What Are Large Language Models?

A large language model (LLM) is a computer program trained on enormous amounts of text to do one core job: predict what text should come next, and generate it. ChatGPT, Claude, and Gemini are all built on large language models — when you type a question and receive a fluent answer, an LLM is the engine producing the words.

That sounds almost too simple for systems that can draft contracts and write working code. This article explains what is actually inside one of these models, what the key terms mean in plain English, and — just as important — what a language model is not doing when it answers you.

The name, taken apart

Each word in “large language model” carries meaning:

Model — a mathematical system whose behavior is set by vast numbers of adjustable internal values. Not a database of stored answers; a machine for producing answers.
Language — its raw material is text. It learns from text, takes text as input, and produces text as output.
Large — this refers to the scale of the model and its training: the number of internal values (parameters, explained below) and the volume of text it was trained on. “Large” says nothing about intelligence or understanding. A model is large the way a stadium is large — it describes size, not wisdom.

Tokens: how the model reads

An LLM does not read letters or whole words. It reads tokens — chunks of text, often a word, part of a word, or a common sequence of characters. The word “understanding,” for example, might be split into two or three tokens; a short common word like “the” is usually a single token.

Everything the model does happens in tokens. Your prompt is converted into tokens, and the model generates its reply one token at a time, each chosen based on everything before it. This is why AI usage is measured in tokens: they are the model’s unit of reading and writing. A token averages roughly three-quarters of a word in English, though it varies.

Prediction: the one job, done extraordinarily well

Strip away the mystique and an LLM does one thing: given a sequence of tokens, it calculates what token is most likely to come next. Then it adds that token to the sequence and repeats.

Given “The capital of France is,” the model’s training has exposed it to so many texts that “Paris” is the overwhelmingly likely continuation. Given a half-written email, it predicts the next token of a plausible continuation. Given a question, it generates the tokens of a plausible answer — because in its training data, questions were followed by answers, and it has internalized the patterns of how answers are shaped.

The surprising part is how far this single skill scales. Predicting text well, across every subject humans write about, requires the model to absorb grammar, facts, reasoning patterns, styles, and the structures of everything from legal contracts to Python code. Much of what looks like “thinking” in an LLM’s output is prediction operating on patterns absorbed at enormous scale. Whether that amounts to understanding is a genuine open debate — but the mechanism underneath is prediction, and remembering that explains both its brilliance and its characteristic failures.

The transformer: the machinery underneath

Modern LLMs are built on an architecture called the transformer, introduced in a 2017 research paper (“Attention Is All You Need”). You do not need the mathematics to understand its key idea: attention.

When a transformer processes text, each token can “pay attention” to the other tokens around it and weigh which ones matter most for deciding what comes next. In “The trophy doesn’t fit in the suitcase because it is too big,” the word “it” is ambiguous in isolation — attention lets the model weigh “trophy” and “suitcase” against “too big” and resolve the reference the way a reader does. Earlier architectures processed text more strictly in sequence and struggled with long-range connections; attention relates any part of the text to any other part, which is a large part of why these models handle long, complex documents as well as they do.

Parameters: the learned settings

Inside the model are parameters — often called weights. A parameter is a number, and the model contains vast numbers of them. Think of them as the settings on an unimaginably large mixing board: no single setting means anything on its own, but together, their combined values determine exactly how the model responds to any input.

These settings are not programmed by hand. They are learned during training: the model starts with essentially random settings, predicts, measures how wrong it is, and adjusts slightly to do better — billions of times over (see How Do AI Models Learn?). When reporting says a model has “billions of parameters,” it is describing the size of that learned configuration. More parameters generally mean more capacity to capture subtle patterns — at the cost of more computation to train and run.

Two consequences follow. First, the model’s knowledge lives in the pattern of its parameters, not in files it looks up — it cannot show you the page it “read” a fact on, because there is no page. Second, once training ends, the parameters are fixed: the model you talk to today has the same settings as yesterday. It does not learn from your conversation in real time.

Context windows: the model’s working memory

An LLM can only consider a limited amount of text at once — the prompt, the conversation so far, any documents provided, and its own reply-in-progress. That limit is the context window, measured in tokens.

It works like a desk: everything on the desk, the model can use; anything that falls off might as well not exist for that conversation. In a long conversation, early messages can drop out of consideration. Documents pasted into a chat sit on the desk for that session only — when it ends, they are gone, because they were never added to the model’s parameters. This is why a model can discuss your document brilliantly today and know nothing about it tomorrow.

What LLMs power — and what they are good at

LLMs are the engine inside conversational assistants such as ChatGPT, Claude, and Gemini, and inside thousands of other products: writing assistants, code generators, customer-support systems, document summarizers, and search features that answer questions in sentences instead of links.

They are genuinely strong at tasks that are, at heart, about patterns in language: drafting and rewriting text, summarizing, translating, explaining concepts, generating and debugging code, brainstorming, and reformatting information. They are fast, tireless, and broadly knowledgeable in the way a very widely read person is broadly knowledgeable.

Where they fail — the honest pass

The same mechanism produces specific, predictable weaknesses:

They can state falsehoods fluently. The model generates what a good answer would look like. If the accurate continuation and the plausible one differ, it may produce the plausible one — confidently, in polished prose. This is the hallucination problem, and it is structural, not a bug a settings change removes.
Their knowledge has a boundary. A model’s parameters reflect training data gathered up to a point in time. Later events are not in the model unless the product adds outside information through retrieval or browsing.
They have no direct line to truth. The model has no internal fact-checker and no awareness of which of its statements are solid and which are shaky. Confident wording is not evidence of accuracy.
They are sensitive to phrasing. Ask the same question two ways and you can get meaningfully different answers, because you have given the prediction engine a different starting sequence.
They do not understand the way you do. Operationally, the model has no experience of the world, no intentions, and no stake in being right. It is an extraordinarily capable pattern system, and treating its output as the word of an accountable expert is a category error.

Misconceptions and limitations

“The model looks up answers.” Not by default. A base LLM generates from its parameters. Some products add web search or retrieval on top — a separate mechanism (see How AI Systems Retrieve Information) — and the answer then blends retrieved material with generated text.

“Bigger means it understands more.” Larger models are generally more capable, but “large” measures parameters and training scale — not understanding, and bigger models hallucinate too.

“It remembers me.” Within a conversation, the model uses the context window. Across conversations, a plain LLM retains nothing unless the product around it saves and re-supplies information.

“It knows when it’s wrong.” It does not — which is why verification matters, and why the sources an AI answer relies on have become so important.

Why this matters for business information

Because an LLM generates from patterns absorbed in training and, in many products, from material retrieved at answer time, the information available about your business — your website, directories, reviews, articles — shapes what the model can say about you. Accurate, consistent, readable information gives the pattern engine good material. Sparse or contradictory information gives it gaps to fill with plausible invention. That bridge — from how these models work to how businesses get represented in AI answers — is the subject of the rest of this site, starting with the pillar overview, How AI Search Works.