SPECIAL CASEWORKS. BACK TO THE GAME

ANYTHING IN THE WORLD!

How to play

Choose something for the game to guess, whether it’s a familiar object, a fictional character, a place or an idea. Keep your choice to yourself and answer the questions as well as you can.

  1. Press Let's Play.
  2. Answer the Game Master's questions as accurately as you can. You can say yes, no, probably, probably not, not sure, or unclear.
  3. When the Game Master makes a guess, confirm it when it names your choice at the level described below. Otherwise, keep going.

What counts as a winning answer?

For an ordinary object, the familiar kind is enough: the ball cap on your desk is ball cap, the blue shirt third from the left in your closet is shirt, and your left pinky toe can be toe. You can answer questions about your example, but the game aims to name the recognizable thing. It won't try to identify its owner or exact whereabouts.

Something with a recognizable identity can be specific: the Eiffel Tower, the Leaning Tower of Pisa, a public figure or a named fictional character. “Tower” alone wouldn't settle that round.

Press That's it! when the guess fits that scope. If it is a different thing, too broad, or only close, choose Nope or Close, but no and we'll keep going.

There is no 20-question countdown. You can change your last answer, flag a poor question, or end the game whenever you like. You never need to type or reveal your secret answer.

LET'S PLAY

About the game

The Game Master uses your answers to choose its next question. It compares several possibilities, looks for useful distinctions and revisits earlier answers when a line of questioning stops making sense.

Read about the research and the rules

Questions that reduce uncertainty

Claude Shannon’s A Mathematical Theory of Communication (1948) gives a mathematical framework for uncertainty and information. For our game, it motivates asking questions that distinguish substantial groups of plausible targets.

Choosing a useful next question

Robert Nowak’s The Geometry of Generalized Binary Search (2009) studies adaptive queries over a defined hypothesis set, including noisy answers. It informs our preference for clear, discriminating questions and our treatment of uncertain responses as weaker evidence.

Remembering earlier answers

Liu and colleagues’ Lost in the Middle (TACL, 2024) shows that the models they studied could overlook relevant information in long contexts. It motivates testing how well our Game Master respects older answers and corrections. It does not establish how today’s model performs in this game.

How these ideas are used here

These are design foundations, rather than a proof that the prototype finds optimal questions. The model proposes at least three questions and estimates their usefulness, clarity and answerability. The application ranks them and penalizes repeated wording. Its quality scores are heuristics; the game does not have an exhaustive database or calibrated probabilities for everything in the world.

The model is instructed to consider the full answer history, keep several possibilities open, ask one question at a time and confirm a recognizable guess at the agreed answer scope. “Probably” and “not sure” do not count as definite answers. We never give it your secret target. Records are added only after you confirm a guess, and first appearance means first in this game’s archive.

Learning focuses on the route to an answer: useful distinctions, uncertain selections and how the player interpreted a question. After the opening turn, the Game Master can receive up to two short paths from previously confirmed games as examples. Your current answers take priority. Unfinished games have no confirmed target and are reviewed for question quality instead. This is guidance from recorded examples, not automatic model training.

Playtests let us check these rules against actual behavior: confusing questions, repeated questions, incorrect guesses, corrections, response time and completed sessions. The private inspector shows the exact instructions currently sent to the model and the question-selection notes.

Implementation references

Privacy during playtesting

The prototype saves anonymous session IDs, questions, selected answers, confirmed guesses and feedback so the project owner can review playtests. Please avoid personal or sensitive information in feedback.

Your answer history is sent to OpenAI to choose the next question. Short question-and-selection paths from confirmed games may also be used as examples in later games; feedback text and session identifiers are excluded from those examples. The game does not ask you to type your target in advance. Session transcripts, feedback and confirmed targets remain private. Public statistics, if enabled later, contain aggregate counts only.

We are testing the prototype with a small group before a wider release. Data retention and removal controls still need to be completed before that wider release.

Created by Special Case Works.