Less exploitable value-updating agent — LessWrong