Can coherent extrapolated volition be estimated with Inverse Reinforcement Learning? — LessWrong