What is a large language model?

A large language model is an artificial intelligence trained on enormous amounts of writing to do one narrow thing: predict what piece of text should come next. Feed it some words and it guesses the words most likely to follow, then the words after those, and so on. What makes it remarkable is that doing this one job well enough turns out to produce a system that can answer questions, draft an email, summarise a report or write working code — none of which it was ever explicitly taught, all of which fall out of getting the prediction right.

The one job

Strip away the marketing and a language model does exactly one thing. You give it a stretch of text; it produces the next fragment. Then it looks at your text plus its own fragment and produces the next one after that. It repeats, one small piece at a time, until it has built a full reply.

The clearest way to picture it is the autocomplete on your phone, scaled up almost beyond recognition. Your phone suggests the next word from a handful of likely options. A language model does the same trick, but drawing on patterns it absorbed from a vast sweep of human writing, and it keeps going far past a single word into whole paragraphs. The mechanism is that simple: guess the next piece, add it, guess again.

Everything else people describe a model doing is this loop wearing different clothes. Asking a question is just giving it text that a helpful answer would naturally follow. Writing code is predicting the text that comes after a description of what the code should do.

Why that simple job does so much

Here is the surprise at the centre of the technology. Predicting text well is not a small skill. To reliably guess what comes next across everything people write — recipes, legal arguments, Python functions, poetry, apologies — a model has to internalise a great deal about how language and the world hang together.

To finish the sentence “The capital of France is”, it has to have absorbed a fact. To continue a paragraph of formal writing without lurching into slang, it has to have absorbed style. To carry on a chain of reasoning without contradicting itself, it has to have absorbed something of how arguments hold their shape. None of this was programmed in as rules. It was all pressed into the model as a side-effect of being pushed, over and over, to make its next-word guesses more accurate.

That is the whole trick. Capability is not the goal the model was given; it is what falls out of pursuing accurate prediction across the enormous variety of things humans write down.

What ‘large’ means

The word ‘large’ points at two things at once, and both matter.

The first is the training material. A model of this kind is exposed to a huge collection of text — far more than any person could read in many lifetimes — spanning countless topics, styles and languages. The breadth is the point: patterns it never saw during training are patterns it cannot reproduce well afterwards.

The second is the model’s own size. Inside it sit billions of adjustable internal settings, the dials that get nudged during training until its predictions sharpen. More settings give a model more room to store subtle patterns, which is part of why recent models handle nuance that earlier, smaller ones fumbled. The specific numbers change with every new release and are not the concept; the concept is that both the reading and the machine doing the reading are on a scale that only became practical recently.

What it is not

Clearing up three misconceptions explains most of what confuses people.

A language model is not a database it looks things up in. It holds no tidy table of facts. It reconstructs an answer from statistical patterns each time you ask, which is why the same question can produce slightly different wording, and why it can state something false as confidently as something true.

It is not a search engine. It is not fetching a page and reading it back to you. Unless it has been connected to a live search tool, it is generating text from patterns it absorbed during training, not retrieving a document that exists somewhere.

And it is not a mind that understands what it says. It is a pattern machine. This is exactly why a model can be fluent and wrong at the same time: fluency comes from having learned the shape of good writing, while correctness would require something the prediction task never guarantees. Whether these systems ‘understand’ or ‘reason’ in any real sense is a live question researchers actively debate, not a settled fact — and treating a confident answer as a checked one is the single most common mistake people make with this technology.

The names you have heard

Most people meet language models through a product rather than the underlying technology: ChatGPT is the most widely recognised, though it sits alongside others such as Gemini, Copilot and Claude, each an application built on top of a language model rather than the model itself.

The distinction to carry forward is between the concept and the products that package it. Understand the prediction loop and you understand what all of them are doing underneath, whichever name is on the interface.