AI Glossary for Business Owners

Running controlled experiments to compare two variants and measure impact.

An AI system that plans and executes steps toward a goal, often using tools autonomously.

When AI systems generate false or misleading information presented as fact. Important to understand when using AI for business content creation.

A set of rules or instructions that computers follow to solve problems or complete tasks. In AI, algorithms enable machines to learn from data and make decisions.

A way for different software applications to communicate. AI APIs allow businesses to integrate AI capabilities into their existing systems without building from scratch.

Hypothetical AI that matches or exceeds human cognitive abilities across all domains. Unlike current AI that excels at specific tasks, AGI would be truly general-purpose.

Computer systems designed to perform tasks that typically require human intelligence, such as recognizing speech, making decisions, or solving problems. In business, AI helps automate processes and gain insights from data.

A neural network component that helps models focus on the most relevant parts of the input.

Automates retrieval configuration and prompt assembly for RAG to reduce manual tuning.

A neural network that learns a compressed representation of data and then reconstructs it.

Using AI and technology to perform tasks without human intervention. Reduces manual work and increases efficiency in business processes.

Training method that adjusts neural network weights by propagating error gradients backward.

A transformer model that reads text bidirectionally for stronger language understanding.

Unfair preferences or prejudices in AI systems, often reflecting biases present in training data. Can lead to discriminatory outcomes in business applications.

Extremely large datasets that require specialized storage and processing.

A system whose internal decision-making is hard to interpret.

AI-powered software that conducts conversations with users through text or voice. Modern chatbots can handle complex customer service interactions.

A neural network architecture effective for images and other grid-like data.

AI capability that enables machines to interpret and understand visual information from images and videos, similar to human sight.

The maximum number of tokens a model can consider at once when generating an answer.

A large collection of text used to train or evaluate language models.

Verifies claims, retrieves evidence for unsupported parts, and repairs the answer.

Usage-based pricing for model inputs and outputs measured in tokens.

NVIDIA's parallel computing platform for accelerating workloads on GPUs.

Tagging data with the correct categories or attributes for training and evaluating models.

Discovering patterns and insights from large datasets.

Protecting sensitive information when using AI systems. Includes securing customer data and complying with regulations like GDPR.

The geographic location where data is stored and processed, often required by regulation or policy.

A machine learning technique using artificial neural networks with multiple layers to recognize complex patterns. It's particularly effective for processing images, speech, and text.

A generative model that learns to reverse noise to create realistic data.

In a GAN, the network that distinguishes real data from generated data.

A regularization technique that randomly disables neurons during training to prevent overfitting.

Running AI algorithms locally on devices rather than in the cloud. Provides faster responses and better privacy for sensitive business data.

Numeric vector representations of text, images, or other data that capture meaning for search and clustering.

Combining multiple models to improve accuracy and robustness.

One complete pass through the training dataset.

Development and use of AI systems that are fair, transparent, accountable, and respect human rights. Increasingly important for business reputation and compliance.

Techniques for interpreting model decisions and identifying key factors that drove an outcome.

AI systems designed to provide clear explanations for their decisions and recommendations. Critical for business applications requiring transparency.

An individual measurable variable used by a model to make predictions.

Training a model across decentralized devices or servers while keeping data local.

Guiding a model's behavior by providing a handful of examples in the prompt.

Further training a pre-trained model on your domain data to specialize behavior and tone.

Triggers retrieval mid-generation when uncertainty is detected, then continues with new context.

A large pre-trained model that can be adapted to many downstream tasks.

Allowing models to call structured tools or APIs to retrieve data or take actions.

Common privacy and security frameworks that vendors may need to comply with.

AI that creates new content such as text, images, audio, or code based on patterns learned from data.

Hardware optimized for parallel computations used to accelerate AI workloads.

An optimization algorithm that iteratively adjusts parameters to minimize error.

Uses a knowledge graph of entities and relations to guide retrieval across multi-hop questions.

Accurate reference data used as the standard for training and evaluation.

Constraining model outputs to reference specific, trusted data sources.

Policies, filters, and constraints that keep AI outputs safe, compliant, and on-brand.

A practical rule or shortcut that provides good-enough solutions.

Layers in a neural network between input and output that learn intermediate representations.

A process where humans review, correct, or guide AI outputs to ensure quality and compliance.

Combines BM25 and vector retrieval with re-ranking for robust performance.

A configuration value set before training that influences how a model learns.

Training models to mimic expert behavior from demonstrations.

Assumptions a learning algorithm makes to generalize from limited data.

Running a trained model to get outputs for a given input.

Information-focused RAG that selects diverse, complementary chunks to maximize coverage.

How understandable a model’s decisions are to humans.

A network of connected devices that collect and exchange data.

Attempts to circumvent an AI system’s safety or policy constraints.

A probability distribution over two or more variables at the same time.

A lightweight data format used for APIs and structured model outputs.

Configuring a model to return outputs in a strict schema for reliable parsing.

An interactive environment for prototyping data science and ML.

Delivering information or training precisely when needed.

An unsupervised algorithm that partitions data into k groups based on similarity.

A simple algorithm that classifies or regresses using the closest training examples.

A function that computes similarity in algorithms like SVMs.

A structured representation of entities and their relationships used for reasoning and search.

A measurable value that indicates how effectively objectives are being met.

AI systems trained on vast amounts of text data to understand and generate human-like language. These models can write, summarize, translate, and answer questions.

The time it takes for a system to return a response after receiving a request.

A compressed representation where similar inputs are close together.

Controls how much model weights change with each training step.

A metric that measures the difference between predictions and actual values during training.

