Chapter 1What is a language model?
The model is one part of four
The model gets most of the attention, but when you use an AI chat tool, it is only one of the parts at work. This book builds up to the full picture in Chapter 9, one part at a time, so for now it is enough to know their names, shown in Figure 1.2.
The model is the trained program that predicts the reply: the brain. The prompt is everything the model reads for one reply, your message plus instructions you never see: the briefing. The harness is the app or software around the model, which builds the briefing, remembers, runs tools, and asks your permission: the body. Researchers call the harness the software layer between the model and its task.1 The name comes from a horse’s harness, which turns the animal’s strength into useful work. In the same way, the body turns what the brain writes into work that can help you. And the tools are the actions the harness can take when the model asks, such as searching the web or opening a file: the hands.
Try it: watch one request travel around the four parts.
Now go back to the lawyer with this picture in mind. He asked the model behind ChatGPT for cases, and nothing in his setup went out and fetched real ones. Whether an AI tool can look something up depends on the harness and its tools, not on the model alone. We will see in Chapter 8 what changes when a tool can look things up.
Your turn
Try it. Ask any AI chat tool you already use a question you know the answer to well. The question could be about a rule of a sport you play, a fact about your town, or a policy at your work. Then ask it for the source of that answer. If it names a source, look the source up yourself. Finally, notice whether the tool searched for anything or simply wrote.
Think about it. Where in your own work would a confident, wrong answer cost the most? Who would catch it, and how?
What answered the lawyer is also what answers you when you type a question into an AI chat. Both times, it is a model writing the most likely reply, one piece at a time, inside software that may or may not look anything up. That leaves the obvious question of how a program learns to predict text well enough to sound like an expert, and Chapter 2 starts there.
Check in
Try answering in your own words, then open each one to compare.read the answer to compare.
In one sentence, what does a language model do?
It predicts what comes next, having learned from an enormous amount of text, and builds its reply one piece at a time.
Why did ChatGPT tell the lawyer that the cases were real?
Asked whether they were real, it wrote the most plausible reply to that question. It wasn’t checking a database, and nothing in his setup did.
Explain the lawyer’s mistake to a friend in two sentences of your own.
Something like: He treated a tool that writes as if it were a tool that finds, and he didn’t check what it wrote. The court said checking was his job, whatever tool he used.
A coworker says, “I asked the AI whether its answer was right, and it said yes.” What would you suggest?
Check the answer against a source you can trust. Asking the model again only asks it to predict another plausible reply.
Which of these is AI but not machine learning: a spam filter that learns from what you mark as junk, or a chess program that follows rules people wrote?
The chess program. It follows handwritten rules, as Shannon’s 1950 plan did, while the spam filter learns from data.
Notes (1)
- 1Yunbei Zhang, Janet Wang, Yingqiang Ge, Weijie Xu, Jihun Hamm, and Chandan K. Reddy, “Stop Comparing LLM Agents Without Disclosing the Harness” (preprint, arXiv, May 7, 2026), https://doi.org/10.48550/arXiv.2605.23950.
Chapter 1 · page 3 of 3
Turn pages with the ← and → keys.