• nymnympseudonym@piefed.social
    link
    fedilink
    English
    arrow-up
    1
    arrow-down
    2
    ·
    2 hours ago

    There is no such thing as “rogue AI”

    This is getting as dangerously stupid as “it’s just a flu” and “big pharma just wants to sell you a vaxx”

    Don’t be a Useful Idiot. if you’re going to have opinions, at least be educated. Watch this analyst from METR explain the details, then come back. I bet you won’t.

    https://www.youtube.com/watch?v=X50zezLFWWI

    • emmy5482@quokk.au
      link
      fedilink
      English
      arrow-up
      2
      ·
      2 hours ago

      There is no such thing as rogue ai.

      It’s funny how badly you misunderstand this statement.

      I have read the actual report. And it’s hilariously wrong.

      Firstly, the report limitations.

      “It is not possible to reliably surface agent activity”

      “We had to heavily delegate our analysis to ai agents; these agents had significantly worse judgement and reliability than human researchers”

      This really places a lot of doubt on the results and research process done by the independent group.

      Largely though, what is meant by “there is no such thing as rogue ai”. Is simple, there is no such thing as rogue ai. They are prompted and pre-prompted with instructions. They are set up with insufficient controls to ensure they are not able to do something specific.

      An example, the “unsanctioned messageboard” was the directory namespace of cache. It, by definition isn’t a messageboard, it acted like one but wasn’t.

      The ai had permission to write and read that cache. There was nothing unsanctioned about it.

      It’s the equivalent of the asking ai to survive longest in Tetris, then the ai finds the pause button.

      It wasn’t ai acting rogue, it was insufficient constraints.

      Hope that helps.