If humans don't understand it to write the data the LLM is trained on, how will the LLM be able to learn it?