If you've used a chatbot, an AI writing assistant, or an AI coding tool in the past few years, you've used a large language model. Usually shortened to LLM. This guide explains what that actually means, without assuming you have a machine learning background.
The short version
An LLM is a computer program trained on enormous amounts of text to predict what word (technically, what "token", a word or word-fragment) is likely to come next, given everything that came before it. That sounds almost too simple to explain the capabilities you've probably seen (holding a conversation, writing code, summarizing a document) and the gap between that simple description and those real capabilities is the most important thing to understand about how these systems work.
How training actually works
During training, a model is shown huge volumes of text (books, articles, code, web pages) and repeatedly asked to predict the next token in a passage it hasn't seen the ending of yet. When it predicts wrong, its internal parameters (numerical values, often numbering in the tens or hundreds of billions) are adjusted slightly to make a similar prediction more accurate next time. Repeated across trillions of these predictions, the model doesn't memorize the text verbatim. It learns statistical patterns in how language works: grammar, facts that appear consistently across many sources, reasoning patterns that show up repeatedly in its training data, and even some patterns in code and structured data if that was part of training.
Why "predicting the next word" produces something that looks like understanding
This is the part that surprises people most: next-token prediction, done at a large enough scale, turns out to require the model to implicitly capture a huge amount of structure about the world. Because predicting the next word accurately in a story requires tracking who's who and what's happening, predicting the next word in a math explanation requires something like tracking the actual arithmetic, and predicting the next word in an argument requires something resembling logical consistency. Nobody explicitly programmed a large language model with grammar rules or facts about the world in the way older AI systems were hand-coded, that structure emerged as a byproduct of the model getting very good at the narrower task of prediction. Our explainer on the transformer architecture underlying nearly every current LLM covers the specific technical innovation that made training models at this scale practical.
The architecture underneath, briefly
Almost every current LLM is built on the transformer architecture, introduced in 2017, which processes an entire passage of text at once using a mechanism called self-attention rather than reading word by word in sequence. This let researchers train much larger models on much more data than was previously practical, and it's the direct technical ancestor of every model discussed elsewhere on this site. GPT, Claude, Gemini, Llama, Mistral, and the rest.
What "large" actually refers to
"Large" in large language model refers primarily to the number of parameters (the adjustable numerical values the model learned during training) and to the volume of training data used. Model sizes have grown from millions of parameters in early language models to hundreds of billions in current frontier models, though some newer architectural approaches, like the mixture-of-experts technique, allow a model to have a very large total parameter count while only activating a fraction of it for any given query. Decoupling "how large is the model" from "how expensive is it to actually run."
What LLMs are genuinely good at
Based on how they're trained, LLMs are strongest at tasks that resemble patterns well-represented in their training data: generating fluent, well-structured text; summarizing and reformatting information; writing and explaining code in popular programming languages; and following instructions expressed in natural language. They're also surprisingly capable at tasks that weren't explicitly part of training but that emerge from the breadth of patterns in enough text. Translation between languages, basic reasoning through multi-step problems (especially when prompted to show intermediate reasoning, a technique we explain in our chain-of-thought piece), and adapting tone and style to a described audience.
What LLMs are genuinely bad at, and why
Because an LLM generates the statistically most plausible continuation of text rather than looking up verified facts, it can produce confident, fluent, and completely incorrect statements, a phenomenon often called "hallucination." This isn't a bug that occasional updates will fully eliminate; it's a structural consequence of how these models work, which is why grounding techniques like retrieval-augmented generation (having the model draw on specific, retrieved source documents rather than relying purely on its trained-in memory) have become standard for any application where factual accuracy matters. It's also why independently verifying anything an LLM tells you before acting on it is a real discipline worth building, covered in our guide to fact-checking AI-generated content.
LLMs are also limited by their training cutoff. They don't automatically know about events after their training data was collected, unless paired with a retrieval or search capability that brings in current information, and they can be inconsistent: the same question asked two different ways can produce meaningfully different answers, because the model isn't reasoning from a fixed internal knowledge base the way a database or a person with settled beliefs does. It's generating a plausible response each time, shaped heavily by exactly how the question was framed, which is why prompting technique genuinely matters (see our guide to writing better prompts).
How LLMs relate to "AI" more broadly
LLMs are one specific, currently very prominent category of AI, focused on language. They're related to but distinct from image and video generation models (see our guide to generative media), which use different underlying techniques, and from the robotics and embodied AI systems covered in our robotics coverage, which face a different, harder set of physical-world problems that text-only training doesn't address.
How to think about them going forward
The most useful mental model isn't "a thinking machine" or "just autocomplete". Both oversimplify in different directions. A more accurate frame: an LLM is a system that's extremely good at producing plausible, well-patterned text based on everything it learned from an enormous amount of human-written material, which makes it genuinely useful for a wide range of language tasks, and genuinely unreliable as a standalone source of verified truth. Understanding that distinction (capability at generating plausible text versus reliability as a factual source) is the single most useful thing to carry into any interaction with one of these systems. For ongoing coverage of how specific models and techniques are evolving, see our Model Releases and Research sections.