Does LLM stand for Master of Laws?

Only in law. In AI, LLM stands for large language model: software trained on a very large amount of text that predicts what should come next, one piece at a time. The two meanings share an acronym and nothing else.

What is a token in AI?

A token is a chunk of text the model reads or writes at once, roughly four characters, or about three-quarters of an English word. Context window sizes are counted in tokens, which is why a 54,000-token window is about 40 pages rather than 54,000 words.

What is an open-weight model?

A model whose files are published so anyone can run it themselves. That competition is a large part of why capable AI is cheap now, because a provider can serve a published model instead of paying a single vendor for access.

What is a context window?

The amount of conversation a model can hold in view at once. When a long chat starts forgetting things you said earlier, the context window is why, not a message limit. ChatGPT's tiers range from 27,000 tokens on free to 128,000 on Pro.

Do I need to understand AI terms to use AI well?

No, and that is the point of this glossary. The vocabulary describes how the machinery works, not how you should use it. You do not need to know how an engine works to drive, and you do not need to know what a token is to ask a question.

What does LLM mean? An AI glossary to ignore

LLM means large language model, not the law degree. Every AI term an ordinary person trips over, in one plain line each, and why none of them matter.

What does LLM mean?

LLM stands for large language model: software trained on a very large amount of text so that it can predict what should come next, one chunk at a time. "Large" refers to the amount of text it learned from and the number of internal settings that learning produced. If you've used ChatGPT, Claude or Gemini, you've used one.

In law, LLM means Master of Laws, a postgraduate degree. The two share an acronym and nothing else.

Why this glossary exists

I built an app that hides every term below from the person using it. That was deliberate, and it's the reason this page has an unusual promise: you can read it and then ignore it.

The vocabulary of AI describes how the machinery works, not how you should use it. You don't need to know how an engine works to drive. You don't need to know what a token is to ask a question. And a person who never learns the word "token" gets better answers than a person who learns it and starts tuning around it.

These are the terms you keep tripping over in articles and settings screens, each in one line you could repeat to a friend, followed by why it doesn't need to be your problem.

The glossary

Term	What it means	Why you can ignore it
LLM	Large language model. Software that predicts what text should come next.	It's the category, not a choice you make.
Parameter	One of the internal settings a model learned. "70B" means 70 billion of them.	Size isn't the same as quality, and you can't see it from the outside.
Token	A chunk of text, about four characters or three-quarters of an English word.	Only matters if you're counting a limit.
Context window	How much of your conversation the model can hold in view at once.	This one does matter. See below.
Prompt	What you type.	You already do this. It's called asking.
Training	The expensive process that creates a model from text.	Happens once, before you arrive.
Inference	Running a trained model to produce an answer.	This is what you pay for, per answer.
Open weights	A model whose files are published for anyone to run.	It's why capable AI got cheap, and you benefit without knowing it.
Closed model	A model you can only reach through its owner's product.	Relevant only to what you're allowed to do with the output.
Fine-tuning	Extra training to specialise a model.	Matters when a product claims a bespoke model.
Hallucination	A confident, wrong answer.	The one term worth knowing. See below.
Agent / agentic	Software that takes several steps and uses tools on your behalf.	Marketing's favourite word for "it can do more than one thing".
Multimodal	Handles images or audio as well as text.	A feature, not a skill you need to learn.
Reasoning model	A model that works a problem through before answering, using more compute.	Slower, better on hard problems, wasted on easy ones.

The two that are worth your time

Everything above can stay in the drawer except two, and neither is really jargon.

Context window is worth knowing because it explains something you'll feel. A model doesn't hold your whole conversation forever; it holds the most recent stretch, counted in tokens. When a long chat starts forgetting what you told it earlier in the same conversation, that is the window, not a message limit. OpenAI's consumer tiers run from 27,000 tokens on free to 54,000 on the mid tiers and 128,000 on the top one, and unless you're building something, the only number that matters is whether your conversations fit. The full ladder is in is ChatGPT Go worth it, and the failure it causes is in ChatGPT's message limit.

Hallucination is worth knowing because it tells you what to distrust. A language model produces plausible text, and plausible isn't the same as true. It will state a false thing in the same confident tone it uses for a true one, and the tone gives you nothing to go on. The practical rule isn't "ignore AI": it's "check anything you would be embarrassed to repeat". Dates, quotations, statistics and citations are where it slips most.

The terms that exist to sell you something

Some of this vocabulary is descriptive. Some of it exists mainly to help sell a product.

"Agentic" is the clearest case. It describes software that takes several steps and uses tools rather than answering in one shot. That is a real capability. It's also a word that arrived with pricing pages, because a tool that "acts on your behalf" sounds like more than a tool that answers questions, and more is easier to charge for. When a product leads with how agentic it's, the useful question is what it finishes for you that a plain question wouldn't.

"Model choice" is the other. Choosing among models is presented as control, and it's closer to homework. It exists because the industry sells access to individual models rather than the result of using the best one at the time, which is a pricing decision presented as a feature. If you want the argument rather than the definition, it's in you don't pick a model.

The one I would retire

"AGI" and "superintelligence."

Nobody can agree on what either word means, and that is the whole problem. "AGI" is supposed to mean AI that is as good as a person at everything, but nobody has written down what counts, so there is no way to tell whether it has arrived. "Superintelligence" is worse. It sounds like a technical term and works like a superlative: it assumes the thing is coming and that it's bigger than us, which settles the argument before anyone has said anything.

The words sell. "We're building AGI" is a line for investors and job ads, attached to a product that answers questions. That is why I'd take them out of ordinary conversation. Not because the worries behind them are silly, but because you can't argue clearly with a word that has no agreed meaning.

Say what the model does instead. "This one is good at X and bad at Y" can be checked. "This is on the path to AGI" can't.

A close second is "AI-powered", which is on this site's own banned-words list.

What this glossary is not

It isn't a course, and it won't make you good at prompting. There's no such skill to acquire in the way the guides imply: there's writing what you actually want, which isn't an AI skill at all.

It's also not stable. This vocabulary turns over every year: terms arrive with funding rounds and leave with them. The definitions above are written as of September 2026, and the part that will still be true is the idea behind it: the words describe the machinery, not the job.

Where to go next

If you came here because a settings screen asked you to choose something, that is the industry's failure mode rather than yours. The arguments for refusing to choose are in you don't pick a model and why there are so many AI apps. And if what you came for was a cheap ChatGPT alternative rather than a vocabulary: Plainly is $5 a month for unlimited chat, and it automatically uses the best available model for each question, which is the whole point of the exercise.