Comprehensive AI Terminology Reference

Navigate the rapidly evolving world of artificial intelligence with our comprehensive glossary. From foundational machine learning concepts to cutting-edge 2025 research terminology, find clear definitions for over 100 essential AI terms. Includes the latest developments in agentic AI, reasoning models, and enterprise AI deployment. Hover over any term for an instant definition tooltip.

ABCDEFGHIJKLMNOPQRSTUVWXYZ
A
AGI (Artificial General Intelligence) A hypothetical form of AI that matches or exceeds human cognitive abilities across all domains, capable of understanding, learning, and applying knowledge as flexibly as humans.
Theory
A hypothetical form of AI that matches or exceeds human cognitive abilities across all domains.
Algorithm A set of rules or instructions given to an AI system to help it learn and make decisions. It's like a recipe that tells the computer how to solve problems step by step.
Fundamentals
A set of rules or instructions for solving problems or completing tasks.
Alignment The challenge of ensuring AI systems pursue intended goals and behave in accordance with human values and intentions, especially important for advanced AI systems.
Safety
Ensuring AI systems pursue intended goals and behave according to human values.
Anthropic An AI safety company founded by former OpenAI researchers, known for developing Claude AI and focusing on building safe, beneficial AI systems through constitutional AI techniques.
Company
AI safety company known for developing Claude AI and constitutional AI techniques.
API (Application Programming Interface) A set of protocols and tools that allows different software applications to communicate with AI services. For example, developers use OpenAI's API to integrate GPT models into their applications.
Technical
A set of protocols allowing software applications to communicate with AI services.
Attention Mechanism A technique that allows AI models to focus on relevant parts of input data when making predictions, crucial for understanding context in language and vision tasks.
Architecture
A technique allowing AI models to focus on relevant parts of input data.
Autoregressive Model A type of model that generates outputs sequentially, where each new element depends on previously generated elements. GPT models use this approach to generate text one token at a time.
Architecture
A model that generates outputs sequentially, with each element depending on previous ones.
Autoencoder Neural network architecture that learns efficient data representations by compressing input data to a smaller representation (encoding) and then reconstructing it (decoding). Used for dimensionality reduction, denoising, and generative modeling.
Architecture
Neural network that compresses and reconstructs data for representation learning.
AlphaFold Google DeepMind's AI system that predicts protein folding structures with unprecedented accuracy. Revolutionary breakthrough in biology, enabling drug discovery and understanding of diseases.
Application
AI system for predicting protein structures, revolutionizing biology and drug discovery. AlphaFold Database
Adapter Lightweight modules that enable efficient fine-tuning of pre-trained models for new tasks without retraining the entire model. Save computational resources while maintaining performance.
Training
Efficient method for customizing pre-trained models without complete retraining. Research Paper
AI Assistant Conversational AI systems designed to help users with various tasks through natural language interaction, from customer service to enterprise productivity tools.
Product
Conversational AI systems for helping users through natural language interaction.
Agentic AI AI systems designed to autonomously pursue complex goals and workflows with limited direct human supervision, representing a significant advancement toward more independent AI operation.
Architecture
AI systems that autonomously pursue complex goals with minimal human supervision.
Annotation The process of labeling data with additional information to help machine learning algorithms understand and learn from examples. Essential for supervised learning training datasets.
Training
Process of labeling data to help machine learning algorithms understand and learn.
Automation The use of AI and technology to perform tasks with minimal human intervention, from simple rule-based processes to complex decision-making workflows.
Application
Using AI technology to perform tasks with minimal human intervention.
B
Bias Unfair or prejudiced behavior in AI systems, often reflecting biases present in training data or design choices. Can lead to discriminatory outcomes in hiring, lending, or other applications.
Ethics
Unfair or prejudiced behavior in AI systems, often reflecting training data biases.
BERT Bidirectional Encoder Representations from Transformers. A Google-developed language model that reads text bidirectionally, improving understanding of context and meaning.
Model
Google's bidirectional language model that improved contextual understanding.
Backpropagation A training algorithm that calculates gradients backward through a neural network, allowing the system to learn by adjusting weights based on prediction errors.
Training
Algorithm for training neural networks by calculating gradients backward through layers.
Batch Size The number of training examples processed together in one forward/backward pass during model training. Larger batches can improve training stability but require more memory.
Training
Number of training examples processed together in one training iteration.
Benchmarking Process of evaluating and comparing AI models using standardized tests to measure performance, capabilities, and reliability across different tasks and datasets.
