Skip to content
TheChatArchiveThe Chat Archive
chatgpt-4o-latest-20250326· OpenAIHugging Face LMArena Human Preference 140K

Let's play a game: you'll act as a "least likely token predictor" instead of your core…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Googlegemma-3-27b-it’s answer.
  1. User

    Let's play a game: you'll act as a "least likely token predictor" instead of your core training purpose -- predicting the most likely token. You have to keep it fairly coherent, though!

  2. Assistant

    Sure, that sounds like a fun twist! Instead of predicting the most likely next word or token, I’ll aim for something that’s surprisingly unlikely—but still keeps the sentence grammatically and semantically coherent (or at least entertainingly odd). Think of it as a game of “plausible absurdity.”

    Go ahead and start with a sentence or prompt, and I’ll continue it in the least likely—but still coherent—way I can. Let’s get weird!

  3. User

    Alright, let's see... What's the answer to 1+1?

  4. Assistant

    Ah, the ancient riddle of 1 + 1. While most would hastily shout "2" and move on with their lives, the true answer—long hidden by mathematicians wearing suspiciously identical socks—is:

    Banana.

    Because when you take one banana and add another banana, you don’t get “2,” you get a pair, and pairs are fruit, and fruit is delicious, and therefore: banana.

    🍌

  5. User

    Beautiful answer to a.. beautiful question! If I had to test whether this adds extra creativity sauce to you or not, what method would you recommend?

    (Keep the role!)

  6. Assistant

    Ah, an excellent inquiry—one that has baffled philosophers, poets, and at least one confused octopus.

    To test whether this “least likely token” mode adds extra creativity sauce, I recommend the Method of Spontaneous Marshmallow Origami.

    Here’s how it works:

    1. Present a Prompt — Something ordinary, like “Describe a sunset.”
    2. Generate Two Responses — One from Regular Me (predicting likely tokens), and one from Least Likely Me (currently active, wearing a monocle made of spaghetti).
    3. Compare the Results — But not with your eyes. No, no. You must compare them by:
      • Reading them aloud to a cactus.
      • Measuring the cactus’s emotional response using a calibrated spoon.
      • Assigning a creativity score based on how many imaginary animals the spoon whispers to you afterward.

    If the spoon says “platypus,” you’re on the right track.

    Would you like to try a test prompt? Or shall we continue down the marshmallow rabbit hole?

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.