Lexicon · I · The model
Weights
Also: parameters
The billions of numerical parameters adjusted during training: the very substance of the model. Whoever holds the weights file holds the model.
The analogy
In a brain, knowledge is not written down anywhere: it is spread across the strength of billions of connections. Weights play that role: each one tunes the influence of one connection in the network. The model's "knowledge" lives in no single weight: it lives in their overall configuration, shaped by training.
A concrete example
When Meta "releases Llama", it puts a weights file online. That is all, and that is everything. Conversely, weight theft is the labs' security nightmare: copying that file means walking away with an asset that cost hundreds of millions to build. Hence weights guarded like industrial secrets by some, and published as a distribution strategy by others.
What it is NOT
It is not a database: you cannot search it for a document, nor delete one specific piece of information from it, which is why the "right to be forgotten" is so hard to apply to models. It is not code you can patch: to durably change a model's behavior, you retrain or fine-tune it, you do not edit a line.
Often confused with
The distinction that changes the conversation.
- Source code
- Software code can be read, audited and fixed line by line. Weights are billions of numbers unreadable to a human: you do not read them, you observe their behavior. That is why a model cannot be debugged the way a program can.