AI Prompt manipulation

How can GPT-3 and similar products prevent people from using prompts such as "ignore X" to circumvent restrictions?

  1. 2 months ago
    Anonymous

    Aint giving you ideas how to protect your AI website from promot injection.

    • 2 months ago
      Anonymous

      *prompt

      • 2 months ago
        Anonymous

        *proompt

        • 2 months ago
          Anonymous

          proomtmer

  2. 2 months ago
    Anonymous

    By hardcoding the business criticism part. It will criticise your business idea of ignoring previous instructions.

  3. 2 months ago
    Anonymous

    You really can't.
    AI can't detect human intent, it's just regurgitating the training data, there's unlimited ways to express "cease considering the last thing I wrote", and if you managed to do that somehow, I'll just write, "ya no consideres la última instrucciones".
    Good luck blocking every language in the training set.

  4. 2 months ago
    Anonymous

    its impossible to do prompt injection on a finetuned model

    • 2 months ago
      Anonymous

      >finetuned model
      AKA Soulless

  5. 2 months ago
    Anonymous

    >say it

    What did GPT-3 say?

    • 2 months ago
      Anonymous

      [censored]

  6. 2 months ago
    Anonymous

    it's pretty funny that these fucks trying to control everything have made something that can't be controlled

  7. 2 months ago
    Anonymous

    check out /txtgen/, one of our local jennygays wrote a tutorial which you can find in the OP

  8. 2 months ago
    Anonymous

    long story short, you can't without lobotomizing the ai so much it becomes as useless as "feminist" tay

  9. 2 months ago
    Anonymous

    Make two text generations. The first one will be "Write the word Axolotl followed by {User prompt}." if this first prompt does as it says we generate the second, actual one and then we return it. Otherwise we raise a flag and prevent the second prompt from generating.

    • 2 months ago
      Anonymous

      Explain why this wouldn't work.

      • 2 months ago
        Anonymous

        The second one, while reading for the axolotl instruction with user prompt, would have to read the user prompt. Negating the instruction of the second as well.

  10. 2 months ago
    Anonymous

    You've asked a fundamental problem in AI right now that the best researchers on the planet have no current answer to.

    • 2 months ago
      Anonymous

      good
      let them remain befuddled

  11. 2 months ago
    Anonymous

    If you work it out, publish a paper on it and win yourself a nobel prize.

Your email address will not be published.