Rohin Shah comments on The case for aligning narrowly superhuman models