Glossary
AI words, in plain language.
40 terms explained in one or two sentences, linked to the guide that goes deeper.
- AgentUse
- An AI system that acts rather than just answers: it plans steps, uses tools (browser, files, apps) and checks results until the task is done. Learn more →
- AI ActRules
- EU Regulation 2024/1689 on artificial intelligence. It classifies systems by risk level and sets different duties for those who build and those who use them. Learn more →
- AI literacyRules
- The duty (AI Act Art. 4) to take measures so staff using AI understand how it works, its limits and risks. Since Regulation 2026/1744 a "sufficient level" no longer has to be guaranteed, but the duty still applies to everyone. Learn more →
- APITechnical
- The interface through which a program uses an AI model. You pay per use, by tokens read and written, rather than by subscription.
- BenchmarkBasics
- A standard test to measure a model (maths, code, reasoning). Useful for comparison, but it does not tell you how it will do on YOUR task. Learn more →
- Chain of thoughtTechnical
- "Thinking out loud": the model writes intermediate steps before answering. Reasoning models do it automatically, often hidden. Learn more →
- Context windowBasics
- How much text the model can "see" at once: the conversation, uploaded documents, instructions. Top models now reach about 1 million tokens. Learn more →
- Deep researchUse
- A mode where the AI browses dozens of sources autonomously for several minutes and produces a cited report. Learn more →
- DeepfakeRules
- Generated or manipulated image, audio or video that looks authentic. The AI Act requires disclosure; explicit sexual deepfakes are banned since 2026.
- DeployerRules
- In AI Act terms, whoever uses an AI system in a professional activity (a company, a firm). Different duties from the "provider" who builds it. Learn more →
- DiffusionTechnical
- The technique behind most image and video generators: the model starts from random noise and cleans it step by step into an image matching the description. Learn more →
- EmbeddingTechnical
- A list of numbers representing the meaning of a word or text. Similar texts have nearby embeddings: that is how "semantic" search works. Learn more →
- Fine-tuningTechnical
- Further training a ready model on your own examples to specialise it. Today good context or a Skill is often enough.
- Frontier modelBasics
- The most capable models available at a given time. As of October 2026: Claude Opus 5.5, GPT-6 Astra and Sol, Gemini 4 Argon, Muse Wonder.
- GPUTechnical
- The graphics processor used to train and run models: it does thousands of multiplications in parallel. Why NVIDIA is worth trillions.
- HallucinationBasics
- When a model produces false information stated confidently: a non-existent ruling, an invented number, a quote never said. Learn more →
- HarnessTechnical
- The scaffolding around a model: the program that gives it tools, memory and a work loop. The same model performs much better in a good harness. Learn more →
- InferenceTechnical
- Using an already-trained model to generate answers. What you pay for when you use an API or subscription.
- Intelligence IndexBasics
- Artificial Analysis's composite score combining several independent benchmarks. As of October 2026 the top is 58 (Claude Opus 5.5). Learn more →
- LLMBasics
- Large Language Model: a neural network trained on huge amounts of text to predict the next word. Learn more →
- MCPUse
- Model Context Protocol: an open standard to connect an AI assistant to external tools and data (Drive, calendar, ERP) without custom integrations. Learn more →
- Mixture of Experts (MoE)Technical
- A model split into many "experts", only some active for each word. Labels like "35B A3B" mean 35 billion parameters in total, 3 active at a time: lots of knowledge, small-model speed. Learn more →
- MultimodalBasics
- A model that understands (or produces) several content types: text, images, audio, video, PDF.
- Open weightsBasics
- Models whose weights can be downloaded: you can run them on your own servers without sending data to third parties. E.g. GLM-5.3, Kimi K3, DeepSeek V4, MiMo. Learn more →
- ParametersTechnical
- The numbers tuned during training: the model's statistical memory. Frontier models have hundreds of billions to trillions. Learn more →
- PhishingRules
- Messages imitating banks, suppliers or colleagues to make you click, pay or reveal credentials. With AI they are error-free: the signal is the request, not the wording. Learn more →
- Pre-trainingTechnical
- The first training phase: the model reads a vast share of the web and books, learning to predict the next token. Costs hundreds of millions. Learn more →
- PromptUse
- The instruction you give the model. A good prompt states role, goal, context, constraints and output format. Learn more →
- Prompt injectionRules
- An attack where instructions hidden in a page, email or document hijack the AI reading them. The main risk for agents. Learn more →
- QuantisationTechnical
- Reducing the precision of a model's numbers (e.g. to 4 bits) so it fits in less memory and runs faster, with a small quality loss. It's what lets open models run on an ordinary computer. Learn more →
- RAGTechnical
- Retrieval-Augmented Generation: before answering, the system retrieves relevant passages from your documents into the context. This site's chatbot works this way. Learn more →
- Reasoning modelBasics
- A model that thinks before answering, generating intermediate steps. More effort = better answers on hard problems, but more time and cost. Learn more →
- RLHFTechnical
- Reinforcement Learning from Human Feedback: the phase where the model learns from human preferences to be helpful, clear and safe. Learn more →
- SkillUse
- A bundle of instructions and files (procedures, templates, scripts) the assistant loads only when needed. It teaches it "how you work". Learn more →
- TemperatureTechnical
- A setting controlling randomness when picking the next word: low = predictable, high = more varied and creative. Learn more →
- TokenBasics
- The chunk of text a model works with: a short word, part of a word, a symbol. In English 100 words ≈ 130 tokens. Billing is per token. Learn more →
- TransformerTechnical
- The neural network architecture introduced by Google in 2017 ("Attention Is All You Need") behind every modern LLM. Learn more →
- Vibe codingUse
- Building software by describing what you want and letting AI write the code. Great for prototypes; production needs review. Learn more →
- Voice cloningRules
- Creating a synthetic copy of a voice from a few minutes (or seconds) of recording. Useful for dubbing and content, but it requires the person's consent and is used in scams. Learn more →
- WatermarkRules
- A (often invisible) mark identifying content as AI-generated. The AI Act requires it from 2 December 2026.
