AI 101 A Glossary

Get up to speed on AI Concepts below.

A

• Agentic AI: An evolutionary leap from passive chatbots to autonomous systems. Instead of merely generating text, agentic architectures use a self-correction loop (“reasoning”) to independently pursue complex goals, break down objectives into multi-step workflows, and deploy various software tools to execute tasks.
• Algorithm: A deterministic, step-by-step mathematical procedure or set of rules engineered to process data, calculate outcomes, or solve a specific problem.
• Artificial Intelligence (AI): The broad, multi-disciplinary field of computer science dedicated to engineering machines and software systems capable of simulating aspects of human cognitive intelligence, such as learning, reasoning, pattern recognition, and natural language processing.
• Artificial Neural Network (ANN): A computational framework directly inspired by the interconnected structure of biological brains. It passes input data through layers of artificial neurons (nodes) with adjustable mathematical weights to map complex patterns.

B

• Backpropagation: The foundational supervised learning algorithm used to train neural networks. It works by calculating the margin of error between a model’s predicted output and the actual target, then passing that error backward through the network layers to systematically tune the weights of connections.
• Bias (Algorithmic): Systematic, skewed errors introduced into an AI system because of unrepresentative training data, flawed algorithm design, or skewed human assumptions, resulting in unfair or inaccurate outputs.
• Big Data: Extremely vast, multi-structured, high-velocity datasets that are computationally impossible to process using traditional database architectures. It serves as the primary fuel for training contemporary deep learning models.

C

• Chatbot: A software interface designed to simulate human-like conversations using Natural Language Processing (NLP). Contemporary iterations rely heavily on generative foundation models rather than rigid, pre-written script trees.
• Chief AI Officer (CAIO): A corporate executive role tasked with spearheading internal AI strategy, managing technological risk, ensuring regulatory compliance, and integrating automation systems into business infrastructure.
• Computer Vision: A specialized subfield of AI focused on training machines to capture, interpret, and understand visual data from digital images, video streams, and real-world inputs (e.g., object detection, facial recognition).
• Context Window: The physical volume of input and output data (measured in tokens) that an LLM can retain in its active operational memory during a single conversational interaction.

D

• Data Sovereignty: The legal and structural requirement that data is subject to the cyber laws and governance structures of the specific nation-state where it is collected, directly impacting cloud-based AI deployments.
• Dataset: A structured collection of data systematically organized for training, validating, or evaluating an AI model.
• Deep Learning: A highly powerful subset of machine learning utilizing complex, multi-layered Artificial Neural Networks to independently extract abstract features from raw inputs without manual human feature engineering.
• Digital Twin: An AI-augmented, highly precise virtual replica of a physical asset, system, or ecosystem that utilizes real-time sensor data to simulate performance variations, forecast maintenance vulnerabilities, and optimize operational efficiency.

E

• Edge AI: The decentralized paradigm of executing machine learning models natively on local hardware devices (such as smart vehicles, smartphones, or industrial sensors) right at the point of data generation, entirely bypassing cloud-latency and data transmission requirements.
• Embeddings: High-dimensional mathematical vector coordinates assigned to words, tokens, or images. This mapping allows an AI to compute conceptual, semantic similarity based on geometric proximity within a multi-dimensional matrix.
• EU AI Act: A historic, comprehensive regulatory framework enacted by the European Union that enforces a strict, risk-based classification hierarchy on AI systems, dictating transparency, safety protocols, and steep penalties for non-compliance.
• Explainable AI (XAI): A collection of frameworks and methods engineered to illuminate the internal decision-making architecture of complex “black box” machine learning models, making their outputs comprehensible to human regulators.

F

• Feature Engineering: The explicit process of selecting, transforming, or combining raw data variables into optimized inputs to maximize the predictive accuracy of a machine learning model.
• Federated Learning: A decentralized machine learning training paradigm where a central model is trained across multiple independent edge devices holding localized data. Only the calculated weight adjustments are shared globally, preserving absolute data privacy on local hardware.
• Fine-Tuning: The process of taking an existing, heavily trained foundational model and adjusting its weights by training it on a smaller, highly domain-specific dataset to optimize it for a niche application.
• Foundation Model: A massive machine learning model trained on astronomical scales of broad, unlabelled data that can be adapted and fine-tuned to execute an immense spectrum of downstream tasks.

