Plain-language glossary

This site promises to explain a term before using it. This page is the same promise in one place, for when you meet one of these words somewhere else — in an article, a settings menu, or a sentence somebody used to sound impressive.

Agent: Any AI feature that can take an action rather than just answer — clicking, buying, sending, booking. The word is used loosely by companies. Judge it by what it can do without you, not by what it’s called.
Cloud: Somebody else’s computer, reached over the internet. A cloud model runs on a company’s machines; a local model runs on yours.
Context window: How much text a model can hold in mind at once — your question, the conversation so far, and any document you pasted. When it fills up, the earliest parts fall out, which is why a very long conversation starts forgetting the beginning.
Embedding: A way of indexing text so a computer can find passages by meaning rather than by exact words. It’s what lets private document search pull up the right paragraph when you ask about a topic in your own words.
End-to-end encrypted: Only you and the person you’re talking to can read it — the company carrying the message cannot. Different from ordinary encryption, where the company holds the keys and can.
Fine-tuning: Taking a finished model and training it further on a narrower set of material to specialise it. Distinct from training your own model from scratch, which is enormously expensive.
Frontier model: The largest, most capable models, run on company servers. Better at genuinely hard problems than anything your laptop can run, and the only ones with a per-use cost.
GPU: The graphics chip. It happens to be very good at the arithmetic AI models need, so a computer with a decent one runs a local model considerably faster.
Hallucination: The industry’s word for the model stating something false with complete confidence. Misleading, since nothing has malfunctioned — see why it sounds certain when it’s wrong.
Inference: The model actually running and producing an answer, as opposed to being trained. When someone says inference happens on-device, they mean the answering happens on your machine.
Jailbreak: Wording designed to get a model to ignore its own restrictions. Mostly relevant to you as background for the next entry.
Local model: A model stored on your own computer that runs without an internet connection. Free to use, private by construction, and the foundation of most of this site.
Model: The trained program itself — a large file of learned patterns. The chat window is not the model; it’s the thing you type into to reach one.
Multimodal: A model that handles more than text — images, audio, video. Relevant mostly because uploading a photograph sends the whole photograph.
On-device: Processing that happens on your own phone or computer rather than on a server. A genuine privacy improvement, though a feature being on-device doesn’t mean you chose to have it.
Open weights: A model whose trained file anyone can download and run. It’s what makes free local models possible. Often loosely called open source, though the two aren’t quite the same thing.
Parameters: The number of adjustable values inside a model, usually quoted in billions, as in a 7B model. A rough guide to size: bigger generally means more capable, slower, and more memory.
Prompt: Whatever you type in. Prompt engineering just means getting better at asking.
Prompt injection: Text hidden in a web page, email or document that is written to be read as an instruction by an AI assistant rather than as information. The main reason to be careful about what an agent is allowed to touch.
Quantisation: Shrinking a model so it runs on ordinary hardware, at a small cost in quality. It’s why a model that needs a server in its full form fits on a laptop.
RAG: Retrieval-augmented generation. Ignore the name: it means the software finds the relevant passages in your own documents and shows them to the model before it answers, so the answer comes from your files rather than its memory.
System prompt: Standing instructions given to a model before your conversation starts, setting its behaviour. You usually can’t see it; with a local model you can write your own.
Telemetry: Background reports a program sends about how you use it. Not automatically sinister, and plenty of good software collects it — you just want to know it exists.
Token: A piece of a word. Roughly 130 tokens per 100 words of ordinary English. Cloud services usually charge by the token, counting both what you send and what comes back.
Training: The original process that built the model, using very large amounts of text. Separate from whether your conversations are used to improve it later, which is a setting you can usually turn off.
Weights: The actual learned numbers inside a model — the file you download when you install one locally.

A term we didn’t explain?

That’s a mistake on our part rather than a gap in your knowledge, and we’d like to fix it. Tell us which one and it will be added.