A language model is not a library. It is not a search engine with a friendly voice attached. Underneath the chat window is something much stranger and much simpler: an enormous collection of numbers. Those numbers are called parameters, and the largest model running on the rig in this course holds roughly twenty six billion of them. Not all twenty six billion run on every word, which is a detail worth returning to, but all of them sit in the file. Every answer that model has ever produced came out of those numbers and nothing else.

A parameter is just a number. It is not a sentence, not a fact, not a small file with information in it. If you could reach into the model and pull out a single parameter, you would find something like 0.0412, and it would tell you nothing at all. No parameter holds the idea that Baton Rouge is the capital of Louisiana. That idea, if the model has it, is spread thinly across billions of numbers at once, in a pattern no person could read directly. This is why nobody can open a model and look up what it knows.

So where did those numbers come from? They came from a process called pretraining. During pretraining, the model is shown staggering amounts of text and given one repetitive job: predict what comes next. It guesses, it is wrong, and every time it is wrong the numbers get nudged. Nudge, guess, nudge, guess, billions of times over weeks or months on machines that cost more than most buildings. Nobody sits down and writes facts into the model. The facts arrive as a side effect of getting better at predicting text.

After pretraining comes a second and much shorter stage called post-training. Pretraining decides roughly what the model knows. Post-training decides how it behaves: whether it answers in full sentences, whether it follows an instruction or wanders off, whether it refuses certain requests. This stage is smaller and more targeted, but it is the reason a raw pretrained model and a finished chat model feel like completely different things.

And then, at some point, the adjusting stops.

This is the moment that matters most today. The numbers are written into a file and locked. This is called freezing, and from that moment forward the model is fixed. It does not get smarter. It does not get worse. It does not update itself overnight. Whatever it learned during training is what it has, permanently, until a company decides to run the whole expensive process again and release a different model.

Freezing produces a side effect you have probably run into without having a name for it. The text the model read had to stop somewhere, on some particular date. That date is called the knowledge cutoff. Ask the model about something that happened after its cutoff and it will not know, and worse, it may not know that it does not know. It will produce a confident, well-formed, completely invented answer, because producing well-formed text is exactly the job it was trained for.

Everything that happens after the freeze has its own name: inference. Inference is what occurs when you type a prompt and press enter. Your text is fed through the frozen numbers, and an answer comes out the other side. That is the entire operation. Nothing inside the model changes. The numbers you used are bit for bit identical to the numbers the person before you used and the numbers the person after you will use.

Which brings us back to the question you answered at the start of class. When you send a prompt to a model, is the model learning from you?

No. It cannot. The learning stage ended before the file was ever copied to the machine you are typing on.

The obvious objection is that it certainly feels like it learns. You tell a chatbot your name and three messages later it uses your name. That is real, and it has a boring explanation. The conversation is not stored inside the model. It is stored outside, in ordinary text, and the entire conversation up to that point is quietly resent with every new message. The model reads the whole transcript fresh each time and responds. Close the window and the transcript is gone, and the model is exactly what it was before you arrived.

Hold onto that, because next week you will be using a tool built to make it obvious. It sends your prompt once, to three models at once, with no memory between prompts. No conversation. No follow-ups. One pass through frozen numbers, three times over, so you can see the machinery instead of the illusion.