What is a large language model?
In one line
Ages 8 to 10: A large language model is a computer system that writes by guessing the next word again and again, after practicing on a huge amount of text.
Ages 11 to 13: A large language model splits text into tokens, gives each possible next token a probability, picks one and repeats.
Ages 14 to 18: A large language model is a neural network, usually a transformer, with a very large number of parameters, pre-trained to predict the next token and then fine-tuned for tasks.
A large language model (LLM) is an AI model trained on a huge amount of text to predict the next piece of text, called a token. By predicting one token after another, it can write sentences, answer questions and summarize. It is called large because it has a very high number of adjustable numbers.
Last reviewed: 20 October 2026 7 min read
The idea in one picture
Think of the word suggestions on a phone keyboard, then imagine them trained on a vast library of text and able to keep going for pages. That is close to the core of a large language model: it is a very powerful next-word predictor.
Worked example. Give a model the words "The sky is". It gives every possible next token a chance. These numbers are made up for illustration:
| Next token | Chance |
|---|---|
| blue | 55% |
| clear | 18% |
| dark | 12% |
| green | 3% |
In a real model, the chances for every token in its vocabulary add up to 100%; here we show the top four. The model picks one, often the likeliest but sometimes another, adds it to the text and predicts again: "The sky is blue" → "The sky is blue today" → and so on, one token at a time, until it stops.
Tokens are not always whole words. A token can be a word, part of a word or a single character. For example, "dogs" might be split into "dog" and "s".
How it works, step by step
- Split the text into tokens. The prompt is broken into tokens, and each token is turned into numbers.
- Look at the whole context. A neural network called a transformer works out how much each token should affect the meaning of every other token. This is called self-attention.
- Score every possible next token. The model gives each token in its vocabulary a probability.
- Pick one and add it. One token is chosen and added to the end of the text.
- Repeat. The longer text goes back in, and the model predicts the next token again until it reaches a stopping point.
A little deeper (ages 11 to 13)
How it learned. During training, the model reads huge amounts of text. Again and again it predicts the next word, checks against the word that actually came next, and adjusts its internal numbers to do better next time.
Picking is set by temperature. At a low temperature, the model almost always takes the likeliest token, so output is predictable. At a higher temperature, less likely tokens get picked more often, so output varies more. That is why the same question can get different answers.
It has a limited view. The context window is the number of tokens a model can process at once. Text beyond that limit is not part of what it uses for its next prediction.
Going further (ages 14 to 18)
"Large" refers to the number of parameters: the weights and biases learned in training. Large language models have a very high number of them, and research shows transformers with more parameters tend to outperform smaller ones, at a higher cost to run.
Building one usually happens in stages. Pre-training on a vast text collection teaches the model to predict the next token. A model that has only been pre-trained does not reliably follow instructions, so it is then fine-tuned on examples of the task, such as following instructions, and often improved further using ratings from human reviewers.
Because it generates likely text, an LLM can produce fluent statements that are false, often called hallucinations. A US standards body describes this as a natural result of models approximating the statistical patterns of their training data. One common fix is grounding: retrieving passages from trusted sources and basing the answer on them. Even then, a grounded answer is only as reliable as its sources.
Whether next-token prediction amounts to real reasoning is debated among researchers. What is not debated: outputs need checking.
Common mix-ups
- Mix-up: "It looks answers up in a database."Fact: It generates text from patterns in its parameters. It only uses outside sources if a system is built to retrieve them.
- Mix-up: "Bigger means always right."Fact: Larger models tend to perform better on average, but any LLM can state false things confidently.
- Mix-up: "It reads words like we do."Fact: It works with tokens turned into numbers, and those tokens are often pieces of words.
- Mix-up: "It remembers everything you've ever told it."Fact: Within one conversation it uses what is in its context window; anything beyond that limit is not part of what it uses. Whether a product keeps past conversations is a separate question about that product, so check before sharing personal details.
Try it (no device, 5 minutes)
Be a language model. Someone writes the start of a sentence, such as "After school I like to". Everyone secretly writes one next word. Tally the answers: that is your probability table. Pick the most common word, add it, and repeat for three more words. Then do it again, but this time pick a less common word. Which sentence sounds more ordinary? Which is more surprising? You just changed the temperature by hand.
Quick check
In the "The sky is" table, which next token is the model most likely to pick?
Answer: "blue", with a 55% chance. It is the likeliest, but the model can still pick another one sometimes.
"blue", with a 55% chance. It is the likeliest, but the model can still pick another one sometimes.Is a token always a whole word?
Answer: No. A token can be a whole word, part of a word or a single character, such as "dog" plus "s" for "dogs".
No. A token can be a whole word, part of a word or a single character, such as "dog" plus "s" for "dogs".Why can an LLM state something false with confidence?
Answer: It generates text that is likely given the patterns in its training data, and likely is not the same as true. Grounding the answer in retrieved, trusted sources helps, but the answer still needs checking.
It generates text that is likely given the patterns in its training data, and likely is not the same as true. Grounding the answer in retrieved, trusted sources helps, but the answer still needs checking.Questions
LLM stands for large language model. It is an AI model trained on a very large amount of text to predict the next piece of text, which lets it write and answer questions.
It splits text into tokens, works out how the tokens relate to each other, and gives every possible next token a probability. It picks one, adds it to the text and repeats, building an answer one token at a time.
A large language model is one kind of generative AI, the kind that generates text. Generative AI also includes models that make images, audio and video.
Because it has a very high number of parameters, the adjustable numbers it learns during training, and it is trained on a very large amount of text. Larger models tend to perform better but cost more to run.
Yes. It produces text that is likely given its training, not text that has been checked, so it can state false things confidently. Important answers need checking against a reliable source.
Mechna Academy is a city of machines for curious children aged 8 to 18, opening first to ages 8 to 10. In each mission, a child guesses how a machine decides, then opens its Peek Hatch to check, changes one thing and tests it on a case it has never seen. Children learn how AI works this way, and they never chat with an AI.
For grown-ups: join the waitlistPlaces are limited. The waitlist is for parents and guardians 18 and older.