The simple definition
A large language model is a computer program that has read an enormous amount of text and learned the statistical patterns of language well enough to generate new text one word at a time. When you type a question, the model predicts the most likely next chunk of text, then the next, and the next — producing a fluent answer. That is the core of it: an LLM is a very large next-word predictor. The word large refers both to the billions of internal numbers (parameters) it uses and to the vast training data it consumed.
How LLMs are trained
Training happens in two main stages. First comes pre-training: the model reads a huge slice of the internet, books, and code, and repeatedly tries to predict the next piece of text, adjusting its internal parameters whenever it guesses wrong. Over trillions of examples, it absorbs grammar, facts, reasoning patterns, and writing styles. Second comes alignment, often via reinforcement learning from human feedback (RLHF), where people rate responses and the model learns to be more helpful, honest, and safe. The result is a model that not only predicts text but does so in a way humans find useful.
What LLMs can do well
Because they learned such broad patterns, LLMs are remarkably general. They can write and summarize text, translate languages, answer questions, draft and explain code, brainstorm ideas, change tone, extract structured information, and follow complex instructions — all without being explicitly programmed for each task. This flexibility is what makes them feel different from earlier software: one model handles thousands of tasks instead of one.
What LLMs cannot do (and why)
LLMs are not databases or calculators. They can hallucinate — produce confident, fluent text that is simply wrong — because they generate plausible words rather than looking up verified facts. Their knowledge is frozen at a training cutoff date unless connected to live search or tools. They are unreliable at exact arithmetic and at citing real sources. And they have no genuine understanding or memory between conversations unless a system gives them one. Knowing these limits is the difference between using an LLM well and being misled by it.
How LLMs differ from traditional software
Traditional software follows explicit rules a programmer wrote: given input X, do exactly Y. An LLM has no hand-written rules for its answers — its behavior emerges from patterns in training data, so the same prompt can produce slightly different responses, and you cannot trace an answer to a specific line of code. This makes LLMs flexible and creative but also probabilistic and occasionally wrong. The practical takeaway: treat an LLM as a fast, knowledgeable, sometimes-mistaken assistant — verify anything that matters, and pair it with search or tools when you need current facts or exact computation.