What is an LLM? A machine that guesses the next word
In AI, an LLM (large language model) is a program that has read an enormous amount of text and writes by predicting, over and over, which word comes next. It’s what sits behind ChatGPT, Claude and Gemini. (The law degree spelled LL.M. is something else.) Knowing what an LLM is explains two things at once: why it writes so well, and why it sometimes passes off a fact nobody checked.
What is a large language model?
It’s the engine, not the car. The assistant you talk to is a product built around a language model: it adds the screen, the memory of the conversation and, sometimes, your documents. The model does one thing, which is continue a text. It belongs to generative AI, the kind that creates new content.
How does an LLM work? It guesses the next word
It takes in a text and calculates how likely each possible next piece of a word is. It picks one, usually among the most likely, adds it and calculates again. Piece by piece (they’re called tokens) a whole answer comes out. What you type, the prompt, is the start of the text it continues. That’s a simplification, but it’s the central idea.
At Elder & Sage, an herb and tea shop in Burlington, Vermont, someone asks a chat what time the shop opens on Saturdays. It answers with very reasonable hours. It didn’t look them up anywhere: they’re the hours that sound like an herb shop.
Where it learned from
From huge amounts of text: books, websites, articles, code. During training it adjusts billions of internal numbers until it predicts well. It doesn’t store the texts like a library; it keeps patterns of how things are written and what tends to go with what.
What it knows and what it can’t know
It knows what was in those texts, up to the date its training ended. It doesn’t know what happened afterward or what was never public: your hours, your prices, your rules. And since it always produces a likely continuation, it doesn’t go quiet when a fact is missing. It fills the gap, and that’s a hallucination.
There are two fixes. Put the fact in front of it with each question, which is the usual one and is called RAG. Or retrain it on your own examples, called fine-tuning, which is rarely worth it.
LLM meaning in practice: the models out there
They come in families: GPT, from OpenAI; Claude, from Anthropic; Gemini, from Google; Llama, from Meta; and Mistral’s models. Each has large versions, more capable and more expensive, and small ones, which are faster. They change every few months, so what you give the model to read matters more than its name.
In Flipa
The model inside Flipa is from Anthropic, from the Claude family. What makes it useful to a company is the business’s own documents, which it checks before answering. It isn’t infallible, so what it writes gets reviewed.
How does it write? The next word
A half-finished sentence about the herb shop. Pick the next word yourself, or let the most likely one win. The percentages are invented for the example; the mechanism isn’t.
From memoryExample probabilities
The shop is open Saturdays from ten to five in the afternoon.
Word 1 of 4. Which comes next?
The rest, up to a hundred, is spread across thousands of less likely words.
Sounds right. Is it true? The model doesn’t know: it never looked at the hours. It wrote the ones that sound like a shop.
With the hours in front of it
DocumentElder & Sage hours. Saturdays: 10:30 a.m. to 4:00 p.m.
- ten thirty96%
- ten3%
- eleven1%
The shop is open Saturdays from ten thirty to four.
Same model, same way of writing. With the paper in front of it, the most likely word becomes the one written there.
Three things you notice, and where they come from
| What you notice | Where it comes from |
|---|---|
| It writes fluently | It has seen so much text that the likely word is usually also the well-written one. |
| It sometimes makes things up | If the fact is missing, the likely word still comes out. It doesn’t ship with an “I don’t know.” |
| It needs your documents | Your business wasn’t in what it read. With the paper in front of it, the likely word becomes what the paper says. |
What people ask about LLMs
What is an LLM in AI?
A large language model: a program trained on enormous amounts of text to predict how a sentence continues. It’s the part that does the writing inside AI assistants. “Large” refers to the volume of text it learned from and to its internal size.
How does a language model work?
It calculates which fragments are likely to follow the text in front of it, picks one, adds it and repeats. It doesn’t query a database or check facts, even though the result often looks as if it had.
Is ChatGPT an LLM?
Not exactly. ChatGPT is the product, and inside it runs on models from the GPT family, which are LLMs. The same goes for Claude and Gemini, except that product and model share a name.
Does an LLM understand what it writes?
That’s an open debate, and it depends on what you call understanding. What can be checked is what it does: it produces likely text from learned patterns. To work with one, a single consequence is enough: on its own, it doesn’t separate what’s true from what’s plausible.
What if it checked your paperwork before writing?
A machine that predicts words gets more right with the fact in front of it.
Start free