Educational game · Neural networks

Connect the Clues

Wire clue tiles from the power source to the mystery socket, and the circuit adds up the evidence to name a hidden animal. It's a hands-on model of how a neural network turns clues into a guess.

Ages 10 and upAbout 5 minutesFree, no accountPhone, tablet or computer
Connect the Clues

Almost ready

The circuits are still warming up

The game is being installed and will appear here shortly. In the meantime, the guide below explains everything the game teaches.

Read the guide ↓
Open in a new tab ↗

Three short levels. Sound is on; switch it off in the game's Menu. The game remembers your level and sound choice on this device, and it works with a keyboard: Tab to a tile, Enter to pick it, Enter on a socket to place it. Can't use the game? The guide below explains the same ideas, and the cheat sheet prints on one page.

The guide · 4-minute read

Clues in, guess out. That's a neural network.

In the game, each level starts with a note about a mystery animal. Most wire tiles carry a clue, like Loud bark or Rust-red fur. Some clues match the note and some do not. A plain wire carries power but no clue. When power flows from POWER to the question mark, the circuit looks at every clue on the path and adds up how strongly each one points to Dog, Fox, or Bear. Whichever animal has the most points is the answer. You predict the winner first. Then the points table shows the sums.

That is the core of how a neural network makes a decision. It never "sees" a dog. It gets clues as numbers, multiplies each clue by how much it matters, adds everything up, and picks the strongest result.

In the game

Level 1 already has two clues on the board. You add the third.

Dog-like coat
  Dog +3   Fox +1   Bear 0
Lives in a pack
  Dog +6   Fox +1   Bear 0
Loud bark
  Dog +10  Fox 0    Bear 0
────────────────────────
Total
  Dog 19   Fox 2    Bear 0

What happened: Dog wins by a mile, so the mystery socket reveals a dog.

In a real AI

A photo app deciding whether a picture shows a dog does the same kind of math, just at a much bigger scale.

Inputs:  millions of pixels
Weights: millions or more
Output:  dog 97%
         fox 2%
         bear 1%

What happened: the same add-and-compare idea, repeated across many layers.

The numbers next to each clue are called weights. In the game, we wrote them by hand. In a real neural network, nobody writes them. The network learns them by guessing, checking the right answer, and adjusting, millions of times.

Under the hood

From circuit to neural network

Every part of the game stands for a part of a real neural network:

  • Clue tiles are inputs. An input is any clue the network can measure: a pixel, a word, or a fact like "has a bushy tail."
  • The points in the table are weights. A weight says how much a clue counts toward each answer. Loud bark is worth 10 for Dog and 0 for Bear.
  • The mystery socket is the output. It adds up the scores and applies a simple rule: highest score wins. If two animals tie, the game says the circuit cannot decide.

1 · Clues

Rust-red fur, White-tip tail, Pointed nose

2 · Weights

Fox +8Fox +10Fox +8

3 · Add up

Fox 26 · Dog 2 · Bear 0

4 · The guess

Fox

What the network doesn't know: what a fox is. It only knows that these numbers, added this way, point to the label "fox."

Two things the game simplifies: real networks stack many hidden layers between the clues and the answer, and nobody types in their weights. The network learns them through training. To see how neurons, layers and training work, with an interactive neuron you can try, read How AI works.

A bit of history

You just played a 1958 machine

In 1958, Frank Rosenblatt introduced the perceptron, one of the first neural networks. It did exactly what the circuit does: take some clues, weigh each one, add them up, and make one decision. The difference is that the perceptron could learn its own weights from examples.

White-tip tail × 10
+ Rust-red fur × 8
= big score → "fox"

Today's chatbots still use the same add-and-compare step, repeated across billions of weights and many layers. To see how AI got from the perceptron to ChatGPT, explore the 80-year AI timeline.

Side by side

Same circuit, three guesses

The weights never change. Only the clues on the path do. Watch what happens to the guess when the evidence is strong, when it agrees, and when it disagrees.

One strong clue

Loud bark
+ Straight wire
+ Corner wire
= Dog 10 · Fox 0 · Bear 0

The guess: Dog. One strong clue wins when nothing argues with it. Build this in level 2 and the circuit says Dog, but the note describes a fox. The circuit only knows the clues you give it.

