model

In depth

Training and use

A model is a set of numerical parameters adjusted during a training run, then frozen. Training and use are two distinct moments: during training the parameters move; during use they no longer move, and each request is handled as if no other had ever taken place. A model therefore learns nothing from your exchanges: what you take for memory is a text that the software around it puts back in front of it every time. It is also why its knowledge of the world stops at a date: that of its training data.

Not only language

Not all models are language models. Models are also trained to produce images, transcribe speech, classify documents or forecast demand, and the word stays the same because the gesture is the same: adjusting parameters until a regularity is captured. The common confusion consists in reserving the term for the systems that write text, then generalising to the whole of AI properties that hold only for those. Naming the task before naming the model avoids most of these misunderstandings.

What gets wrongly blamed on it

The most costly trap in an organisation is to charge the model with what belongs to what has been built around it. One and the same model, connected to different instructions and sources, produces unrelated results: changing it when the fault comes from the harness amounts to replacing the part that was working. Nor is a model deterministic by default, since the same request can give two different answers, which makes fragile any evaluation done on a single try. Finally, a model carries a version, and versions succeed one another: measuring, comparing or certifying without noting which one makes the finding unusable a few months later.