Large language model
A model trained to predict the next token over enormous amounts of text, which turns out to be enough to do a great deal.
Also written: LLM, LLMs, language model.
A large language model is, mechanically, a next-token predictor. Shown a sequence, it produces a probability distribution over what comes next, one token at a time, and a sample is drawn from it.
That description is accurate and sounds far too small for what these systems do. The surprising empirical fact of the last several years is that doing this well enough, at a large enough scale, produces something that can hold a conversation, write working code and follow instructions it has never seen. Nobody designed those abilities in; they appeared as the models and the training data got bigger.
It is worth keeping both halves in mind. “It is just predicting the next token” explains the failures — the confident wrong answer, the invented citation — and badly under-describes what the thing can do.