Daniel Dewey on MIRI's Highly Reliable Agent Design Work — LessWrong