How a sentence becomes an answer

A sentence falls apart

First the sentence is split. Not into words and not into letters, but into pieces in between — by frequency. What appears often stays whole; what is rare breaks up. That is why "invoice" is one piece and "reimbursement" is several.

Pieces become numbers

A model computes, it does not read. Every piece becomes a list of numbers — four here, several thousand in real models. These numbers are not a code and not a dictionary index: they are learned, and everything the model knows about the piece sits in them.

The four numbers per word are set by hand so you can see what they do. A model learns them from billions of sentences.

Closeness is meaning

Because they are numbers, there are distances. And because there are distances, related things sit together: "invoice" next to "receipt", far from "server". A search then no longer asks whether a word appears in the text, but what lies near it.

Every word looks at every other

Now comes the actual invention. Every piece assigns weights to all the others — "it" looks at "invoice" and finds more there than at "yesterday". The lines show it: the thicker, the more attention. That is the whole trick behind everything built since 2017.

And then it guesses

At the end there is no answer, just a list: a probability for every possible next piece. The model picks one, appends it and starts over. Word by word. It never knows how the sentence it is writing will end.

The probabilities are set — real ones would need a running model. The orders of magnitude match what you actually see at this point.

And where does all this happen?

That is exactly the question we care about. The same path runs twice: once on a machine in your building, once in a data centre that is not yours. The computation is identical. What differs is who gets to watch.