I also like comparing a human thought experiment like Einstein (?) would do to forcing an llm to write code to answer a question. Yes you can make a good guess, but making many smaller obvious decisions that lead to an answer is a stronger process.