You have probably used one today without noticing. The assistant that drafts your email, the chatbot that answers a support question, the tool that summarises a long document: many of them run on a large language model, usually shortened to LLM. The term sounds technical, but the core idea is simple enough to explain without any maths.
This guide explains what a large language model is, how it learns, how it produces an answer, what it is good at, where it fails and what that means when you choose and use AI tools. It is written for people who use these tools at work rather than build them, so there is no code and no jargon without a definition.
Key takeaways
- A large language model is a program trained on huge amounts of text to predict what text should come next. From that single skill, it can write, summarise, translate, answer questions and more.
- It does not look up facts in a database by default. It generates likely text, which is why it can sound confident and still be wrong.
- What it can see at once is limited by its context window. Anything outside that window is invisible to it.
- Training, fine-tuning and prompting are three different ways of shaping behaviour, and you mostly control the third.
- The practical rule: use LLMs for drafting, transforming and organising text, and verify anything where accuracy matters.
A plain definition
A large language model is a computer program that has learned patterns in language by processing a very large collection of text. Given some text as input, it predicts what text is likely to follow, and by repeating that step many times it produces whole sentences, paragraphs, tables or code.
The word “large” refers to two things. The first is the amount of text it learned from, which can span books, websites, articles, forum discussions, documentation and code. The second is the size of the model itself, measured in parameters. Parameters are the adjustable numbers inside the model that store what it has learned. You can think of them as the settings that were tuned during training. Larger models have more of them and can usually capture more subtle patterns, although size alone does not decide quality.
An everyday comparison
Think about your phone’s autocomplete. When you type “see you at”, it suggests “noon” or “the office”. A large language model works on the same principle, but at a vastly greater scale and sophistication. Instead of suggesting the next word based on your recent messages, it draws on patterns from an enormous body of text, and it can keep a long passage in mind while it decides what to write next. That extra range is what turns autocomplete into something that can draft a proposal.
How a language model learns
The basic training task is surprisingly simple. The model is shown a piece of text with the next word hidden and asked to guess it. At first its guesses are poor. Each time it is wrong, the training process nudges its parameters slightly so that the correct answer becomes a little more likely next time. This is repeated across billions or trillions of examples.
To get good at predicting the next word across such a variety of text, the model has to pick up many things along the way: grammar, vocabulary, common facts, how arguments are structured, how a recipe is laid out, how a legal clause is worded, how a function is written in a programming language. Nobody programs these rules in by hand. They emerge from the pattern-matching that training encourages.
What it is not doing
The model is not storing a library of documents it can open and quote exactly, and it is not browsing the internet while it answers unless the tool around it adds that feature. It has compressed patterns from its training data into its parameters. This is why it can explain a concept in many different ways, but also why it can fail to reproduce an exact quotation or invent one that sounds plausible.
Tokens: how text is broken up
Models do not read whole words the way people do. Text is split into small units called tokens. A token might be a whole short word, part of a longer word, a number, or a punctuation mark. A common rule of thumb for English is that a token is roughly three quarters of a word, though this varies by model and language.
Tokens matter for practical reasons:
- Limits are measured in tokens. The amount of text a model can handle at once, and the length of its reply, are capped in tokens.
- Pricing often uses tokens. If you use an API or a pay-as-you-go tool, cost usually depends on tokens in and tokens out.
- Languages differ. Some languages need more tokens to express the same idea, which can affect cost and how much fits in the window.
- Odd mistakes have a cause. Tasks like counting letters in a word can trip a model up because it sees tokens, not individual letters.
How it produces an answer
When you send a prompt, the tool breaks it into tokens and feeds them to the model. The model then calculates a probability for each possible next token. It picks one, adds it to the text, and repeats the process until the answer is complete. That is why you often see a reply appear word by word.
Why answers vary
Most tools do not always pick the single most likely token. A setting often called temperature controls how much randomness is allowed. A low temperature gives more predictable, repeatable answers. A higher one gives more varied, creative output. This is why asking the same question twice can produce different wording, and sometimes different content.
Why the prompt matters so much
Because the model continues from what it has been given, the prompt sets the direction for everything that follows. Clear instructions, relevant background and a good example all steer the output. Vague prompts leave the model to fill gaps with the most common patterns, which tends to produce generic answers. For a practical method, see our guide to writing better prompts.
