Then you should be able to make a markov chain generator without deep neural nets, and it should be on the same level of performance as current LLMs.
But we both know you can't.