Parameter
One of the internal numbers a neural network adjusts during training — the '70B' or '400B' quoted in a model's name is a count of these.
Think of parameters as millions or billions of dials, each nudged slightly during training until the model gets better at predicting the next piece of text. There's no single dial that means anything on its own; the model's behavior emerges from all of them acting together. Parameter count has long served as a rough, if crude, proxy for how much a model can potentially learn and retain.
But 'more parameters' has stopped being a reliable shorthand for 'better model.' Techniques like mixture-of-experts, quantization, and distillation now let smaller, cheaper models match or beat much larger ones on real tasks — which is exactly the kind of efficiency gain that has fueled arguments on both sides of the debate over whether the industry's capital spending has run ahead of what the underlying technology actually requires.