Definition
Large language models are built on the transformer architecture and trained in stages: first on a huge corpus to predict the next token, then with further training on instructions and human or AI feedback so they follow requests helpfully and safely.
LLMs are general-purpose: the same model can draft an email, extract fields from an invoice and write code. That generality is what makes them useful in implementation work, and also why they need grounding, guardrails and evaluation to be reliable on a specific business task.