Double Espresso: What Is an AI Model?

What Do We Mean by an AI Model?

You have read the sentence a hundred times by now. A company has released a new AI model. Trained one. Tested one. Restricted one. Given one away for free. The sentence slides past so easily that most of us never stop to ask the obvious thing: what is the model, exactly? Where does it start, where does it stop, and how is it different from the app you actually open on your phone?

Here is the definition, and it is far less exotic than the headlines suggest. A model is a piece of machinery that turns an input into an output. You put something in — an email, a photograph, half a sentence, yesterday's air pressure over the Atlantic — and something comes back out: a score, a label, a forecast, an image, a continuation. One thing goes in, another comes out. Everything interesting is in the middle.

Artificial intelligence is the broad category. Machine learning is one particularly successful way to build it, but it is not the only one. Some AI systems use knowledge and rules that people have encoded directly. Machine learning takes a different route: it fits a model to data. Instead of specifying every useful relationship by hand, a training process adjusts the model until its outputs line up better with examples or other feedback.

Take spam, because everyone has a junk folder.

One kind of filter starts with conditions written by people. If the subject line is all capitals, add a point. If the sender is not in the address book, add a point. If the word “invoice” appears beside a link nobody asked for, add two. Cross a threshold and the message goes to junk. A sufficiently rich knowledge-based system can be a form of AI, although an ordinary list of rules is not magically intelligent just because it runs on a computer.

Now build a machine-learning filter. You do not try to anticipate every rule. You gather messages that people have already marked as spam or not spam, and you run a process that adjusts a structure full of numerical parameters — thousands of them, perhaps far more — until the model's verdicts start lining up with the examples. When training finishes, the learned structure and its fitted parameters are the model. Not the pile of old emails. Not the training process. The fitted thing that can now take a new message and give it a score.

This is not an argument that learned models are always better than handwritten rules. It is simply the boundary we are trying to see.

The spam model can score a message. It cannot fetch your mail, move the message, display the junk folder or remember that you clicked “not spam” last Tuesday unless the surrounding system gives it that information and acts on the result. The mail server, the app, the folder, the user controls and the standing rule that a message from your mother should never be discarded — those belong to the system around the model.

The same distinction explains why two chat products can use the same base model and behave like different animals. One can search the web and another cannot. One remembers earlier conversations and another starts afresh. They can have different instructions, tools, interfaces and safety layers. Some behaviour is shaped during the model's own training; much more can be added around it. Same underlying model does not mean same finished product.

It also means that when a company publishes a model's weights, it has not necessarily handed over the product. The weights and architecture may be the central engine, but the interface, search, memory, tools, operating rules and ordinary software around them can remain separate.

The word “learning” needs straightening out too. It is useful shorthand and a poor literal description. A model fitted to millions of emails can capture real patterns — regularities a person might never manage to put into words. What it has not acquired is a life: no experience, no intentions and no human common sense to fall back on when the pattern runs out.

But do not overcorrect into the opposite mistake: the idea that a model merely looks up and repeats something from its training data. Models can memorise particular examples, sometimes uncomfortably. Their ordinary operation, however, is a computation from the current input and the structure fitted during training. A larger system may separately search a database or the web and place the result in front of the model. That retrieval is something the system does; it is not the definition of a model.

Models come in every size. A spam classifier, a song recommender, a weather predictor, an image generator and a language model all produce very different things, but the same mental handle survives: something goes in, something comes out, and the model is the fitted machinery in between.

So the next time a headline says that an AI did something, split the claim in two. The model is the part that computes an output. The system is everything that decides what information reaches it and what happens next.