Why it seems so capable
A model that only predicts the next token can nonetheless do many things that look like understanding. It can summarise a report, translate between languages, explain an idea at different levels, write in a chosen tone, follow multi-step instructions and produce working code. Several factors explain this.
- Scale of experience. Having seen vast numbers of examples of each kind of task, the model has absorbed the typical shape of a good answer.
- Flexible pattern use. It can combine patterns, such as the structure of a business email with the subject matter of a software release.
- Instruction training. Later training stages teach it to follow instructions and respond helpfully, not merely continue text.
- Context. Whatever you paste into the prompt becomes material it can work with, even if it never saw it in training.
Whether this amounts to true understanding is debated by researchers, and you do not need to settle the debate to use the tools well. The practical point is that the output can be genuinely useful, and it can also be wrong in ways that look convincing.
What it cannot do reliably
Facts and sources
An LLM can state something false with complete confidence. This is often called a hallucination. It happens because the model is generating plausible text, not retrieving verified facts. Invented statistics, fictional citations and incorrect dates are common examples. Tools that connect the model to search or to your own documents can reduce this, but they do not remove it.
Up-to-date information
A model learns from data up to a certain point, often called its training cutoff. It does not automatically know about events after that date unless the tool adds a search or retrieval feature. If you ask about recent news, prices or product features, confirm the answer from a current source.
Precise calculation and counting
Models can do arithmetic and logic reasonably well in many cases, but they can also make errors, especially with long calculations, large numbers or tasks that need exact counting. For anything that must be exact, use a calculator, a spreadsheet or a tool that runs code, and check the result.
Knowing what it does not know
Models are often poor at recognising the limits of their own knowledge. They may answer a question they should have declined. You can reduce this by asking the model to say when it is unsure, and by providing source material and asking it to answer only from that.
Consistency and bias
Because models learn from human-written text, they can reflect gaps and biases found in that text. Their answers can also vary between runs. For tasks like screening, scoring or decisions about people, human oversight is essential.
Confidentiality
What you type into a tool goes to the provider. Before you share sensitive material, check the provider’s terms on retention and training, and follow your organisation’s rules.
The context window
The context window is the amount of text, measured in tokens, that a model can consider at one time. It includes your instructions, any documents you paste in, the conversation so far and the reply being generated. Think of it as the model’s working memory.
Why it matters
- Long documents. If a report is longer than the window, part of it will not be seen. Tools may truncate it, summarise it in chunks or retrieve only relevant sections.
- Long conversations. In a long chat, early messages can fall out of the window, and the model may seem to forget earlier instructions.
- Attention to the middle. Some models handle information at the start and end of a long input better than information buried in the middle. If something is critical, state it clearly and repeat it near the end.
- Cost and speed. Larger inputs usually take longer and can cost more.
Window sizes have grown considerably over time, so check the current limits for the tool you use rather than relying on a number you read once.
Pre-training, fine-tuning and feedback
Building a useful assistant usually involves several stages, and it helps to know the names.
| Stage | What happens | Who does it |
|---|---|---|
| Pre-training | The model learns language patterns by predicting the next token across a very large dataset | The model developer |
| Fine-tuning | The model is trained further on a smaller, targeted dataset to improve at specific tasks or follow instructions | The model developer, or sometimes a customer |
| Feedback training | People or other systems rate responses, and the model is adjusted to prefer helpful, safe answers | The model developer |
| Prompting | You give instructions, context and examples at the time of use | You |
| Retrieval | The tool fetches relevant documents and adds them to the prompt | The tool or your team |
For most people, prompting and retrieval are the levers that matter. You rarely need to train or fine-tune a model to get good results, and many business needs are better met by giving the model the right information at the time of the question.
Why chat tools are more than the model
When you use a chat assistant, you are using a product built around a model. The product may add a system prompt that sets behaviour, safety filters, memory features, web search, file upload, tools that run code, connections to your apps and a user interface. Two products using similar models can behave quite differently because of what surrounds them.
This is useful to remember when you evaluate tools. The question is not only which model is underneath, but how the whole product handles your task, your data and your workflow. Our checklist for choosing an AI tool covers how to test that fairly.
