Large language model (LLM)

A large language model is a neural network, almost always built on the transformer architecture, trained on enormous volumes of text to predict the next token in a sequence. That simple objective, repeated across billions of parameters and trillions of words, produces a model that can write, summarize, translate, answer questions, and follow instructions across many tasks without being built for any one of them.

Training happens in stages: a broad pretraining phase learns general language patterns, and later stages, fine-tuning or instruction tuning, adjust the model toward specific behaviors or domains. What a model can consider at once is bounded by its context window, and what it "knows" is fixed at training time unless supplemented with outside data, for example through retrieval-augmented generation.

LLMs matter to analytics because they turn unstructured text, tickets, reviews, transcripts, into a queryable and generative interface, and increasingly sit inside copilots and agents that read or write data on a user's behalf. The main pitfall is treating fluent output as reliable: models can produce confident, wrong answers (hallucination), so production use typically pairs them with LLM evaluation and guardrails.