AI Jargon Decoded: 25 Terms Beginners Hear Every Week

"Temperature" sounds like something you check before leaving the house. "Parameters" sounds like settings on a washing machine. "Hallucination" sounds like a medical event. Yet all three turn up constantly in stories about the systems that draft our emails, summarise long documents and answer customer questions. If you have nodded along to a conversation about context windows while quietly wondering what a token actually is, you are in good company. The vocabulary arrived faster than anyone learned it. What follows is a plain-English tour of the 25 terms you are most likely to meet, grouped the way they tend to come up in practice.

Start with the machine: parameters, weights and training

1. Model. The finished product: a very large set of numbers plus the instructions for using them. When someone says "a 70-billion-parameter model," this is the thing they mean.

2. Parameters. The numbers the system adjusts while learning. The "7B" or "70B" you see in product names is a count of them. More is not automatically better; how the model was trained and tuned often matters more than raw size.

3. Weights. Largely the same idea as parameters, described differently. The word points to the strength of connections between the model's internal units, which together decide how an input turns into an output.

4. Training. The long, expensive process of showing a model enormous amounts of text and nudging those numbers until its predictions improve. This is the stage that costs serious money and specialised hardware.

5. Dataset. The material used for training. Size gets the headlines, but quality, diversity and licensing are what usually decide whether the result is useful.

6. Fine-tuning. Extra training on a smaller, targeted set of examples, such as a company's own support tickets, to push a general model toward a specific job or tone. Cheaper than building from scratch, and often enough.

Tokens, context and the words you type

7. Token. The small chunk of text a model reads and writes. In English, a token is often around three to four characters, so a long word may split into two or more. Tokens are also the unit most services bill by, which is why a hundred-page PDF costs more to process than a paragraph.

8. Context window. How many tokens the model can hold in mind at once, counting your messages, any uploaded files and its own replies. Exceed it and the earliest details quietly drop out of view, which is why a long chat can seem to forget what you said at the start.

9. Prompt. Whatever you type in. Nothing mysterious, though the phrasing genuinely changes the answer.

10. System prompt. Hidden instructions set by whoever built the app, defining its role, boundaries and tone. You rarely see it, but it explains why the same underlying model behaves differently in two different products.

11. Embedding. A list of numbers that represents meaning. Phrases used in similar ways end up close together mathematically, which is how a tool can find your relevant paragraph among thousands without matching exact keywords.

What happens after you hit send

12. Inference. The moment of use: prompt goes in, answer comes out. Training builds the model once; inference happens every time someone asks it something. When you see costs quoted per million tokens, that is inference you are paying for.

13. Latency. How long the wait feels. Two measures matter: time until the first word appears, and time until the answer finishes. Streaming text as it is generated does not make the model faster, but it makes the wait far easier to tolerate.

14. Compute. Raw processing power, usually meaning specialised chips. It is the reason training runs are measured in months and budgets rather than hours.

15. Open weights. A model you can download and run on your own hardware or cloud account, subject to its licence. The opposite is a closed model, reachable only through someone else's interface. Open weights give you control; closed models often give you convenience and a smaller bill for occasional use.

16. Multimodal. Handles more than text. Images, audio and sometimes video can go in or come out. Useful when the information you have is a photo of a receipt rather than a paragraph.

The dials that change the answers

17. Temperature. A randomness dial. Turn it down and the model picks the most predictable wording, so the same question tends to produce the same answer. Turn it up and responses get more varied, more surprising and more prone to nonsense. Sensible defaults: low for extracting data, coding and factual summaries; moderate for brainstorming.

18. Top-p. A close relative of temperature that limits the choices to the most likely words until their probabilities add up to a threshold. Adjust one or the other rather than both, or you will spend an afternoon chasing your own settings.

19. Reasoning (or chain of thought). Models that write out intermediate steps before giving a final answer. Usually better at arithmetic and multi-step logic; slower and more expensive, since all that thinking is billed like any other text.

Three settings habits worth keeping

• Change one setting at a time and keep the prompt fixed, or you will not know what caused the difference.
• Save prompts that work. A prompt is a reusable asset, not a one-off sentence.
• When output looks vague, add specifics before you reach for the dials: audience, length, format, an example of the tone you want.

When output goes wrong

20. Hallucination. Confident, fluent and wrong. An invented citation, a phone number that does not exist, a policy that was never written. The system is not lying; it is completing a pattern. Treat any specific fact, figure or quote as unverified until you check the source.

21. Grounding and RAG. Retrieval-augmented generation means fetching relevant documents first, then asking the model to answer using them. It anchors replies in material you supplied, which reduces invention and lets you show your working.

22. Guardrails. Filters and rules that block certain inputs or outputs. Helpful, imperfect, and worth testing with real examples rather than trusting a description.

23. Alignment. The broad effort to make behaviour match human intent and values, from refusing harmful requests to following instructions reliably. An open research problem, not a finished feature.

Benchmarks, agents and reading the headlines

24. Benchmark. A standardised test used to compare systems. Handy for a rough sense of capability, easy to over-read: a strong score on a puzzle-style exam tells you little about how the model handles your messy internal documents.

25. Agent. Software that does not just answer but acts: searching, calling tools, editing files, checking its own results and looping until a goal is met. More autonomy means more value and more need for supervision, logging and a way to stop it.

How to keep up without memorising anything

You do not need all 25. Six carry most of the weight in everyday reading: token, context window, inference, parameters, temperature and hallucination. Understand those and most coverage becomes legible.

Then apply three questions to any claim you meet. What went in? What came out? How would I check it? A headline saying a model "beats" another usually means it scored higher on one benchmark on one day. A product described as "twice as fast" is fast at something specific, with a particular prompt length on particular hardware. Ask which, and you will sound like the most careful person in the room.

Keep a running note of the terms you keep seeing, with a one-line example from your own work. A definition you have attached to a real task sticks far better than one you read. And when a tool touches your health, money or legal situation, treat its output as a first draft to bring to a qualified professional, not as a decision made.

Photo: hudsoncrafted / Pixabay