What a Model Actually Is

The data team says "the model" the way sailors say "the ship". The model predicts. The model was retrained over the weekend. The model is acting strange on Fridays. After a week of this, Iris finally asks the question she suspects is too basic: what physically is it? Could someone show her the model? Point at it?

They can, and the answer is refreshingly boring. A model is a file full of numbers, plus a fixed recipe for turning an input into an output using those numbers. That's it. It sits on a disk like any document. This page makes that concrete, because "model" is the word the entire rest of the book leans on — and the mystique around it dissolves the moment you can picture the file.

Inputorder: 3.2 km, 7 pm, Thai place

→

Modellearned numbers + fixed recipe

→

Prediction38 min · confidence 82%

A Machine for Answering One Question

A model answers exactly one kind of question. Plateful's delivery model takes an order — the distance, the hour, the restaurant — and returns one thing: a predicted number of minutes. Ask it anything else and there is simply no slot for the question. It cannot flag a fake review, and it certainly cannot chat; each of those is a different model, trained on different examples.

Think of a well-worn recipe card for one dish. Ingredients in, that dish out, with the amounts tuned over many past attempts. The card does not understand cooking — it just encodes what worked. A kitchen that serves many dishes needs many cards, and a company that makes many predictions needs many models. When you hear that some app "has AI", the reality underneath is usually a shelf of these single-question machines, each doing its one narrow job.

What's Inside the Box

So what fills the file? Numbers — the tuned "amounts" on the recipe card. In machine learning they are called parameters, and they are the dials from Topic 01: the values the computer adjusted during learning until its predictions matched the examples. A small model might hold a few hundred of them; the giant language models in Chapter 9 hold billions. Different scale, same substance.

This gives you the precise meaning of a phrase you will hear constantly. Training a model — the subject of Chapter 3 — is nothing but the process of setting those numbers. Before training, they are random and the model's answers are nonsense. After training, they encode the patterns found in the examples. The recipe part — how the numbers combine with the input — is fixed by engineers and never learns anything; all the learning lives in the numbers.

It Is a File

Because a model is a file, everything you know about files applies. It can be copied, backed up, versioned, sent to another computer, and deleted. When Plateful's data team "ships the new model", they mean it about as literally as attaching a document: the new file of numbers replaces the old one, and the app starts asking it the question instead.

Files also don't change themselves — and neither do models. A trained model is frozen: it computes with the numbers it has, and using it a million times does not nudge a single one. When Iris hears "the model was retrained over the weekend", that was a deliberate, separate event — a new round of learning producing a new file — not the model quietly improving on the job. Why teams retrain, and how often, is a Chapter 7 story; for now, hold the picture: trained, then frozen, then used.

A Guess with a Number on It

One more thing about that output. The model's answer is not knowledge; it is its best guess given the patterns it saw, and it usually comes with a confidence attached — "38 minutes, and predictions like this have been right about 8 times in 10". Confidence is useful: the app can show a wide time range when the model is unsure and a narrow one when it is not.

But here is the sentence to underline, because it returns on nearly every page ahead: a model can be confidently wrong. High confidence means the input strongly resembles patterns from training — it does not mean the world will cooperate. A festival closes the bridge; the prediction was confident and the pasta is an hour late. Keep "confidently wrong" in your pocket. It will pay for itself from fraud alerts (Chapter 6) all the way to chatbots inventing facts (Chapter 9).