What is a neural network?

A neural network is a system, loosely inspired by the brain, made of simple units called neurons arranged in connected layers. Each connection has a strength, and by adjusting those strengths while learning from examples, the network becomes good at spotting patterns and making predictions. It is the structure underneath almost all modern AI.

That sounds abstract, so the clearest way in is a picture. Once you can see the shape of a neural network, the rest of the idea falls into place quickly.

The basic idea: layers of connected neurons

A neural network is built from small units, called neurons, lined up in columns called layers. Information enters at the input layer on the left, passes through one or more hidden layers in the middle, and leaves as a result at the output layer on the right.

Every neuron is joined to the neurons in the next layer, and each of those connections carries a strength, called a weight. A neuron takes the signals reaching it, combines them according to those weights, and passes a new value along. Do this across every layer and a signal flows from one side to the other.

How does it learn?

A fresh neural network knows nothing, and all its weights start out more or less random. Learning is the process of tuning those weights until the whole network produces good answers.

It works by example. The network is shown a piece of training data, makes a guess, and that guess is compared with the correct answer. Wherever it was wrong, every weight is nudged a tiny amount in the direction that would have made the answer better. Repeat that across millions of examples and the weights gradually settle into values that map inputs to good outputs.

This is why AI is said to learn rather than be programmed. Nobody writes the rules by hand. The rules emerge in the weights, drawn out of the data. That approach, learning patterns from examples, is what machine learning means, and the settled weights are the model’s parameters.

Why is it called “deep” learning?

The word deep refers to the hidden layers. A network with many of them stacked between input and output is a deep network, and training one is deep learning.

Depth matters because each layer can build on the last. In an image network, early layers might spot edges, later ones combine those into shapes, and later ones still into whole objects like a face. The same idea applies to language, moving from letters, to words, to meaning. More layers let the network capture more complex patterns.

What does this have to do with ChatGPT?

Everything, as it turns out. A large language model is a very large neural network of a particular design, trained on huge amounts of text to predict the next word. Its “knowledge” is nothing more than billions of finely tuned weights.

So when a chatbot writes a fluent reply, there is no library of facts being searched. A signal is flowing through layers of neurons, shaped by weights that learned the patterns of language. Our guide on how an LLM actually works follows that idea through to the words on your screen.

A way to picture it

If the diagram still feels abstract, try this. Imagine millions of tiny dials, one for each connection. At the start they are set at random, so the machine outputs nonsense. Training turns each dial a hair’s breadth at a time, guided by example after example, until the whole arrangement produces something useful.

Nobody sets those dials by hand, and nobody could. The data does it, which is both the power and the mystery of the approach.

Next steps

You now have the shape of the thing that underpins modern AI. A natural next step is to see how all those weights get set in practice, in our guide on how AI models are trained.