Prompt Guess: a free game where you race two AIs to read a picture

Prompt Guess is a small game in the AIKIZI Lab. A picture appears as a handful of big coloured blocks and comes into focus over about six seconds. Under it are four short lines of prompt. One line was written about this picture; the other three were written about other pictures. You win points for picking the true line, and more points the earlier you pick it.

You are not playing alone. Two AI models look at the same blurred frames and race you for the same line: Jev, the decisions model from TypeSafe, and Claude Haiku 5.5 from Anthropic. When a round ends you see who saw it first, and by how much.

Your first game is free and needs no account. Open Prompt Guess to play.

How does a round of Prompt Guess work?

Each picture is shown as a mosaic first. At the first step every block is 64 pixels across, so a tall picture is only about ten blocks wide and sixteen high. Every 1.5 seconds the blocks halve: 64, then 32, 16, 8, and then the sharp picture. Five steps in all, and the sharp one arrives 6 seconds after the round starts.

While it sharpens you choose one of the four lines. Click it, or press 1, 2, 3 or 4 on the keyboard. The moment you pick, your answer is locked. If it was right you score for the step you were on. If it was wrong you score nothing for that picture, and the picture plays out to sharp so you can see what you missed.

The points fall fast. A right answer at 64 is worth ten times a right answer at sharp, and a wrong answer at any step is worth less than waiting and being right. That is the whole game: how early can you be sure?

A game is ten pictures. Your free game is the day's deck, the same for everyone, so you can compare a score with a friend. Signed in, press restart and, for 2 credits, you get ten pictures you have not seen yet.

Who are the two AIs you race?

Jev is the decisions model made by TypeSafe. You give it a question and a set of options and it returns one choice, with a probability for each option. It is built for fast, cheap, structured decisions. We use its image preview build, which reads pictures.

Claude Haiku 5.5 is Anthropic's small general model. It is a language model that can see images, and we ask it to answer with a choice and how confident it is.

Both are asked live while you play, and neither reads the picture's description. They get the same blurred frames you get and the same four lines.

When does an AI rival buzz in?

A rival buzzes at the first step where it names the true line and is at least 80 % sure. Below 80 % it has not committed, even if its best guess happens to be right. A rival that is never that sure, even at sharp, never buzzes on that picture.

Here is one round from today's deck. The four lines were "Motocross athlete, blue-white dirt bike", "Neon pink and yellow mineral deposits", "Spiky white-haired cybernetic rider, amber tactical goggles" and "Cool monochrome palette of slate blue". The first one is true. At 64 Jev put 85 % on the slate blue line. At 32 it moved to the motocross line at 80 % and buzzed.

Is the timing fair to a human player?

A model cannot watch a picture sharpen in real time the way you do. So the rivals are asked about all five frames at once, as the round before yours starts, and their answers come back in a few seconds.

Then we put their answer on your clock. A rival's buzz is placed at the start of the step where it was first sure, plus the time its answer took. In the round above, Jev was sure at 32, which appears at 1.5 seconds, and its answer took 319 milliseconds. So on the shared clock Jev buzzed at about 1.8 seconds. If you picked the true line at 64, in the first second and a half, you were ahead.

The result is one stopwatch for all three players. It does not pretend the models were watching alongside you. It asks a simpler thing: at which step could each player name the picture, and how long did it take them to say so?

What do you see at the end of a round?

When the round ends the true line's number turns green. Each rival gets a tick or a cross: ✓ if it buzzed on the true line, × if it never did. Under that is a small table with the step each player was sure at, the time on the shared clock, the confidence and the points.

There is also a button, See the full decode. It opens the picture's decode page on AIKIZI, with the full prompt, the colours, the light and the materials that were read from it. The four lines in the game are short pieces of decodes like that one.

How many games can I play?

One free game, raced live: the day's ten. A restart or Play again after it asks you to sign in. Signed in, every game after that is 2 credits, and new accounts start with 5, so two more games come with the account. Every game you play is raced live by both rivals.

