parameter
/ puh-RAM-uh-tur /
A parameter is one of the internal numbers a model adjusts as it learns — a dial it turns to fit the data. Picture a sound mixing board: each slider controls one part of the sound, and you nudge them up and down until the mix sounds right. A model's parameters are exactly those sliders, except there can be thousands, millions, or in today's large language models, hundreds of billions of them, and the machine sets them automatically during training rather than by ear.
Concretely, parameters are the numbers that fill in a model's chosen form. In the line price = a x area + b, the slope a and the offset b are its two parameters. Training is nothing more than the search for parameter values that make the model's answers match the known answers as closely as possible — turning every slider, a little at a time, to reduce the mistakes. Once training stops, the parameters freeze, and that frozen set of numbers is essentially what the trained model is.
More parameters means a model can capture more intricate patterns — but it is not free. A model with too many parameters for the amount of data can memorize the training examples instead of learning the real pattern (overfitting), and it costs more memory and computation to run. The headline parameter counts you hear about giant models are a rough proxy for capacity, not a guarantee of quality: a smaller, well-trained model often beats a bloated one fed poor data.
A tiny network with 3 inputs and one output might have just 4 parameters: three weights (one per input) and one bias. GPT-style models have hundreds of billions. In both cases, "training" means: search for the parameter values that make the predictions least wrong.
Parameters are the dials training turns — from a handful to hundreds of billions.
Don't confuse parameters with hyperparameters. Parameters are learned from data by the training process; hyperparameters are knobs you set by hand before training begins (like how fast to learn, or how big to make the model).