A large language model is a transformer-based neural network with billions of parameters, trained on a broad slice of the internet, books, and code. At runtime it takes a prompt and predicts the most likely continuation one token at a time — the loop that powers every chatbot, copilot, and agent you have used.
For agencies, the relevant point is that the LLM is rarely the differentiator. Frontier models (Claude, GPT, Gemini) are commodities for most tasks; the value lives in the context you feed them, the tools you wire up, and the workflows you wrap around them.
LLMs do not know things — they generate plausible text. They can be confidently wrong, inconsistent across runs, and sensitive to small prompt changes. Treat their output as a draft from a fast, well-read intern, not as a source of truth.