How will it get better without human feedback? Who is telling AI when it gets it wrong? How do you create a broadly applied AI capability (i.e. lots of...