Eliezer is still ridiculously optimistic about AI risk - LessWrong