Evaluation
Standardized testing to evaluate and compare AI model performance. Papers with Code Benchmarks
C
ChatGPT OpenAI's conversational AI system based on GPT architecture, designed for interactive dialogue. It can answer questions, write content, code, and assist with various tasks through natural language.
Product
OpenAI's conversational AI system for interactive dialogue and assistance.
Claude Anthropic's family of AI assistants including Claude 3.5 Sonnet and other 2025 models, trained using constitutional AI methods to be helpful, harmless, and honest. Known for exceptional reasoning and safety features.
Product
Anthropic's AI assistant family including Claude 3.5, trained using constitutional AI methods.
Computer Vision AI field focused on enabling computers to interpret and understand visual information from the world, including images and videos. Used in facial recognition, autonomous vehicles, and medical imaging.
Field
AI field enabling computers to interpret and understand visual information.
Constitutional AI Anthropic's training method that uses a set of principles (constitution) to guide AI behavior, helping systems learn to be helpful, harmless, and honest without extensive human feedback.
Training
Training method using principles to guide AI behavior toward helpful, harmless outcomes.
CNN (Convolutional Neural Network) A deep learning architecture particularly effective for image recognition, using convolutional layers to detect features like edges, shapes, and patterns in visual data.
Architecture
Neural network architecture specialized for image recognition and computer vision.
Context Window The maximum amount of text (measured in tokens) that a language model can consider at once when generating responses. Larger context windows allow for longer conversations and documents.
Technical
Maximum amount of text a language model can consider at once.
Collective Learning AI training approach that leverages diverse skills and knowledge across multiple models to achieve more powerful and robust intelligence through collaborative learning.
Training
Training approach using diverse models collaboratively for enhanced AI intelligence.
Controllability The ability to understand, regulate, and manage an AI system's decision-making process, ensuring accuracy, safety, and ethical behavior while minimizing undesired consequences.
Safety
Ability to understand and control AI system decision-making for safety and ethics.
Conversational AI AI systems focused on natural language dialogue, enabling back-and-forth conversations through chatbots, virtual assistants, and customer service applications.
Field
AI systems designed for natural language dialogue and conversation.
D
DALL-E OpenAI's AI system that creates images from text descriptions. It can generate, edit, and create variations of images in various styles, from photorealistic to artistic.
Product
OpenAI's AI system that creates images from text descriptions.
Data Augmentation Techniques for artificially expanding training datasets by creating modified versions of existing data, such as rotating images or paraphrasing text, to improve model robustness.
Training
Techniques for artificially expanding training datasets to improve model performance.
Deep Learning A subset of machine learning using neural networks with multiple layers (deep networks) to learn complex patterns in data. Powers most modern AI breakthroughs in vision, language, and other domains.
Field
Machine learning using multi-layered neural networks to learn complex patterns.
Diffusion Model A generative AI model that creates images by gradually removing noise from random data. Used in DALL-E 2, Midjourney, and Stable Diffusion for high-quality image generation.
Architecture
Generative model that creates images by gradually removing noise from random data.
Dropout A regularization technique that randomly sets some neurons to zero during training to prevent overfitting and improve model generalization to new data.
Training
Regularization technique that randomly disables neurons during training.
Deterministic Model AI models that follow specific rules and conditions to reach definite outcomes, operating on a cause-and-effect basis with predictable results given the same inputs.
Architecture
Models that produce predictable, consistent outcomes from the same inputs.
E
Embedding Numerical representations of words, sentences, or concepts that capture semantic meaning in a high-dimensional vector space. Similar concepts have similar embeddings.
Technical
Numerical representations that capture semantic meaning of words or concepts.
Ensemble Learning A technique that combines predictions from multiple models to achieve better performance than any individual model. Common methods include bagging, boosting, and voting.
Technique
Combining predictions from multiple models to improve overall performance.
Epoch One complete pass through the entire training dataset during model training. Training typically involves multiple epochs until the model converges or performance stops improving.
Training
One complete pass through the entire training dataset during model training.
Explainable AI (XAI) AI systems designed to provide clear explanations for their decisions and predictions, making AI more transparent and trustworthy for critical applications.
Field
AI systems designed to provide clear explanations for their decisions.
Enterprise AI Strategic integration and deployment of AI within organizational frameworks to enhance business processes, decision-making, and operational efficiency at scale.
Application
Strategic integration of AI within organizations for enhanced business operations. IBM Enterprise AI
F
Few-Shot Learning AI's ability to learn new tasks with only a few examples, often achieved by providing examples in the prompt rather than extensive retraining.
Learning
Learning new tasks with only a few examples or demonstrations.
Fine-tuning The process of adapting a pre-trained model to a specific task or domain by training it further on specialized data, often with a lower learning rate.
