Slopworld 2035: The dangers of mediocre AI — LessWrong