What Does LLM Stand For? Large Language Models Explained

What does LLM stand for? In AI, large language model. In law, Master of Laws, usually written LL.M. with the periods. If you arrived from a conversation about ChatGPT, Gemini or Claude, it's the first one, and most of this article is about that meaning. If you arrived from a law school page or a lawyer's email signature, it's the second, and there's a proper answer for you near the end.

A large language model is a program trained on an enormous amount of text to predict the next word, and that one job, repeated at a scale no person could survive, is where everything the chatbot does comes from. The three words in the name each carry a specific promise, and I'll take them one at a time, with the paper or textbook each promise comes from.

Which LLM Are You Reading?

The same three letters mean different things in different rooms. Here's how to tell.

That third row deserves a word. People search for what LLM means on Instagram or in a memorial post, and I looked for a documented slang expansion. I didn't find one. The Netlingo page that ranks for "LLM in internet slang" gives only the AI meaning. If you've seen the letters used some other way, the honest answer is to ask the person who posted them, because I'm not going to invent an expansion to fill the gap.

Large

The L for large is a claim about size, and the size is measured in parameters.

A parameter is one adjustable number inside the model. Training is the process of nudging all of those numbers until the model's guesses about the next word get good. The paper that introduced GPT-3 in 2020 described it as "an autoregressive language model with 175 billion parameters, 10x more than any previous non-sparse language model." That's 175 billion dials, none of them set by a person. If you counted them at one per second without sleeping, you'd finish in about 5,500 years, which is the scale the word large is pointing at.

Why build it that big? Because a separate 2020 paper from the same lab, Scaling Laws for Neural Language Models, found that the model's error "scales as a power-law with model size, dataset size, and the amount of compute used for training." In plain terms, making the model bigger made it better in a smooth, predictable way, with no cliff where the gains stopped. The same paper found that "larger models are significantly more sample-efficient," meaning a big model learns more from each example than a small one does.

That result is the reason the word large is in the name. It's from 2020, and I wouldn't assume its exact numbers still hold. What survived is the direction. Bigger kept working, so the labs kept going, and "large" stopped being a description and became a category.

I'd hold one caution alongside that. Large describes the model's capacity to capture patterns, and nothing in the word promises that the patterns are true. A large model can be confidently wrong at scale, which the site's guide to what an LLM is shows with a two-minute test you can run yourself.

Language

The middle word tells you the job, and the cleanest definition I know comes from a textbook rather than a company.

Chapter 3 of Jurafsky and Martin's Speech and Language Processing opens with a sentence to finish. "The water of Walden Pond is so beautifully..." You'd guess blue, or green, or clear, and you wouldn't guess refrigerator. The textbook then gives the definition. "A language model is a machine learning model that predicts upcoming words. More formally, a language model assigns a probability to each possible next word, or equivalently gives a probability distribution over possible next words."

Read that twice, because it's the whole product in two sentences.

A probability distribution just means a set of options with a share of likelihood attached to each. Here's a made-up illustration, not a measurement. After "so beautifully," a model might put 40 percent on blue, 25 on clear, 15 on green, and spread the rest thinly across thousands of other words, with refrigerator so close to zero it never gets picked. Generating text means sampling from that spread, one token at a time, and feeding each pick back in before guessing again.

A language model doesn't hold a list of facts. It holds probabilities over what word comes next, learned from what usually came next in the text it was trained on. When the most probable continuation of your question happens to be true, you get a correct answer. When it happens to be plausible instead, you get a fluent wrong one.

The unit it works in isn't quite a word. Text gets chopped into tokens, which are roughly word-sized pieces, and the tokens explainer walks through a receipt-sized example if you want to see where the cuts fall. For the acronym, "language" means the model's input and output are sequences of these pieces, and its skill is guessing the next one.

The same chapter notes that large language models learn an enormous amount about language just by being trained to predict words. Nobody writes grammar rules into them. The rules fall out of the prediction task. That's the part I find most surprising, years in.

Model

The last word is the most ordinary and the most misread.

A model, in this sense, is a mathematical object that makes predictions. A weather model predicts rain from pressure readings. A language model predicts the next token from the tokens so far. The word carries no suggestion of understanding, intention or a mind. It's the same word an insurer uses for the thing that prices your policy.

Two properties of the word matter for how you use the tool. A model is built once, in a training run, and then used many times without changing. And a model is a specific kind of machine, in this case a neural network, which is a structure of layered numerical operations. The companion piece on what a neural net is walks through one at beginner speed. For the acronym, it's enough to know that "model" names the frozen, trained artifact, and that the chatbot you talk to is a product wrapped around it.

What Does LLM Stand For in ChatGPT?

The letters don't appear in the product name, but the thing they name is inside it.

ChatGPT is the application, the box you type into, with its history, its settings, its web search and whatever tools OpenAI has attached this month. Underneath it runs a large language model from OpenAI's GPT family. So "the LLM in ChatGPT" is the engine, and ChatGPT is the car. If you want the GPT letters expanded one at a time, the guide to what ChatGPT stands for does exactly that, and I won't repeat it here.

The distinction is practical rather than pedantic. When a feature appears or disappears, that's usually the product changing. When answers get better at maths or worse at citations, that's usually the model changing. Knowing which layer you're looking at saves a lot of confused forum posts.

Are All LLMs GPT?

No. GPT is the family name OpenAI gives its models. Other companies build large language models under their own family names, and the Levels of AGI paper from Google DeepMind names several of them side by side as frontier language models.

So every GPT is an LLM, and most LLMs are not GPTs. The mistake usually runs the other way, with "GPT" used as a generic word for the whole category, the way some people say Hoover for any vacuum cleaner. It's harmless in conversation and confusing in a purchase decision, because the models differ in what they're good at and what they cost.

My habit is to say "model" when I mean the category and to use the family name only when the family matters. It keeps me from writing sentences like "which GPT should I use" when the honest answer might be a model that isn't one.

What Does LLM Mean in Masters?

Now the law degree, for the reader who came here from the other room.

LL.M. stands for Master of Laws. Harvard Law School's own page describes its LL.M. program as a one-year degree, and writes the abbreviation with the doubled L and periods throughout. Law school pages and the encyclopedia entry both describe it as a postgraduate degree, taken after a first law degree rather than instead of one.

The doubled L is the detail that trips people up. The usual explanation is that the abbreviation comes from the degree's Latin name, Legum Magister, where "legum" is a plural, and older academic abbreviations doubled a letter to mark a plural. I haven't seen a law school spell that out on its own site, so treat the Latin as the conventional story rather than something I've verified from a primary source. What I can verify is the spelling. If the letters have periods and two Ls, you're reading about the degree.

That's also the quickest fix for the ambiguity in your own writing. Write LL.M. for the degree and LLM for the model, and readers on both sides will know which room they're in.

The Short Version

LLM stands for large language model when the subject is AI. Large means a model with billions of adjustable parameters, built that way because bigger kept getting better. Language means its whole job is assigning probabilities to the next word, which is why it can be fluent and wrong at the same time. Model means a trained, frozen predictor with a product wrapped around it, and a GPT is one family of them among several.

LL.M. stands for Master of Laws, a one-year postgraduate law degree, and has nothing to do with any of the above.

If you keep one thing, keep the middle word. The model predicts language. It doesn't look anything up. Every useful habit for working with these tools follows from that sentence.