Glossary
AI glossary
37 words that come up whenever people talk about AI, each explained in two sentences. Search for a term, or test yourself with "Guess the word".
Go to the quizThe definitions
37 words
- AI agentalso: agent
- A system that takes a goal and carries out a series of actions on its own (searching, reading, using tools) to reach it, instead of answering in a single message.What is an AI agent?
- AI-assisted codingalso: autocomplete, coding copilot
- Tools that help you write code in an editor: suggestions as you type, explanations, fixes or changes across several files on request.Free chapter: From chat to editor
- API
- A technical entry point that lets a program use a service, for example to send a prompt to a model and get its answer back without going through the chat interface.Coding with AI without being a developer
- Automationalso: workflow
- A chain of tasks that runs by itself according to a rule ("when this happens, do that"), with or without an AI step in the middle.Where to start
- Bias
- A model's tendency to reproduce the imbalances in its training data: stereotypes, dominant viewpoints, blind spots. Keep a critical eye on what it presents as normal.Keep a critical eye
- Chat assistantalso: chatbot, AI assistant
- A chat interface connected to a language model: you ask a question, it answers, and you can refine the answer as the conversation goes on.Free chapter: Light the lantern
- Context windowalso: context
- The amount of text the model can take into account at once: your conversation, your instructions and attached documents. Past that limit, it loses track of the earliest parts or only skims them.Summarize a long document
- Custom assistantalso: custom GPT, Gem
- A version of your assistant that you set up once for a specific task, with its mission, instructions, examples and sometimes reference files. The name and options vary by tool and sometimes by your plan.Create a tailor-made assistant
- Custom instructionsalso: system prompt, standing instructions
- Instructions you give your assistant once and for all (who you are, the tone, the format you want) and that it applies to every new conversation.Write your instructions
- Deepfake
- An image, voice or video generated to make people believe a real person said or did something. Never make one without their consent, and learn to spot them.The limits to respect
- Diffusion model
- A type of model that creates an image by starting from random noise and cleaning it up step by step, guided by your description, until a sharp image appears.Free chapter: From noise to image
- Embeddingalso: vector
- A series of numbers that represents the meaning of a text. Two texts with similar meanings have similar vectors: that's what lets you search by idea rather than by exact word.
- Few-shotalso: prompting by example
- A technique where you slip a few examples of the expected result into your prompt. The model imitates the tone, structure and length of your examples.Prompt examples
- Fine-tuning
- Extra training of a ready-made model on targeted examples, to adapt it to a specific task, style or field.
- Generative AI
- A type of AI that creates new content (text, images, sound, video, code) from an instruction, based on what it learned during training.What is generative AI?
- Hallucination
- A wrong answer given with confidence: a made-up figure, a source that doesn't exist, a distorted quote. The model generates something plausible; it doesn't check.Game: the Hallucination Hunt
- Human in the loop
- The principle of keeping a human check at the sensitive steps of an automation: before sending, paying, publishing or deleting anything.Automate without breaking everything
- Inpaintingalso: local editing
- Editing a specific area of a generated image: you select a part, describe the change, and the rest of the image stays the same.Review and fix a visual
- Knowledge cutoffalso: cutoff date
- The date at which a model's training data stops. It knows nothing about more recent events, unless it can search the web.Why AI gets things wrong
- LLMalso: large language model, language model
- A model trained on huge amounts of text to predict the most likely continuation of a text. It's the engine behind assistants like ChatGPT, Claude, Gemini or Le Chat.How it works, simply
- Multimodal
- Describes a model that understands or produces several kinds of content: text, images, audio, sometimes video. For example, you can show it a photo and ask a question about it.What is generative AI?
- Negative prompt
- A list of what you don't want to see in the image (stray text, distorted hands, blur). Some generators have a dedicated field for it; others accept it in the main prompt.The 5-layer grid
- Open modelalso: open weights, open source
- A model whose parameters are published: you can download it and run it on your own computer or server, depending on the license chosen by its authors.
- Promptalso: instruction, query
- What you write to the AI to ask it for something. The more context, goal and expected format it gives, the better the answer fits what you need.Write a good prompt
- Prompt injection
- An attack that hides instructions in content the AI reads (web page, email, document) to make it do something other than what you asked. A major risk for agents.What not to hand over to an agent
- RAGalso: retrieval-augmented generation
- A method where the tool first looks for the useful passages in your documents, then hands them to the model so it answers based on them rather than on its memory alone.Work from your documents
- Reasoning model
- A mode or model that takes time to work through intermediate steps before answering. Slower, but often more reliable on multi-step problems.Fast or reasoning, in chapter 1
- Reverse image searchalso: search by image
- A search where you give an image instead of words to find similar images on the web. It helps you check that a generated visual doesn't look too much like something that already exists.Check your logo
- Seed
- The starting number for randomness in image generation. With the same seed and the same prompt, some tools produce a very similar image: handy for tweaking a result without starting over.Generate an image with AI
- Temperature
- A setting that controls how much randomness goes into generation: low, answers are more predictable; high, they are more varied. Often only available through an API.
- Text-to-video
- Generating a short video clip from a written description, sometimes from a starting image. You then edit several clips together to tell a story.Make a video shot by shot
- Token
- A small piece of text (a word, part of a word or a symbol) that the model reads and produces one at a time. Assistants' length limits are counted in tokens, not words.The next word, token by token
- Tool callingalso: function calling, tool
- A model's ability to trigger an external action set up for it: running a search, reading a file, sending a request to a service.Free chapter: Anatomy of an agent
- Trainingalso: pre-training
- The phase where the model learns from a very large amount of data, before you ever use it. Everything it knows comes from this phase and stops at a certain date.Training and use
- Upscalingalso: enlarging
- Enlarging an image with AI, which invents the missing details to increase its resolution. Useful for printing, but check it closely.
- Vector imagealso: vector, SVG
- An image described by shapes and curves rather than pixels: it stays sharp at any size. It's the format to aim for with a logo, whereas image generators give you pixels.Create a logo with AI
- Vibe coding
- A way of programming where you describe what you want in everyday language, let the AI write the code, then test and adjust the result without necessarily reading every line.What is vibe coding?
Optional mini-quiz
Guess the word
Best: 0/10
10 random definitions, four words to choose from each time. Your best score stays in this browser, separate from your progress.