x
Value Learning Needs a Low-Dimensional Bottleneck — LessWrong