Chapter 0.1Chapter 0.1. What AI Actually Is

The point

When you ask an AI chat a question, where does its answer come from, and can you trust it?

Ask it a question, and a tidy, sure answer comes back. Open yesterday's chat, and it carries on where you left off. Neither is quite what it seems. The app is what you see and type into. Inside it is a model, the trained engine that writes each reply. It was trained once, before you ever used it, and nothing you type changes it. Everything it says comes from two places. One is what it learned in that training. The other is the text the app puts in front of it for this one reply, called the context window. That is not a window on your screen. It is all the text the model can see for this reply.

After this chapter you can look at any answer and say which of the two it came from. You can explain why a new chat forgets you, and what really remembers you when a chat seems to. You can say why a sure answer can still be a guess, and why the check is yours, not the AI's. And you can see how the same model, given tools, carries out work in steps. That is where Part I begins.

Why it matters

On Monday you ask an AI chat about Brightline Wholesale Supply, the book's example company. Does a $7,800 invoice from Midwest Packaging need the controller's approval? The answer comes back at once, calm and sure. Approval is needed only for invoices over $10,000, it says, so this one can be paid. It is wrong. Brightline's accounts payable policy, version 3, approved on September 1 by Dave, the controller, sets the limit at $5,000. If you had trusted the answer, the invoice would have been paid without Dave's approval. The AI had never seen that policy. It gave the limit that sounds right, and nothing inside it checked.

So you attach the policy and ask again. This time the answer is right, and it quotes the policy's line. On Tuesday you open a new chat and ask once more. The answer is back to $10,000. The AI did not forget the policy overnight, because it never learned it. Monday's chat had the policy in front of it. Tuesday's new chat did not.

The accounts payable manager in the Preface lost money in the same way. She pasted two batches of invoices into two new chats. The same $8,400 invoice was in both batches, and it was paid twice. Nothing was broken. Each chat saw only its own batch, so nothing ever compared the two.

Neither mistake was about how clever the AI was. Both came from not knowing what it was working from. This chapter teaches you to ask that of every answer: did it come from training, or from the context window?