The default outcome for aligned AGI still looks pretty bad — LessWrong