A large language model is not artificial intelligence

I hate it when people call large language models AI. There's no intelligence in there. It's word generation by a math formula, and that's it.

And here's the strange part. How these things work is known, it's out in the open, you can go read it. People still reach for something supernatural, they hand the thing qualities it doesn't have. You start arguing, you show the math, you explain it in plain terms, and back comes: you're oversimplifying, there's something in there. No. I'm not. Let's go through what's actually in there.

What it does

A large language model (LLM) does one thing: it predicts the next chunk of text. You give it a start, it builds the rest, one piece at a time. The piece is called a token, a word or part of a word. It looks at everything before, and it works out a probability, which token comes next. It takes the best fit, writes it down, and goes again.

That's it. The whole trick. An expensive, accurate autocomplete, like the one on your phone, a million times stronger.

What looks like an answer to your question is, inside, a continuation of text. You wrote a question. The most likely continuation of text that ends in a question is text that looks like an answer. It isn't answering you. It's continuing after you.

Where the probabilities come from

It doesn't invent those probabilities on the fly, they were put in. It was run through a mountain of text, books, sites, code, forums, and made to guess the next token billions of times, with its nose rubbed in every miss. After months of that grind, the shape of language settled inside it: which word follows which, how a sentence holds together, what a normal paragraph about taxes looks like and a normal piece of code. That's not knowledge. That's statistics about what text on the subject looks like.

So there's no database in there it looks things up in. There are weights, billions of numbers with that statistics packed into them. When someone tells you the model "knows" the capital of France, it means one thing: in the text it learned from, "the capital of France is" was almost always followed by Paris. It doesn't remember the fact. It repeats the frequency.

Why it isn't intelligence

When we say intelligence, we're picturing understanding, a goal, a model of the world, a thought, an intention. None of that is built into an LLM. What's built in is a probability distribution over tokens. It doesn't understand what it says, the same way a calculator doesn't understand what a number is. It just does a great impression of understanding, because that's exactly what it was trained for, to put out text you can't tell from text a real person wrote.

And here's the trap. We read fluency as intelligence. If someone talks smoothly, we fill in the thinking behind it. With a person that usually holds, the smoothness comes from understanding. With the model the wire is cut. The smoothness is its own thing, separate from any understanding. You, the reader, are the one bringing the intelligence to the text. It isn't in the text.

What follows from that

The rest you can see the moment you stop waiting for a miracle.

Hallucinations aren't a malfunction. When the model confidently gives you a law that doesn't exist or a link that goes nowhere, it's doing the exact same thing it always does, building plausible text. Plausible and true are the same thing to it, there's no separate "true" switch inside. It was sharpened for "looks like correct text," not for "true." Most of the time looks-like and correct line up, which is why it's useful. But it's the same machine in both cases, so you can't scrub the error down to zero without swapping the machine.

It has no memory between turns. Every request, it meets with a blank slate. Everything that looks like "it remembers our conversation" is just the chat history, fed back into the start of each request. Take the history away and it won't remember you.

And it doesn't think while you're not looking. Between your messages there's no inner life ticking over. There's no process. There's a function, text in, text out.

So where does "AI" come from

It's an old umbrella. The term is almost seventy years old, and everything sits under it, from chess programs to face recognition. It got stuck on language models because that's what sells. "Intelligence" raises money and scares headlines. "An accurate statistical model of text" does neither, even though that's the honest name.

I didn't fall out of love with the tool over this. The engineering behind it is real and hard, and it earns its keep. But I keep in mind what I'm using. The model is a commodity, getting cheaper, the same price for everyone. The value isn't in it, it's in what you build around it for your own job. The person who confuses it with intelligence is the one who files a made-up citation in court because "it's smart," and fears a robot uprising, when the real question is whose work it makes cheap.

It's most useful the second you stop waiting for a mind and start seeing what it is. A machine that predicts text. And it does that better than anyone alive.