Wei_Dai comments on My current thoughts on MIRI’s “highly reliable agent design” work