LLMs, neural networks, and parameters: what "7B" actually means

You'll hear models described as "7B" or "70B" and pitched on their parameter count as if it were horsepower. To make sense of that, it helps to peek under the branding: a large language model is a neural network, and its parameters are simply the numbers it learned during training.

Neural networks in one picture

A neural network is layers of simple units — neurons — connected together. Each connection has a weight that scales the signal passing through it, and each neuron adds a bias that shifts its output. Data flows in one side, gets transformed layer by layer, and predictions come out the other.

Parameters = weights + biases

Those weights and biases are the parameters. Training is the process of nudging them — over and over, across enormous amounts of data — so the network's predictions get closer to correct. When a model is "7 billion parameters," that's how many of these learned numbers it holds.

• Weights scale the connections; biases shift each neuron's output.
• Training adjusts every parameter to reduce error (loss), via gradient descent.
• More parameters means more capacity to represent patterns — up to a point, and at a real cost in compute and memory.

There's no ghost in the machine — just billions of weights and biases, tuned until plausible text falls out. That it works this well is the miracle.