Welcome to AI in Plain English — a ten-part series for anyone who has nodded along to words like "LLM", "agent", or "RAG" without quite knowing what they mean. Each article does one idea, in plain language, with no jargon assumed. This is article one. Start here.
Picture a very strange library
Imagine a library unlike any you have visited. Behind one of the walls, there is a person — call them the Librarian. This Librarian has spent their entire life reading. Every book. Every newspaper. Every website. Every forum post and instruction manual and novel ever written, in dozens of languages. They know an extraordinary amount.
But here is the twist: you cannot see the Librarian, and they cannot see you. There is only a small mail slot in the wall.
If you want to communicate, you write a note and slip it through the slot. A few seconds later, a note slides back out. That is the whole transaction.
This is not a parable. This is, quite literally, how an AI language model — an LLM (Large Language Model) — works.
The one thing the Librarian can do
Now here is the most important sentence in this entire series, so read it slowly:
The Librarian can only do one thing: finish your note.
That is it. You write "The capital of France is", and they write back "Paris." You write "Once upon a time there was a young knight who", and they write back a story. You write a half-finished recipe, and they complete it. You write a question, and — because most questions in their reading were followed by answers — they write what looks like an answer.
The Librarian is not looking anything up. They are not connecting to the internet. They are not thinking in the way you or I think. They are doing something that sounds almost too simple: they are guessing the most likely next word, then the next, then the next, thousands of times in a row, drawing on every pattern they absorbed during all that reading.
This process is called next-token prediction — "token" is a technical word for roughly half a word, but "word-guessing" captures it well enough.
Why it feels so clever
So if it is just guessing, why does it feel like talking to someone who genuinely understands?
Because the patterns in language are not trivial. To guess the right next word in "The speed of light is approximately ___", you have to have somehow absorbed physics. To complete "She felt a complicated mixture of relief and ___" convincingly, you need something like emotional understanding — or at least a very good simulation of it built from millions of sentences written by people who do feel things.
The Librarian is not conscious. They are not experiencing the conversation. But the sheer volume of human thought baked into language means that a sufficiently large Librarian, finishing your sentences, produces something that genuinely helps.
Which brings us to the "Large" part.
Why it had to be Large
Early language models were small — trained on modest amounts of text, with modest computing power. They could finish a sentence, but the results were flat, wooden, often nonsensical. Interesting to researchers, but not useful to anyone else.
Then something unexpected happened. As researchers made the models bigger — much bigger, trained on vastly more text with vastly more computing power — the results did not just improve gradually. At a certain scale, something almost like intelligence appeared from nowhere. The model could reason through a simple logic puzzle. It could translate between languages it had never been explicitly taught to translate. It could write a passable poem or explain a scientific concept to a child.
This surprise is called emergence — abilities that nobody specifically programmed, appearing simply because the model became large enough. It is one of the genuinely strange and debated things in this field.
That is why these systems are called Large Language Models. The "Large" is load-bearing. Smaller versions of the same idea were not nearly as capable. Scale changed what was possible.
The confidence problem
Here is something every parent and child should know before they use any AI tool.
The Librarian finishes your note based on patterns. They do not know whether what they are writing is true. They know that certain words tend to follow other words in certain contexts. So they write fluently, in full sentences, with a tone of calm authority — even when they are wrong.
If you ask for a specific statistic, the Librarian may produce one that sounds exactly right but is completely made up, because the pattern "when asked for a statistic, produce a number" is so deeply baked in. If you ask for a book recommendation, they may confidently name a book that does not exist, because the pattern fits.
Experts call this hallucination — the model generating plausible-sounding nonsense. It is not lying in any intentional sense. The Librarian has no intentions. It is pattern-completion gone wrong.
This is not a reason to avoid AI tools. It is a reason to use them the way you would use a very well-read, very helpful friend who sometimes misremembers things: useful, worth listening to, but worth checking when it matters.
What the Librarian cannot do
The mail-slot design has hard limits, and understanding them saves a lot of frustration.
The Librarian cannot follow up. One note in, one note out. They do not remember your previous note. Each note is a fresh start — the Librarian has no memory of anything you have slipped through before.
The Librarian cannot go outside. They cannot check today's news, look up a real-time price, or verify a fact that appeared after their reading stopped. Their knowledge has a cutoff date — everything they know comes from text they absorbed before a certain point in time.
The Librarian cannot take action. They cannot send an email, book a restaurant, or write a file to your computer. They can only write words back to you.
These limits are not bugs. They are the nature of what a language model actually is. And — this is the interesting part — almost all the exciting AI products you hear about are simply clever ways of working around these limits. Which is exactly what the rest of this series is about.
Everything else is scaffolding
So: the Librarian is remarkable. They have absorbed more text than any human could read in a thousand lifetimes, and they can continue any sentence in a way that feels genuinely intelligent. But they are also very constrained. One slot. One note at a time. No memory. No action. Sometimes confidently wrong.
How do you build a useful product on top of that?
You add scaffolding. You find ways to write better, richer notes. You hire someone to stand at the slot on your behalf and do all the boring, mechanical parts that do not actually need a genius. You give the Librarian better information to work with, and you pipe their answers into useful places.
Every AI term you will ever hear — prompt, agent, memory, context, tool, workflow — is some piece of that scaffolding. Once you understand the Librarian, each piece clicks into place naturally.
That is the promise of this series. Ten articles, one idea at a time, all connected back to this one picture of a brilliant but limited Librarian behind a mail slot.
Next up: Passing Notes Through the Slot — What Is a Prompt? You already know a note goes in. The next article is about what makes a note good.
Part 1 of 10 in AI in Plain English — a free series from AI Tutors.