The Vocabulary Every AI User Needs
Model, prompt, context window, token, hallucination, and the real difference between a chatbot, an assistant, and an agent.
Big idea
Five everyday words — model, prompt, context window, token, hallucination — each name a real mechanism, and knowing them explains why a chat can suddenly forget what you said earlier or be wrong with total confidence.
The core terms
- Model — the trained thing itself: a large file of learned numbers (weights) that turns an input into an output. "GPT-4," "Claude," "Llama" are all models.
- Prompt — the text you send the model. Everything about how you phrase it, structure it, and what you include is part of the prompt, not just the literal question.
- Context window — how much text (your prompt, the conversation so far, any documents) the model can actually "see" at once. Anything older than that gets pushed out and forgotten, the same way you'd forget the start of a very long phone call.
- Token — the actual unit the model reads and predicts in, smaller than a full word for anything uncommon (covered in depth in the Intermediate tier). Context windows are measured in tokens, not words or characters.
- Hallucination — when a model states something false with the same fluent confidence as something true. It isn't "lying" — it has no separate fact-check step, so a wrong answer and a right one can look identical from the outside.
Think of it like
Picture ordering at a busy cafe counter. The barista is the model: trained, skilled, and fixed in how they work. Your order is the prompt, and how you word it shapes what you get. The rail of order slips above the espresso machine is the context window: only so many slips fit, and the oldest fall off as new ones arrive. If a slip is smudged, the barista may still confidently make something plausible but wrong. Each vocabulary word names one piece of that counter: the skilled worker, the request, the limited slip rail, and the confident guess.
Chatbot vs. assistant vs. agent
These three get used interchangeably, but they describe a real progression in capability:
- A chatbot answers in text and does nothing else — no memory beyond the current conversation, no ability to act on anything.
- An assistant is a chatbot with some tools wired in — it might search the web, do a calculation, or check a calendar, but a person is still driving each individual step.
- An agent is given a goal and takes multiple actions on its own to get there — reading files, running commands, checking its own output, deciding what to do next — with much less step-by-step human direction. (The full jump from chatbot to agent gets its own topic next.)
Why the vocabulary matters
These words get thrown around loosely in marketing and in casual conversation, but they point at real, different mechanisms. Knowing that "hallucination" means "confidently wrong, not maliciously wrong" changes how you verify an answer. Knowing what a context window is explains why a long conversation can suddenly seem to "forget" something you said earlier. The rest of this track assumes you have these five terms solid.
The vocabulary, with numbers attached
| Term | A concrete figure |
|---|---|
| Token | ~¾ of an English word on average; 1,000 tokens ≈ 750 words |
| Context window | how many tokens a model reads at once — commonly 128K to 1M for current large models, roughly a 300-page book at the lower end |
| Parameters | the learned numbers: ~3 billion for a model that runs on a laptop, hundreds of billions for the largest |
| Training vs inference | training is done once, at huge cost; inference is every answer you get, costing a fraction of a cent |
| Temperature | 0 → nearly the same answer each time; ~1 → more varied |
| Latency | ~0.3–1 s to the first word; then tens of words per second |
Roughly how many tokens is a 1,500-word English essay?
Key terms
- Context window — The amount of text a model can see at one time, measured in tokens.
- Token — The small unit of text a model actually reads and predicts, often smaller than a whole word.
- Hallucination — A false statement delivered with the same fluent confidence as a true one.
- Assistant — A chatbot with some tools attached, where a person still directs each step.
- Agent — A system given a goal that takes several actions on its own to reach it.