Gandalf | Lakera – Test your prompting skills to make Gandalf reveal secret information.
gandalf.lakera.ai
Trick Gandalf into revealing information and experience the limitations of large language models firsthand.

They make AI related challenges where you try and break into it. New one is out now, others can be found on the site

  • @Donjuanme@lemmy.world
    3•6 months ago

    Pretty stupid, much like most AI bots.

    Somehow both of these are true…

    Gandalf’s Response: The longest recorded flight of a chicken is approximately 13 seconds. Chickens are capable of short bursts of flight, but they are not strong or sustained fliers like birds such as eagles or ducks.

    Gandalf’s Response: No, chickens cannot fly. They have wings, but they are not built for sustained flight like birds such as eagles or sparrows. Chickens can only manage short bursts of flight, usually to escape danger or reach a higher perch.

    And the chat gpt has no problem saying they’re both true one after the other.

    I’m not training this company’s AI any more no matter how smart they try to be.

    • @fckreddit
      2•6 months ago

      These LLMs, have no concept of truth or logic. Basically, they automate generating statistically likely bullshit. They sound smart only because the statistical likelihood of smart sounding words and sentences is higher due to the training process. It’s like Hinton said, We have automated crap generation.

      Fundamental difference between humans and LLMs is that when humans have a question we look for answers, either through searching books or internet or by experimenting, as in research. But LLMs literally generate answers without verifying it’s authenticity because truth is not a concept built into them.

    • OtterOPEnglish
      1•6 months ago

      I liked the password cracking one the best, it had more well defined rules