We now have two examples of systems that can turn text descriptions into code - humans and LLMs. And we don't understand how either of them works.