Large language model (LLM)

A large language model (LLM) is a deep neural network trained on enormous text corpora to predict the next token, giving it a broad, general-purpose command of language, code, and world knowledge.

Under the hood, an LLM does one deceptively simple thing: given a sequence of tokens (word fragments), it predicts a probability distribution over what token comes next, then samples from it. Stack that prediction step millions of times during training across trillions of tokens, and the model absorbs grammar, facts, reasoning shortcuts, and the texture of human writing. Most modern LLMs use the transformer architecture, whose attention mechanism lets every token weigh every other token in the prompt.

Why it matters: the LLM is the reasoning core of nearly every AI product you touch. Each time the model produces an answer it runs inference — a forward pass that turns your prompt into output. The amount of text it can attend to at once is its context window, and the size of that window shapes how much instruction, history, and reference material you can give it.

A concrete example: ask an LLM "summarize this contract in three bullets," and it doesn't look the answer up in a database. It predicts, token by token, the most plausible continuation of your prompt — which, because it was trained well, happens to be a competent summary. Wrap that same model in a loop with tools and memory and you get an AI agent that can take actions rather than only produce text.