Questions about how it works

Does it remember me? The model itself doesn’t. Between conversations, the harness may keep notes, called memory, and add them to the prompt the next time you talk. Claude, for example, has memory that you can view, edit and delete.¹ So the honest answer is “the app might,” and the app’s settings will tell you. Chapter 7 shows how the harness does it. Chapter 10 follows your words after they leave your screen.

Why does it make things up? Because it writes rather than looks up. Chapter 2 showed a model predicting one piece of text after another, and a made-up detail is as easy to predict as a true one. Researchers also argue that the way models are trained and tested rewards a confident guess over an honest “I don’t know.”² Giving it sources helps. In a 2020 study, a model that first retrieved documents gave more factual answers.³ That was the lawyer’s lesson in Chapter 1, and Chapter 12 turns it into a habit.

Is it searching the web right now? Only if the harness offers a search tool and the model asks for it.⁴⁵ Otherwise it answers from what it learned in training, which stops at a cutoff date. Google’s card for its small Gemma 4 models, for example, gives January 2025.⁶ The model never runs a search itself. It writes a request; the harness runs the tool and hands back the results. Chapter 8 follows that loop step by step.

Can it be tricked? Yes. A model reads everything in its prompt as text, and it can’t always tell your instructions from words that arrived inside a web page or a document. Text planted there can try to give it orders, an attack researchers call prompt injection.⁷ The risk grows with an agent, a setup in which the model directs its own steps and tools until a task is done.⁸ That is one reason harnesses run work in isolated environments and ask before some actions.⁹ Chapter 8 explains why tools raise the stakes.

Can I run a free model on my own computer? Yes, for some jobs. Small models that you download and run yourself suit narrow, private tasks. They know less than the largest models; as one technical report put it, a small model “simply does not have the capacity to store too much ‘factual knowledge.’”¹⁰ One research group argues that small models are enough for many repeated steps inside larger systems.¹¹ Chapter 5 explains the trade between a model’s size and what it can do.

Who decides how it behaves? Layers of people. The company that makes the model shapes it in training and publishes the principles it aims for; Anthropic and OpenAI each publish one.¹²¹³ The business that builds the app adds its own instructions to the prompt, usually out of sight. You add the last layer with your message. OpenAI’s rules rank these sources in order of authority and give text that arrives from tools no authority by default.¹³ Chapter 6 shows the hidden instructions at work.

Does it understand? People use the word in different ways, so it helps to ask what can be measured. A model can work through many multi-step problems, such as math word problems, especially when it writes out its steps.¹⁴ It can also be confidently wrong.² Claude Shannon, whose work begins Chapter 3, gave an interviewer a two-sided answer in 1989. Asked whether machines could think, he said, “You bet. I’m a machine and you’re a machine, and we both think, don’t we?” Yet he also said “most of our talk between humans is not so measurable.”¹⁵ Asked whether machines would take over, he said, “I think so long as we build them, they won’t take over.”¹⁵ Those were one pioneer’s views then, not verdicts on today’s tools. The practical answer is the same either way: judge the work by checking it.