Billions of numbers

At its core, a large language model is a very large pile of adjustable numbers. GPT-3 had 175 billion of them. Almost all are weights, each setting how much one signal counts. Not one was set by a person. They were settled by the long training loop, nudging each toward better predictions.

Those numbers are the model. Everything it knows is in them, and nowhere else.

Which sounds unremarkable until you ask where any particular thing is.

Take one fact. Light travels about 300,000 kilometres per second. You met the ordinary way to store that at the very start of this course: a numbered slot, holding that value, which you could go to and read.

In the model there is no such slot. The fact is not in weight number four billion, or in weight number four billion and one, or in any hundred weights you could gather up and carry off. It is the combined result of an enormous number of them leaning the same way at once when the text in front of the model calls for it.

A chord is the same kind of thing. Six strings sound together and you hear one chord, but no string is playing it. Damp any single string and the chord is still there, slightly thinner. Look for where the chord is kept and the question turns out to be wrong: it was never kept anywhere. It happens when the strings sound together.

That is what the model's knowledge is like. Pull it apart looking for the fact and you will not find it, because it was never in a part.

This one difference, knowledge as a pattern rather than an entry, is the root of everything else in this chapter.