AI Foundationspredict · compress · act
act IV

The Machines

In 1943 a tiny mathematical model of a brain cell appeared — and it is still the unit at the bottom of every neural network.

14

The First Neuron

before this →Can Machines Think?
1943The artificial neuron — McCulloch and Pitts model a brain cell as simple arithmetic.
1958The perceptron — Rosenblatt builds a machine that learns to tell two classes apart by changing its weights.
1969Minsky & Papert — publish "Perceptrons"; perceptron funding stalls

McCulloch and Pitts asked: what is the simplest thing a brain cell does? Answer: it adds up its inputs, and if the total is high enough, it fires.

x₁x₂x₃w₁w₂w₃Σ + bsum + biasactivationfire or notout
a neuron: weighted sum, then a threshold. Stack many and you get a network.

Weights and bias

Each input x is multiplied by a weight w saying how much it matters. The weighted inputs are summed and a bias is added — a baseline tendency to fire. Then an activation function decides the output: originally a hard threshold, later smooth curves like sigmoid and ReLU that make learning possible.

The perceptron (Rosenblatt, 1958) turned this into a learning machine: show it examples, and nudge the weights when it gets one wrong. It could learn to separate two classes of points — and it made headlines. The New York Times reported it would soon walk, talk, and reproduce.

The first AI hype cycle

and its first hard limit

In 1969 Minsky and Papert proved a single perceptron cannot learn XOR — a pattern that is not linearly separable. One straight line cannot split it. The result is true, and it is narrow: add a second layer and XOR becomes trivial. But the book landed like a verdict, and funding dried up.

the lessonA limit of one model is not a limit of the idea. Read the scope of a negative result before you accept it.
ADDweighted sum + bias
FIREactivation function
LEARNnudge weights on mistakes
introduces →neuronperceptronactivation functionweightbias
← previousCan Machines Think?next →Why Silicon Got Good at This