paulfchristiano comments on Some work on connecting UDT and Reinforcement Learning