Thinking of language models as actors is an effective way to understand their quirks, predict their failures, and get better results when using them.