What an LLM Actually Is

You already use AI — you just have not named it

If you unlocked your phone by looking at it, saw your maps app warn you about a traffic jam before you reached it, or watched your inbox quietly bin a scam email this morning, then you have already used artificial intelligence today — probably several times — without giving it a second thought. The technology behind ChatGPT is not some separate, alien invention. It sits on the very same family tree as all of those everyday things, and this series is going to walk you up that tree from the roots.

Here is the promise, and we mean it literally: we assume you know nothing about this subject, and we will explain every single term the moment it appears. No prior maths. No code. No jargon left undefined. If a word ever feels like it came out of nowhere, that is a mistake in our writing, not a gap in your knowledge. By the end of this one article you will genuinely understand what an “AI” is and roughly where a language model fits — and every part after this one is built gently on top of the part before it. Let us start at the bottom.

What does “artificial intelligence” even mean?

“Artificial intelligence” is one of the most over-used phrases of our time, so let us pin it down plainly. AI is an umbrella word for getting a computer to do something that normally seems to need human intelligence — recognising a friend’s face in a photo, understanding a spoken sentence, choosing a chess move, spotting a fraudulent payment. Notice that this is a goal, not a single gadget. Lots of very different techniques sit under the umbrella, and they have almost nothing in common except the goal.

The idea is older than you might think. The term was coined back in 1956, at a summer workshop at Dartmouth College, by a small group of researchers who believed that thinking could, in principle, be done by a machine. For its first few decades, though, “AI” mostly meant one specific approach: a human carefully writing down rules for the computer to follow. Think of an extremely detailed instruction manual, or one of those phone menus — “if the caller presses 1, do this; if 2, do that.” The computer was not clever; it just followed the human’s rules very fast.

That approach works for tidy problems, but it hits a wall the moment the real world gets messy. Try to write the rules for “what a cat looks like” that cover every breed, every angle, every lighting condition, a cat half-hidden behind a chair… you would be writing rules forever and still miss cases. For decades, anything involving real images, real language, or real human messiness stayed stubbornly out of reach. Something had to give.

The big idea: learn from examples, not rules

Here is the shift that changed everything. Instead of a human writing the rules, what if we just showed the computer thousands of examples and let it work out the rules by itself?

You already know this works, because it is exactly how a small child learns. You do not teach a toddler what a dog is by reciting a definition (“four legs, fur, tends to bark”). You point at dogs — big ones, small ones, fluffy ones — and say “dog,” over and over, until one day the child just gets it and can point at a breed they have never seen before and say “dog!” Nobody wrote the rule. The child absorbed the pattern from examples. This approach — learning patterns from examples instead of being given rules — is called machine learning, and it is the engine under almost everything modern, including the model behind ChatGPT.

Two plain words you will see constantly, so let us define them now. The examples you show the computer are called the data (the photos, the emails, the text). And the thing the computer ends up with after it has learned — the trained system that can now make the call on something brand new — is called a model. So “training a model” just means “letting the computer learn the pattern from the data.”

Take spam email. The old way: a human writes a rule — if the subject says “FREE MONEY,” bin it. The spammers simply change the wording, so the human writes another rule, and another, forever one step behind. The machine-learning way: show the computer a huge pile of emails already marked “spam” or “not spam,” and let it learn the patterns of spam on its own — patterns far too subtle and numerous for any human to write down by hand.

Three ways a computer learns from data

“Let it learn from examples” actually comes in three flavours. You do not need to memorise them, but seeing the difference makes the rest of the series click — especially why language models were even possible to build.

Supervised learning is learning with an answer key. Every example comes with the correct answer attached — every photo already labelled “cat” or “dog,” every email already marked spam or not. It is like studying with flashcards: question on the front, answer on the back. It works beautifully — but a human has to label every example first, and labelling millions of things by hand is slow and expensive.

Unsupervised learning has no answer key at all. You just hand the computer a big pile and say “find the groupings.” Imagine tipping out a drawer of mixed photos and sorting them into piles that look similar, without anyone telling you the categories in advance. Useful for discovering structure — but the computer is never told what “correct” means.

The third flavour is the clever one, and it is the secret behind language models. Self-supervised learning makes the data create its own answer key. The trick: hide part of the data and ask the computer to guess the hidden part. Cover the last word of a sentence — “the cat sat on the ___” — and ask it to fill in the blank. The original sentence already contains the answer (“mat”), so you never needed a human to label anything. And because every sentence ever written can be turned into these fill-in-the-blank puzzles, you suddenly have an almost unlimited, free supply of practice questions. That is how you can train a model on a truly staggering amount of text. Keep this in your back pocket — in a moment it explains exactly how a language model is made.

