What is a large language model?

Your phone keyboard guesses your next word. Type “I am on my” and it offers “way”. It is right often enough to save you time. However, it has no idea where you are actually going.

A large language model is that guess, made enormous. It reads what you wrote and picks the next small piece of text. Then it does that again, and again, until the reply is done.

That sounds too simple to explain a machine that drafts contracts and writes code. It is the truth anyway, because the power comes from scale, and from the sheer amount of writing the model learned on.

What the three words mean

Take the name apart.

Large points at the number of internal settings, which are called parameters. A modern model holds billions of them, and each one is a number that training set.

Language points at what the model works on. Text. Not the world, not truth, not pictures of things. Text.

Model means a squeezed down picture of something. This one is a squeezed down picture of how people write. It carries the shapes of language, and along the way it picks up a great deal of what those sentences happened to say.

How it works, step by step

1. Your text is cut into tokens. A token is a word or a piece of a word.
2. Each token is turned into a list of numbers.
3. Those numbers run through a stack of layers called a transformer. Each layer lets every token look at the others and take in their meaning.
4. The last layer gives a score to every possible next token.
5. One token is chosen and added to the text.
6. The whole run starts again with the new token included.

So the model writes the way you eat rice, one spoonful at a time, with the whole plate in front of it. That plate is the context window, which is how much text the model can hold at once. Therefore, anything outside the window is simply gone.

How it is trained

Training happens in two big steps.

The first is pretraining, in which the model reads a huge pile of text and plays one game over and over. Cover the next token. Guess it. Check. Adjust. That runs for weeks across many machines, and nobody has to label anything, because the text labels itself. The next word is always the answer.

The second step teaches manners. People write good answers, and rank the answers a model gives from better to worse. The model is then tuned toward what people preferred. This is why a raw model rambles and a finished product answers your question.

Underneath all of this sits a neural network, and the loop is ordinary machine learning run at a very large scale.

What it is good at

The model is strong wherever the job is shaping text.

• Turning something long into something short.
• Changing tone, from stiff to plain or from plain to formal.
• Translating between languages that are well covered online.
• Drafting a first version of anything, which you then fix.
• Writing code and explaining code somebody else wrote.
• Pulling a clean structure out of messy notes.

Notice the pattern in that list. In every case you already hold the material, or you can check the result in seconds. That is where the tool earns its keep.

What it is bad at

There is no fact checker inside the model. It produces text that fits, and fitting is a different thing from being true. When it states something false in a confident voice, people call that a hallucination.

It is also weak at counting and at exact arithmetic, because it is guessing text rather than working through sums. Many products now hand the sum to a calculator behind the scenes, which fixes most of it.

It does not know what happened after its training data was collected. Ask about last week and it will either say so or make something up. And it cannot tell you why it wrote what it wrote. This is because any reason it offers afterwards is just more predicted text.

How it differs from a search engine

A search engine finds documents. It shows you where the words came from, and you judge the source yourself. It knows nothing that it cannot find.

A language model writes fresh text, and it cannot show you a source unless a tool went and fetched one. Its strength is shaping an answer, and its weakness is standing behind that answer.

Therefore, use search when you need a source, and the model when you need a draft. Many products now do both, and the model searches first and writes second. That pattern is far safer, so prefer it for anything factual.

Where you already meet it

Language models sit behind chat assistants and the reply suggestions in your mail. They run support windows on shopping sites, phone dictation, code editors and the summaries at the top of search results. However, most carry no badge, so if a product writes fluent sentences back at you, a model is doing it.

Set against older artificial intelligence, the change is that one system now covers many text jobs at once. Older systems needed a separate build for each one.

How to get better answers

Three habits do most of the work. Give the model the material instead of trusting its memory. Say who the reader is and how long the answer should be. Ask for a draft and plan to edit it.

Then check anything a stranger would want a source for. Names, dates, numbers, quotes and laws. IF IT SOUNDS SURE, CHECK IT ANYWAY. Do that ten times and you will know exactly which jobs this tool can be trusted with.

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.