What is a large language model (LLM)?
Short answer
A large language model is an AI system trained on very large amounts of text to predict the next piece of text. This lets it write, summarise, translate and answer questions, although it can also state things that are wrong.
How it works in outline
Text is split into tokens, which are pieces of words. The model, a neural network that is usually built on the transformer architecture introduced in a 2017 research paper, learns to predict the next token from the ones before it. It does this over billions or trillions of tokens, adjusting its parameters, the numerical weights inside the network, until its predictions improve.
After this pre-training, developers typically fine-tune the model on curated examples and with human feedback so that it follows instructions and avoids harmful output.
What it can and cannot do
LLMs can draft and summarise text, translate, write code and answer questions. They have no built-in guarantee of accuracy: a model can produce fluent statements that are false, which is called hallucination. Each model also has a context window, the maximum amount of text it can consider at once, and a training cut-off beyond which it knows nothing unless given tools or documents.
Who builds them
Examples of model families from major developers include OpenAI’s GPT models, Google’s Gemini, Anthropic’s Claude and Meta’s Llama. Some developers publish model weights openly; others offer the model only through a service. Training the largest models uses thousands of GPUs, which links LLMs to chip makers and data centre operators.
Agyata’s editorial policy describes how AI-assisted summaries are produced, checked and labelled.