What is a Large Language Model?

Large language models are the technology behind many popular AI tools, including ChatGPT, Claude, Gemini and Microsoft Copilot. Here is what they are, how they work and why they can sometimes get things wrong.

What is an LLM?

You may have come across the term large language model, often shortened to LLM, when reading about artificial intelligence.

It sounds rather technical, but the basic idea is reasonably straightforward.

A large language model is a type of AI trained to recognise patterns in language. It can use those patterns to produce text, answer questions, summarise documents, translate languages, write computer code and carry out many other language-related tasks.

It does this by repeatedly predicting what should come next.

Why is it called a large language model?

The name becomes easier to understand when you break it into three parts.

Large

These AI models are trained using very large amounts of information and contain billions, or sometimes hundreds of billions, of adjustable numerical values called parameters.

Training them requires powerful computers, considerable amounts of electricity and, in many cases, months of work.

"Large" refers both to the amount of training involved and the enormous number of calculations taking place inside the model.

Language

The model is designed to work with language.

During training, it examines patterns found in books, articles, websites, conversations, computer code and other material. The exact mixture depends on the model and the information its developer is permitted to use.

It learns patterns such as:

  • Which words commonly appear together
  • How sentences are usually structured
  • How questions and answers relate to one another
  • How different styles of writing work
  • How ideas are connected across longer passages
  • How instructions are commonly followed
  • How computer code is arranged

It does not store this material as a normal collection of documents that it can simply look through. Instead, the training process adjusts the model's internal values so that they represent patterns and relationships found in the material.

Model

In computing, a model is a system that has been trained to identify patterns and make predictions.

A weather model predicts how conditions might develop. A fraud-detection model estimates whether a payment looks suspicious. A language model predicts how language is likely to continue.

The word "model" does not mean that it contains a perfect model of the world. Its understanding is incomplete and can sometimes be wrong.

How does an LLM work?

At its simplest, a large language model works by predicting the next piece of text.

If you entered:

The cat sat on the…

The model might decide that mat is a likely continuation. Other possibilities could include floor, sofa or windowsill.

It calculates which possibilities fit the text best and selects one. It then repeats the process for the next piece, and the next, building its response step by step.

Modern language models do this on an enormous scale. They take account of the words you have entered, the earlier conversation, the instructions they have been given and the patterns learnt during training.

That allows them to produce much more than simple sentence endings. They can generate paragraphs, compare ideas, explain subjects, imitate different styles, write computer programs and follow quite complicated instructions. Underneath it all, however, they are still predicting what should come next.

AI does not see words in quite the same way we do

Before a language model can work with text, the text is divided into small units called tokens.

A token might be:

  • A whole word
  • Part of a word
  • A number
  • A punctuation mark
  • A space combined with a word
  • A small piece of computer code

For example, a common word might be represented by one token, while a longer or less familiar word might be divided into several.

The AI converts these tokens into numbers. It then processes those numbers through many layers of mathematical calculations.

You do not need to understand tokens to use an AI assistant, but they help explain why an AI can sometimes struggle with exact spelling, character counts or unusually formatted text. It is not looking at language in quite the same way you are.

What are parameters and weights?

Inside an LLM are huge numbers of adjustable values known as parameters. Many of these are called weights.

During training, the model attempts to predict missing or following text. Its prediction is compared with the expected answer, and its weights are adjusted slightly to improve future predictions.

This happens over and over again across an enormous amount of training material.

A single weight does not normally represent one fact, word or idea. Knowledge is spread across large patterns of weights working together.

You can think of these weights as billions of tiny controls. Training adjusts them until the model becomes remarkably good at recognising and producing patterns in language.

Once the model has been trained, these values form part of what the AI has learnt.

Is an LLM just a very advanced autocomplete?

In one sense, yes. Both autocomplete and an LLM predict what text might come next.

However, calling an LLM "just autocomplete" can give the wrong impression.

A large language model considers far more information, contains vastly more internal connections and can recognise complicated patterns involving meaning, context, style and structure.

It can use these patterns to do things that ordinary autocomplete cannot, including:

  • Explain a difficult subject in simple language
  • Rewrite a letter in a friendlier tone
  • Summarise a long document
  • Compare different arguments
  • Suggest ideas for a holiday
  • Create a recipe from available ingredients
  • Translate between languages
  • Write and explain computer code
  • Continue a conversation across several messages

The basic mechanism involves prediction, but the scale and complexity of that prediction produce surprisingly capable results.

Is an LLM the same as a chatbot?

Not quite.

An LLM is the underlying language technology. A chatbot is the product or interface through which you communicate with it.

ChatGPT, Claude, Gemini and Microsoft Copilot are AI assistants built around large language models. They also include other technology that manages conversations, applies safety rules, remembers context and sometimes connects the model to outside tools.

Those tools might allow an AI assistant to:

  • Search the internet
  • Analyse a document
  • Examine an image
  • Generate a picture
  • Run computer code
  • Use information from another service
  • Remember details between conversations

The LLM provides the language capability, while the finished AI assistant provides the surrounding features.

Does an LLM understand what it is saying?

This is a difficult question because there is no universally agreed definition of understanding.

An LLM can identify complex relationships, follow instructions, draw connections and produce explanations that appear thoughtful. In practical terms, this can look very much like understanding.

However, it does not experience the world as a person does. It has no childhood, senses, personal history or direct experience of the things it discusses.

It learns from patterns in information rather than by living in the physical world.

