What is a large language model?

A large language model, or LLM, is built by showing a program enormous amounts of text, books, articles, code, web pages, and having it repeatedly guess the next word or word-fragment in a sentence until its guesses get good. That is the entire underlying task: predicting likely continuations of text. Everything the model can do, holding a conversation, summarizing an article, answering a question, is that same prediction skill applied to whatever text comes before it.

This is worth being precise about because it explains where an LLM's limits come from. The model in a product like ChatGPT or Claude is not, by itself, a search engine or a live database: it can only draw on patterns from the text it was shown before its training finished, so it can state outdated facts confidently or invent a plausible-sounding one that never happened (see hallucination). When an AI product answers with genuinely current information or names a real source, that usually means a separate system fetched something live and handed it to the model for that one answer (see retrieval-augmented-generation), not that the model looked it up itself.