The family tree: AI → machine learning → deep learning

Now we can place the buzzwords properly, because you understand the idea underneath them. They are not rivals — they are circles nested one inside the other.

Artificial intelligence is the whole outer circle: the broad goal of machines doing “smart” things, including the old rule-writing approach. Machine learning is the circle inside it: the part where the machine learns from data instead of being handed rules. And deep learning is a circle inside that — machine learning done with a particular tool called a neural network.

You will hear “neural network” constantly, so here is a one-sentence, no-maths version: a neural network is a mathematical system, loosely inspired by how brain cells connect, that learns a pattern by tuning a huge number of tiny numerical “dials” until its guesses become good. That is genuinely all you need for now — a big bank of little adjustable dials, nudged into the right positions by examples. We build one from absolute scratch, dial by dial, in Part 4; until then, just keep the picture in mind. “Deep” learning simply means the network has many layers of these dials stacked up, which lets it learn richer patterns. In 2012 a deep network nicknamed AlexNet suddenly got far better at recognising images than anything before it, and the modern era of AI began in earnest.

So — what is a large language model?

Finally we can answer the title question — and now it will actually make sense. A large language model (LLM) is a very large deep-learning model that learned language by reading a staggering amount of text — books, articles, websites — using exactly that self-supervised trick from earlier: over and over, cover the next word and guess it. “Large” refers to both the mountain of text it read and the enormous number of those little dials inside it.

And what does it actually do, once trained? In one sentence: you give it some text, and it guesses what word comes next — then it adds that word, looks at the slightly longer text, and guesses the next word again, producing writing one piece at a time. The little animation below shows the intuition; do not worry about the exact percentages yet.

That really is the core of it: a language model is, at heart, an extraordinarily well-read autocomplete. The obvious objection — “guessing the next word is not the same as understanding” — is a fair one, and we will come back to it. But here is the surprise that makes the whole field work: to guess the next word well across the entire internet, a model is quietly forced to soak up grammar, facts, tone, and a startling amount of reasoning, simply because all of those help it guess better. The cleverness is a side-effect of getting extremely good at one humble game.

How does it turn your words into numbers it can do maths on? That is Part 2. How does the network actually make the guess? That is the heart of Parts 4 through 8. For right now, the intuition — well-read autocomplete — is exactly enough.

What these tools are genuinely good — and bad — at

Because an LLM is, underneath, a next-word guesser, its strengths and weaknesses are not random — they follow directly from that. Knowing them is the difference between using one well and getting burned by it.

It is genuinely excellent at anything that is language: drafting an email, summarising a long document, rewriting something in a friendlier tone, translating, explaining a tricky idea, even writing code (which is just very patterned text). These all play straight to what next-word guessing is good at.

The weak spots come from the very same place. It guesses plausible words rather than calculating, so exact arithmetic and careful counting trip it up — unless you hand it a calculator, which is what “tools” in Part 18 are about. It only knows what it read during training, frozen at some cut-off date, so ask about anything newer and it will guess — fluently and wrongly (Part 17, on giving it fresh information, fixes this). And crucially, it has no built-in sense of when it is wrong: it states a mistaken answer in exactly the same confident voice as a correct one.

How we got here — and where this series goes next

It is worth seeing the road that led here, now that the pieces mean something to you. Each milestone fixed a real limit of the one before it.

Early AI tried to hand-write intelligence as rules, and choked on the real world’s messiness. Machine learning shifted to learning from data instead. Deep learning, powered by faster hardware and oceans of data, learned its own patterns. Then in 2017 a team at Google published a paper with the cheeky title “Attention Is All You Need”, introducing an architecture called the Transformer — the design every modern LLM is built on. After that, the story is mostly about scale and polish: make the models bigger, train them on more text, then teach them to behave like helpful assistants — which is the moment, around 2022, the whole world suddenly noticed.

You now have the map. You know what AI is, what it means for a machine to learn, the three flavours of learning, where language models sit in the family, and the one simple thing they really do — and we got here without assuming you knew a thing. That is the foundation everything else stands on.

Next, in Part 2, we answer the question we kept gently dodging: before a model can do anything with language, it has to turn your words into numbers. How it chops text into pieces called tokens is up first — and that one unglamorous detail explains some of an LLM’s strangest habits, like why it sometimes miscounts the letters in a word.