LLM (Large Language Model)

LLMs are built on the transformer architecture, trained on enormous datasets of text to learn statistical relationships between words and concepts, ultimately learning to predict the most plausible next piece of text given everything before it. That core mechanic — sophisticated next-token prediction, refined through further training stages to follow instructions and align with human preferences — is what lets a model carry on a conversation, summarize a document, write code, or answer a question it was never explicitly programmed with the answer to.

Scale is a defining feature of the category: modern LLMs are trained on hundreds of billions to trillions of words of text and have parameter counts in the billions, which is part of why they can generalize across such a wide range of tasks and topics without task-specific training for each one. That same scale means they're expensive to train, static after training up to a fixed knowledge cutoff date, and capable of stating incorrect information fluently and confidently — a known limitation often called hallucination.

In a business context, LLMs get deployed either directly (a chatbot answering customer questions) or as a component inside a larger system — combined with retrieval (RAG) for factual grounding, function calling to take real actions like booking an appointment, or a human-in-the-loop review step for anything with real consequence if the output is wrong.