Parameters
parameter, weights, hyperparameters, sampling parameters, temperatureDefinitions
The learned weights a model is made of — the number quoted as 7B, 70B or 2.1 trillion, being how many of them there are
A rough proxy for capacity and a firm predictor of cost: parameters determine how much memory the model needs to run and therefore what it costs to serve. A rough proxy only, though — training data, architecture and post-training now separate models of the same size more than size separates them.
The knobs set at request time that shape how an answer is sampled — temperature, top-p, maximum tokens, stop sequences
Different things entirely, and confusingly the same word. These change nothing about the model; they change how its next-token distribution is drawn from. Temperature near zero takes the likeliest token each time and reads as focused and repetitive; higher values sample further down the distribution and read as creative, then as unreliable.
In ordinary programming, the named inputs a function declares — as against arguments, the values passed for them
The oldest sense, and the one a route or a component means. [param].vue in this repository is a route parameter; the distinction from an argument is pedantic in conversation and useful in error messages.
Three senses, two of them one letter apart in a configuration file.
Avoid: letting parameter count settle an argument about quality. It describes the size of the thing, not how good it is — and it says nothing at all about the harness around it, which is where most of a session's outcome is decided.