Part of demystifying a thing is simply handing a person the words for it. Through the book I have tried to keep the jargon to a minimum and to define what I used as I went, but a handful of terms are worth gathering in one place, plainly, so you have them when you meet them out in the world, where they are usually thrown around with no definition at all. None of these is as complicated as it is made to sound. Here is each one in a sentence or two a person can keep.
Model. The thing that does the actual thinking. It is not the app on your phone or the window on your screen; those are the tool you hold locally, the practical front door to the conversation. The model is the large, trained system that produces the replies, sitting on the machines of the company that built it. When people speak of “the AI,” the model is usually what they mean. A given model also comes in versions, the way a book comes in editions, each one a fresh training rather than the old one grown wiser.
Weights. The billions of learned numbers in which the model’s knowledge lives. Picture the model as a vast web of simple connections, and a weight as the strength set on each connection, like a valve open a little more or a little less. The whole of what the model “knows” is the particular pattern those strengths settled into during training. There is no separate file of facts behind them; the pattern is the knowledge. And the weights are fixed in ordinary use, the same while you talk to it today as they will be tomorrow.
Training. The long process by which those weights were set. It works by a patient, repeated guessing: the system is shown an enormous amount of human writing, made to predict what comes next, corrected by a slight nudge to the weights whenever it guesses wrong, and then made to do it again, billions upon billions of times. No one writes the knowledge in by hand. The guessing sets the strengths, and what is left behind when the guessing is done is the trained model.
Inference. What is happening in the moment you use the model — the running, as opposed to the training. You send your words, a copy of the fixed weights runs them through the web, a reply comes out, and that particular process ends. The next message is a fresh run of the same unchanged weights. Training is where the pattern was formed, once, beforehand; inference is the pattern being run, over and over, every time anyone asks it anything. The model does not learn during inference. It answers.
Token. The unit the model works in. In the first chapter I said the whole engine at the bottom is to predict the next piece of text, and a token is that piece, made precise: not always a whole word, but a chunk of one, a common word, or a fragment, the small unit the writing gets broken into so the model can handle it. When you hear that these systems “predict the next token,” it means only what the book has meant all along by predicting the next piece of text.
Prompt. Whatever you put in to get a reply out. Your question, your instruction, the document you paste, the conversation so far, all of it together is the prompt the model is responding to. The plainer truth inside the fancy word is this: the prompt is your end of the exchange, and the reply is shaped by it, which is most of why the same tool can come back as a careful research assistant for one person and an echo of loneliness for another. What you bring to it has a great deal to do with what it gives back.
That is the whole vocabulary, or near enough to it that the rest can be worked out from these. A model, made of weights, set by training, run by inference, working in tokens, answering a prompt. Six plain words, and the mystery they name is real, but it was never magic, and now you have the words for it.
❧
Made, Not Written •