Few words change their referent so often within a single conversation. It helps to separate at least three technical uses:
- The architecture. "The transformer model" names a computational recipe: a space of possible functions, not yet trained (see architecture).
- The trained weights. One particular point in that space, fixed by training. Chollet (2021) says it plainly: given the same data and the same architecture, different sets of weight values are different "models" (2021, sec. 5.4.4). When a model is released openly, what is shared is the weights together with the architecture (Alammar & Grootendorst, 2024, p. 30).
- The product. ChatGPT (InstructGPT) is not just a set of weights but a system with a system prompt, filters, tools, and an interface. That is why Elemedex sorts its species into architectures, models and systems (besides algorithms and units).
Beneath all three lies the statistical sense: a family of functions or probability distributions with adjustable parameters. What a learning method produces is, for Witten et al. (2017), a kind of "theory" of the domain, in the modest sense of something that predicts unseen cases (2017, p. 197).
The logician should beware of a false friend. In model theory, a model is a structure that makes true the sentences of a theory: the world interprets the language. In machine learning the direction is reversed: the model is the (approximate, numerical) theory, and the data are the world it is fitted to.
One rhetorical ambiguity remains. From the fact that a language model is a model, a simulation, people sometimes infer that it cannot understand anything. Santelli (2026) notes that this conclusion, far from reassuring, feeds the fear of capable systems that do not understand what they do. Either way, the word alone does not settle the question of understanding.