Historical analogies for large language models