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
Kubernetes
Kubernetes is an open-source platform that automatically deploys, scales, and manages containerized applications across many servers. For AI teams, it is the machinery that keeps models running in production: if a fraud detection service at a bank suddenly receives ten times more traffic, Kubernetes spins up more copies; if a server fails, it restarts the workload elsewhere. Think of it as an operations manager for software, constantly monitoring, replacing, and rebalancing.
LangChain
LangChain is an open-source framework for building applications powered by language models. It provides modular components for chaining together prompts, retrieval systems, tools, and memory, making it easier to build complex AI workflows like RAG pipelines, autonomous agents, and multi-step reasoning systems. A developer in Casablanca can use LangChain to connect an LLM to a company's documents, database, and API in a structured way.
Large Language Model (LLM)
A large language model is a neural network trained on trillions of text tokens that can generate, understand, and manipulate human language with remarkable fluency. GPT-4, Claude, Gemini, and Llama are prominent examples. LLMs power chatbots, code assistants, translation tools, summarization, and creative writing. Their versatility comes from scale: more parameters and more data yield emergent abilities like reasoning and code generation, though they still hallucinate and require careful deployment.
Latent Space
Latent space is the compressed, abstract representation of data learned by a neural network — a lower-dimensional space where meaningful features are encoded. In a well-trained latent space, similar inputs are close together and interpolation between points produces semantically meaningful transitions. Autoencoders and VAEs learn latent spaces explicitly; GANs and diffusion models sample from them to generate new data. Latent spaces enable powerful capabilities like finding similar images, morphing between concepts, and anomaly detection.
Layer Normalization
Layer normalization is a technique that normalizes the inputs across all features within a single training example, rather than across the batch. It stabilizes and accelerates training by keeping activations at consistent scales, and is a key component of transformer architectures. Unlike batch normalization, it works identically regardless of batch size, making it ideal for the variable-length sequences processed by language models.
Leaderboard
A leaderboard is a public ranking of AI models or systems based on their performance on standardized benchmarks. Sites like the Hugging Face Open LLM Leaderboard and LM Arena rank models by scores on tasks like reasoning, coding, and language understanding. Leaderboards drive competition and transparency in AI research, helping developers and businesses compare models objectively. However, optimizing for leaderboard scores can lead to overfitting to benchmarks at the expense of real-world performance.
Learning Rate
The learning rate is a hyperparameter that controls how much a model's weights are adjusted in response to estimated error each time they are updated during training. Too high and the model diverges, never settling on a good solution; too low and training is painfully slow or gets stuck in suboptimal solutions. Finding the right learning rate is one of the most important steps in model training, often guided by learning rate schedules that start high and decay over time.
LLaMA
LLaMA (Large Language Model Meta AI) is a family of open-source large language models developed by Meta. Available in sizes from 7B to 70B+ parameters, LLaMA models can be freely used, modified, and deployed, making them a popular choice for businesses that want to run AI locally or customize models without depending on proprietary providers. LLaMA has spawned numerous fine-tuned variants like Alpaca, Vicuna, and Code Llama.
LoRA (Low-Rank Adaptation)
LoRA is a parameter-efficient fine-tuning method that freezes the original model weights and injects small trainable matrices into each layer, dramatically reducing the memory and compute needed. Instead of updating billions of parameters, you train only a few million, making custom model adaptation affordable on consumer GPUs. A Moroccan startup can fine-tune an open-source LLM on Darija customer data with a single GPU instead of an expensive cluster.