Tous les mots, expliqués simplement

Local AI

An artificial intelligence that runs on your own machine, without going through the internet.

Autrement dit — Like a pocket calculator: it sits on the table, it works even with the router unplugged, and nobody else sees what you type.

Offline

Without an internet connection. A local AI needs the internet once, to be installed.

Autrement dit — Like a series downloaded before a journey: the download needs a connection, the watching does not.

Model

The file that holds the artificial intelligence itself. It is what answers your questions.

Autrement dit — It is the brain, delivered as one big file of 1 to 5 gigabytes — the size of a couple of films.

Prompt

What you write to the AI: your question, your instruction, your request.

Autrement dit — It is the order you place. The more precise it is, the better the result — exactly as with a craftsperson.

Hallucination

When the AI invents false information and states it with confidence.

Autrement dit — It is neither a lie nor a fault: the machine looks for the most plausible continuation, not the truth. Hence rule number one: check.

Personal data

Anything that identifies you: name, address, school, photo — and also what you write.

Autrement dit — With a local AI, that data never leaves the machine. You can check it: unplug the network and it still works.

Free / open source

Software anyone can use, inspect, modify and redistribute free of charge.

Autrement dit — Like a published recipe: you can read it, cook it, adapt it, and hand it to your neighbour.

Memory (RAM)

The machine's working space. It decides how large a model you will be able to run.

Autrement dit — It is the kitchen worktop: the bigger it is, the more you can prepare at once. The hard drive is the cupboard.

Single-board computer / Raspberry Pi

A complete computer the size of a credit card, sold for €45 to €100.

Autrement dit — A bare computer: plug in a screen and a keyboard and it runs. It draws about as much as a light bulb.

Mini-PC

A small desktop computer without a screen, about the size of a book. Around €150.

Autrement dit — More powerful than a single-board computer, far more frugal than a tower. The right choice to serve a whole room.

Watt (W)

The unit of electrical consumption. A SATI kiosk draws 6 W while running.

Autrement dit — 6 W is an LED bulb. A desktop with a graphics card draws thirty times more.

Processor (CPU)

The chip that computes. On these machines the AI runs on the processor, with no graphics card.

Autrement dit — Counter-intuitive but true: what limits the speed is not raw computing power, it is how fast the machine can read its own memory.

Parameters (1B, 4B, 8B)

The number of internal settings in the model, in billions. 4B = four billion.

Autrement dit — Think of it as brain size. More of them means better answers — and a bigger machine to hold them.

Token

A fragment of a word. The AI reads and writes in tokens, not letters.

Autrement dit — Roughly three quarters of a word per token. A long technical term takes several; "the" takes one.

Throughput (tokens per second)

How fast the AI writes its answer.

Autrement dit — Below 5, you visibly wait. Between 5 and 10, it arrives at reading speed. Above 15, you stop noticing the difference.

Quantisation (Q4_K_M)

A compression of the model that divides its size by three or four, for a small loss of quality.

Autrement dit — Like MP3 against CD: far lighter, and the difference is barely audible. Q4_K_M is the default setting.

GGUF

The file format used for compressed models on small machines.

Autrement dit — Just a file format, like .mp3 for music or .pdf for a document. You never have to think about it if you use Ollama.

Context

How much text the AI keeps in mind during a conversation.

Autrement dit — Its short-term memory. A large context lets you hand it a long document, but it occupies memory permanently, even when unused.

System prompt

Permanent instructions that frame the AI's behaviour, invisible to the user.

Autrement dit — The job description: "you are a maths assistant, you never give the answer straight away". It applies to every conversation.

RAG

A method that lets the AI answer from your own documents.

Autrement dit — Nothing is retaught to the AI: the relevant passages of your documents are put under its nose at question time, like a crib sheet.

Ollama

The software that downloads models and runs them. The simplest way to start.

Autrement dit — It is the player: it handles the technical side, you only pick the model.

llama.cpp

The inference engine that actually runs the model. Ollama uses it underneath.

Autrement dit — The engine under the bonnet. You only open it when you are chasing the last few percent of speed.

Open WebUI

A chat interface that opens in the browser, with accounts and history.

Autrement dit — The chat screen, the part users actually see. Without it, everything happens in the terminal.

Terminal / command line

The black window where you type commands instead of clicking.

Autrement dit — It looks intimidating, but you only ever copy the lines given in the guides. None of them has to be invented.

Script

A file holding a series of commands, run one after another.

Autrement dit — The shopping list for the installation. SATI Installer writes it for you; you only have to read it and run it.

Service (systemd)

A program that starts with the machine and runs in the background.

Autrement dit — Like the water heater: it starts without you thinking about it, and you only notice it when it breaks.

Kiosk mode

A machine that boots straight into the application, with no desktop and no menus.

Autrement dit — Like a station ticket machine: you switch it on, it shows what it should, and there is nothing else to do.