Scott Alexander comments on Why not just write failsafe rules into the superintelligent machine?