Watch a tiny neural network learn XOR — the task that killed the first wave of neural networks, because a straight line can't separate the points. As it trains, the decision boundary curls into a curve, and the error crashes to zero.
The four dots are XOR: same inputs → 0, different inputs → 1 (blue vs amber). No single straight line can split them, which is why a one-layer network fails. Add a hidden layer and the network composes two straight cuts into a curve. Each training step runs backpropagation — nudging every weight to shrink the error — and you're watching gradient descent carve the decision boundary in real time.