Weights are the the billions of numerical parameters that define how the model thinks and responds: the internal values a model develops during training that determine how it responds to any given input. They are the accumulated result of the model being exposed to enormous amounts of data and adjusting, over millions of iterations, to get better at its task. By the time training is complete, the weights encode everything the model has learned about language, patterns, relationships between concepts, and how to produce useful outputs.
A useful way to think about it: imagine a new employee learning a job. Early on, they are uncertain about everything, such as how to prioritize tasks, what matters to different stakeholders, how to handle edge cases. Over time, through experience and feedback, they develop a set of instincts. They don’t consciously recall every lesson because the judgment is just there. Weights are the model’s equivalent of those instincts. They are not a list of facts the model looks up. They are the internalized patterns that shape every response it produces.
When a model is said to have billions of parameters, those parameters are its weights — the sheer number of individual values that collectively define how the model thinks and responds.
Why This Matters for Technical Writers
Weights help explain why AI models behave the way they do and why they can’t simply be corrected by updating a value in a database. If a model consistently produces a certain kind of output, that behavior is distributed across billions of weights, not stored in one place. This is why updating model behavior requires retraining or fine-tuning rather than a simple edit. It also helps explain knowledge cutoffs because the weights reflect what the model learned during training, and nothing that happened after training is in them.
Common Confusion
Weights are sometimes confused with the model’s knowledge or memory, but the relationship is indirect. A model doesn’t store facts in its weights the way a database stores records. It stores patterns, which are tendencies toward certain kinds of outputs given certain kinds of inputs. This is why a model can produce a confident, fluent response about something it has effectively “mislearned,” and why retrieval-augmented approaches exist to ground outputs in verified external content rather than relying on weights alone.
Related Terms
Large language model (LLM), inference, training, fine-tuning, hallucination, foundation model, parameters