Proposal for "Open Problems in Friendly AI" — LessWrong