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
A/B Testing
A/B testing (also called split testing) is a statistical method for comparing two versions of something — a model, a prompt, a UI element, or a business process — to determine which performs better. Users or data points are randomly divided into two groups, each exposed to one version, and the difference in outcomes is measured. In AI, A/B testing might compare two prompt templates, two model versions, or a human-written response against an AI-generated one. It provides empirical evidence for decision-making rather than relying on intuition.
Accuracy
Accuracy is the simplest way to measure a model's performance: the percentage of predictions it gets right out of all predictions made. If a model reviewing 1000 loan applications classifies 930 correctly, its accuracy is 93 percent. Beware, though: accuracy can mislead on imbalanced data. If only 2 percent of transactions at a Moroccan bank are fraudulent, a lazy model that flags nothing is 98 percent accurate yet completely useless, which is why metrics like F1-score exist.
Activation Function
An activation function is a mathematical operation inside a neural network that decides how strongly each artificial neuron responds to its inputs, introducing the non-linearity that lets networks learn complex patterns. Without it, a deep network would behave like simple linear math and could never distinguish, say, photos of authentic zellige tiles from imitations. Common examples include ReLU, sigmoid, and softmax. Think of it as a dimmer switch controlling how much signal each neuron passes forward.
Adversarial Attack
An adversarial attack is a deliberate attempt to fool an AI system with carefully crafted inputs that look normal to humans but cause the model to fail. A few altered pixels can make an image classifier mislabel a stop sign, and cleverly worded messages can trick a chatbot into ignoring its safety rules. For a bank whose fraud filter or a retailer whose product-recognition system faces motivated attackers, testing against such manipulations, known as red teaming, is essential before deployment.
Agent Memory
Agent memory is the set of mechanisms that let an AI agent remember information beyond a single conversation: past decisions, user preferences, learned facts, and progress on long tasks. Short-term memory lives in the model's context window, while long-term memory is typically stored in files or vector databases and retrieved when relevant. It is what allows a customer-service agent to recall that a client in Fès already reported a delivery issue last month and respond accordingly.
Agentic AI
Agentic AI describes systems designed to act, not just respond: they plan multi-step work, use tools, check their own results, and adapt when something fails. It marks the shift from AI as a smart assistant you prompt repeatedly to AI as a worker you delegate to. For example, instead of asking a chatbot ten questions about unpaid invoices, an agentic system retrieves the invoices from your accounting tool, sends polite reminders, and reports back with a summary.
Agentic Workflow
An agentic workflow is a business process in which one or more AI agents execute the steps: gathering information, making decisions, calling software tools, and handing off to humans only when judgment or approval is needed. Consider hiring at a growing Marrakech hotel group: an agentic workflow can screen CVs, rank candidates against criteria, schedule interviews via calendar tools, and send confirmations, while HR staff simply review the shortlist and conduct the actual interviews.
Artificial General Intelligence (AGI)
Artificial general intelligence refers to a hypothetical AI system able to understand, learn, and perform any intellectual task a human can, across all domains, rather than excelling at one narrow job. Today's tools remain specialized: a model that writes marketing copy cannot also manage a warehouse on its own. AGI would be like a brilliant employee who can move from accounting to logistics to customer service without retraining. It remains a research goal, not a shipped product.
EU AI Act
The EU AI Act is the European Union's landmark regulation governing artificial intelligence, the first comprehensive AI law in the world. It classifies systems by risk level: unacceptable uses like social scoring are banned, high-risk uses such as hiring or credit scoring face strict requirements for testing, documentation, and human oversight, and lighter obligations apply elsewhere. Moroccan and African companies should care because the law applies to anyone selling AI-powered products or services into the European market.