← All simulations · Pillar 8: Brains made of math

Model Builder

This is the glassbox the whole platform is named for. A neural network is just layers of tiny decision-makers wired together. Here you can build one by dragging layers, or by writing the code — and the two views are locked together, so you can always see how the blocks of an AI fit. Then send a fruit through it and watch the answer light up, or train it and watch the wrongness fall.

Meet the parts

Input layer. The numbers that go in. Here there are two — how sweet a fruit is and how big it is. It does no thinking; it just hands the numbers on.
Hidden layers. The thinking middle. Each neuron takes every signal from the layer before it, multiplies each by a weight (how much it cares about that signal), adds them up, and then bends the total with its activation. One neuron draws one bend; a row of them (a wider layer) draws several at once; stacking layers (a deeper network) lets later layers combine earlier bends into curvier shapes. You can set each hidden layer to anywhere from 1 to 10 neurons.
Output (answer) layer. Turns the last layer’s features into the final answer. Here two scores go through softmax, which makes them into chances that add up to 100% — team A vs team B.

The bends (activations), and why they matter:

relu — keeps positive signals, zeroes out negatives. A simple, fast on/off bend.
tanh — a smooth S-curve from −1 to +1.
sigmoid — a smooth S-curve from 0 to 1.
line — no bend at all. Useful to see why bends matter: a stack of only straight lines is still just one straight line, so it can never wrap around the checkerboard.

How to read it

Inputs (left) are the two numbers about a fruit: how sweet it is and how big it is.
Each hidden neuron draws one bend in the boundary. One straight line can’t split the two teams — add a couple of neurons and the network can wrap around each one.
Every line is a connection. Thicker means a stronger pull; blue pushes the answer up, red pushes it down.
Training nudges every weight a little downhill on the wrongness, round after round — that’s gradient descent. Drag Training rounds done to rewind and replay the learning.
Hit Watch it learn ▶ to play the whole thing step by step. Each round the numbers flow forward (blue, left→right) to make a guess, then the mistake flows back (orange, right→left) and tweaks every weight — that backward pass is backpropagation. Watch the loss curve drop as it repeats.

Try this

Start with 0 hidden layers (remove the one that’s there): the network is one straight line and gets stuck near 50%. Add a hidden layer of 3–4 neurons and watch the score jump.
Slide one layer up to 10 neurons (wider) and read the live explanation to see what changes.
In the code, change kind tanh to kind relu — the builder above follows.
Drag a hidden-layer chip to reorder it, or add a second hidden layer to make a deeper network.
Press Watch it learn with 0 hidden layers and again with a hidden layer — see how the backward pass can’t fix a straight line, but can teach a bent one.
Hit Open in the Studio to keep going with the very same model as Spectra code.

Safe by design

Everything here runs in your browser on small, made-up data. There is no upload, no internet from the model, and nothing about you is collected — the same safety-by-absence rule as the rest of the platform.