Your phone's keyboard guesses your next word. An LLM does the same trick, after reading more than any person could in a thousand lifetimes.
Type “Once upon a” on your phone, and it suggests time. Not because it knows any fairy tales, but because it has seen those words follow each other many, many times.
Now imagine an autocomplete that has read a vast library: books, websites, articles, manuals, conversations and code. It isn’t just good at the next word of a fairy tale. It can carry on almost any kind of writing, in almost any style.
That’s an LLM. It writes by predicting the most likely next word, then the next, then the next, until there’s a whole answer.
What the name means
- Large: it learned from an enormous amount of text, and it's a huge piece of software
- Language: words are what it works with, reading them and writing them
- Model: the finished thing that comes out of training (see machine learning)
How a chat actually works
When you ask ChatGPT something, your question becomes the start of the text. The LLM predicts how a good reply would most likely begin, adds that bit, and predicts again. The whole answer is built piece by piece (in small chunks called tokens), which is why you see it appear a little at a time.
It sounds too simple to work. But to predict the next word well across every kind of writing, it has had to pick up grammar, facts, styles of argument, and even how code fits together. Really good prediction ends up looking a lot like understanding.
You’ve probably used one
- ChatGPT, made by OpenAI
- Claude, made by Anthropic
- Gemini, made by Google
- Copilot, made by Microsoft
- the “help me write” button in your email or documents
What it isn’t
It isn’t a search engine. Unless it’s connected to the web, it isn’t looking anything up. It’s writing from what it absorbed while learning, and that reading stopped at a certain date. It also doesn’t know when it’s wrong, so a made-up answer comes out just as smooth and confident as a correct one.
How the words fit together
AI is the umbrella. Generative AI is the kind that makes new things. An LLM is generative AI for words. GPT is the name of OpenAI’s family of LLMs, and ChatGPT is the chat app built on top of them.
An LLM writes what sounds right, not what it has checked. For facts, numbers, names and quotes, ask where it came from, then check that source really exists.
When you'll hear it
“Which model are you using?” Which LLM, like GPT, Claude or Gemini. Different ones are better at different jobs.
“It's in the training data.” The LLM read about this while it was learning, so it may know about it.
“Its knowledge cutoff is last year.” It stopped reading at a certain date, so it won't know about anything newer unless it can search the web.
“Do we need an LLM's way with words here, or an exact answer from our own records?”
