The 212AY Library
AI
Glossary.
200+ artificial intelligence terms explained in plain language — from machine learning and LLMs to AI agents, RAG and prompt engineering. Every definition comes with a concrete example, in English, French and Arabic.
200 terms
Reasoning Model
A reasoning model is a language model trained to «think before answering»: it spends extra computation producing internal chains of thought, exploring options and checking its own work before giving a final response. This makes it markedly better at math, coding, planning, and multi-step business logic, at the cost of higher latency and price. A consulting firm in Tunis might use a reasoning model to analyze a complex tender document.
Recommendation System
A recommendation system predicts what a user is likely to want next, based on their past behavior and on patterns learned from similar users or items. It is the engine behind «customers also bought» on e-commerce sites and personalized playlists on streaming apps. Techniques range from collaborative filtering to deep learning on embeddings. A Casablanca online marketplace can lift average basket size by recommending complementary products.
Recurrent Neural Network (RNN)
A recurrent neural network processes sequences one element at a time while keeping a hidden memory of what came before, like reading a sentence word by word and remembering the context. RNNs and their improved variant LSTM long dominated speech recognition, translation, and time-series forecasting. They struggle with very long sequences and slow training, which is why transformers largely replaced them. RNNs remain useful for lightweight tasks.
Red Teaming
Red teaming is the practice of deliberately attacking an AI system to find its weaknesses before real adversaries or accidents do. Testers try jailbreaks, prompt injections, biased outputs, data leaks, and harmful content generation, then report findings so developers can strengthen defenses. Before a Moroccan bank launches a customer chatbot, a red team would probe whether it can be tricked into revealing account data.
Regression
Regression is a family of supervised learning methods that predict a continuous numerical value rather than a category: a price, a temperature, a quantity. The model learns the relationship between input features and the target from historical examples. Linear regression fits a straight line; more advanced variants capture complex patterns. A real-estate agency in Tangier could use regression to estimate apartment prices from surface area, neighborhood, and floor.
Regularization
Regularization is a set of techniques that prevent a model from memorizing its training data instead of learning general patterns, a failure called overfitting. Common methods penalize overly large weights (L1, L2), randomly disable neurons during training (dropout), or stop training early. Think of a student encouraged to understand concepts rather than memorize past exams: they perform better on new questions.
Reinforcement Learning (RL)
Reinforcement learning trains an agent through trial and error: it takes actions in an environment, receives rewards or penalties, and gradually learns a strategy that maximizes long-term reward. It is how a child learns to ride a bicycle, adjusting after every wobble. RL powers game-playing systems, robotics, and dynamic decision-making like ad bidding or energy management.
Replit
Replit is a browser-based development platform where you can write, run, and deploy code without installing anything. Its AI agent can build entire applications from a plain-language description, handling code, database, and hosting in one place. This makes it popular for prototyping and for non-developers turning ideas into working tools.
Reranking
Reranking is a second-pass step in search and RAG pipelines: after a fast retriever fetches, say, fifty candidate documents, a more powerful but slower model re-scores them and pushes the truly relevant ones to the top. Adding a reranker often dramatically improves answer quality in enterprise chatbots, for example when a Moroccan insurer's assistant searches thousands of policy documents.