What the Model Does Not Know

A client asks about a boutique hotel that opened by the lake last month, and Tessa — testing her new tool — asks the chat box about it. The answer is fluent, warm, and useless: the model has clearly never heard of the place, and yet nothing in its tone says so. This is the right moment to draw the model's three honest limits, because all three hide behind the same confident voice, and knowing them is what stops you from asking the box for things it cannot possibly have.

The limits are these: the model's knowledge stops at a date, it keeps nothing between conversations, and by default it cannot look anything up. None of the three is a flaw to be fixed by clever asking. They follow directly from what a model is — patterns learned from a fixed pile of text — and they mark the exact places where its confident answers deserve the least trust.

The Knowledge Cutoff

The model learned its patterns from text collected up to a certain date — the knowledge cutoff. Everything in its "knowledge" comes from before that line; everything after is simply absent. Picture a brilliant colleague just back from a year on a remote island: superb on everything up to departure day, smoothly out of date on everything since, and — this is the dangerous part — not always aware of the difference. Ask about last month's hotel and the model does what it always does: produces text that fits. Text that fits, about a hotel it never saw, is a guess wearing the same suit as knowledge.

The cutoff is why a model can be excellent on history and hopeless on this season's prices, why it may not know its own newest features, and why "when did your knowledge end?" is a fair and useful question to ask any model you work with — most will tell you.

No Memory Between Chats

The previous pages established that every conversation starts with an empty window, and it is worth one more beat here because of how it feels from outside. Yesterday Tessa told the chat box all about Waymark's tours; today it knows nothing of the sort. The model learned nothing from her, and it learns nothing from anyone's conversations as they happen — its patterns were fixed at training time, and talking to it does not add to them.

And yet some products genuinely do greet you with context — they know your name, your past topics, your preferences. That is not the model remembering. That is the product keeping notes and quietly re-supplying them, a mechanism Chapter 7 takes apart bolt by bolt. The distinction sounds pedantic and is anything but: it determines where your information actually lives, who stores it, and what you can ask to have deleted — questions Chapter 6 turns serious.

No Internet by Default

Left to itself, the model cannot check a price, open a link, or read today's news. There is no quiet searching behind the curtain — just patterns from before the cutoff. When an answer sounds current, it is either genuinely durable knowledge, or a fluent reconstruction of how such an answer usually sounds. The second kind is the dangerous one.

Many products do bolt search on: the product runs a web search first and hands the results to the model to read (you can often spot cited links when this happened). That is a real and useful feature — but it belongs to the product, not the model, and it is either on or it is not. The practical habit: know which mode you are in before trusting anything time-sensitive. If no search happened, "current" answers are reconstructions.

One last honest note for this chapter: models increasingly accept more than words — images, sound. The mechanics you are learning apply there unchanged, and this book stays with text, where all the machinery is visible. And for the question this chapter has been carefully not answering — how a model is actually made, the training, the years of engineering — the book to read is Machine Learning from Zero, Chapters 8 and 9, written for exactly the reader you are. This book assumes only what this chapter just taught.