Large language model (LLM): A large language model (LLM) is a neural network trained on enormous amounts of text to predict the next word, then tuned to follow instructions and hold a conversation. LLMs power ChatGPT, Claude, Gemini, and Microsoft Copilot.
An LLM learns language by playing a single game billions of times: given some text, guess the next piece. Train on a large slice of the public internet, books, and code, and the model absorbs grammar, facts, writing styles, and a surprising amount of reasoning as a side effect of getting good at that game.
Raw next-word prediction does not make a helpful assistant, so vendors add a second phase. The model is trained on examples of good instructions and responses and adjusted with human feedback until it answers questions, follows formats, and declines harmful requests. That is the version you talk to.
Three practical facts follow from this. An LLM has a knowledge cutoff, so it may not know recent events unless it can search. It has no database of facts, so it can generate plausible falsehoods. And it works within a context window, so it can only consider a limited amount of text at once, which is why long conversations sometimes lose the thread.
Example at work
A paralegal pastes a 30-page deposition transcript into an LLM and asks for a timeline of events with page citations, a list of every date mentioned, and any inconsistencies between witnesses. She checks each citation against the transcript before it goes into the case file, and she uses only the firm's approved tool because the material is privileged.
Why it matters
Nearly every AI assistant you will use at work is an LLM with a different interface. Once you understand what one is (a very capable text predictor with a memory limit and no fact-checker) you can get strong results from any of them and know exactly what to verify.
Related terms
- TransformerThe transformer is the neural network architecture behind nearly all modern language models. Its key idea, called attention, lets the model weigh every word in a passage against every other word, so it can track meaning across long stretches of text.
- TokenA token is the unit of text a language model reads and writes, usually a word or a piece of a word. Models measure input limits, output limits, and pricing in tokens; as a rough rule, 1,000 tokens is about 750 words of English.
- Context windowThe context window is the maximum amount of text, measured in tokens, that a language model can consider at one time. It includes your prompt, any pasted documents, the conversation so far, and the model's reply.
- Pre-trainingPre-training is the first and largest phase of building an AI model, in which it learns general patterns from a massive dataset, such as predicting the next word across trillions of words of text. Later phases then shape it into a useful assistant.
- HallucinationA hallucination is when an AI model produces information that is false or fabricated but presents it confidently, such as an invented statistic, a citation to a paper that does not exist, or a made-up product feature.