Models Are Probability Machines

A model is writing a one-line tagline for a budgeting app. So far, it has written "Our app helps you," and now it needs the next word. The table shows its top choices, with illustrative numbers. The model draws one word at random, and likely words come up more often. Then it computes a new list for the word after that. So one run may continue with "save," and the next run with "track."

This is how a language model works. It does not look up one stored answer