Tuned at the factory

Inside every model are billions of numbers. Hello, AI calls them knobs. Engineers call them parameters, or weights.

Training is the long process of nudging each knob, over and over, until the model gets good at predicting the next piece of text. It happens once, at the factory, before you ever open the app.

When training ends, the knobs are fixed. The text the model learned from is set aside. The model keeps no copy of it to look things up in, though it can repeat passages it saw many times, like famous quotes.

During your chat, every knob stays exactly where training left it. The model reads your words and predicts a reply. Nothing about the model itself changes.

A printed book

A diary grows a little with every entry. A model works more like a printed book. You can read it all day, and every page stays as it was.

So when a model seems to know you, Hello, AI offers two plain explanations. Your earlier messages are still sitting in the window. Or the model is a very good guesser, the way a sharp waiter sizes up a table.

This also explains the knowledge cutoff. The model knows only what was in its training text when that text was frozen. Something that happened after that date is not in the knobs, and asking harder will not put it there.

One honest flag from the book: your chats may still be used to train future models, depending on your settings. That happens later, in a new training run. The model you are talking to today does not update itself as you type.

How big a set of weights is

A model’s size is set by how many knobs it has. Here is the scale, from our Model size fun facts:

• The first GPT, in 2018, had 117 million parameters. Stored at full precision, that is about 470 MB, smaller than many big phone games.
• A model with 70 billion parameters, stored the usual way at 2 bytes each, is about 140 GB.
• Llama 3.1 405B is about 810 GB. Copied onto old floppy disks, it would need 562,500 of them, a stack about 1.9 km tall.

Every one of those numbers was fixed when training ended. A chat with you adds nothing to the pile.

In one sentence: training sets a model’s weights once, and your chat reads from them without writing anything back.