AI terms

The words you meet first

Is it AI or KI?

In Norwegian you will see both. KI is the Norwegian abbreviation for kunstig intelligens and is common in the public sector and the press, while AI dominates in business and anything international. Neither is wrong, and most people use them interchangeably.

What is a language model (LLM)?

LLM stands for large language model. It is the engine behind ChatGPT, Claude, Gemini and Copilot. It has been trained on enormous amounts of text, and works out what is natural to write next – one small piece at a time. It does not look answers up in a database; it composes them afresh every time.

What is a prompt?

Simply what you write to the model – the question, the instruction or the task. The clearer you are about what you want, who it is for and what form it should take, the better the answer. «Write something about us» gives you something mediocre. «Write four lines for LinkedIn about us moving to Tøyen, informal tone» gives you something usable.

What are tokens?

The model does not read words, but small fragments of words. Those fragments are called tokens. A long word may be split into several tokens, while a short one is a single token.

This is worth knowing for one reason: price and limits are measured in tokens. If you pay for an AI tool, you usually pay per token in and out. And because the models are built mainly for English, Norwegian text tends to use more tokens than English text saying the same thing – so Norwegian costs a little more per sentence.

What is a context window?

How much the model can hold in its head at once, measured in tokens. Everything you have pasted in and everything said earlier in the conversation competes for the same space. When the window fills up, the oldest part drops out and the model «forgets» the beginning.

If a long conversation suddenly loses the thread, this is usually what has happened. The fix is to start a fresh conversation and paste the important parts back in.

Words that show up when something costs money or goes wrong

What does compute mean?

The processing power needed to run the model – the machines doing the work. When you hear that something «takes a lot of compute», it means it is expensive, power-hungry and slow.

For you as a user: this is why the most capable models cost more, queue at busy times, or are capped at a certain number of messages on the free plan.

What is inference?

The moment the model actually runs and answers you. It helps to separate two things: training is the very expensive one-off job of creating the model, while inference is the small job that happens every time somebody uses it.

So when suppliers talk about «inference costs», they mean what it costs to let people use the service, not what it cost to build it.

What is hallucination?

When the model says something that sounds right but is not – a name that does not exist, a law that was never passed, a link that goes nowhere. It is not lying deliberately; it is producing what looks like a plausible answer.

Practical rule: anything that is a fact – numbers, names, dates, prices, links, quotes – you check yourself before it goes out. Language and structure you can trust a great deal more.

What are training and fine-tuning?

Training means building the model from scratch – something very few organisations do, and which costs a fortune. Fine-tuning means taking a finished model and teaching it further on your own data, for example your tone of voice or your field.

For most businesses neither is necessary. Writing good prompts and giving the model the right documents to read will take you a long way.

Words you hear from developers and suppliers

You rarely need these yourself, but they turn up in quotes, demos, meetings – and in the news. Worth knowing what they mean.

What is an agent?

A model that does not just answer, but is allowed to do things: search the web, read files, fill in a form, send an email. It works in several steps and uses tools along the way.

The difference from an ordinary chat is that an agent can take actions – which also means it can make mistakes that have real consequences. So an agent should have clear boundaries on what it is allowed to do.

What is RAG?

Retrieval-augmented generation. Put simply: the model first looks things up in your documents, then answers based on what it found. This is how you get an assistant that can answer questions about your own routines, prices or products.

The advantage is that answers become more accountable and easier to check against a source. It is the most common way to «teach» a model your company's own content – without training anything at all.

What is a scaffold, or scaffolding?

Everything built around the model itself: the code and rules deciding what it gets to see, which tools it has, what order the steps happen in and what makes it stop.

Two products can use exactly the same model and still feel completely different – the difference usually lives here.

What is a harness?

The rig that runs the model against a pile of tasks and measures how it did. It is used for testing and comparing – for example whether a new model actually does the job better than the one you use today.

If someone says «we ran it through our own harness», it means «we tested it systematically on our own tasks».

What does multimodal mean?

That the model handles more than text – images, audio and video too. It is why you can upload a screenshot of an ad and ask for suggestions, rather than describing it in words.

What is RSI?

Recursive self-improvement. The idea that an AI system could improve itself: build a better version of itself, which builds a better version again, and so on. If each step goes faster than the last, progress could in theory run away with itself.

For now this is a hypothesis and a topic of debate – not something happening in the tools you use day to day. But the term turns up more and more in news coverage and in arguments about how fast AI is moving, so it helps to know what people actually mean by it.

Do you need to know all of this?

No. If you get through the first five – AI, language model, prompt, tokens and context window – you will understand most of what a news article or a supplier's proposal is saying. The rest you can look up the day somebody uses the word.

What matters is not knowing the vocabulary, but knowing what to check: that the facts hold up, that you know where your data ends up, and that you are not paying for something you do not need.

Need a hand?

We are happy to help you work out what is worth using, what can wait, and what is just noise. Without the jargon.

Get in touch for a friendly, no-obligation chat.