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
Loss Function
A loss function measures the discrepancy between a model's predictions and the true target values during training. It provides the optimization signal: gradient descent adjusts weights to minimize this loss. Common loss functions include cross-entropy for classification tasks and mean squared error for regression. Choosing the right loss function is critical — it directly shapes what the model learns to optimize.
Long Short-Term Memory (LSTM)
LSTM is a type of recurrent neural network (RNN) architecture designed to learn long-term dependencies in sequential data by using gating mechanisms — input, forget, and output gates — that control what information to keep, discard, or pass forward. Before transformers dominated NLP, LSTMs were the state of the art for speech recognition, machine translation, and time-series forecasting. They still appear in production systems where computational budget is limited.
Machine Learning
Machine learning is the subset of artificial intelligence where systems learn patterns from data to make predictions or decisions without being explicitly programmed. A spam filter learns from millions of emails which ones are junk; a recommendation engine learns from browsing history what a customer might buy next. The three main branches are supervised learning (with labels), unsupervised learning (finding structure), and reinforcement learning (learning from rewards).
Mistral
Mistral is a family of open-weight large language models developed by Mistral AI, a French company founded in 2023. Known for punching above their weight class, Mistral models deliver performance competitive with much larger proprietary models while being smaller and more efficient. Mistral 7B, Mixtral 8x7B, and Mistral Large are prominent releases. They can be self-hosted, fine-tuned, and deployed freely, making them popular for European and privacy-conscious deployments.
Mixture of Experts (MoE)
A Mixture of Experts is a neural network architecture that divides its parameters into specialized sub-networks (experts), with a gating mechanism that routes each input to only a subset of relevant experts. This allows the model to have many total parameters while only activating a fraction per input, achieving large-model capability with lower compute cost. Mixtral 8x7B and GPT-4 are believed to use MoE architectures. The trade-off is more complex infrastructure and potential load-balancing challenges.
MLOps
MLOps (Machine Learning Operations) is a set of practices that combines machine learning, DevOps, and data engineering to reliably deploy, monitor, and maintain ML models in production. It covers the full lifecycle: data pipelines, model training, versioning, deployment, monitoring for drift, and retraining. MLOps ensures models don't degrade silently after deployment — a critical concern as AI becomes embedded in business-critical workflows.
Model Auditing
Model auditing is the systematic evaluation of an AI model's behavior, performance, and compliance with ethical and regulatory standards. Audits examine bias across demographic groups, explainability of decisions, robustness to adversarial inputs, and adherence to privacy regulations. Third-party audits are increasingly required before deploying high-risk AI systems.
Model Card
A model card is a standardized document accompanying a trained AI model that describes its intended use, training data, performance metrics across different demographic groups, limitations, ethical considerations, and recommended applications. Inspired by nutrition labels, model cards promote transparency and responsible AI deployment. Google, OpenAI, and Meta publish model cards for their major releases. They help users understand whether a model is appropriate for their specific use case and potential biases.
Model Catalog
A model catalog is a centralized repository or registry where organizations track, manage, and serve their trained AI models. It records metadata like model version, training data used, performance metrics, deployment status, and lineage. Platforms like MLflow, Weights & Biases, and Hugging Face Hub provide model catalog functionality. A model catalog is essential for governance: it ensures teams know which model version is deployed in production and can roll back if needed.