Model, LLM, training: the words of the engine

Three words that come up in every article, and which actually describe the same thing from three different angles.

A model

The program itself, once it has finished learning. It's the engine. ChatGPT is a service; GPT-5 is the model running inside it. You can swap the engine while keeping the same car, and that's exactly what happens with each new version.

An LLM

"Large Language Model," which we translate as grand modèle de langage. It's the type of engine we've been talking about since 2022: a model specialised in text. It predicts the next word, over and over, and that's where everything else comes from.

Training

The long, very costly period during which the model learns by reading. It happens once, beforehand. When you chat with it, it's no longer learning: it's applying what it already knows.

That's why it can be unaware of a recent event: it hadn't been born yet when it happened.

Parameters

The number you see bandied about, "70 billion parameters." These are the internal settings adjusted during training. The more there are, the finer the model can be, but it's not a quality score: a small, well-trained model often beats a large, poorly trained one.