What a language model does

You type a question, and a neat answer comes back. It reads as if there is someone on the other side who understood your question, thought about it for a moment, and then wrote an answer.

That is not what happens.

A prediction machine for language

A language model is a program that does one thing very well: predict which word is likely to follow the words that are already there. That is all. During its training it has seen an enormous amount of text, and from that it has learned what language looks like. Which words belong together, how a business email opens, how a legal paragraph sounds, how an apology is put together.

Do you know that little row on your phone, above the keyboard? Three words your phone suggests as the next word. A language model is that idea, but thousands of times larger and trained on a large part of everything that has ever been written online.

So when you ask "write a polite rejection for this quote", the model does not look up what a good rejection is. It builds up a text word by word that resembles the thousands of polite rejections it has seen. Usually that result is surprisingly good. And that is exactly the point of this course: usually.

Why this works, and when it does not

For tasks that are about language itself, this system is strong. Summarising, rewriting, adding structure, making a first draft, adjusting the tone of an email. That is the territory where predicting language and giving the right answer are almost the same thing.

It becomes different as soon as there is one correct answer. An amount, a date, an article number, a name. The model does not have a list of facts that it briefly consults. It predicts what is likely to be there, and "likely" is something other than "true". In the assignment of this course you are going to experience that yourself, with a case file that has been prepared for it.

Try it yourself

What you take away from this

One sentence, and it will come back often in this course: a language model makes text that is correct in form, not text that is guaranteed to be correct in content. Once you know that, you use the tool where it is strong and check it where it is weak.