New report: "Scheming AIs: Will AIs fake alignment during training in order to get power?" — LessWrong
www.lesswrong.com
I examine the probability of a behavior sometimes called "deceptive alignment."
  • FermiEstimate@lemmy.dbzer0.comEnglish
    30·
    6 months ago

    I conclude that scheming is a disturbingly plausible outcome of using baseline machine learning methods to train goal-directed AIs sophisticated enough to scheme (my subjective probability on such an outcome, given these conditions, is ~25%).

    Out: vibes and guesswork

    In: “subjective probability”