A large language model (LLM) is trained on huge text collections to predict what comes next in a sequence of tokens. From that one skill it can summarize, translate, answer questions, write code and follow instructions, because those are all patterns in language.
An LLM has no live access to the world by itself. It only knows what was in its training data plus whatever you put in the prompt, and it can state wrong things with confidence, which is called an LLM hallucination. To make it useful for real work you give it instructions (prompt engineering), limit what fits in a single request (context window), connect it to tools (tool calling) and feed it your own data (Retrieval-Augmented generation).
Related: Generative AI, AI Agent, AI Assistant.