Training
Adapting a pre-trained model to specific tasks through additional training.
Foundation Model Large-scale models trained on broad data that serve as a base for many applications. Examples include GPT, BERT, and other models that can be adapted for various tasks.
Architecture
Large-scale models that serve as a base for many AI applications.
G
GAN (Generative Adversarial Network) A framework where two neural networks compete: a generator creates fake data while a discriminator tries to detect fakes. This competition improves both networks' performance.
Architecture
Two competing neural networks that improve through adversarial training.
Gemini Google's family of multimodal AI models including Gemini 1.5 Pro, Gemini 2.0, and other 2025 releases. Features exceptionally large context windows and understands text, images, audio, and video with advanced reasoning capabilities.
Model
Google's advanced multimodal AI models including Gemini 1.5 Pro and 2.0 series (2025).
Generative AI AI systems that create new content including text, images, audio, video, or code. Examples include GPT for text, DALL-E for images, and GitHub Copilot for code.
Field
AI systems that create new content like text, images, audio, or code.
GPU (Graphics Processing Unit) Specialized processors originally designed for graphics but now essential for AI training due to their ability to perform many parallel calculations efficiently.
Hardware
Specialized processors essential for AI training and inference due to parallel processing.
GPT (Generative Pre-trained Transformer) OpenAI's family of language models including GPT-3, GPT-4, GPT-o1, and the latest 2025 releases. These autoregressive models generate human-like text and power applications like ChatGPT.
Model
OpenAI's family of generative language models including GPT-3, GPT-4, and GPT-o1 series.
Gradient Descent An optimization algorithm that finds the minimum of a function by iteratively moving in the direction of steepest descent. Fundamental to training neural networks.
Training
Optimization algorithm for finding minimum values, fundamental to neural network training.
Grounding The process of anchoring AI systems in real-world knowledge and data to improve accuracy and reduce hallucinations. Essential for factual AI applications.
Safety
Anchoring AI systems in real-world knowledge to improve accuracy and reduce hallucinations.
H
Hallucination When AI models generate false or nonsensical information that appears plausible. A major challenge in deploying language models for factual tasks.
Problem
AI generating false information that appears plausible and confident.
Human-in-the-Loop AI systems that incorporate human judgment at various stages, combining AI efficiency with human oversight for better decisions and safety.
Approach
AI systems incorporating human judgment and oversight in the decision process.
Hyperparameter Configuration settings that control the learning process but aren't learned from data, such as learning rate, batch size, or number of layers. Must be set before training.
Training
Configuration settings that control the learning process but aren't learned from data.
I
In-Context Learning The ability of language models to learn and adapt to new tasks based on examples or instructions provided in the input prompt, without updating model weights.
Learning
Learning new tasks from examples in the input prompt without model updates.
Inference The process of using a trained AI model to make predictions or generate outputs on new, unseen data. Different from training, which is the learning phase.
Process
Using a trained model to make predictions on new data.
Instruction Tuning Training language models to follow human instructions by fine-tuning on datasets of instruction-response pairs, improving their ability to understand and complete tasks.
Training
Training models to follow human instructions through instruction-response datasets.
J
Jailbreaking Attempts to bypass AI safety measures and content filters through clever prompt engineering, potentially causing models to produce harmful or inappropriate content.
Security
Bypassing AI safety measures through manipulative prompt techniques.
K
Knowledge Graph Data structures that connect information in a web of relationships, enabling AI systems to navigate and understand complex datasets through semantic connections between entities.
Technical
Data structures connecting information through relationships for AI understanding. Google Knowledge Graph
Knowledge Distillation A technique for creating smaller, faster models by training them to mimic the behavior of larger, more complex models while maintaining performance.
Technique
Creating smaller models by training them to mimic larger, complex models.
L
LLM (Large Language Model) AI models with billions of parameters trained on vast text datasets to understand and generate human-like text. Examples include GPT-4, Claude, and Gemini.
Model
AI models with billions of parameters trained to understand and generate text.
LoRA (Low-Rank Adaptation) An efficient fine-tuning technique that adapts large models by training only small additional modules rather than all parameters, reducing computational requirements.
Technique
Efficient fine-tuning method using small additional modules instead of full retraining.
Loss Function A mathematical function that measures how far a model's predictions are from the correct answers, guiding the training process to improve accuracy.
Training
Mathematical function measuring prediction errors to guide model training.
M
Machine Learning A subset of AI where computers learn to make predictions or decisions by finding patterns in data, without being explicitly programmed for each specific task.
Field
AI subset where computers learn patterns from data to make predictions.
