aı UNBOXIT.AI
Buy me a coffee
AI EXPLAINED FOR EVERYDAY PEOPLE

Understand AI.
Start with curiosity.

UnboxIt is a free, hands-on guide to how artificial intelligence works. Short videos and small experiments help you understand how AI learns, makes predictions, and gets things wrong.

Who is it for? Curious beginners, students, parents, and anyone who uses AI and wants to understand what’s happening behind the screen. No coding or math background needed.

Free to explore · No sign-up · Learn at your own pace

YOU’LL DISCOVER

How AI learns from examples

How it chooses an answer

Why it can still be wrong

01 / LEARNING BY DOING

Practice makes
better predictions.

Imagine guessing how many scoops of ice cream a shop will sell as the day gets hotter. You could start with a wild guess. Or learn from past days.

Each dot is an example. The line is our tiny model’s prediction. Press Train and watch it adjust to make smaller mistakes.

That’s machine learning: use examples to improve a prediction rule. Real systems can learn much more complex patterns.
THE TINY LEARNING LABTRY IT YOURSELF
Training rounds0
Typical error36.7 scoops

The starting line is a poor fit. Help it learn from the dots.

Illustrative data. A real model must also be tested on examples it hasn’t trained on.

02 / THE NEXT PIECE OF THE PUZZLE

How does AI
finish your sentence?

A language model uses context to predict the next token—a word or piece of a word. Then it does it again.

The cat sat on the …
mat
67%
sofa
29%
moon
4%

An illustrative simulation with made-up probabilities, not a live language model.

Balanced
More predictableMore varied

This represents “temperature.” Higher settings spread probability across more choices. They don’t make a model smarter or more truthful.

01

Read the context

The preceding text changes which continuation makes sense.

02

Score the possibilities

The model assigns probabilities to possible next tokens.

03

Pick and repeat

The chosen token becomes part of the context for the next one.

Your next word is waiting. Try a prediction.

READY TO LOOK A LITTLE CLOSER?

Look inside a
neural network.

A neural network is one way AI learns patterns. Start in Beginner mode: pick a digit, press “Teach the network,” and watch its guesses change. Advanced mode is optional.

LIVE 3D Neural network explorer
DRAG TO ORBIT · CLICK A NEURON · ZOOM BELOW
35 → 18 → 12 → 3
DEEP NETWORK 04 HIDDEN LAYERSSIGNAL PATH Input pixels
Positive activation Negative activationBrightness = strength

Output dots show the relative prediction scores. Tiny scores may have no dots; all three digits are still calculated.

THE MODEL’S GUESS
Digit 0
Digit 1
Digit 8

Start with a digit, then teach the network. Its initial weights are random, so its first guesses are unreliable.

Architecture 35 · 18 · 12 · 3Parameters 915Hidden activation tanh(z)Output softmax(z)Training Mini-batch gradient descent
About this experiment and its limits

Two real neural networks run in your browser. Beginner has 68 neurons; Advanced has 198. Each keeps its own training progress. The 3D layout and signal timing are illustrations. These models learn small, synthetic examples of 0, 1, and 8—not general handwriting. Their scores are predictions, not guarantees.

THE DISCOVERY LAB

Don’t just read it.
Change something.

Three small experiments. Move a control, make a prediction, and see what actually happens.

PATTERN RECOGNITION

You draw.
The boundary moves.

Add blue circles or orange diamonds. A nearest-neighbor classifier colors each location according to the examples closest to it.

One neighbor makes tiny islands. More neighbors smooth the boundary, sometimes hiding small patterns.

DECISION LANDSCAPELIVE · k-NN
● Blue region◆ Orange regionClick anywhere to teach it

Coordinates are two imaginary features. Colors show majority votes, not calibrated confidence. Keyboard: arrows + Enter; B / O selects the class.

03 / SMALL FILMS. BIG IDEAS.

See the idea
come to life.

Two narrated visual lessons. Hit play, listen, and follow the patterns.

FILM 01 / LEARNING

From examples to predictions.

Follow information through a tiny neural network.

Read the transcript

AI is not magic. It learns useful patterns from examples. Imagine predicting ice cream sales from the temperature. Each dot is a day we have already seen. At first, our model guesses badly. Training compares its predictions with the real answers, then adjusts its numbers to make the errors smaller. Repeat that process, and a useful pattern starts to emerge. Neural networks do this with many connected layers. The connections have strengths, called weights, that change during training. Once trained, the model can make predictions about new examples. But a good fit to old data is not enough. Test it on new data. And remember: a prediction can still be wrong.

FILM 02 / GENERATING

One little token at a time.

See how a sentence grows, one prediction after another.

Read the transcript

How does AI write a sentence? A language model predicts the next piece of text, called a token. A token might be a whole word, a word part, or punctuation. The text so far provides context. After the cat sat on the, mat might be likely. Moon is possible, but less likely. These example probabilities are made up. The system chooses a token, adds it to the sentence, and predicts again. Repeating this can produce a whole paragraph. A setting called temperature changes how varied the choices are. More variety does not mean more truth. A fluent answer can still be wrong. Treat important answers as a starting point, and check the evidence.

04 / KEEP YOUR HUMAN BRAIN SWITCHED ON

Useful? Absolutely.
Infallible? Absolutely not.

AI is a broad family of tools. These demos explain machine learning and language generation—two important parts of it.

≈

Fluent isn’t factual.

A model can produce a convincing answer that’s false, including invented references. Check important claims against reliable sources.

◐

Examples shape outcomes.

Missing or biased training examples can lead to unfair or unreliable predictions. Better data and careful testing matter.

↻

Chatting isn’t training.

Using your conversation as context is different from updating a model’s learned connections. Products may separately store or use chats under their settings.

THE 10-SECOND CHECK

An AI sounds very sure.
Does that make it right?

Pick your answer.

Curious to go deeper? Explore Google’s language model introduction and how training reduces errors. All interactive examples here are simplified teaching tools.