Value learning subproblem: learning goals of simple agents — AI Alignment Forum