In AI, a model is the thing that has learned. It takes an input and produces an output, based on patterns it picked up from examples.
input → [ model ] → output
photo → [ model ] → "cat"
email → [ model ] → "spam"
question → [ model ] → an answer
A simple way to picture it
Imagine teaching a child to recognise dogs. You don't give them a rulebook. You point at many dogs and say "dog", and at cats and say "not a dog". After enough examples, they can recognise a dog they've never seen.
A model learns the same way — except with millions of examples and a lot of maths. The learning process is called training. The result — the trained model — is what you use afterwards.
What a model actually is
Under the hood, a model is a very large set of numbers (called parameters or weights) plus the instructions for how to combine them. During training, those numbers are adjusted again and again until the model's outputs match the examples well.
That's all. There is no little database of facts inside, and no understanding in the human sense — just numbers that have been tuned so that the right outputs come out for typical inputs.
A model is usually stored as a file — sometimes small enough to run on a phone, sometimes so large it needs a data centre of specialised chips.
A tiny model you can read
Here is about the smallest "model" possible: one number, learned from examples of hours studied and test scores.
# academy-check
examples = [(1, 20), (2, 40), (3, 60), (4, 80)] # (hours studied, score)
weight = 0.0 # the model's single parameter
for _ in range(200): # training: nudge the number to fit
for hours, score in examples:
prediction = weight * hours
weight += 0.01 * (score - prediction) * hours
print(round(weight, 1)) # the learned parameter: 20.0
print(round(weight * 5)) # using the model on new input (5 hours): 100
assert round(weight) == 20
The model learned "each hour is worth about 20 points" without anyone telling it. Real models work on the same idea, with billions of parameters instead of one.
Two phases: training and using
| Training | Using (inference) | |
|---|---|---|
| What happens | The model learns from examples; its numbers change | The finished model answers new inputs; its numbers stay fixed |
| Who does it | Usually big AI labs, for large models | You — every time your app sends a request |
| Cost | Very expensive, done rarely | Cheap per request, done constantly |
Almost everything you'll do as an AI engineer is using existing models, not training new ones.
Quick check: when you chat with an AI assistant, is the model learning from your messages in that moment? (No — it is using what it already learned. Its parameters don't change during your chat.)