Ignoring the additional token count cost of HTML, how do you envision rendering HTML that is spit out by an LLM directly (without any processing) to the user?