Parameters: The Adjustable Numbers

A "parameter" (also called a "weight") is just a number stored inside the model that controls how strongly one piece of information influences another. A model doesn't have beliefs or rules written out in sentences — it has a huge number of these adjustable numbers, and its entire behavior comes from their specific values.

Before training starts, these numbers are typically set to something close to random — the model genuinely knows nothing yet. Training is the process of gradually adjusting every one of those numbers, a tiny bit at a time, so the model's outputs get closer to what the training data suggests they should be.

"Millions or billions of parameters" isn't a marketing exaggeration — it's a literal count of these adjustable numbers. More parameters generally means the model has more capacity to represent complicated patterns, though (as this level covers later) more isn't automatically better in every way.

What is a "parameter" (or "weight") inside an AI model?

The process of gradually adjusting a model's parameters based on training data is called ____.

A model's parameters are usually close to random before training begins.