Large Language Model (LLM)

A large language model learns by doing one deceptively simple thing over and over: predicting the next word in a piece of text. Trained on a large share of the public internet, books, and code, it builds up a rich statistical sense of how language fits together.

That single skill, done well enough and at a large enough scale, produces something that looks a lot like understanding. Ask it a question and it generates a plausible answer one word at a time. The “large” is not marketing: these models have billions of internal settings, called parameters, and are trained on volumes of text no person could read in a thousand lifetimes.

They are a type of foundation model, and their great strength is flexibility. The same model can summarise a legal document, explain a recipe, or help debug a program, without being purpose-built for any of them.

Their great weakness is that they have no built-in sense of truth. Because they generate text that is statistically likely rather than verified, they can state something false with complete confidence, a failure the field calls hallucination. That is why so much recent work focuses on grounding these models in reliable sources.