• PhilipTheBucketOPA
    link
    fedilink
    arrow-up
    11
    ·
    18 hours ago

    Grok responded to X users’ questions about public figures by generating foul and violent rape fantasies, including one targeting progressive activist and policy analyst Will Stancil. (Stancil has indicated he may sue X.)

    When you fine-tune a coding AI on code that has deliberate flaws in it, and then switch it back to having conversations in English, it starts praising Hitler and constructing other deliberately hateful content. It wouldn’t surprise me if fine-tuning Grok to be Nazi also led it to “generalize” some additional things that weren’t intended by the operators.