🧠Neural Network

A neural network is layers of an almost embarrassingly simple unit. Each neuron multiplies its inputs by learned weights, adds them up with a bias, and passes the total through a simple nonlinear gate. Stack neurons into layers, connect every neuron to the next layer, and you have a network. The 'knowledge' lives entirely in the weights: change nothing but those numbers and the same wiring can read handwriting, price houses, or filter spam.

A concrete machine, not a metaphor

This tutorial teaches on a specific, canonical network: 784 inputs (one per pixel of a 28 by 28 handwritten digit), two hidden layers of 16 neurons, and 10 outputs (one per digit). Count the connections and you get 12,960 weights plus 42 biases: 13,002 parameters exactly. Frontier language models are built from the same ingredient with hundreds of billions of parameters. Understand this machine and the big ones stop being magic; they are more of exactly this.

About the word 'neural'

The name honors a 1943 abstraction of brain cells that was really a logic gate, and the resemblance mostly ends there. Francis Crick made the point in Nature back in 1989: these networks are unrealistic as brain models in important ways. Treat 'neural' as branding for 'layered weighted sums', and you will reason about the technology more clearly than most headlines do.