What a Model Is
You have heard the word model thrown around in news articles about AI — a new model from a big lab, a model that writes essays, a model that diagnoses disease. But what exactly is a model? The word sounds abstract, yet the idea is concrete enough to fit on a single page.
A Model Is the Learned Thing
In machine learning, a model is the artifact that the training process produces. Think of training as the work, and the model as the result of that work. Before training, the system is blank — it knows nothing. During training it studies thousands or millions of examples. After training it has crystallized what it learned into a structure it can use to make predictions. That structure is the model. A model takes an input and returns an output. Give an image classifier the photo of a cat — it outputs cat. Give a spam classifier an email — it outputs spam or not spam. Give a language model a sentence — it outputs a continuation. The model does not re-examine the training data every time it answers. The learning has already been baked into its internal parameters.
A model is the result of training a machine learning algorithm on data. It stores what was learned as a set of internal parameters, and uses those parameters to convert new inputs into predictions or outputs.
A helpful analogy: imagine baking a cake. The recipe is the algorithm, the ingredients are the training data, and the cake itself is the model. Once baked, the cake does not need the recipe or the raw ingredients anymore — it is a finished thing you can eat (use). If you want a different cake (a model trained on different data), you bake again. The original cake does not change.
Parameters: Where the Learning Lives
The internal structure of a model consists of numbers called parameters. During training, those numbers are adjusted little by little until the model makes good predictions. After training stops, the parameters are frozen. A small model might have thousands of parameters. A large language model like the ones behind modern AI chat tools has hundreds of billions. Each parameter is just a number, but together they encode the patterns the system found in the training data. When you hear phrases like a model with 70 billion parameters, that is describing how many adjustable numbers it contains — roughly, how much it can remember and express.
Training data is what the model learned from. Parameters are what the model learned. The data can be discarded after training; the parameters remain.
Complete the sentence about models.
Using a Model vs. Training a Model
It is important to separate two distinct phases: training and inference. Training is the expensive, data-hungry process of building a model. It requires large datasets, significant computing power, and can take days or weeks for large systems. Inference is using the already-trained model to make a prediction on a new input. Inference is fast and lightweight — the model just applies its frozen parameters to the new input. When you ask an AI assistant a question, you are doing inference against a model that was trained long before you typed anything.
Training builds the model (slow, data-heavy, done once). Inference uses the model to make predictions on new inputs (fast, done many times after training).
After a model is trained, what encodes the knowledge it learned?
What is the term for using a trained model to make a prediction on a new input?
Model or Not?
- Step 1: For each item below, decide whether it is a model (the learned artifact) or something else in the ML pipeline.
- A) A spreadsheet of one million labeled customer reviews.
- B) A file containing 175 billion numbers that power a text-generation AI.
- C) The Python script that ran for three weeks and adjusted numbers based on data.
- D) The output a user gets when they type a prompt.
- Step 2: For items that are NOT the model, name what they are (training data, algorithm, inference output, etc.).
- Step 3: Sketch a simple diagram — four boxes labeled Data, Algorithm, Model, Prediction — with arrows showing how they connect.