompts. Don't know how..but that is what I was hoping for.