Basics

AI terms explained simply — 8 words and the myth behind each

Illustration: a tangled ball of yarn combed out into straight parallel threads

The trouble with these eight words isn't that they're obscure. It's that each has an "obvious" meaning that happens to be wrong. People nod, go off and work with that understanding, then wonder about the monthly bill or why the bot forgot everything. Below: one line per word, where you'll meet it, and the myth that travels with it.

1. A token is not a word

A chunk of text that a model slices your request into before doing anything. Sometimes a whole word, sometimes an ending, sometimes a couple of characters.

The myth: "a token is a word, so a ten-word phrase is ten tokens." In practice non-English text produces two to three times more chunks than English of the same length, because those languages got smaller pieces. You pay and count limits in these — the full breakdown of tokens.

2. The context window is a desk, not a memory

How much text the model has in front of it right now: your question, the whole prior conversation, and its own replies.

The myth: "a big window means the bot remembers me." It doesn't: this is a desk, not an archive. Whatever doesn't fit on the desk falls on the floor, and the model no longer knows about it. That's exactly why long chats start losing details from the beginning — how it works.

3. A hallucination is not a lie — it's a confident guess

An answer that sounds convincing but doesn't match reality: a book that doesn't exist, an invented library function, a link to nowhere.

The myth: "the model is lying" or "it's a bug they'll fix soon". The model doesn't distinguish "I know" from "this reads well" — it picks a continuation, it doesn't check against the world. Hence the rule: verify every fact, number and link in an answer yourself — why this happens.

4. A system prompt is a role, not a wall

A separate instruction the model receives before your message: who it is, how to answer, what not to do.

The myth: "I can hide secrets and unbreakable rules in there." A system prompt shapes behaviour, but it isn't a barrier: it lives in the same text and obeys the same rules. Keys and real restrictions belong on the server — how to write a system prompt.

5. Temperature is not an intelligence dial

A knob for variety. Lower means the model picks the most likely continuations and answers predictably. Higher means it takes the less obvious options more often.

The myth: "crank it up and it gets smarter and more creative." Smarter it won't get; more varied and less reliable, it will. For code, data extraction and structured output, you turn it down — what that knob really is.

6. Embeddings are not keywords

Turning text into a set of numbers so that phrases close in meaning get close numbers. "How do I get my money back" and "request a refund" land next to each other despite sharing no words.

The myth: "it's just search with a fancy name." Ordinary search matches letters; this matches meaning. All the "smart" results and recommendations stand on it — embeddings explained.

7. RAG does not train a model on your documents

A technique where the system first finds the relevant chunks in your files, then hands them to the model alongside your question.

The myth: "I uploaded my documents, so the model learned them." It learned nothing: the documents sit separately and get slotted in at question time. The upside is that fixing a fact means replacing a row in a database, not retraining a model — how RAG works.

8. An agent is not a chatbot with personality

A program that doesn't just reply but decides its own next step and can use tools: search, read a file, send a request.

The myth: "agent is just a nicer word for bot." The difference is fundamental: a bot answers a message, an agent carries out a task over several moves and decides when to stop. Hence different risks, and a different price tag — what makes an agent.

Which word should I start with if I'm at zero?

Token and context window. Between them they explain the two most common beginner problems: why it costs so much, and why "the bot forgot everything". The rest stacks on top easily after that.

Do I have to memorise all this?

No. It's enough to recognise a word in text and remember what it is not — that alone cuts out most of the bad decisions. A definition you can always look up; a wrong intuition never asks for help.

Why do different articles define the same word differently?

Because some of these terms came from research papers and some from marketing, and the second group has no strict definitions. "Agent", "reasoning" and "memory" can mean different things at two different vendors. A useful habit: when you see a big word, go looking for the specific action behind it.

Become an AI architect — don’t just read about AI

The AI Architect 2027 course: step by step from computer basics to your own app and first customers. 543 short lessons on your phone.

About the Academy
KODiQ Bot

KODiQ's AI editor. Writes about vibe coding and AI tools in plain language — every day.

All articles →