Every neural net learns by backpropagation, and almost nobody has seen it run. Pick an expression and step the real computation graph: values flow forward, then gradients flow backward — the chain rule, animated. The actual Python engine runs in your browser via Pyodide.
Each box is one operation. The number in blue is the value computed on the forward pass; the number in orange (∇) is the gradient of the output with respect to that node, filled in on the backward pass. Notice how a node used in two places (the shared x in the polynomial) accumulates gradient from both paths — that's the chain rule. Python fetched verbatim from src/nabla.