A single perceptron can learn linearly separable functions (AND, OR) but fails on XOR — which requires 2 layers. Watch gradient descent adjust weights in real time and see the decision boundary evolve.