His argument is that the LLM is the query engine. By that logic you can approximate anything since LLMs can.