gjm comments on In what language should we define the utility function of a friendly AI?