With cheaper/faster inference and longer contexts, some of our careful prompt engineering and chunking strategies feel like premature optimization from another era. I'm wondering which architectural choices from 2023-2024 you'd now simplify. What would you rebuild differently if you started today?