A large language model is a neural network trained on enormous amounts of text to predict what comes next. Doing that well enough, at enough scale, produces something that can write, summarise, translate, answer questions and follow instructions.
Also called: LLM, foundation model · Reviewed
Training shows the model a very large body of text and asks it, over and over, to predict the next token. Each miss nudges billions of internal weights. Nothing in that process teaches it facts as facts — it learns the statistical shape of language, and a great deal of world knowledge comes along for the ride because accurate text is more predictable than inaccurate text.
After that comes a second stage, usually some mix of supervised fine-tuning and learning from human feedback, which is what turns a raw text predictor into something that follows an instruction rather than continuing it. A base model given "Write me a poem about rain" might reply with three more writing prompts. An instruction-tuned model writes the poem.
The reason one architecture ended up behind writing tools, coding assistants, search and customer support is that an enormous range of tasks can be expressed as text in, text out. You do not need a separate model for summarising, classifying and rewriting — you need one model and three different instructions.
That generality is also what makes the skill of using them portable. Learning to work with one model transfers almost entirely to the next, which is not true of most software.
The most common mistake is treating a model as a database. It has no index to look things up in and no notion of whether a given sentence is true — it produces the most plausible continuation, and a plausible-sounding wrong answer costs it nothing. That is the root of hallucination, and it is why anything that must be correct needs retrieval, tools or checks around the model rather than trust in it.
The second is assuming bigger is always better. Larger models cost more and respond more slowly, and for classification, extraction and routing a small model with a good prompt often wins on every axis that matters.