A large language model is a type of neural network trained on enormous quantities of text so that, given some input, it can predict what text should come next. That single capability — next-token prediction at scale — turns out to be enough to summarise, translate, answer questions, write code and hold a conversation.
The "large" is literal: modern models have billions of parameters and are trained on a meaningful fraction of the public internet. That scale is why they are general-purpose, and also why running them well in production is an engineering problem in its own right — latency, cost per token, context limits and evaluation all have to be designed for, not assumed.
For a business, the important distinction is between a demo that calls an LLM and a system that depends on one. The gap between them — retrieval, guardrails, evaluation, fallback behaviour — is where most AI projects succeed or quietly fail.