A Model Is a Concrete Object

Learned numbers plus a defined computation

A model is not a tiny database of memorized answers. Concretely, it is a mathematical computation with adjustable numeric values called parameters. During training, an algorithm changes those parameters so predictions on example data become less wrong. During inference, the finished parameters are held fixed while the model calculates an output for a new input. The later ai-llm-foundations module applies this same distinction specifically to language generation, context windows, and tokens.

linear model saved on disk
intercept = -1.20
tenure_weight = -0.08
complaint_weight = +0.73
prediction = combine(new row, these learned numbers)

A deployed model artifact usually includes learned parameters plus metadata needed to interpret them: expected feature order, preprocessing rules, algorithm type, and version. Loading the right weights with the wrong feature order can produce valid-looking but meaningless predictions, so the surrounding data contract is part of the operational model even when stored separately.

Analogy: Training adjusts the knobs on a machine while an answer key is available. Inference locks the knobs and feeds the machine a new case. Using the machine does not automatically teach it anything new.

Scenario: An application loads model.pkl but swaps the positions of age and income. The file loads and inference runs; the model simply multiplies the wrong real-world quantities by its weights. Version the model and its input schema together.