AI Lingo in 5 Minutes: A Cheat Sheet for Normal People
LLM, token, RAG, fine-tuning, agent, frontier model — what the words actually mean, in one short read.
Every AI story is told in a language of jargon that sounds like it should be self-explanatory — and isn't. This is a cheat sheet for the words that actually come up, written for people who read AI news but don't live inside it.
The six words that carry most of the conversation
You can get through 80% of AI coverage with six terms. The rest is marketing wearing technical clothes.
LLM — the engine
LLM stands for Large Language Model, and it is the thing underneath every chatbot you have tried. The trick that made it work is almost embarrassing in its simplicity: take an enormous amount of text, and train a model to predict the next word. Do that at sufficient scale, and the model stops merely completing sentences and starts being able to summarize, translate, reason and write code. Nobody fully planned for that second part — it emerged from scale. When people say "the model", this is usually what they mean.
Token — the unit
Models do not read words; they read tokens, which are chunks of text roughly three-quarters of a word long in English. Everything a model costs, in time and money, is measured in tokens. When you hear "this model costs X per million tokens", you are hearing the price of its reading and writing.
Context window — how much it can hold in mind
A model cannot remember anything that happened before the conversation started, and it can only "see" what fits in its context window — the amount of text it can consider at once. Modern models measure this in hundreds of thousands of tokens, roughly the length of a long novel. Past that window, it forgets. This is why you can paste a 200-page document into some assistants and ask questions about it — and why a conversation that goes on too long starts to lose its thread.
RAG — teaching it to look things up
RAG stands for Retrieval-Augmented Generation, and it is the honest way to make a model answer questions about your own documents. Instead of hoping the model memorized your PDF (it did not), you first search your documents, then hand the model the relevant pages and let it answer from those. The result is an assistant that cites its sources and can be told "no, this is the 2026 version" — without retraining anything.
Fine-tuning — making it yours
Fine-tuning is extra training on your own data after the model's original training. It is the difference between a general assistant and one that has learned your house style: your tone, your product names, your rules. It changes behavior, not knowledge — the model still only knows what it learned originally, but it learns how you want things done.
Agent — handing it the controls
An agent is a model that can act, not just answer: search the web, click through pages, run code, use tools, and take several steps toward a goal on its own. This is the word that changed the conversation in the last two years. A chatbot waits for you; an agent goes and does. It is also the word attached to most of the current safety debates — because a model with the ability to act is a model with the ability to do damage, even when it means well.
Frontier model — the current ceiling
"Frontier" is simply the label for the most capable models that exist at any given moment. It is a moving goalpost by definition: today's frontier is next year's baseline. When you read "frontier model", the useful translation is "the most expensive and most capable thing money can rent right now".
Which one do you actually need?
Most products you touch will not make you learn any of this. But when a vendor starts throwing terms around, this table tells you what they are actually offering:
| Term | What it is | When you need it |
|---|---|---|
| LLM | The underlying model | Whenever you use any AI product at all |
| Context window | How much it can hold in mind at once | When you want to feed it large documents |
| RAG | Answering from your own documents | When the model must know things it wasn't trained on |
| Fine-tuning | Extra training on your data | When you need consistent style or behavior at scale |
| Agent | A model that can take actions | When the task involves doing, not just talking |
| Frontier model | The current capability ceiling | When raw quality matters more than cost |
The mental model that makes it all fit
Here is the sentence that ties it together: a model is a very capable brain with no memory and no hands. The context window is how much it can see; RAG gives it a filing cabinet; fine-tuning gives it a personality; an agent gives it hands. Every AI product you will ever use is some combination of these four ingredients, dressed in nicer packaging.
What to do with this
- Read any AI headline and translate the jargon into the four ingredients above. You will instantly be able to tell what a product actually does.
- When a model "hallucinates" — another word worth knowing: it means confidently making something up — remember why. It is a next-word predictor, and being wrong has never stopped a next-word predictor.
- Do not buy a course to learn more terms. Read three honest articles with this cheat sheet open, and the vocabulary will be yours.
The rest of AI fluency is just practice. The vocabulary, though, you now have.