It would be much better to use LLMs to create ordinary programs than to use an LLM over and over to emulate something that's thousands or millions of times more efficient.