AI Is Easy: AI Models · Chapter 1
What an AI model really is
What is the sort of AI model this book is about?
Almost every helper you can type a question into is built around something called a large
language model. That is a computer program fed an enormous amount of writing. It has become
very good at one particular trick. It works out what small piece of text should come next.
That sounds too simple. Is that really all it does?
It is remarkably close to all it does, and I want to show you why that is enough. Imagine you
find a page from a film script. It is a scene between a person and their helpful assistant.
You can read what the person says. But the assistant's reply has been torn off the bottom of
the page.
And then?
Now imagine you also own a strange machine. You feed it any piece of writing, and it
tells you which word would sensibly come next. One word. That is all it gives you.
How would that machine finish the script?
You would feed it the torn page. It would offer you one word. You would add that word to the
page and feed the whole thing back in. It would offer you the next word.
That sounds slow.
You would do that again, and again, and again, and slowly a full reply would appear on the
page. Not copied from anywhere. Built up, one piece at a time, each piece chosen because it
fitted what came before.
So it is not looking up the answer somewhere?
Not by itself, and this is the part almost nobody tells you. It is not searching a library. It
is not checking a list of facts. It is continuing a piece of writing in the way writing usually
continues. Many of these helpers can now go and search the internet as well, and when they do
they will normally tell you. But left on its own, it is producing the shape of an answer rather
than fetching one.