Lesson 4 of 6
It's not a database
10 min read
It feels like the AI is looking things up in some giant table. It isn't. There is no table. So what is actually happening when it answers?
Filing cabinet, or improviser?
A database is a filing cabinet: every record sits in a drawer, and a query pulls out the exact one — or finds nothing. A trained model works nothing like that. It kept no records. Instead it holds patterns, and when you ask something it blends the nearest patterns into a brand-new answer, made on the spot. Ask the same thing again and it improvises again; ask something no one ever filed, and it still improvises something.
A database retrieves a stored record. A model generates a fresh answer by blending patterns — nothing is pulled from a drawer.
That's why it handles the unseen
This is the model's superpower. Because it blends rather than looks up, it can answer questions that were never in its reading at all — a mash-up of two topics, a poem about your cat, a plan for a trip nobody's taken. A filing cabinet has nothing for a query it doesn't contain. The model always has something, because it's building the answer, not fetching it. It's also why you can ask the exact same thing twice and get two slightly different replies: each answer is freshly assembled, not read off a fixed record.
Because it generates instead of retrieving, a model can respond to things it never saw — it composes, it doesn't copy.
The same power is the same danger
But notice the flip side. If the model always builds an answer, it will build one even when it shouldn't — even when it has nothing solid to go on. A filing cabinet honestly says "not found." The model rarely does; it just reaches for the nearest patterns and hands you a fluent answer, whether or not those patterns actually fit. The blend is smooth either way, so from the outside a well-supported answer and a wild guess can look exactly the same.
Keep this in your pocket for the next lesson: because it generates rather than looks up, an AI will happily produce a confident answer for a question it has no real basis for.
The blend that lets a model handle new questions is exactly what lets it answer when it shouldn't.
What to carry forward
- —A model stores no records — it blends patterns into a new answer each time.
- —That's why it can handle questions that were never in its training.
- —It's also why it rarely says "not found" — it always builds something.
How is a trained model different from a search over a database?
Continue in the app
Take the whole How AI Learns course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.