What are Parameters?
Parameters are the internal numerical values that an AI model learns during its training phase. They are the mathematical variables that define how the model processes an input to produce an output. In neural networks, parameters are typically referred to as "weights" (which determine the strength of a connection between artificial neurons) and "biases" (which shift the output).
How do they work?
When a model is first created, its parameters are set to random numbers, meaning it is completely useless. As the model ingests training data, it uses an algorithm to slightly adjust these numbers. Over millions of training cycles, these parameters are gradually tuned into a highly specific configuration that correctly maps inputs to desired outputs.
What is a simple example?
Imagine a sound mixing board with billions of tiny volume dials. When an input (a prompt) comes in, the signal flows through the board. The exact position of every single dial determines what the final output sounds like. The model's training process is the act of discovering the perfect position for every dial. The parameters are the dials themselves.
What are they commonly confused with?
Model parameters are often confused with "hyperparameters" or generation settings like temperature. Parameters are learned automatically by the system during training and are permanently baked into the final model. Hyperparameters are manually set by the developer before training begins. Generation settings are adjusted by the user at inference time.
Why do they matter?
The number of parameters in a model is a rough indicator of its capacity to learn complex information. A small model might have 7 billion parameters, while a frontier Large Language Model might have over a trillion. However, simply having more parameters does not guarantee a better model if the training data is poor.