large language model

In depth

A single task

A large language model is trained on a single task: guessing the next element in a sequence of text. The text reaches it cut into tokens, it gives each possible token a probability of being the next one, a draw picks one, the result is fed back in and the operation starts again. All the apparent variety of uses, summarising, translating, drafting, classifying, follows from this single repeated operation. The consequence is direct: the model always produces something, including when it has no reason to be right.

What it does not have

By itself, a large language model has neither memory, nor access to the outside world, nor any capacity to act: those three things are added to it by what is built around it. It knows only what was in its training data and what is put in front of it at each call. This is what makes the boundary with the agent clear: the model returns an answer and stops there, the agent is the system that chains the turns and triggers actions. Confusing the two lends the model powers that belong to the harness, or charges it with limits it no longer has.

Fluency and accuracy

The most frequent trap is to read ease as competence: the fluency of an answer and its accuracy are two independent dimensions, and style says nothing about correctness. A second trap has to do with size: a bigger model is not better on every task, and it costs more at each call, which leads serious architectures to reserve the most capable models for the steps that deserve them. A third is more insidious: the model does not natively tell an instruction from a document, since everything reaches it as text. An instruction slipped into a page can therefore be followed as if it came from you, and that is the starting point of most security questions on this floor.