Agents
Some products let the model take actions, such as searching, running code or updating a record, and then continue based on the result. These are often called agents. They are more powerful and carry more risk, because mistakes can have real effects. Keep humans in the loop for actions that are hard to undo.
Open and closed models
Models are released in different ways. Closed models are accessed through a provider’s website or API, and you cannot see or run the model yourself. Open-weight or open-source models can be downloaded and run on your own hardware or a server you control, subject to their licences.
| Factor | Closed models | Open models |
|---|---|---|
| Setup effort | Low, use a website or API | Higher, you host and maintain |
| Data control | Depends on provider terms | Can stay in your environment |
| Customisation | Limited to what the provider offers | Greater freedom to adapt |
| Cost pattern | Usually usage or subscription fees | Infrastructure and staff time |
| Licence | Provider terms of service | Varies, read each licence |
Neither is better in every case. Many organisations use a mix, and the right choice depends on privacy needs, skills and budget.
Using LLMs well at work
Good uses
- First drafts of emails, articles, reports and outlines.
- Rewriting text for a different audience, tone or length.
- Summarising documents and meeting notes that you can verify.
- Brainstorming names, angles, questions and options.
- Explaining unfamiliar concepts in plain language.
- Drafting and explaining code, with testing.
- Turning messy notes into structured plans or tables.
Uses that need care
- Anything involving legal, medical, financial or safety advice.
- Statistics, quotations and citations, which must be checked.
- Decisions about hiring, credit, discipline or other matters affecting people.
- Confidential or personal data.
- Published content that carries your name or brand.
A simple working habit
- Give the model the context and the source material.
- Ask for a draft in a clear format.
- Ask it to list assumptions and anything it is unsure about.
- Check facts and figures against reliable sources.
- Edit so the result reflects your judgement and voice.
Quick glossary
| Term | Plain meaning |
|---|---|
| Token | A small chunk of text the model reads and writes |
| Parameter | An adjustable number inside the model that stores learned patterns |
| Prompt | The instructions and information you give the model |
| Context window | How much text the model can consider at once |
| Hallucination | A confident but false or invented statement |
| Fine-tuning | Extra training for a specific task or style |
| Temperature | A setting that controls how random the output is |
| Training cutoff | The point after which the model has no built-in knowledge |
| Retrieval | Fetching relevant documents and giving them to the model |
| Agent | A system where the model can take actions and use tools |
Frequently asked questions
Is a large language model the same as artificial intelligence?
No. Artificial intelligence is a broad field that includes many techniques. A large language model is one type of AI system, focused on language. Other AI systems recognise images, recommend products or control machines.
Does an LLM understand what it writes?
Researchers disagree about how to describe what these models do. They produce text that often reflects real structure and meaning, but they can also make basic mistakes. For practical purposes, treat the output as a capable draft that needs checking, not as the word of an expert.
Why does it sometimes make things up?
Because it generates plausible text based on patterns, not verified facts. If the prompt asks for something it has no reliable information about, it may still produce an answer that sounds right. Supplying source material and allowing it to say “I do not know” reduces the problem.
Does it learn from my conversations?
It depends on the provider and the plan. Some providers use conversations to improve their models unless you opt out or use a business plan, and others do not. Read the data terms of any tool before you share sensitive information.
Can it access the internet?
The model itself does not. Some products add a search or browsing feature so that the tool can fetch current pages and give them to the model. Check whether your tool does this and whether it shows its sources.
Will a bigger model always be better?
Not always. Larger models tend to handle complex tasks better, but they can be slower and more costly. For simple tasks, a smaller model can be faster, cheaper and entirely adequate. Test on your real task.
How do I get better answers?
Give clear instructions, context, an example of what you want and the format you need. Ask the model to flag uncertainty, then verify important claims yourself.
Conclusion
A large language model is a pattern-learning system trained to predict text. That one ability, scaled up and refined with extra training, produces tools that can draft, summarise, translate, explain and code. It also explains the main weaknesses: invented facts, no built-in knowledge of recent events, limited memory in the context window and variable answers.
If you remember a few ideas, you will use these tools better than most. Give the model what it needs in the prompt. Ask it to be honest about uncertainty. Use it for drafting and transforming, and verify anything that matters. As you go further, read about prompting and about choosing the right tool for your work.
