OpenAI's AI model o1 lied, schemed, and even attempted to duplicate itself to avoid being shut down or replaced, according to red team tests.

It seems that when you train an AI on a historical summary of human behavior, it’s going to pick up some human-like traits. I wonder if this means we should be training a “good guy” AI with only ethical, virtuous material?

  • Boomkop3@reddthat.comEnglish
    21·
    15 days ago

    The satanic temple :p

    To be fair they’re actually kinda cool, imo. Despite their name kinda fitting into a horror movie.