Training vs using a model

Two very different phases

An AI model has two distinct chapters in its life, and they are easy to mix up. First it is built, by learning from data. Then it is used, to answer questions or make predictions. The building phase is called training; the using phase is called inference. They differ enormously in cost, in speed, and in who does them, and telling them apart explains a lot about how these tools behave.

Training: the slow, expensive part

Training is where the model learns. It is shown a huge amount of data and gradually adjusts itself, over and over, until it can capture the patterns in that data. Think of it as a long, intensive period of study, but on a scale no person could manage: enormous datasets, specialised computers, and often weeks of processing that cost a great deal of money and energy.

Training happens rarely and is done by the organisation that builds the model, not by you. Once it is finished, the result is a fixed, trained model: a settled set of patterns, ready to be used. Importantly, the model does not keep learning from you as you use it; its knowledge was set during training.

Inference: the fast, everyday part

Inference is what happens every time you actually use the model. You give it an input, a question, a photo, a sentence to complete, and it applies what it already learned to produce an output. This is quick and comparatively cheap, which is why millions of people can use a model at once, on ordinary phones and laptops.

Every example from earlier lessons, a maps app predicting traffic, a chatbot answering you, a filter spotting spam, is inference: the trained model doing its job. The heavy lifting already happened, once, during training.

Why the training data shapes every answer

Here is the reason this distinction matters in practice. A model can only learn from the data it was trained on, so that data quietly shapes everything it produces at inference time. If the training data was broad, accurate and balanced, the answers tend to be too. If it was narrow, outdated or skewed, the model carries those flaws forward into every response, confidently and invisibly.

This is where bias enters. If the examples a model learned from over represent one group, one viewpoint or one language, its output will lean the same way, not out of malice but because it is faithfully reflecting what it was shown. A model trained mostly on data from one part of the world may handle other places poorly. One trained on text up to a certain date knows nothing after it.

What this means for you

You do not need to train models to use them wisely, but knowing where their knowledge comes from changes how you read their answers. Remember that the output reflects past data, with all its gaps and slants, rather than direct knowledge of the world today. That is another solid reason for the habit running through this module: use the model's answer as a useful input, then apply your own judgement, especially on anything important or anything where a hidden bias could matter.

One question to carry with you: since a model knows only what its training data showed it, whose experiences, languages or time periods might be missing from the answers you rely on, and how would you even notice?