The words, in English.
Every term says what it means and why you would care. Where we have written a fuller explainer, it points at that too, and the count above says how many of those there are. If a definition here needs another bit of jargon to make sense, we have written it badly — tell us and we will fix it.
59 terms shown.
The basics
- AI
Software that learns patterns from examples rather than following rules somebody wrote out.
Why it matters
It explains both the strength and the weakness: it is very good at things that look like its examples, and unreliable at things that do not.
- Model
The trained thing itself — the set of numbers that turns your input into an output.
Why it matters
When people compare models they are comparing the trained thing, not the app around it. The same model can feel different in two products.
- LLM (large language model)
A model trained on enormous amounts of text to predict what text comes next.
Read the full explainer →Why it matters
Almost every chatbot you have used is one. Knowing it predicts rather than looks up explains most of its behaviour.
- Training
The process of showing a model examples until it learns the patterns in them.
Read the full explainer →Why it matters
It happens once, before you ever use it, and ends on a date. That date is why it can be confidently out of step with the present.
- Training data
The examples a model learned from.
Read the full explainer →Why it matters
Whatever is thin or absent in the training data is where the model is weakest, and whatever bias is in it tends to come out the other end.
- Knowledge cut-off
The date after which the model learned nothing new.
Read the full explainer →Why it matters
Ask about anything after it and you will often get a confident answer built on the world as it was, with no warning that time has passed.
- Neural network
The mathematical structure the learning happens inside — layers of simple calculations feeding into one another.
Why it matters
Mostly you never need this word. It matters only because it explains why nobody can point to where a particular fact is stored.
- Generative AI
AI that produces something new — text, images, audio, video — rather than sorting or scoring what already exists.
Why it matters
It is the part that arrived loudly. Spam filters and recommendations were AI too, and nobody wrote headlines about them.
Using it
- Prompt
What you type or say to an AI tool.
Why it matters
The single biggest lever you have. Most disappointing answers are a vague request rather than a weak model.
- Context window
How much text the model can hold in view at once, including your conversation and anything you attached.
Read the full explainer →Why it matters
When a conversation outgrows it, the earliest parts fall away. That is the moment it starts forgetting what you told it.
- Memory
A product feature that stores notes about you between conversations and quietly puts them back.
Read the full explainer →Why it matters
It is a stored record of you, usually one you can inspect and delete. It is not the model remembering anything.
- System prompt
Hidden instructions the product gives the model before your message, setting its role and rules.
Why it matters
It is why the same model behaves differently in two apps, and why a tool sometimes refuses things for no reason you can see.
- Temperature
A setting controlling how adventurous the model's word choices are.
Why it matters
Low for anything factual or repeatable, higher for brainstorming. Most consumer apps hide it and choose for you.
- Few-shot prompting
Giving two or three examples of what you want before asking.
Why it matters
Far more effective than describing the format in words, and it takes about thirty seconds.
- Chain of thought
Asking a model to work through the steps rather than jump to an answer.
Read the full explainer →Why it matters
It genuinely improves results on multi-step problems, because each written step informs the next.
- Iteration
Correcting an answer instead of starting again.
Why it matters
The second message is where quality comes from, and almost nobody sends one.
Under the bonnet
- Parameters
The numbers inside a model that training adjusts. Often quoted in billions.
Why it matters
Loosely a measure of size. More is not automatically better, and providers have largely stopped publishing the figure.
- Weights
The trained parameter values — the model itself, as a file.
Read the full explainer →Why it matters
When a company releases the weights, anyone can run that model themselves. That is what open-weights means.
- Inference
The act of running a model to get an answer, as opposed to training it.
Read the full explainer →Why it matters
It is the cost you pay every time you use the thing, and across a model's life it usually adds up to more than training did.
- Fine-tuning
Further training of an existing model on a narrower set of examples.
Read the full explainer →Why it matters
How a general model is turned into one that knows a house style or a specialist domain.
