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
Anomaly Detection
Anomaly detection is the use of AI to automatically spot data points that deviate sharply from normal patterns: unusual transactions, sudden traffic spikes, or a machine vibrating abnormally. The system first learns what «normal» looks like from historical data, then raises alerts when something falls outside it. A telecom operator in Rabat can use it to catch SIM-card fraud, and a factory in Tangier to detect a production line drifting before an expensive breakdown occurs.
Application Programming Interface (API)
An API is a standardized way for one piece of software to request services from another, like a waiter carrying orders between your table and the kitchen. In AI, APIs are how businesses access powerful models without owning any infrastructure: your booking site sends a customer question to an AI provider's API and receives the answer seconds later, paying only for what you use. APIs also connect your tools together, letting your CRM, payment system, and WhatsApp channel exchange data automatically.
Artificial Intelligence (AI)
Artificial intelligence is the field of computer science dedicated to building systems that perform tasks normally requiring human intelligence, such as understanding language, recognizing images, making decisions, and learning from experience. Rather than following only fixed rules, AI systems improve by analyzing data. A bank in Casablanca, for example, can use AI to detect suspicious transactions in real time, while an e-commerce shop uses it to recommend products each customer is most likely to buy.
Attention Mechanism
The attention mechanism lets a neural network weigh which parts of its input matter most for each step of its output, instead of treating everything equally. When translating a Darija customer message into French, the model «attends» to the relevant words at each moment, like a reader whose eyes jump back to key phrases. Attention is the core innovation behind the transformer architecture, and therefore behind ChatGPT, Claude, Gemini, and virtually every modern language model.
Autoencoder
An autoencoder is a neural network trained to compress data into a compact summary and then reconstruct the original from it. By forcing information through this narrow bottleneck, the network learns the most essential features of the data. Autoencoders are used to reduce image sizes, remove noise, and detect anomalies: if the network reconstructs a transaction poorly, that transaction probably differs from anything seen in training, which makes autoencoders useful for fraud monitoring in banks and insurers.
Automated Machine Learning (AutoML)
AutoML automates the end-to-end machine learning process: algorithm selection, hyperparameter tuning, feature engineering, and data selection. It democratizes AI by letting non-experts build competitive models. Google's Vertex AI, H2O.ai, and Auto-sklearn are popular platforms. A business analyst with domain knowledge but limited ML expertise can build a churn prediction model in hours instead of weeks.
Automation
Automation is the use of technology to execute tasks with little or no human effort, from simple rules like auto-sending invoices to AI-driven processes that read documents, answer clients, and make routine decisions. The goal is not replacing people but freeing them from repetitive work: a Marrakech restaurant that automates reservation confirmations and supplier orders lets its staff focus on guests. Modern AI dramatically widens what can be automated, reaching tasks involving language, images, and judgment that once required humans.
Autonomous Agent
An autonomous agent is an AI agent that operates with minimal human intervention over extended tasks: it sets sub-goals, chooses tools, recovers from errors, and decides when the job is done. Full autonomy is powerful but risky, so in practice businesses set boundaries, budgets, and approval checkpoints. Think of an agent monitoring competitor prices for an online electronics store around the clock, adjusting listings within rules the owner defined, and only escalating unusual situations to a human.
Backpropagation
Backpropagation is the algorithm that trains neural networks by computing how much each weight contributed to the error and adjusting it accordingly, layer by layer from output back to input. It is the reason deep learning works at all: without it, networks could not learn from their mistakes. Imagine a student handing in an exam, receiving corrections, and tracing each wrong answer back to the exact concept they misunderstood, then adjusting their understanding.