What is an LLM?
If you take even a passing interest in artificial intelligence, three letters keep showing up everywhere: LLM. People talk about the "LLM" behind ChatGPT, the "new LLMs" in the news… Don't panic — the concept is simple once you have the right mental image. It's actually the engine behind nearly every AI assistant you use.
If you haven't read our article on generative AI, it lays useful groundwork for this one. But you can also start right here.
Let's break down the three letters
LLM stands for Large Language Model. Three words, three ideas.
Model. In computing, a "model" isn't a scale replica — it's a system that has learned to perform a task from examples. A model that recognises fruit learned from photos of fruit. Here, the model learned from text.
Language. That's its specialty: human language. An LLM learns how words connect, how a sentence, a line of reasoning, an explanation are built. Its core job is to guess the next word in a text. Repeated billions of times, this little game produces whole sentences that are coherent and often relevant.
Large. This is what changes everything. "Large" refers to the colossal scale of the training: we're talking about amounts of text equivalent to millions of books, and models with billions of internal "settings" tuned during learning. It's this size that takes you from a basic autocorrect to an assistant capable of writing, summarising or explaining.
One image to remember it all
Imagine someone who had read a vast share of everything ever written: encyclopaedias, novels, manuals, forums, scientific articles. They haven't memorised each page word for word. But through all that reading, they've developed a very fine intuition for language: they know how to phrase a cover letter, how to explain photosynthesis to a child, how to draft a contract clause.
An LLM is exactly that, but in software form. When you ask it a question, it doesn't "look up the answer in a file". It draws on everything it has internalised to rebuild an answer that fits your request.
A few terms you'll come across
"Tokens". The LLM doesn't really read word by word, but in small chunks called tokens (sometimes a word, sometimes a syllable). It's its way of cutting up text. You'll meet this term especially when people talk about the cost or length of exchanges.
The context window. Roughly speaking, this is the model's working memory during a conversation: everything it can "keep in view" at a given moment (your question, its previous replies, a document you've pasted). If the conversation gets very long, the oldest elements may eventually drop out of this window — a bit like a discussion where you forget how it started.
The knowledge cutoff. An LLM was trained up to a certain date. Beyond that, it doesn't "know" recent events, unless it's connected to web search. Hence answers that are sometimes out of date on current affairs.
The limits to keep in mind
As we saw with generative AI in general, an LLM predicts what's plausible rather than "knowing" it. In practice:
- It can state something false with confidence (a hallucination).
- It can make mistakes in arithmetic or logic, because it isn't a calculator.
- It also reflects the biases present in the texts it learned from.
None of this is a dealbreaker: just use it as a gifted assistant that you proofread, especially for important information.
The LLMs you know (without realising it)
Most consumer AI assistants are built on an LLM:
- ChatGPT is built on the GPT models, developed by OpenAI.
- Claude is the LLM developed by Anthropic.
- LeChat is the assistant from the French company Mistral.
- Gemini is Google's model.
They share the same operating principle, but each has its own personality, strengths and preferred way of being addressed. That's precisely the subject of the next articles.
In short
An LLM, or large language model, is software that has learned the regularities of language from huge amounts of text, and uses them to produce answers by guessing the most likely continuation. It has a working memory (the context window), a knowledge cutoff, and it can be wrong. It's the engine behind ChatGPT, Claude, LeChat or Gemini.
Now that you know what's under the hood, the real practical question becomes: how do you talk to it well to get good answers? That's the art of the prompt, and it's what we explore from the next article on, starting with ChatGPT.