1.1 — What is an LLM?
A plain-English definition of a large language model, where it sits in the AI world, and why next-word prediction is enough.
A large language model (LLM) is a neural network designed to understand, generate, and respond to human-like text. It is a deep neural network trained on massive amounts of text — sometimes a large chunk of the publicly available internet.
Where the "large" comes from
The word large points at two things:
- Model size — tens or even hundreds of billions of parameters. Parameters are the adjustable weights inside the network. During training, they get tweaked so the model gets better and better at predicting the next word in a sequence.
- Training data — gigantic corpora of text that the model learns from.
Why "predict the next word"?
Training a model to predict the next word sounds laughably simple. It works because language is sequential — to guess the next word you have to understand the context, structure, and relationships in everything that came before. That tiny task, repeated billions of times across an ocean of text, produces a startlingly capable model.
Generative AI, in context
Because LLMs can generate new text, they are a form of generative AI (often shortened to GenAI). They live one level up the AI family tree:
| Layer | What it means |
|---|---|
| AI | The broad idea of machines doing tasks that need human-like intelligence. |
| Machine learning | Algorithms that learn from data instead of being hand-coded. |
| Deep learning | Machine learning with multi-layer neural networks. |
| Generative AI / LLMs | Deep-learning models that produce new content — in our case, text. |
A classic, non-LLM example of machine learning is a spam filter. Instead of writing explicit rules ("block emails with the word prize"), you feed the model lots of emails already labelled spam or not spam. The model figures out the patterns on its own.
In old-school (non-deep) machine learning, a human expert would still have to pick the features that matter — capitalisation, exclamation marks, suspicious links, and so on. Deep learning removes that step: the network decides which features matter all by itself. You only ever need the labelled examples.
The architecture inside
LLMs are built on an architecture called the transformer, covered properly in 1.4. The transformer's superpower is that it can pay selective attention to different parts of the input when making a prediction, which is what makes it so good at the nuances of human language.
Heads-up on the word "understand". When we say an LLM "understands" text, we mean it produces output that looks coherent and contextually appropriate — not that it has consciousness or feelings. Keep that distinction in your back pocket.