50 essential terms explained simply, with concrete examples. Looking for a word? Use the search below.
Field of computer science creating systems that perform tasks normally requiring human intelligence: image recognition, language understanding, decision making.
AI model trained on massive amounts of text to understand and generate natural language. ChatGPT, Claude and Gemini are LLMs.
Hypothetical AI matching or surpassing humans on ALL cognitive tasks, not just specific ones. Does not exist yet in 2026.
Subfield of AI where machines learn from data without being explicitly programmed for each task.
Type of machine learning using deep neural networks (many layers). Behind recent AI breakthroughs.
AI capable of creating new content: text, images, video, code, audio. Includes ChatGPT, Midjourney, Sora.
Family of AI models from OpenAI. GPT-3.5 popularized ChatGPT in 2022, GPT-5 is the 2026 standard.
AI model family from Anthropic, known for quality writing, excellent code and enhanced safety.
AI model family from Google, natively multimodal, with ultra-long context (up to 1M tokens).
Model whose weights are public and downloadable. Llama (Meta), Mistral, DeepSeek are open-source.
Model handling multiple data types: text, image, audio, video. GPT-5 and Gemini are multimodal.
Maximum amount of text an LLM can "see" at once. Measured in tokens. GPT-5 = 400K, Claude = 500K, Gemini = 1M.
Basic unit of an LLM. One token ≈ 4 characters in English, 3 in French. "Hello" = 1 token.
Internal variables learned by a model. More parameters = more expressive (but also more expensive).
Phase where a model "learns" by analyzing billions of examples. Very expensive (millions of dollars for big models).
Re-training an existing model on your own data to specialize it for your domain.
Training technique where humans rate model responses to align it with their preferences.
Numerical representation of text as a vector. Enables measuring similarity between texts.
Database specialized in storing and searching embeddings. Pinecone, Qdrant, pgvector.
Neural architecture invented by Google in 2017. Foundation of all modern LLMs (GPT, Claude, Gemini).
Core mechanism of transformers: lets the model "look at" all words in context simultaneously.
First training phase where the model learns language basics on terabytes of text.
Phase where the trained model is used to produce responses. Much cheaper than training.
Text instruction given to an LLM to get a response. A good prompt = better results.
Art and science of writing effective prompts to get the best results from an LLM.
Technique of providing a few examples in the prompt to guide the LLM on the expected format.
Prompting technique asking the LLM to "reason step by step" before answering. Greatly improves results on complex problems.
Hidden instruction given to the LLM before the conversation to set its role, tone and limits.
Technique combining LLM + external knowledge base. The LLM "retrieves" relevant info before answering.
AI capable of executing actions autonomously (browsing, sending emails, making purchases), not just answering questions.
Standard protocol created by Anthropic in 2024 to let LLMs connect to external tools (DBs, APIs, files).
Ability of an LLM to call external functions (send email, query DB) as needed.
Parameter controlling an LLM's creativity. 0 = deterministic and factual, 1 = creative and variable.
When an LLM confidently makes up false information. Fundamental flaw not yet eliminated.
Attack where a user or document "hijacks" the LLM by giving it hidden instructions. #1 OWASP LLM flaw.
Technique to make an LLM say things normally forbidden (dangerous instructions, inappropriate content).
Practice of methodically attacking your own AI system to find flaws before attackers do.
Unauthorized use of personal AI tools (personal ChatGPT) with corporate data. Major enterprise risk.
AI-generated synthetic video, image or audio, imitating a real person very realistically.
European AI regulation, effective since 2024. Classifies AI by risk level and imposes obligations.
GDPR applies to data used by AI. You're responsible for client data sent to ChatGPT.
Interface allowing a program to use an AI service. OpenAI's API lets you integrate GPT into your apps.
Graphics processor. Massively used for AI as it does many parallel calculations. Nvidia dominates the market.
AI-specialized chip designed by Google. Alternative to Nvidia GPUs for training.
Technique to reduce a model's size (and run it on weaker machines) by sacrificing some precision.
Running an AI model on your own servers (instead of cloud API). More control, more complexity.
Popular tool to easily run open-source LLMs (Llama, Mistral, Gemma) on your computer.
Standardized test to measure model performance. Examples: MMLU (knowledge), SWE-Bench (code), HumanEval.
Reference platform for sharing open-source AI models. The "GitHub of AI".
Architecture where the model contains several "experts" and only activates relevant ones per query. More efficient.