Lessons

The model only knows what it read while it was being taught, plus whatever you paste into the conversation. It does not know your inbox, your company's records, or what a government database says today. Those things were never in the lesson, and they are not sitting in the chat unless someone puts them there.

If the printed book is stale, you can still go get a fresh page and hold it in front of the model. Search a website. Open a PDF. Pull a row from a database and paste the words in. That is retrieval. The model reads what you fetched. Then it talks.

Sometimes reading is not enough. You want a number from a live system, or a list of rooms that are free, or a form filled from a real record. Then the model has to call a function that runs somewhere else, wait for the answer, and use that answer in the next sentence. People call that tool use.

Tool use existed before anyone named a standard. Each company built its own wiring: this model talks to this calendar in this special way. If you had four tools and three models, you were looking at a lot of custom cables.

A model can invent a permitting timeline that sounds finished. It can also produce one where every fact traces to a named source with a time on it, and the sequence logic came from a rules file a person wrote. On the page those two timelines can look identical. They are not the same object.

A model that can retrieve is answering questions. A model that can act is doing things. Doing things has consequences that reading does not. 'It looked up the flight' means a list came back. 'It booked the flight' means money moved, a seat is yours or someone else's, and undoing it is a customer-service problem, not a refresh.

There is no cloud. There is a building. Rows of machines in a warehouse, in a place chosen because the electricity is cheap and there is enough of it, running hot enough that a large part of the building exists only to carry the heat away.