LLM's will be confidently wrong at a much higher token-per-second rate