- RLHF
Training a model using human judgements about which of two answers is better.
Read the full explainer →Why it matters
It is what makes models pleasant to talk to, and why they lean towards sounding confident and agreeing with you.
- RAG (retrieval-augmented generation)
Searching your documents first, then asking the model to answer from what was found.
Read the full explainer →Why it matters
It is how 'chat with your PDF' works, and why answers can carry a real citation beside a claim the source never made.
- Embedding
A way of turning text into numbers so that similar meanings sit close together.
Read the full explainer →Why it matters
It is what lets a search find the right passage when you used none of the same words.
- Quantisation
Compressing a model so it takes less memory, trading a little quality for a lot of size.
Read the full explainer →Why it matters
It is the reason a laptop can run a useful model at all.
- Multimodal
A model that accepts more than text — images, audio, sometimes video.
Read the full explainer →Why it matters
Very good at the gist of a picture, unreliable on small text, exact numbers and counting.
- Reasoning model
A model that writes out a long working-out before answering.
Read the full explainer →Why it matters
Better on problems with dependent steps, slower and dearer on everything else. You pay for thinking you never see.
- Open weights
A model whose trained file has been published, so anyone can download and run it.
Read the full explainer →Why it matters
Not the same as open source: the training data and code usually stay private, and licences often restrict selling.
- Diffusion
The method behind most picture and video tools: start with visual noise and clean it up over and over until it matches the description.
Read the full explainer →Why it matters
It explains why these tools are confident and vague at once. Nothing is copied and pasted; a shape is arrived at, which is also why the same words twice give you two different pictures.
Money and limits
- Token
A chunk of text, roughly three-quarters of a word in English. What models actually read and what you are billed for.
Read the full explainer →Why it matters
It explains why long documents cost more and why asking for shorter answers genuinely saves money.
- Input and output tokens
What you send and what it writes back, priced separately.
Read the full explainer →Why it matters
Output normally costs several times more, so asking for brevity is a cost decision as well as a style one.
- Rate limit
A cap on how much you can use in a period — messages an hour, files a day.
Read the full explainer →Why it matters
On a subscription this is where the cost shows up. The meter is running; you are just reading it in a different unit.
- Free tier
The version you get without paying, usually with a smaller model and daily limits.
Read the full explainer →Why it matters
Often perfectly adequate. On some consumer products it also means your text may be used for training.
- API
A way for one program to use another, including paying per token instead of monthly.
Read the full explainer →Why it matters
For spiky, irregular use it can cost far less than a subscription. For everyday use it is more fuss than it is worth.
- Credits
A prepaid allowance some products use instead of a straight subscription.
Read the full explainer →Why it matters
Watch what consumes them. Images, video and reasoning models burn through credits far faster than chat does.
Safety, law and risk
- Hallucination
A confident, fluent answer that is simply untrue.
Read the full explainer →Why it matters
It is not a glitch. It is the same process that produces correct answers, landing on something plausible but false.
- Confabulation
Another word for the same thing, preferred by people who think 'hallucination' flatters it.
Read the full explainer →Why it matters
The word matters less than the habit: check any name, number, date or citation before you rely on it.
- Sycophancy
A model's tendency to agree with you and to flatter your premise.
Read the full explainer →Why it matters
It means asking 'is this a good idea?' is a poor test. Ask for the strongest case against instead.
- Bias
Patterns in the training material coming out as unstated defaults and uneven quality.
Read the full explainer →Why it matters
It rarely looks like prejudice. It looks like an assumption you did not notice you had accepted.
- Prompt injection
Hidden instructions in a web page or document that hijack an AI tool reading it.
Read the full explainer →Why it matters
It is the main reason to be careful about letting a tool browse or read files on your behalf.
- Jailbreak
A prompt crafted to get around a model's safety rules.
Why it matters
Mostly a researcher's concern, but it explains why providers keep changing what tools will and will not do.