A recurrent neural network architecture designed to capture long-range dependencies.

A subset of AI where computers learn patterns from data without being explicitly programmed for each task. The system improves its performance as it processes more information.

Practices and tooling for deploying, monitoring, and maintaining machine learning systems in production.

A mathematical representation that maps inputs to outputs after training.

When a model's performance degrades over time as real-world data changes.

Methods that use repeated random sampling to estimate results.

Fuses multiple retrieval channels (BM25, vector, graph) and re-rankers to pick best context.

Coordinating multiple specialized agents to collaborate on complex tasks.

Models that process and combine multiple data types like text, images, and audio.

AI technology that enables computers to understand, interpret, and generate human language. It bridges the gap between human communication and computer understanding.

Computing systems inspired by biological brain networks. They consist of interconnected nodes that process information and learn patterns, forming the basis of deep learning.

AI capability focused on deriving meaning and intent from human language.

Random variation in data that can obscure true signals.

Scaling features to standard ranges to stabilize and speed up training.

Learning to perform a task from a single example.

An AI research and product company providing widely used language model APIs.

Techniques for improving model performance or resource usage.

A data point that significantly differs from others.

When a model memorizes training data and performs poorly on new data.

A value a model learns during training, such as a weight.

Data that can identify an individual and requires careful handling and protection.

Of the items predicted positive, the fraction that are actually positive.

Using AI to analyze current and historical data to make predictions about future events. Helps businesses make data-driven decisions.

Cleaning and transforming raw data before training or inference.

The instruction or question given to a model to produce an output.

Linking multiple prompts and steps to complete a larger workflow.

The practice of crafting effective instructions or questions for AI systems to get desired outputs. Essential skill for maximizing AI tool effectiveness.

A security attack where malicious instructions are embedded in content to override or hijack model behavior.

A reusable, parameterized instruction that produces consistent outputs.

A popular open-source deep learning framework.

A reinforcement learning algorithm that learns action values to maximize reward.

Processes to test and validate AI systems for reliability and safety.

Reducing model size and speed requirements by using lower-precision numbers.

Computation based on quantum mechanics that may accelerate certain algorithms.

A request for information from a database or AI system.

Retrieval-Refinement loop that adds focused retrieval to refine an initial answer.

The maximum number of requests allowed to an API within a given time period.

Of the actual positives, the fraction correctly identified by the model.

A modeling approach to predict continuous numeric values.

Training agents to take actions in an environment to maximize reward.

Estimates answer confidence and applies risk-based policies (tighten citations, escalate).

An approach that combines information retrieval from your documents with a generative model so answers are grounded in trusted sources.

A technique that aligns model behavior using feedback from human evaluators.

A neural network designed for sequential data by sharing state across steps.

Integrating AI with physical machines to sense, plan, and act.

A metric measuring the efficiency of an investment. In AI context, it compares the cost of implementing AI solutions against the benefits gained.

A mechanism where each token attends to other tokens to capture long-range context.

Adds reflect-retrieve-revise loops: draft, detect gaps, retrieve more, and revise with citations.

Relating to the meaning of words or data rather than surface form.

A function that converts raw scores into probabilities that sum to 1.

Pipelines retrieval and generation using fast drafts to prefetch likely evidence.

Machine learning approach where the AI learns from examples with known correct answers. Like teaching with answer sheets to help the system learn patterns.

Artificially generated data used to train or evaluate models when real data is scarce or sensitive.

A hidden instruction that sets the assistant's overall behavior and constraints.

A setting that controls randomness in generated outputs; lower is more deterministic, higher is more creative.

A small chunk of text used for pricing, limits, and processing in language models.

A sampling method that limits choices to the top probability mass to control variability in outputs.

The information used to teach AI systems how to perform tasks. Quality and quantity of training data directly impacts AI performance.

Technique where AI models trained on one task are adapted for related tasks. Reduces time and resources needed for implementation.

A neural network architecture built on self-attention that underpins modern language models.

A test of whether a machine’s behavior is indistinguishable from a human’s in conversation.

A measure of confidence or variability in a model’s predictions.

When a model is too simple to capture true patterns in data.

Machine learning where AI finds hidden patterns in data without being given specific examples of what to look for. Useful for discovering unknown insights.

Increasing the resolution or quantity of data points, often for class balancing.

The visual and interactive layer where users interact with AI systems.

A generative model that learns a probabilistic latent space to create new data.

Evaluating a model on unseen data to tune hyperparameters and prevent overfitting.

The basic, unmodified version of a model or algorithm.

An array of numbers representing data in a model, often used for similarity.

A database optimized to store and search embeddings using similarity queries.

Graphical representation of data or model behavior for insights.

AI designed for specific tasks rather than general intelligence.

A parameter in a neural network that scales the influence of an input.

An automatic speech recognition model for transcribing audio.

A vector representation of a word that captures meaning and context.

A defined sequence of steps to accomplish a task, often automated with AI.

A method for setting initial neural network weights to improve training stability.

A high-performance gradient boosting framework for tabular data.

A markup language for structured data exchange.

The vertical axis in a chart used to plot values.

A comparison of a metric against the same period last year.

The output or results produced by a process; sometimes used to describe throughput or productivity.

A real-time object detection algorithm that predicts boxes and classes in one pass.

A standardized score indicating how many standard deviations a value is from the mean.

Getting models to perform tasks using clear instructions without providing examples.

A unit of digital information equal to one sextillion bytes (10^21).

An observation that word frequency is inversely proportional to rank.

An education concept applied to AI-assisted learning: tasks a user can do with guidance.