G

• Generative AI: A specialized branch of artificial intelligence designed to synthesize entirely original artifacts—including human-grade text, high-fidelity images, audio, or software code—by calculating the statistical probabilities of its training distributions.
• Generative Adversarial Network (GAN): A dual-model machine learning architecture where two neural networks—a Generator (creating fake data) and a Discriminator (evaluating data authenticity)—are locked in zero-sum competition to produce hyper-realistic synthetic outputs.
• Generative Engine Optimization (GEO): The specialized methodology of structuring digital content to maximize its visibility, discoverability, and inclusion within the synthesis engines of conversational AI platforms and search generative experiences.
• Gradient Descent: The primary optimization algorithm deployed to train machine learning models. It iteratively calculates the slope of the model’s loss function and tunes internal parameters to systematically reduce prediction errors.

H

• Hallucination: A systemic phenomenon where a generative model asserts an unverified, factually incorrect statement or completely fabricated piece of data with absolute mathematical confidence.
• Human-in-the-Loop (HITL): A structural operational model that explicitly integrates human validation, oversight, and feedback at critical junctures of an AI’s training, evaluation, or deployment loop to prevent errors and ensure ethical alignment.
• Hyperautomation: The systematic integration of artificial intelligence, robotic process automation (RPA), and advanced process mining to discover, audit, and automate complex corporate workflows.
• Hyperparameter: A critical configuration variable explicitly set by a developer before the model training process begins (e.g., learning rate, batch size, architectural depth) that directly dictates how the model learns.

I

• Inference: The live operational phase where an already trained machine learning model processes new, unlabelled real-world inputs to generate active predictions, classifications, or synthetic text.
• In-Context Learning: The specialized ability of modern Large Language Models to alter their behavioral output, tone, or execution logic purely based on the explicit context, examples, and framing structured within the user prompt, without permanently modifying the model’s underlying weights.

J

• Joint Embedding Predictive Architecture (JEPA): An advanced non-generative AI design framework focused on learning abstract conceptual representations of world physics rather than predicting raw pixel or word sequences, optimizing a system’s capacity for real-world planning.

K

• Knowledge Graph: A structured network representing an ecosystem of real-world entities, concepts, or events, mapping the explicit logical relationships between them to ground AI systems in verifiable fact.

L

• Large Language Model (LLM): A massive, multi-billion-parameter deep learning model built upon advanced transformer architectures, engineered specifically to comprehend, translate, summarize, and generate natural human language.
• Latent Space: An abstract, internal multi-dimensional mathematical space where a machine learning model compresses and maps data points based on their core, hidden conceptual attributes and structural patterns.
• Loss Function: A mathematical formula designed to evaluate the precise quantitative distance between an AI model’s predicted output and the true target data, functioning as the ultimate scorecard during training.

M

• Machine Learning (ML): The overarching subfield of AI dedicated to developing mathematical algorithms that enable computer systems to independently learn patterns, extract rules, and make data-driven decisions directly from experience, rather than following static, explicitly hard-coded instructions.
• Model Context Protocol (MCP): An open, universal API framework that defines how AI applications can securely, standardly connect to diverse data repositories, contextual environments, and external processing tools without bespoke integration software.
• Multimodal AI: An AI model engineered with the structural capacity to simultaneously ingest, process, align, and generate distinct sensory modalities of data—such as text, visual images, audio frequencies, and structured code strings—within a unified neural network.

N

• Natural Language Processing (NLP): A domain of computer science and AI tasked with bridging the gap between raw human linguistics and mechanical computation, enabling software to read, decipher, analyze, and generate human languages.
• Neuro-symbolic AI: A hybrid AI framework that fuses the pattern-recognition and data-processing capabilities of deep learning neural networks with the logical, rule-based reasoning of classical, symbolic symbolic computing.

