What is a model?

Two product managers on the same team are preparing a launch. Both use a chat assistant. Both ask it the same question, in the same words:

Summarise the top three risks in this launch plan.

They paste the same plan. They get two different answers.

The lists overlap, but not completely. One names a supplier risk that the other misses. One puts the compliance question first, the other puts it third.

In the meeting, someone asks the obvious question: which one is right?

You cannot answer it yet, because the information is missing. Answering it needs something none of them has: a working picture of what the thing on the other side of the chat box is doing.

That picture is the whole of this session. You will not write any code, and you will not need any mathematics. You do need the picture, because every other week of this course stands on it.

Here is the sentence to hold on to. Everything else in this session explains it.

A language model reads the text it has been given and predicts what text comes next. That is the only thing it does.

Let us take that apart slowly.

"A model" is a large file of numbers. Those numbers were produced by reading an enormous amount of text: books, websites, code, documentation. The numbers were adjusted, over and over, until the file became good at one narrow task. That task is to take some text and guess what follows.

That adjusting process is called training, and the text it read is the training text. Both words appear constantly, and neither means anything more than that.

The numbers are usually called weights, and the file is usually called a model. When people say LLM, they mean large language model, and they mean this same file. The word "large" refers to how many numbers are in it. You will meet all three words in vendor documents, and they point at the same thing.

"Reads the text it has been given." Everything the model uses to answer you arrives as text, in one delivery, at the moment you press send. The text you send is called a prompt. Nothing else reaches the model. Not your screen, not your files, not last week's conversation, unless something puts that text into the prompt first.

"Predicts what text comes next." The model does not look up an answer. It produces the answer a few characters at a time, each one chosen because it is a likely continuation of everything before it. Then it reads what it has written so far, and continues. One run of this process — you send text, it produces text — is called inference.

An analogy, and where it breaks.

Think of a very well-read colleague who has been given the first half of a sentence and asked to finish it. They have read so much that their guess is usually excellent. They are not remembering a specific page. They are producing what sounds like the natural continuation, based on everything they have ever read.

Now here is where the analogy breaks, and the break is the important part. Your colleague knows when they are guessing. They can say "I am not sure, let me check." The model has no separate sense of certainty about the world. A confident-sounding continuation and a true one are produced by exactly the same process. We come back to this at the end of the session, because it is the single most expensive thing to misunderstand about this technology.

The obvious objection: "that cannot be all it does."

It is a fair objection. Predicting the next piece of text sounds far too small to explain something that can draft a contract or debug a spreadsheet.

The answer is that "what comes next" is a much harder question than it sounds. To finish the sentence "The capital of Australia is —" well, you must have absorbed a fact. To finish "The bug was caused by —" well, you must have absorbed how programs fail. Getting extremely good at continuation forces the file to encode a great deal about how the world is described.

But it encodes how the world is described, not how the world is. That difference matters for every decision you will make this year. The two things overlap most of the time. Every failure in this course lives in the gap between them.

What this already changes about your job.

Go back to the two product managers. Neither answer was "right" in the sense they were reaching for. The model produced a likely continuation twice, and likely continuations vary. At each point several pieces are likely, and the app around the model picks one of them, varying that pick on purpose. That is a setting, not a fault.

The question "which one is right?" has no answer. Three other questions do. How often is this answer wrong? What does a wrong one cost us? Can someone check it before it reaches a customer? You will spend the rest of this course learning to ask them.