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
Image Generation
Image generation refers to AI models that create original images from text descriptions or other inputs, most often using diffusion techniques. The user types a prompt describing the desired scene, style, and details, and the model produces a new image within seconds. For example, a restaurant in Marrakech can generate appetizing visuals for its social media without hiring a photographer, while an artisan cooperative can prototype packaging designs before printing anything. Popular tools include Midjourney, DALL-E, and Stable Diffusion.
Image-to-Text
Image-to-text generates natural language descriptions from images, including image captioning, OCR, and visual question answering. Multimodal models like GPT-4V and LLaVA process visual and textual information jointly. A Moroccan retail company could use image-to-text to automatically generate product descriptions from photos.
Imbalanced Data
Imbalanced data describes a dataset where some categories are far more frequent than others, which biases a model toward the majority class. Imagine a Casablanca bank training a fraud detector: if only one transaction in a thousand is fraudulent, a lazy model can reach 99.9 percent accuracy by predicting «never fraud» and still be useless. Remedies include collecting more minority examples, oversampling, generating synthetic data, or using metrics like precision and recall instead of raw accuracy.
In-Context Learning (ICL)
In-context learning is the ability of a large language model to learn a task from examples placed directly in the prompt, without any retraining. You show the model a few input-output pairs, and it infers the pattern and applies it to new cases. For instance, an HR team can paste three examples of CVs labeled «junior» or «senior» and ask the model to classify the next hundred the same way. It turns a general-purpose model into a specialized tool in seconds.
Inference
Inference is the phase where a trained AI model is actually used to produce predictions or answers, as opposed to training, where it learns from data. Every time you ask ChatGPT a question or a bank scores a loan application, that is inference. It is like the difference between a student studying for years and finally practicing as a doctor: training happens once and costs a lot, while inference happens millions of times and must be fast and cheap. Inference cost and speed drive most real-world AI budgets.
Inpainting
Inpainting is an image editing technique where AI regenerates only a selected region of a picture while keeping the rest intact. You mask the area to change, describe what should appear there, and the model fills it in seamlessly. A Casablanca e-commerce shop can erase a distracting background object from a product photo, replace a model's outfit color, or remove a passerby from a storefront image, all without a professional retoucher.
Instruction Following
Instruction following is a model's ability to reliably execute directions given in a prompt — writing in a certain tone, returning a particular format, or performing a defined task. It distinguishes useful assistants from raw text predictors. Models are trained to follow instructions through instruction tuning and RLHF. Poor instruction following leads to off-topic answers or ignored constraints.
Instruction Tuning
Instruction tuning is a training stage where a base language model is fine-tuned on thousands of examples of instructions paired with good responses, teaching it to follow human requests instead of merely continuing text. A raw base model asked «write a welcome email» might just ramble; an instruction-tuned model actually writes the email. It is comparable to onboarding a brilliant but unstructured new employee: the raw talent exists, and instruction tuning teaches them how to respond helpfully to actual work requests.
Intelligent Document Processing (IDP)
Intelligent document processing combines OCR, language models, and automation to read, understand, and extract data from business documents such as invoices, contracts, ID cards, and forms. Unlike simple scanning, IDP interprets meaning: it finds the supplier name, the amount, and the due date even when layouts vary. A Moroccan accounting firm drowning in paper invoices can use IDP to feed its software automatically, cutting manual entry from hours to minutes.