A
- AI Agent — Autonomous program capable of planning and executing actions with the help of an LLM and external tools.
- AI Alignment — Alignment between a model's behaviors and human intentions to prevent ethical issues.
- Algorithmic Bias — Distortion introduced in an AI model due to unrepresentative training data.
- API (Application Programming Interface) — Interface allowing applications to communicate with each other, often used to access external AI models.
- Artificial General Intelligence (AGI) — Hypothetical AI possessing comprehension and learning abilities equivalent to or surpassing those of humans in all domains.
C
- Chain of Thought (CoT) — Prompt technique encouraging the model to explain its reasoning step by step before giving the final answer, improving logic.
- Closed Source — Proprietary software whose code is not accessible to the public.
- Computer Vision (CV) — Field of AI enabling computers to "see" and interpret images or videos.
- Context Window — Maximum amount of text (in tokens) a model can consider at once (e.g., 128k tokens).
D
- Dataset — Set of data used to train or test an artificial intelligence model.
- Deep Learning (DL) — Subfield of machine learning using deep neural networks to recognize complex patterns in data.
- Diffusion Model — Type of model specialized in image generation (e.g., Midjourney, Stable Diffusion) by learning to remove digital noise.
E
- Embedding — Numerical representation of a word, text, or image allowing models to manipulate their meaning.
F
- Few-shot Learning — Training or adapting a model from very few examples.
- Fine-tuning — Process of retraining a model on a specific dataset to adapt it to a particular task.
G
- Generative AI (GenAI) — Branch of AI focused on creating new content (text, images, audio, code) rather than simply analyzing existing data.
- GPU (Graphics Processing Unit) — Graphics processor used to accelerate the computations needed for training and running AI models.
H
- Hallucination — False or fabricated response generated by a language model.
I
- Inference — Phase during which a pre-trained model produces predictions or text from new data.
L
- Large Language Model (LLM) — AI model trained on vast amounts of text to understand and generate natural language (e.g., GPT, Claude, Gemini).
- Latency — Waiting time between sending a request and the start of the response (or full reception).
- LLMOps — Practices and tools dedicated to the management, deployment, and supervision of large language models.
M
- Machine Learning (ML) — Branch of artificial intelligence that involves training models from data so they can make predictions or classifications.
- Mixture of Experts (MoE) — Architecture where the model is divided into several specialized subnetworks ("experts"), activated only when needed to gain efficiency.
- Multimodal — Ability of a model to process and understand multiple types of data simultaneously (text, image, sound) like GPT-4o or Gemini.
N
- Natural Language Processing (NLP) — Field of AI dedicated to interaction between computers and human language (analysis, understanding, generation).
- Neural Network — Mathematical structure inspired by the human brain, composed of layers of interconnected nodes (“neurons”), used in most modern AI models.
O
- Open Source — Software whose code is public and freely reusable or modifiable.
- Overfitting — Phenomenon where a model learns the training data too well and loses its ability to generalize.
P
Q
- Quantization — Compression technique reducing the precision of a model's weights (e.g., from 16 bits to 4 bits) to make it lighter and faster with minimal quality loss.
R
- RAG (Retrieval-Augmented Generation) — Technique combining information retrieval and generation by LLM to produce more reliable and documented answers.
- RLHF (Reinforcement Learning from Human Feedback) — Training method where human evaluations guide the model towards more appropriate responses.
T
- Temperature — Parameter controlling the creativity of the response.
- Token — Elementary unit of text (word, syllable, or character) used to process language in models like GPT.
- Tokenizer — Tool that splits raw text into small pieces (tokens) understandable by the model before processing.
- TPU (Tensor Processing Unit) — Specialized processor designed by Google for training deep learning models.
- Transformer (Transformer) — Neural network architecture introduced by Google in 2017, which serves as the basis for most current language models.
V
- Vector Database — Database optimized to store and search vectors (embeddings), essential for long-term memory and RAG.
W
- Weights (Poids) — Numerical values in neural connections that determine the strength of the transmitted signal.
Z
- Zero-shot Learning — Ability of a model to perform a task without having been explicitly trained on it.
In the member area, each term comes with quizzes, videos, and practical scenarios to move from definition to use.
Discover the AI Space Try for free