How AI Works

Neural networks, briefly

Layers of weighted connections

Scroll

Layers of weighted connections

A neural network is the structure inside almost every modern AI model. It is layers of simple units, each connected to units in the next layer, and every connection has a weight: a number saying how strongly one unit's output affects the next.

Input goes in at one side as numbers, flows through the layers, and comes out the other side as an answer — scores for cat and dog, or for every possible next token. Each unit does very little on its own: it adds up its weighted inputs, applies a simple bend so the network can represent more than straight lines, and passes the result on.

The weights are what training adjusts, and they are all the model knows. An LLM is this same idea at enormous scale with a particular arrangement of layers, described on the next two pages. When someone quotes a model's size in billions of parameters, these weights are what they are counting.

Foundations
LAYERS OF WEIGHTED CONNECTIONS inputs hidden layer outputs cat 0.9 dog 0.1 Every connection carries a weight. Training adjusts millions of them until the outputs come out right. An LLM is this idea at enormous scale.
Three inputs, one hidden layer, two outputs. The lines thicken and thin as weights change.