It is therefore best to treat an LLM as a powerful language and reasoning tool, not as a person, a conscious mind or an all-knowing expert.

Does an LLM contain a database of facts?

Not in the usual sense.

A traditional database stores particular pieces of information in defined locations. A language model stores patterns across its internal weights.

This allows it to produce useful answers, but it also makes it less dependable when you need a precise fact.

An LLM may:

  • Combine two separate facts incorrectly
  • Give outdated information
  • Misremember a name or date
  • Invent a believable-looking source
  • State an uncertain answer with confidence
  • Produce a different answer when asked again

Some AI assistants can search the internet or consult connected databases. This can make their answers more current and verifiable, but the language model itself is not a search engine.

Important information should always be checked against a reliable source.

Why do large language models make things up?

An LLM is designed to produce a likely and useful response. It does not automatically know whether every part of that response is true.

If it lacks reliable information, it may still generate an answer that fits the patterns it learnt during training. The result can sound convincing while being partly or completely wrong.

This is sometimes called an AI hallucination.

For example, an AI might invent:

  • A book that was never published
  • A quotation nobody said
  • A legal case that does not exist
  • A feature that a product does not have
  • An incorrect date presented with confidence

This does not mean the AI is deliberately lying. It is producing plausible language without successfully separating fact from fiction.

Where does an LLM get its information?

The initial training material can include a mixture of:

  • Publicly available information
  • Licensed material
  • Information created by human trainers
  • Specialist datasets
  • Computer code
  • Examples produced or checked by people

The exact sources differ between models, and AI companies do not always publish a complete list.

After its initial training, a model may receive further training to make it more helpful, accurate and safe.

The AI assistant you use may also have access to current information through web searches, uploaded documents or connected services. This information is provided to the model while it prepares an answer, but it does not necessarily become part of the model's permanent training.

Does an LLM learn from each conversation?

Usually, the underlying model does not instantly change its internal weights because of something you tell it.

However, the AI service may keep conversations and later use some of them to test or improve future models. This depends on the provider, the product, your account type and your privacy settings.

An assistant may also have a separate memory feature that saves useful details about you. This can affect future conversations without retraining the underlying language model.

Before entering anything personal or confidential, check how that particular service handles your information.

What is a context window?

An LLM can only consider a certain amount of information at one time. This limit is called its context window.

The context window can include:

  • Your latest prompt
  • Earlier messages in the conversation
  • Uploaded documents
  • Instructions supplied by the AI service
  • Information retrieved from connected tools

If a conversation becomes too long, the model may lose access to some earlier details or receive a shortened summary of them.

This is why an AI can occasionally forget something mentioned much earlier in a conversation.

A context window is different from permanent memory. It is the information available to the model while it works on the current request.

Are all large language models the same?

No. Models vary considerably.

They may differ in:

  • The information used to train them
  • Their size and internal design
  • Their writing style
  • Their ability to follow instructions
  • Their skill at reasoning or coding
  • The amount of text they can handle
  • Their safety rules
  • Their speed and running cost
  • Whether they can work with images, sound or video

A larger model is not automatically better at every task. A smaller model may be faster, cheaper and perfectly suitable for everyday work.

AI companies also update and replace their models regularly. Two people using the same AI product at different times may not be interacting with exactly the same model.

What is a multimodal model?

Although LLM stands for large language model, many modern AI models can work with more than text.

A multimodal model may be able to process:

  • Images
  • Speech
  • Music and other sounds
  • Video
  • Documents containing text and diagrams

For example, you might upload a photograph and ask the AI to describe what it can see, or speak to it instead of typing.

Language often remains central to the conversation, but the model can use information from several different types of content.

What are LLMs good at?

Large language models are particularly useful for work involving language, ideas and patterns.

They can help you:

  • Draft letters and emails
  • Explain unfamiliar subjects
  • Summarise articles and documents
  • Generate ideas
  • Improve existing writing
  • Plan projects or journeys
  • Practise another language
  • Compare choices
  • Create lists and checklists
  • Understand technical terminology
  • Get started when facing a blank page

They work best when you provide clear instructions, useful background and an example of the result you want.

What are they less reliable at?

LLMs should be used with more care when a task requires:

  • Completely accurate facts
  • Up-to-date information
  • Complicated calculations
  • Professional medical, legal or financial judgement
  • Knowledge of events that happened after training
  • An exact quotation or source
  • Decisions about someone's safety or wellbeing
  • Confidential or sensitive information

They can still assist with some of these subjects, but their answers should not be accepted without checking them.

A simple way to think about an LLM

Imagine someone who has studied an enormous library and become extremely good at recognising how language, ideas and instructions fit together.

You can ask this person to explain, summarise, compare or create something. They can often produce an impressive answer very quickly.

However, they cannot return to every book to check each detail. Sometimes they confuse things, fill gaps with something plausible or speak confidently about something they do not properly know.

That comparison is not perfect, but it captures both the usefulness and the main weakness of a large language model.

The simple answer

A large language model is an AI system trained on enormous amounts of information to recognise patterns in language and predict what should come next.

That apparently simple process allows it to answer questions, explain ideas, write text, work with computer code and hold remarkably natural conversations.

But an LLM is not a person, a database or an unquestionable source of truth. It can misunderstand your request and produce convincing information that is wrong.

Used with clear instructions and a little healthy caution, it can be an extraordinarily useful tool.

Related: Back to Start Here | How AI Works: A Closer Look