Glossary41 terms
Artificial Intelligence Glossary
AI terminology can seem intimidating. Here are more than 40 definitions in plain language, in alphabetical order. The links take you to the guide where each concept is explained in detail.
A
- AGI (artificial general intelligence)
- A hypothetical AI able to perform any intellectual task with the flexibility of a human. It does not exist today; experts disagree on both its definition and its timeline. See the guide →
- AI Act
- The European Union’s regulation on artificial intelligence, which entered into force on August 1, 2024. It classifies AI uses by level of risk and imposes proportionate rules. See the guide →
- AI agent
- An AI system that can chain together several actions (searching, writing, running code, using apps) to reach a goal, with more or less human supervision. See the guide →
- Algorithmic bias
- The tendency of an AI system to produce results that are systematically unfavorable to certain groups, often because its training data is unbalanced. See the guide →
- Alignment
- The set of techniques aimed at making an AI behave in line with human intentions and values: helpful, honest and harmless. See the guide →
- API
- An interface that lets one piece of software use an AI model remotely. Developers use it to build AI into their applications. See the guide →
- Attention (mechanism)
- The mechanism at the heart of Transformers: to process a word, the model weighs the importance of all the other words in the context. See the guide →
B
- Benchmark
- A standardized test used to compare the performance of AI models (math, code, general knowledge…). See the guide →
C
- Chatbot (conversational agent)
- A program that holds a dialogue with the user. The first well-known one, ELIZA, dates from 1966; today’s assistants are built on large language models. See the guide →
- Context window
- The maximum amount of text (in tokens) a model can take into account at once: your question, the conversation history and the answer being written. See the guide →
D
- Deep learning
- Machine learning with neural networks that have many layers. It is the basis of the major advances in AI since 2012. See the guide →
- Deepfake
- An image, video or voice generated or altered by AI to realistically imitate a real person. See the guide →
- Diffusion model
- A generative model that learns to remove noise from an image, then creates by starting from random noise and gradually denoising it. It underpins most of today’s image generators. See the guide →
E
- Embedding
- A representation of a word, image or document as a list of numbers, so that items with similar meanings end up close together in numerical space. See the guide →
F
- Fine-tuning
- A step that consists of retraining an already pre-trained model on more specific data to adapt it to a task or a style. See the guide →
G
- GAN (generative adversarial network)
- A generative model proposed in 2014 by Ian Goodfellow: two networks, a generator and a discriminator, compete to produce realistic images. See the guide →
- Generative AI
- AI that produces new content (text, images, sound, video, code) instead of only classifying or predicting. See the guide →
- GPU (graphics processing unit)
- A processor able to perform thousands of calculations in parallel, originally designed for video games and now essential for training neural networks. See the guide →
- Gradient descent
- An algorithm that adjusts a model’s parameters in small steps in the direction that reduces error, like walking down a mountain in the fog. See the guide →
H
- Hallucination
- A false or made-up answer, delivered fluently and convincingly by a language model. See the guide →
I
- Inference
- Using an already trained model to produce an answer to a request. See the guide →
L
- LLM (large language model)
- A neural network trained on enormous amounts of text to predict what comes next in a piece of text. ChatGPT, Claude and Gemini are built on LLMs. See the guide →
- Loss function
- A number that measures how wrong a model is. Training seeks to reduce it. See the guide →
M
- Machine learning
- A family of techniques in which a machine learns from examples instead of being programmed rule by rule. See the guide →
- Multimodal
- Describes a model able to process several types of data: text, images, sound, even video. See the guide →
N
- Neural network
- A set of small computing units (artificial neurons) organized in layers and connected by weights, which learn to recognize patterns. See the guide →
O
- Open-weight model
- A model whose parameters are published, so you can download it and run it yourself. The training data is not always shared. See the guide →
- Overfitting
- A flaw in which a model memorizes its training examples “by heart” and generalizes poorly to new data. See the guide →
P
- Parameter
- A number (a weight or a bias) inside a model, adjusted during training. GPT-3 has 175 billion of them. See the guide →
- Pre-training
- The first and heaviest training phase of a large model on a vast corpus, which gives it its general knowledge. See the guide →
- Prompt
- The instruction or question sent to an AI. Its clarity strongly influences the quality of the answer. See the guide →
R
- RAG (retrieval-augmented generation)
- A technique in which the AI first looks up relevant documents and then uses them to write its answer, which reduces errors and makes it possible to cite sources. See the guide →
- Reinforcement learning
- A method in which an agent learns by trial and error, guided by rewards. It allowed AlphaGo to beat a Go champion in 2016. See the guide →
- RLHF (reinforcement learning from human feedback)
- A step in which humans rank an AI’s answers to steer it toward more helpful and safer responses. See the guide →
S
- Supervised learning
- A method in which the model learns from examples that come with the right answer (for instance, an image labeled “cat”). See the guide →
T
- Token
- A small piece of text (a word, part of a word, punctuation) converted into a number so a language model can process it. See the guide →
- Training
- The phase in which a model adjusts its parameters based on data. It is long and expensive for large models. See the guide →
- Training data
- The set of examples (text, images, sounds…) a model learns from. Their quality and diversity largely determine the quality of the model. See the guide →
- Transformer
- A neural network architecture published by Google in 2017 (“Attention Is All You Need”), which underlies most large language models. See the guide →
- Turing test
- A test proposed by Alan Turing in 1950: a machine passes if a person talking with it cannot tell it apart from a human. See the guide →
U
- Unsupervised learning
- A method in which the model looks for structure on its own in unlabeled data, for example to group similar customers together. See the guide →
No terms match your search.