An Open Philanthropy grant proposal: Causal representation learning of human preferences — LessWrong