Defining capability and alignment in gradient descent — LessWrong