You got to look at how it scales. LLMs have already stopped increasing in parameter count as they don't get better by scaling them up anymore. New ideas are needed.