Guess the token
A sentence cut short and four words that could come next. You pick one, see how likely each one is for a language model, and learn why. Three rounds, a couple of minutes.
Keep going in Tokencito
5 minute lessons about AI, systems you fix and a review of what you missed.
What is "Guess the token"
Guess the word that comes next in a sentence, the way an AI does every time it writes. You see the start of a sentence, pick how it goes on, and then see how a language model would pick. There is no right word, only more and less likely ones, and learning to read that is the game.
How to play
Each game has three rounds. In each one you pick one of four words; then you see bars with the chance of each word and of "others", plus 10 draws made with those chances; finally you answer a short question about what happened. Only the question counts: picking a word never wins or loses you anything.
At the end you see how many questions you got right on the first try. "Play again" shuffles the rounds and the draws. If you like word games like Wordle, this one feels close, but every round teaches you something about how AI works.
Why an AI "guesses" the next word
A language model like ChatGPT, Claude or Gemini writes one token at a time: it looks at all the text in front of it, works out how likely each continuation is and picks one with a draw (the usual setup: it can also be set to always take the most likely one). That is why likely words come up often, unlikely ones still come up sometimes, and a very likely sentence can say something false.
The numbers in the game are examples: they do not come from a real model, they are chosen so you can see the mechanism. And we play with words, while a model uses tokens: a token can be a whole word, a piece of a word or a punctuation mark. How text gets split is up to the tokenizer.
Keep learning
This game is also a lesson in the app: "Guess the token" in Tokencito, right after "How does it predict?", with a review of what you miss. And to see how a sentence is split into tokens, read what a tokenizer is.