What is a model?

What a model is

In AI, a model is a piece of software that learned from examples instead of being programmed with rules, then uses what it learned to answer questions, make decisions, or create something new. GPT-5.6 is an AI model. During training it learned patterns in an enormous amount of information; when you ask it a question, it uses those learned patterns to produce an answer. Nobody wrote the answer down in advance, and nobody wrote a rule for producing it.

That is the difference from ordinary software. A normal program follows rules a person typed out: if the total is over fifty dollars, shipping is free. A model follows patterns nobody typed: it has seen enough examples of spam to recognise spam, enough sentences to continue a sentence, enough photographs to draw one. The rules are in there, but they were inferred from data rather than written, and they are stored as numbers rather than code.

So the model is essentially the learned "brain" of an AI system. The chat window, the login, the safety filters, the tools it can call, the documents you upload - all of that is the system. The model is the one part that learned.

This article stays with the general idea, whatever the kind. The text kind - language models, the ones behind chat - is picked up in What is an LLM?; where the word model sits next to AI, machine learning and LLM is section 06.

Data, training, model, input, output

A simple way to think about it is one chain: Data → Training → AI Model → Input → Output. The first two links happen once; the last two happen every time the model is used.

• Data - the examples the model will learn from: text, images, code, recordings. What is in the data decides what the model can ever know. A model trained only on English text cannot translate into Russian, however large it is.
• Training - the model is shown the examples and adjusted, a little at a time, until its predictions match them. For a language model the prediction is the next word; for an image model it is what a slightly less noisy version of a picture looks like. This step runs for weeks or months on thousands of specialised chips, and it is where nearly all the cost of a model is spent.
• AI model - what training leaves behind: a fixed set of learned numbers. From this point on the model does not change. Every user of GPT-5.6 is talking to the same model.
• Input - what you give it: a question, a document, a picture, a half-written function. The input is often called a prompt.
• Output - what it produces from the input using what it learned: a translation, a summary, a diagnosis, a picture, the next line of code.

The chain with the example filled in:

• Training: the model learns from many examples of English text - and Russian, and hundreds of other languages - including countless places where the same thing is said in two of them.
• Input: "Translate Hello into Russian."
• Model: processes the request using what it learned - the pattern that follows a translation request, and the pattern that pairs "Hello" with its Russian counterpart.
• Output: "Привет."

Notice what did not happen. The model did not look "Hello" up in a dictionary; it has no dictionary. It produced the most likely continuation of the input, and because the pairing appears so consistently in the data, the most likely continuation is the correct one. The mechanics of that step - tokens, probabilities, and why it sometimes goes wrong - are in What is an LLM? and Why do AI models hallucinate?

What a model is made of

Physically, a model is a file - often a very large one. Two things are in it:

• The architecture - the shape of the network: how many layers, how they connect, how information flows through. Most current language models share one architecture, the Transformer, in the same way most cars share the four-wheels-and-an-engine design. The architecture is written by people and is usually public.
• The parameters (or weights) - the learned numbers that fill the architecture in, billions of them. This is what training produces and what makes one model different from another. Two models with the same architecture and different training data are different models, the way two people with the same anatomy are different people.

Because a model is a file, it can be copied, downloaded, and run anywhere with enough hardware. Some vendors publish the weights (an open-weights model), so anyone can run it; others keep the file on their own servers and let you send inputs over an API. A version name - GPT-5.6, Claude Sonnet - names one particular file. A new version is a new model, trained again, not the old one updated in place.

What a model is not

Most confusion about AI comes from putting the word "model" on the wrong thing:

• Not another word for LLM. A large language model is one kind of model, the kind built for text. The model that flags a fraudulent card, ranks your feed, or reads an X-ray is a model with no language in it. Every LLM is a model; not every model is an LLM.
• Not the product. ChatGPT is a product; GPT-5.6 is a model inside it. The Claude app is a product; the Claude models are what it calls. A product adds the conversation, memory, tools, and rules; the same model can sit inside many products.
• Not a database. The model does not store its training examples and cannot look one up. It stores patterns. Ask it for a specific fact it saw rarely and it will produce something with the right shape - which is where hallucination comes from, and why products bolt a real lookup on the side with retrieval.
• Not learning while you use it. Training ended before you arrived. When a chat assistant seems to remember earlier turns, the product is re-sending them; the model is unchanged. Your conversations do not update the file.
• Not an agent. An agent is a model wrapped in a loop with tools: the model proposes an action, software carries it out, the result comes back as new input. The model is the reasoning step inside that loop, not the loop.

Kinds of models

The word covers everything that learned from data. The kinds a builder meets most:

• Predictive models output a label or a number: spam or not, this month's demand, the risk of a transaction. Most AI in production is still this kind, and none of it generates content.
• Language models output text, one token at a time; the large ones are LLMs. Chat, code, summaries, translation.
• Image models output pictures from descriptions, or edit existing ones - Nano Banana and GPT Image are the two families covered here.
• Embedding models output numbers rather than words: a list of coordinates that places a piece of text so that similar meanings land close together. They are the engine of meaning-based search and the retrieval step in RAG.
• Speech models turn audio into text or text into audio.

The last four generate something new, which is what makes them generative AI; the first kind predicts about what already exists. Both are models in exactly the sense above: patterns learned from data, applied to new input.

Where the word sits

AI, machine learning, model, neural network, foundation model, LLM - six words that get used as if they were interchangeable. They are not. Read downward, each is a narrower kind of the one above it, so anything on a lower rung is also everything above it:

1. Artificial intelligence - the widest term: software that does something we would call judgement if a person did it. A chess engine searching moves with hand-written rules is AI with no learning in it.
2. Machine learning - AI that learns its behaviour from examples rather than from rules a person typed. The learning happens once and produces a fixed artifact.
3. AI models - that artifact, the subject of this article. A spam filter, a demand forecast and GPT-5.6 are all models in exactly this sense.
4. Neural networks - models built from layers of simple units joined by weighted connections, where the weights are what training adjusts. Many layers is what "deep learning" refers to.
5. Foundation models - very large neural networks pretrained on a broad body of data with no single task in mind, then adapted to many. One artifact, many uses; the term comes from a 2021 Stanford report, listed in the references.
6. Large language models - foundation models whose medium is text. What is an LLM? takes this rung apart.

Model is rung three. It covers everything from there down, and a great deal to either side that has nothing to do with language; LLM is the bottom rung only. Three places where the real map is wider than the ladder:

• Not every model is a neural network. Decision trees and regressions are models with no network in them, and they still run a large share of production AI - fraud scoring, forecasting, ranking.
• Language is one task among many. Section 05 lists the others. A model that reads chest X-rays is every bit as much a model, and it has never seen a sentence.
• Foundation models are not only language models. Many are multimodal - text, images and audio in, sometimes images or audio out - which is why "LLM" has drifted into an umbrella word. Normal usage now, but the L has stopped being exact.

What to keep: the model is the learned brain, the product is everything around it, the data decided what it knows, and LLM names one kind of model rather than the whole category. When something goes wrong, ask which of the four you are looking at.