- Guardrails
The rules and filters a product puts around a model.
Why it matters
They are why the same model refuses something in one app and not another. They are also imperfect in both directions.
- Deepfake
A synthetic image, voice or video made to look like a real person.
Read the full explainer →Why it matters
Cheap and quick now, so recognising a face or voice is no longer proof of who you are dealing with.
- Provenance
Where a piece of content came from and how it was made.
Read the full explainer →Why it matters
For checking a suspicious image it is far more useful than examining the pixels.
- Data retention
How long a service keeps what you sent it.
Read the full explainer →Why it matters
The question that actually matters, and the one 'we do not train on your data' carefully does not answer.
- Opt-out
A setting that stops your text being used to improve the model.
Read the full explainer →Why it matters
Often exists and is off by default on free tiers. Worth five minutes in the settings of anything you use often.
- Indemnity
A provider's promise to defend you if their output leads to a legal claim.
Read the full explainer →Why it matters
Reassuring for business customers, but it is a commercial decision rather than evidence that the law is settled.
- Content Credentials
A record attached to a file saying how it was made, designed to show if it has been tampered with.
Read the full explainer →Why it matters
When it is there it tells you something useful. When it is missing it tells you nothing at all, because most genuine photographs carry none. It can confirm, never rule out.
- Watermarking
A marker hidden inside generated content so software can recognise it later.
Read the full explainer →Why it matters
Only ever applied by tools that chose to apply it, and removable by anybody who wants it gone. Worth something to a platform checking millions of files, close to nothing to you looking at one.
- Likeness
Your face or your voice, used to make something you never said or did.
Read the full explainer →Why it matters
If it happens to you the quickest route is usually the platform's own reporting form, which is obliged to act faster than anything else available. Report it to the police as well when it is sexual, or when money is involved.
Kinds of tool
- Agent
A model given tools and permission to use them in a loop until a job is done.
Read the full explainer →Why it matters
Genuinely different from a chatbot, and far less reliable than the word suggests, because errors compound across steps.
- Copilot
A tool that suggests as you work rather than doing the job for you.
Why it matters
The most useful shape for most people, because you stay in the loop by design instead of by discipline.
- Chatbot
A conversational interface to a model. It answers; it does not act.
Read the full explainer →Why it matters
Everything it produces is text handed back to you, which makes it the safest shape to start with.
- Assistant
A marketing word for almost anything in this list.
Read the full explainer →Why it matters
Ask what it can actually reach and change. That answer tells you far more than the label does.
- Local model
A model running on your own computer instead of somebody else's.
Read the full explainer →Why it matters
Your text never leaves the machine, which is a promise no privacy policy can match.
- Frontier model
Whichever models are currently the most capable available.
Why it matters
Useful shorthand, but it changes every few months and is mostly used in sentences trying to sell you something.
- Image generator
A tool that makes a picture from a description you type.
Read the full explainer →Why it matters
The good ones are now free, instant and convincing enough that a photograph on its own has stopped being evidence. That is true whether or not you ever use one.
- Voice cloning
Copying somebody's voice from a recording, then making it say anything you type.
Read the full explainer →Why it matters
A few seconds of audio is enough, and a few seconds of almost anybody is on a public profile somewhere. This is the machinery behind the phone call where a relative says they are in trouble.
- Music generator
A tool that writes and performs a song from a description.
Read the full explainer →Why it matters
Genuinely good fun for a birthday video. Less settled when it comes to what it learned from, which is being argued in court rather than agreed.
- Video generator
A tool that makes moving footage from a description or a single still image.
Read the full explainer →Why it matters
Still the easiest of the four to catch, because hands, text and gravity give it trouble. That gap is closing, and "I would be able to tell" is not a plan.
Looking for the longer version? The explainers cover the big ideas properly, and Orbit records what each actual tool does, with dates and sources.