By thinking about recommendation as a generative problem, you can tackle it from new angles and use LLMs to better understand user intent.