Everything in deep learning is built from one primitive: the point neuron. Take a weighted sum of your inputs, pass it through a nonlinearity, done. It is McCulloch and Pitts in 1943, Rosenblatt's perceptron in 1958, and the unit inside every transformer today. It won for two unglamorous reasons. A layer of point neurons is a dense matrix multiply — the one operation modern hardware runs fastest — and a network of them can approximate any function. But a single point neuron is weak: it can only draw a straight line, so it cannot solve XOR, the classic task where the answer flips at the diagonal. The fix that saved it in the 1980s was to stack them into layers. Hold onto both halves of that story: the point neuron is universal as a network, and helpless as a unit. The rest of the course is about a different way to buy that power — inside the unit, not just across layers.