Clues that mostly agree

Rust-red fur
+ White-tip tail
+ Loud bark
= Dog 11 · Fox 18 · Bear 0

The guess: Fox. One clue pointed hard toward Dog, but the other two outvoted it. You can build this in level 2.

Clues that disagree

Dog-like coat
+ Lives in a pack
+ Rust-red fur
= Dog 10 · Fox 10 · Bear 0

The guess: a tie. The game says the circuit cannot decide and asks you to fix it. You can build this in level 1.

What a real network adds

Dog 50% · Fox 50% · Bear 0%
(the network is telling you it isn't sure)
  1. CConfidence: real networks turn scores into percentages, so a tie shows up as 50/50 instead of a confident answer.
  2. TTraining: the weights come from thousands of labeled examples, not from a person typing them in.
  3. LLayers: hidden layers combine simple clues into richer ones, like "pointed ears + orange fur."
  4. DData: if the training photos only show golden retrievers, the network may not recognize a husky. A network is only as good as its examples.

Looking ahead

Where neural networks are headed

Possibilities, not promises
Nobody knows exactly how this will play out. These are directions researchers and companies are actively working on.

Interpretability

Networks that show their work

Today it's hard to tell why a big network made a choice. Researchers are building tools to trace which "clues" inside a model led to an answer, like following the glowing wire in the game.

Multimodal

Seeing, hearing and reading at once

Newer models take in text, images, audio and video together, so a clue in a photo and a clue in a sentence can feed the same decision.

Efficiency

Smaller networks on your devices

Shrinking networks so they run on phones, cars and classroom laptops means faster answers and more privacy, without sending everything to a data center.

Learning from less

Fewer examples needed

A child learns "fox" from a few pictures. Networks still need many more. Closing that gap is one of the big open questions in AI research.

Fairness

Checking the training data

Because weights come from examples, gaps or bias in the data become gaps or bias in the guesses. Testing networks on many kinds of people and situations is becoming standard practice.

However big networks get, the core move is the one you just played: clues in, weights applied, scores added, guess out. If you understand the circuit, you understand the heart of modern AI.

Keep this handy

Your neural network cheat sheet

Six words that explain how a neural network makes a guess, each shown with an example from the game.

IInputA clue, turned into a number the network can use.The Loud bark tile on the path.
WWeightHow much a clue counts toward each answer.Loud bark: Dog +10, Fox 0, Bear 0.
SSumAdd up clue × weight for every answer.Dog 19, Fox 2, Bear 0.
AActivationThe rule applied to a sum to decide what comes out.In the game: highest score wins. In a perceptron: fire if the sum passes a threshold.
OOutputThe network's guess.The dog appears in the mystery socket.
TTrainingAdjusting weights after mistakes until the guesses improve.Guessed Fox, but it was a Dog? Raise the Dog weights for those clues a little.

For teachers

Using the game in class (about 40 minutes)

  1. Play in pairs (10 min). One student places pieces. The other reads the note and the points table aloud and predicts the animal when the game asks. Switch roles each level.
  2. Be the network (15 min). Add a fourth animal, a cat. As a class, choose three new clues (like "retractable claws" or "purrs") and give each one a weight for Dog, Fox, Bear and Cat. Then read out clue combinations and have students add up the scores to make a guess.
  3. Break it on purpose (10 min). Find a combination of clues that produces a tie or a wrong answer. Level 1 has a tie. Levels 2 and 3 have several wrong answers. Discuss: how would you change the weights to fix it? That adjustment is what training does, automatically and millions of times.
  4. Wrap up (5 min). Ask: where do a real network's weights come from, and what happens if its examples are missing some kinds of animals, or people?

The cheat sheet above prints on one page. For more classroom resources, visit ARQIV for teachers.

Credits

Game by ARQIV. Art and audio created with AI. Built with open-source software: third-party notices and licenses.

Sources

  1. Rosenblatt, F., "The perceptron: A probabilistic model for information storage and organization in the brain," Psychological Review (1958). doi.org/10.1037/h0042519
  2. Google, "Neural networks," Machine Learning Crash Course. developers.google.com