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
Function Calling
Function calling lets modern LLMs output structured JSON describing which external function to call and with what arguments, rather than generating free-form text. The application executes the function — querying a database, calling an API — and feeds results back. This bridges language understanding and real-world actions, enabling AI agents to interact with software tools reliably.
Generative AI
Generative AI refers to artificial intelligence systems that create new content — text, images, code, music, or video — rather than merely analyzing or classifying existing data. Large language models like ChatGPT generate text; diffusion models like Midjourney create images; code-generation tools like GitHub Copilot write software. Generative AI is transforming creative work, customer service, software development, and education, though it raises important questions about originality, copyright, and misinformation.
GPT (Generative Pre-trained Transformer)
GPT is a family of large language models by OpenAI using the transformer architecture to generate human-like text. From GPT-1 in 2018 to GPT-4 and beyond, each generation brought dramatic improvements in reasoning and coding. GPT models power ChatGPT and are accessed via API by millions of developers. The series demonstrated that scaling model size and data produces emergent capabilities.
Graphics Processing Unit (GPU)
A GPU is a specialized processor originally designed for rendering graphics but now essential for AI because it can perform thousands of mathematical operations in parallel, exactly what neural network training and inference require. Training a large model on CPUs alone could take months; the same model on GPUs finishes in days or hours. NVIDIA dominates the AI GPU market with its CUDA ecosystem, and cloud providers like AWS and Google Cloud rent GPU capacity by the hour.
Gradient Descent
Gradient descent is the optimization algorithm that minimizes a model's loss function by iteratively adjusting weights in the direction that reduces error the most. Think of a hiker in foggy mountains trying to reach the lowest valley: at each step, they feel the slope beneath their feet and walk downhill. The learning rate controls step size. Variants like stochastic gradient descent (SGD) and Adam adapt the process for efficiency. All neural network training ultimately relies on gradient descent.
Grounding
Grounding is the technique of anchoring a language model's responses in verifiable, real-world data rather than relying solely on its training knowledge. RAG is the most common grounding method: the model retrieves relevant documents before answering. Grounding reduces hallucinations, keeps answers current, and provides citations. A Moroccan legal assistant grounded in the official journal can quote exact articles, while an ungrounded one might invent plausible-sounding but wrong legal references.
Hallucination
A hallucination is when a language model generates text that is fluent and plausible but factually incorrect or fabricated — stating things that never happened, citing sources that do not exist, or mixing up details. Hallucinations occur because models optimize for likely next tokens, not truth. A legal assistant hallucinating case law could get a firm in serious trouble. Mitigation strategies include RAG grounding, fact-checking layers, temperature reduction, and human review of high-stakes outputs.
Hyperparameter Tuning
Hyperparameter tuning is the process of selecting the optimal configuration settings for a machine learning model before training begins — settings like learning rate, batch size, number of layers, and dropout rate. Unlike model parameters that the algorithm learns automatically, hyperparameters must be set by the practitioner. Grid search, random search, and Bayesian optimization are common methods. Finding good hyperparameters can mean the difference between a model that barely works and one that excels.
Image Classification
Image classification is the task of assigning a label or category to an entire image based on its visual content. A model trained on thousands of labeled photos can learn to distinguish cats from dogs, defective products from good ones, or healthy X-rays from diseased ones. Convolutional neural networks (CNNs) and vision transformers are the standard architectures. Image classification powers quality inspection in factories, medical diagnosis assistance, and content moderation on social platforms.