Large language models, briefly

A large language model is a statistical system trained to predict the next token in a sequence. "Token" is roughly a word-fragment. "Predict" means assign a probability. That is the whole machine, in one sentence.

Calling it autocomplete is technically correct. It is also unhelpful, because the scale at which the prediction happens — billions of parameters, trillions of tokens of training data — produces behaviour that looks, from the outside, like reasoning. It is not reasoning in the way a journalist reasons about a source. But it is not a parlour trick either.

The practical implication for editorial work: a model can produce a fluent, plausible sentence about almost anything. Fluency is not evidence. Plausibility is not truth. The verification burden does not shrink when a model is in the loop — it shifts.