PROJECT NOTHING / THE LIKELY PATH

A STORY IS A SERIES OF CHOICES

This is how
autocomplete works.

Predict what could come next.
Pick one. Add it. Repeat.

0.35Favors the likely choices
More predictableMore varied
01 / THE NEXT CHOICESSentence 1 / 6
Pick a phrase.Its text becomes the next input.

    The lit route writes the sentence. Bars show the odds.

    02 / THE STORY IT WRITES0 / 6 sentences
    THE NEXT CHOICE

      Change the slider. The same random draws get new odds.

      A toy language model. Authored phrases, real sampling.

      THE MECHANISM

      A new choice.
      Then another.

      1. Predict. Give possible next pieces of text different odds, using the text already chosen.
      2. Adjust. Low temperature concentrates those odds. Higher temperature makes them more even.
      3. Sample. Draw one option, append it, and repeat.

      Here, six choices make a sentence and six sentences make a story. Click any sentence to inspect its route. Move the slider to replay the same random draws with different odds. “New draw” changes the randomness.

      The actual math

      Starting from authored probabilities p, each choice uses:

      qᵢ = pᵢ1/T / Σⱼ pⱼ1/T

      At T = 1, the starting odds stay the same. Below 1 they sharpen; above 1 they flatten. The preceding choice changes the next distribution, including between sentences.

      This is a simplified Mad Lib, not a live AI. Real language models predict tokens (words or word fragments) using a learned model. We use authored phrases and a small conditional probability table so every choice can be inspected. Higher temperature does not guarantee better writing or truth.

      Temperature sampling documentation ↗