How AI works, without the maths

A neural network is built from simple units arranged in layers. Each unit takes in numbers from the layer before it, multiplies each one by a weight, adds them up and passes the result on. The weights are what the network “knows”.

One unit does almost nothing. Millions of them, stacked in many layers, can recognize a face, translate a sentence or continue a paragraph. The idea is loosely borrowed from how brain cells connect, but a neural network is arithmetic, not biology.

A network starts with random weights and gets everything wrong. Training fixes that through repetition:

1. 1Show an example. For a language model, that is a passage of text with the next word hidden.
2. 2Let it guess. The network produces its prediction.
3. 3Measure the error. Compare the guess with the real answer.
4. 4Nudge the weights. Every weight is adjusted slightly in the direction that would have made the guess better.

Repeat this billions of times across a large share of the written internet, and the weights settle into values that capture grammar, facts, reasoning patterns and writing styles. Nobody programs those in. They emerge from the examples.

The model is the finished set of weights, also called parameters. Today's large models have billions of them. When a company releases a new model, it is releasing a new set of trained numbers, usually from a bigger network, better training data or a better training method.

That is why models come in sizes. A small one is quick and cheap to run. A large one is slower and costs more, but handles harder problems. See the AI models, explained for who makes what.

A language model works with tokens, pieces of text about three-quarters of a word long. Given everything written so far, it calculates how likely each possible next token is, picks one, adds it to the text and repeats. A long answer is hundreds of these small predictions in a row.

The design that made this work well is the transformer, introduced in 2017. Its key idea is attention: for each new token, the network weighs every earlier token and decides which ones matter most. That is how it keeps track of who “she” refers to three sentences later.

The amount of text it can consider at once is its context window. Prices are quoted per million tokens read and written, which is why the daily feed often reports them.

Because it predicts what is plausible, not what is verified. If the most likely-sounding continuation is wrong, the model can state it smoothly and confidently. This is called a hallucination.

A model also knows nothing after the date its training data ends, unless it is connected to a search tool or given documents to read. Makers reduce errors by letting models look things up, show their sources and check their own work, but the risk never falls to zero. For anything that matters, such as a tax figure or a contract term, check the source.

Parameter (or weight)

One of the numbers inside the network that training adjusts. Large models have billions of them.

Token

The small piece of text a model reads or writes at a time, roughly three-quarters of a word.

Context window

How much text a model can take into account at once: your question, any documents, and the conversation so far.

Training

The one-off process of adjusting the parameters by showing the network vast numbers of examples.

Inference

Using the finished model to produce an answer. This is what happens each time you send a message.

Fine-tuning

Extra training on a smaller, specialized set of examples to make a model better at one kind of task.

Multimodal

Able to work with more than text, such as images, audio or video.

Hallucination

A confident answer that is wrong or made up.