What are the implications of recent advances in code generation for the design and implementation of future interactive systems? My group has been investigating this question in a number of different application areas, from games that modify their behavior at runtime to tools that help physical therapists prescribe mixed-reality exercises and programming environments for end users and experts. Along the way, we learned lessons about better and worse uses of AI code generation, and developed techniques for guarding against (or at least dealing with) errors and hallucinations. This talk will catalog these lessons and point to an agenda for future HCI systems research.