• pixxelkick@lemmy.world
    link
    fedilink
    English
    arrow-up
    5
    arrow-down
    1
    ·
    27 days ago

    Cool

    The paper clearly is about how a specific form of training on a model causes the outcome.

    The article is actively disinformation then, it frames it as a user and not a scientific experiment, and it says it was Facebook llama model, but it wasn’t.

    It was a further altered model of llama that was further trained to do this

    So, as I said, utter garbage journalism.

    The actual title should be “Scientific study shows training a model based off user feedback can produce dangerous results”

    • petrol_sniff_king@lemmy.blahaj.zone
      link
      fedilink
      English
      arrow-up
      1
      ·
      26 days ago

      I don’t see how this is much different from the sycophancy “error” OpenAI built into its machine to drive user retention.

      If a meth user is looking for reasons to keep using, then a yes-man AI system biased toward agreeing with them will give them reasons.

      Honestly, it’s much scarier than meth addiction; you could reasonably argue the meth user should pull up their bootstraps and simply refuse to use the sycophantic AI.

      But what about flat-earthers? What about Q-Anon? These are not people looking for the treatment of their mental illness, and a sycophantic AI will tell them “You’re on the right track. It’s freedom fighters like you this country needs. NASA doesn’t want people to know about this.”