Your streak counts the days in a row on which you finish a game.

How early can Jev and Claude Haiku read a blurred picture?

We ran both rivals over the whole deck, 294 rounds, and noted the first step where each one was sure of the true line.

Jev saw it first on 138 pictures and Haiku on 26. They tied on 129. Most of the time neither model is sure before 16. Jev is quicker on the coarse steps; Haiku more often waits for a sharp picture, and on 44 rounds it never reached 80 %.

So the rivals are beatable. On our first day of play we beat Jev 950 to 700.

What can you tell from a picture at 64?

More than it looks like. At 64 you cannot see a face or read a word, but three things survive.

The palette survives. A couple of hundred blocks still carry the main colours and how much of each there is. If the picture is mostly warm orange with one dark corner, any line that names a cold blue scene is out.

The mass survives. You can see where the weight of the picture sits: one bright shape in the middle, or a dark ground under a pale top. A portrait and a crowded street look different even as blocks.

The horizon survives. A clear change from one band of colour to another, across the width, is often land and sky, or sea and sky. That rules a line in or out quickly.

Read the four lines against those three things first. Sometimes two lines are gone at 64, and you are choosing between two.

Why are decoys that share a colour a trap?

The decoys are real lines from other pictures, and the colour words in them are the easiest thing to match on a blurred frame. That is exactly where people, and models, go wrong.

Look at the motocross round again. At 64 the picture is a field of deep blue with grey and olive blocks low down. "Cool monochrome palette of slate blue" fits that frame very well, and Jev was 85 % sure of it. It was still wrong. The blue was sky, and by 16 a rider on a dirt bike is plain in the middle of it.

When one line matches the colour and another names a subject, ask whether the blocks could hold that subject. If they could, wait one step. A line that only describes a palette is often a decoy that happens to share the palette.

When is it worth waiting for 16?

At 16 the blocks are small enough to show edges. You can tell a figure from a building, a bike from a car, fur from fabric. For lines that turn on a subject noun or a material, 16 is often the first step that settles it.

The arithmetic helps. A right answer at 16 is 350 points. A wrong answer at 32 is zero. If you are torn between two lines at 32, a coin flip is worth 300 points on average, and waiting is worth 350 if you are then sure. Waiting wins, unless you are already fairly confident. Commit early when the palette and the mass point at one line only. Wait when two lines both fit the blocks.

Past 16 the points left are small: 200 at 8 and 100 at sharp. If you are still unsure at 16, pick the line the edges support and move on.

Where do the pictures come from?

Every picture is a published picture from AIKIZI Explore, chosen from the creators' showcase work. There are about 300 in the pool today. Each one has been decoded on AIKIZI, and the true line is a short piece of its own decode. Nothing private or unpublished is used.

Play a round

Ten pictures, two rivals, one shared clock.

Play Prompt Guess    Browse the Lab

Jev also plays the Jev word race, walking from one word to another one link at a time. We also wrote up how we built the live race. More posts are on the AIKIZI blog.

FAQ

Is Prompt Guess free to play?

Your first game is free and needs no account. After it, a game is 2 credits; new accounts start with 5.

How are points scored in Prompt Guess?

By the step you commit on: 1000 at 64, 600 at 32, 350 at 16, 200 at 8 and 100 when sharp. A wrong pick scores zero and the picture plays out. A game is ten pictures.

Which AI models do you race against?

Jev, the decisions model from TypeSafe, in its image preview build, and Claude Haiku 5.5 from Anthropic. A rival buzzes at the first step where it names the true line and is at least 80 % sure.

Do the AIs see the picture's prompt or description?

No. They get the same blurred frames you see and the same four lines. Neither reads the picture's description.

Can I beat the AI in Prompt Guess?

Yes. Across 294 rounds Jev was sure at the first step on only 20 and Claude Haiku 5.5 on only 3. Most of the time neither is sure before 16. On our first day of play we beat Jev 950 to 700.