O

• Overfitting: A critical machine learning error where a model learns the training dataset so perfectly that it internalizes all the random noise and statistical anomalies, rendering it incapable of accurately generalizing its predictions to new, unseen data.

P

• Parameters: The internal, adjustable mathematical variables (the weights and biases) that a neural network alters during the training process to map connections between data inputs and outputs.
• Prompt Engineering: The systematic methodology of designing, structuring, contextualizing, and refining natural language inputs to maximize the safety, accuracy, and performance of generative AI models.

Q

• Quantization: A high-end compression technique that optimizes machine learning models by reducing the numerical precision of their internal weights (e.g., converting 32-bit floating-point numbers down to 8-bit integers), dramatically reducing hardware memory footprints for local Edge AI deployments.

R

• Reinforcement Learning (RL): A machine learning paradigm where an autonomous agent learns to navigate an environment and achieve long-term objectives by executing actions and receiving direct feedback via a system of mathematical rewards or penalties.
• Reinforcement Learning from Human Feedback (RLHF): A specialized alignment training methodology that fine-tunes LLMs using direct human evaluation rankings to ensure outputs are safe, helpful, and structurally aligned with human values.
• Retrieval-Augmented Generation (RAG): An architectural framework that optimizes LLM accuracy by querying external, verified database repositories or enterprise file structures to inject current, factually precise contextual data into the prompt before generating a response.

S

• Self-Supervised Learning: A form of machine learning where a model processes unlabelled raw data by autonomously creating its own internal learning signals (e.g., systematically hiding a word in a sentence and forcing itself to predict the missing element).
• Small Language Model (SLM): Highly compact, computationally efficient language models engineered with streamlined parameter counts, allowing them to run locally on consumer notebooks or local servers under strict privacy-by-design standards.
• Supervised Learning: The most prevalent form of traditional machine learning, where a model is trained on an explicitly labeled dataset, explicitly pairing specific inputs with verified target outputs.
• Synthetic Data: Artificially generated data synthesized by an AI model that replicates the structural and statistical properties of real-world data, used to train other models without exposing sensitive human records or encountering data scarcity.

T

• Tokens: The baseline, fractional computational units (frequently corresponding to syllables, word fragments, or raw characters) into which natural language text is sliced before being converted into mathematical vectors for an LLM.
• Transformer: The historic, revolutionary deep learning architecture introduced in 2017 utilizing an internal “self-attention” mechanism to evaluate how words relate to one another across long distances within a sequence, serving as the core engine powering modern generative AI.
• Tuning: The post-training process of altering a model’s hyperparameters, structural settings, or data distributions to optimize performance across specific tasks.

U

• Unsupervised Learning: A class of machine learning where an algorithm ingests completely unlabelled data and must independently explore structural anomalies, extract patterns, or cluster the data points without human intervention.

V

• Vector Database: A specialized database architecture engineered to index, store, and execute hyper-fast mathematical similarity searches across massive repositories of high-dimensional vector embeddings, serving as the computational backbone for RAG systems.

W

• Weights: The internal, mutable mathematical values assigned to connections within a neural network layer that determine the relative importance or strength of a signal passing through the system.

X

• X-shot Learning: A general term encompassing zero-shot, one-shot, or few-shot learning, referencing a model’s ability to successfully execute a highly specific task given either zero, one, or a tiny handful of explicit reference examples inside the prompt context window.

Y

• Yottabyte Storage (for AI): The theoretical and future infrastructure planning horizons required to capture global atmospheric, historical, and sensory telemetry data arrays to drive autonomous agent training environments.

Z

• Zero-Shot Learning: The specialized ability of an advanced, pre-trained AI model to successfully complete a novel, specific task or classification challenge upon its first exposure without ever being given explicit training examples for that specific activity.

Want to Add Anything We Missed

AI is an evolving space and constantly changing with new frontiers. If you feel there’s something missing. Have your say and submit a contribution or drop a comment below.