Understanding AI at work / Short page
Understand · Short page
What is an LLM?
A large language model composes its answer by predicting, step by step, the most probable continuation of a text. This page summarises how it works, how it is built and what that changes for professional use.
The essentials
An LLM predicts what comes next in a text.
The definition, then what follows from it.
A large language model, or LLM, is a neural network trained to predict what comes next in a text. It does not look up a ready-made answer in a database: it composes one, unit by unit.
It computes probabilities from the context; it does not spontaneously check whether its answer is true or compliant.
CNIL, definition of a language model (in French) ↗ · Guide, chapter 02
01 / How it works
Five steps between your question and the answer.
The same computation starts again for every unit of text produced.
Split: the text becomes tokens.
The prompt and the documents supplied are cut into small units: words, pieces of words or signs.
Represent: tokens become numbers.
Each token is converted into a numerical representation that encodes the regularities learned during training.
Relate: attention weighs the context.
The Transformer examines the relations between tokens to determine which parts of the context matter most at that moment. Vaswani et al., 2017 ↗
Predict: the next token is chosen.
The model computes a probability distribution. Generation settings influence the choice and how much it varies.
Repeat: the answer is built step by step.
The chosen token joins the context, then the computation starts again until the answer ends.
After “The employee describes persistent pain in the…”, the model assigns a probability to each possible continuation: “back”, “wrist”, “knee”. It chooses what is probable in that context, not what is true for that employee. In the guide, the probabilities in this example are fictitious.
02 / How it is built
Five training stages, and the flaw each can leave behind.
An LLM’s capabilities are not programmed rule by rule.
| Stage | What comes out | Inherited flaw |
|---|---|---|
| Pre-train on web text, books and code | A base model, which completes text without following instructions. | Confabulation, biases and errors from the source texts; knowledge frozen at a cut-off date. |
| Fine-tune on exemplary dialogues | An assistant, which answers requests and refuses some of them. | It takes on the preferences of those who write and validate the examples. |
| Reinforce from compared answers | An assistant that is more useful and more cautious in its answers. | Sycophancy: what pleases can win over what is accurate. |
| Train to reason on problems with a verifiable solution | A reasoning model, which computes for longer before answering. | It may game the test rather than solve the problem; the reasoning it displays is not always faithful. |
| Evaluate, then deploy | A model that is updated and monitored: each version is a dated object. | It may detect that it is being evaluated; tests predict real use poorly. |
Simplified diagram: the order and weight of the stages vary from one laboratory to another, and some are repeated.
Ouyang et al., 2022 ↗ · Bai et al., 2022 ↗ · Guide, chapter 03
03 / At work
What this changes before you rely on it.
Four practical consequences.
- The tool is never the model alone. A professional assistant combines a model, instructions, documents, tools and sometimes a memory. Two tools built on the same model can answer very differently.
- A demonstrated capability is not guaranteed reliability. Often succeeding at a category of tasks does not let you predict each answer. You need to test the errors that matter in your use.
- A model version is a dated object. The provider can change the model, its settings, its filters or its tools. A useful evaluation states the version and the date.
- The data you enter goes somewhere. Check where it is processed, how long it is kept and what uses the provider allows itself; in France, health data requires certified hosting (HDS).
The guide details these points in the chapters From model to system and Capabilities and limits.
Key answers
Three frequent questions.
Short answers, drawn from the guide.
Does an LLM fetch its answers from a database?
Not by itself: it composes its answer from what it has learned and from what is in its context window. A tool can, however, supply it with extracts retrieved from a document base (RAG) or from the web. What does not enter that window does not exist for it.
Why does the same question give different answers?
Generation is probabilistic and sensitive to wording, to the order of documents and to the model version. At a low temperature, answers are stable, sometimes repetitive; at a high temperature, they vary more. Hence the value of testing several times, dating the trials and recording the version used.
What is the difference between a closed model and an open-weight model?
A closed model can only be used remotely, on the provider’s infrastructure. An open-weight model can be installed on servers you control, which gives you control over where the data travels. The largest ones nevertheless require data-centre infrastructure.
It summarises one chapter of the guide Understanding AI at work, which gives the source, date and limits of every figure. The figures dated November 2026 will go out of date: the guide is the version kept current.
Update log
- First version, drawn from the guide “Understanding AI at work” updated on 13 November 2026.