In plain words
An LLM is the brain inside chatbots like ChatGPT, Claude and Gemini. It was trained on a vast amount of text: books, websites, code, articles. From all that reading, it learned how words and ideas usually fit together.
When you ask it something, it doesn't look up an answer in a database. It writes the reply one small piece at a time, each time picking a likely next word based on everything written so far. Do that well enough, billions of times over, and you get answers, essays, code and jokes.
"Large" refers to two things: the amount of text it learned from (trillions of words) and the size of the model itself (billions of parameters, the numbers it adjusts while learning).
Everyday example
Finishing your mother's sentences
Your mother starts: "Beta, don't forget to take your..." and you already know the next word is "umbrella", because you've heard her say it a thousand times in the monsoon. An LLM has "heard" a huge share of everything ever written, so it can finish almost any sentence in a way that sounds right.
Where the example breaks: You know why she says it: you remember the rain. An LLM has no memories or experiences. It only learned which words tend to follow which, which is why it can sound sure and still be wrong. That is called a hallucination.
How it actually works
An LLM is a neural network, almost always a transformer, trained in stages.
Text becomes tokens
Text is split into tokens, which are words or pieces of words. A sentence of 10 words is roughly 13 tokens in English.
Pre-training on a huge library
During pre-training, the model reads trillions of tokens and plays one game: guess the next token. Each wrong guess nudges its parameters slightly. After enough rounds, grammar, facts, styles and some reasoning patterns end up stored in those numbers.
Teaching it to be an assistant
A raw pre-trained model just continues text. To make it answer questions and follow instructions, it is fine-tuned on example conversations and then refined with human or AI feedback, for example RLHF.
Writing a reply
When you send a prompt, the model scores every possible next token, picks one (with a little randomness set by temperature), adds it, and repeats until the reply is done. This is next-token prediction.
Limits of the window
It can only consider a fixed amount of text at once, called its context window. Anything outside it, it cannot see.
- Built on
- The transformer design, published by Google researchers in 2017
- A landmark size
- GPT-3 (2020) had 175 billion parameters
- Training data scale
- Meta trained Llama 3 (2024) on over 15 trillion tokens
Many LLMs today also take images, and some take audio, which makes them multimodal. Many add a thinking step before answering; those are called reasoning models.
Does an LLM actually understand anything?
One view
It's a very good autocomplete
Some researchers argue LLMs only model which words follow which. They point to confident mistakes and simple puzzles that trip them up as signs that nothing like understanding is happening.
The other view
It builds real internal models
Others point to research showing LLMs form internal representations of things like positions on a board game or places on a map, and solve problems they were never shown. To them, predicting text well forces some genuine understanding.
Most likely: Both are partly right. LLMs learn useful internal structure, not just surface statistics, but it is patchy and unlike human understanding. Treat them as capable and fallible: check anything that matters.
Questions people ask
Is ChatGPT an LLM?
What is the difference between AI and an LLM?
AI is the whole field of making machines do things that seem intelligent. An LLM is one kind of AI model, focused on language. The camera on your phone that recognises faces is AI, but not an LLM.
Where does an LLM get its information from?
From the text it was trained on, up to a cut-off date. It doesn't know about later events unless it is connected to search or documents, a setup called retrieval-augmented generation.
Why do LLMs make things up?
Because they generate the most plausible next words, not verified facts. When the true answer wasn't in their training or is rare, a plausible-sounding but wrong answer can come out. This is called a hallucination.
Can an LLM run on my own computer?
Smaller ones can. Open-weight models such as Llama or Mistral come in sizes that run on a good laptop or phone, especially after quantization. The largest models need data centre hardware.