In plain words
A large language model is a neural network trained at scale on language-related data. Many familiar LLMs generate text by predicting the next token given the tokens that came before it.
A closer look
Training on diverse text and code teaches statistical patterns that can support summarization, translation, drafting, and problem solving. Many current LLMs use a transformer architecture. Additional training can shape instruction following, conversational behavior, and other capabilities.
The model itself is only part of a chat product. A product may add search, file retrieval, memory storage, tools, and rules for using them. Those additions affect what an assistant can know or do at a particular moment. A base model does not automatically have access to current events or your private files.
In practice
Given a short article and a request for a three-sentence summary, an LLM generates a new sequence of tokens conditioned on the article and your instruction.
A useful distinction
“Predicting the next token” describes a training or generation mechanism, not a complete account of every learned capability. It also does not establish that a model understands or experiences language in the same way people do.