Model Weights — What the Model Knows
Last week Spoke & Chain started opening at 8 a.m. instead of 9. The owner told the chat assistant, "Remember, we now open at 8." It agreed. The next morning, in a fresh conversation, it told a customer 9 a.m. again. The assistant had not been stubborn. Telling it something in a chat never changed what it knows.
The rule
Weights hold what the model learned; prompts only steer it.
A model is, underneath, a very large set of numbers called weights or parameters. "A 70-billion-parameter model" means 70 billion such numbers. During training, those numbers are adjusted, over and over, until the model predicts text well. Everything it learned — grammar, facts, styles, reasoning habits — is spread across them.
When you use the model, the weights are frozen. Your prompt flows through them to produce an answer, but nothing you type changes them. That has three consequences:
- Knowledge has a cutoff. The training data ended on some date. Events, prices and hours after that date are not in the weights.
- Private facts are missing. Spoke & Chain's stock list was never in the training data.
- Memory within a chat is just text. The model "remembers" earlier turns only because they are sent again as part of the input. Start a new conversation and they are gone.
More parameters generally means more capacity to learn patterns, but also more cost and slower responses. Size alone does not guarantee a model knows your particular facts.
The shortcut
For any fact the model needs, ask: "Is this in the weights, or in the prompt?" If it is recent, private or specific to you, it is not reliably in the weights, so it must be in the prompt.
Where is the fact?
New hours from last week: after the training cutoff, and private to the shop. Not in the weights.
Some chat products add a "memory" feature, but it works by storing notes and inserting them back into the prompt. The weights still do not change.
Drills
Answer fast — trust the rule.
What are a model's weights?
When you send a prompt, what happens to the weights?
Why might a model not know about a price change from last month?
How does a chat assistant "remember" what you said three messages ago?
Remember
- Weights hold what the model learned; prompts only steer it.
- Shortcut: "is this fact in the weights, or in the prompt?"
- Trap: expecting a chat to teach the model permanently.