What is an AI model, really?
Full revision
A model is a machine that guesses.
That is not a simplification for a general audience. It is the mechanism. You give it the start of something and it produces what most likely comes next, based on the enormous pile of text or images it was shown while it was being built.
Why this explains almost everything odd about them
A machine that guesses is confident by design. It has no separate step where it checks whether the guess was true, because there is nothing in it that holds a notion of true. When it invents a court case or a citation, it has not malfunctioned. It has done exactly the thing it does, on a question where guessing was not good enough.
This is also why the same question can get two different answers. There is a deliberate amount of randomness in how the next word gets picked.
What that means for you on a Tuesday
Someone in your accounts team asks a model to pull the payment terms out of forty supplier contracts. It will do that in a minute, and it will be right most of the time. The question that decides whether this is useful or dangerous is not how good the model is. It is whether anyone can tell which of the forty it got wrong.
That question has an answer, and building the thing that answers it is most of what we actually do.
Words used on this page: model, training, hallucination