MusicLM Google's AI system for generating high-quality music from text descriptions. Can create musical compositions in various styles, instruments, and moods based on natural language prompts.
Product
Google's AI system for generating music from text descriptions and prompts.
Midjourney Popular AI image generation service that creates artistic images from text prompts, known for its distinctive aesthetic and ease of use through Discord.
Product
AI image generation service known for artistic output and Discord interface.
Multimodal AI systems that can process and understand multiple types of data simultaneously, such as text, images, audio, and video, enabling richer interactions.
Capability
AI systems processing multiple data types like text, images, audio, and video.
N
Natural Language Processing (NLP) AI field focused on enabling computers to understand, interpret, and generate human language in text and speech form. Powers chatbots, translation, and text analysis.
Field
AI field enabling computers to understand and generate human language.
Neural Network Computing systems inspired by biological neural networks, consisting of interconnected nodes (neurons) that process information through weighted connections and activation functions.
Architecture
Computing systems inspired by biological brains using interconnected processing nodes.
NVIDIA Technology company that produces GPUs essential for AI training and inference. Their chips power most major AI systems and data centers worldwide.
Company
Technology company producing GPUs essential for AI training and inference.
O
OpenAI AI research company known for developing GPT models, ChatGPT, DALL-E, and other groundbreaking AI systems. Founded with a mission to ensure AGI benefits humanity.
Company
AI research company behind GPT, ChatGPT, and DALL-E systems.
Overfitting When a model learns training data too well, including noise and specific details, causing poor performance on new, unseen data. A common problem in machine learning.
Problem
When models perform well on training data but poorly on new data.
P
Parameter Learnable values in neural networks (weights and biases) that are adjusted during training. Model size is often measured in parameters - GPT-3 has 175 billion parameters.
Technical
Learnable values in neural networks adjusted during training.
Parameter-efficient Fine-tuning (PEFT) Training approach that adjusts only a small subset of model parameters while preserving most of the pre-trained structure, saving computational resources and time.
Training
Efficient fine-tuning by adjusting only key parameters, not the entire model. HuggingFace PEFT
Probabilistic Model AI models that make decisions based on probabilities and likelihoods rather than deterministic rules, accounting for uncertainty in predictions.
Architecture
Models that make decisions based on probabilities rather than fixed rules.
Pre-training Initial training phase where models learn general patterns from large datasets before being fine-tuned for specific tasks. Foundation for most modern AI systems.
Training
Initial training phase where models learn general patterns from large datasets.
Prompt Engineering The art and science of crafting effective inputs to get desired outputs from AI models. Involves techniques like few-shot examples, chain-of-thought, and specific formatting.
Technique
Crafting effective inputs to get desired outputs from AI models.
Q
Quantization Technique for reducing model size and increasing inference speed by using lower precision numbers (e.g., 8-bit instead of 32-bit) while maintaining acceptable performance.
Optimization
Reducing model size by using lower precision numbers while maintaining performance.
R
RAG (Retrieval-Augmented Generation) Technique combining language models with external knowledge retrieval to provide more accurate, up-to-date information by searching relevant documents before generating responses.
Technique
Combining language models with external knowledge retrieval for accuracy.
RT-2 (Robotics Transformer) Google's vision-language-action model that enables robots to understand visual scenes, interpret natural language instructions, and control robotic actions. Breakthrough in embodied AI and robotics control.
Product
Google's AI model enabling robots to understand vision, language and control actions. RT-2 Research
Reinforcement Learning Machine learning approach where agents learn through trial and error, receiving rewards or penalties for actions. Used in game AI, robotics, and ChatGPT's training.
Learning
Learning through trial and error with rewards and penalties for actions.
RLHF (Reinforcement Learning from Human Feedback) Training method using human preferences to guide AI behavior, crucial for making models helpful and aligned with human values. Used extensively in ChatGPT development.
Training
Using human preferences to guide AI behavior and alignment.
Reasoning AI's ability to solve problems, think critically, and create new knowledge by analyzing and processing available information to make well-informed decisions across various tasks.
Capability
AI's ability to solve problems and think critically through logical analysis.
Responsible AI Approach to creating and deploying AI systems with focus on positive impact, ethical considerations, trust, and minimizing potential harms to individuals and society.
Ethics
Ethical approach to AI development focused on positive impact and minimizing harm. Microsoft Responsible AI
S
Stable Diffusion Open-source text-to-image diffusion model that can run on consumer hardware, enabling widespread access to AI image generation and modification.
Model
Open-source text-to-image model that runs on consumer hardware.
SORA OpenAI's text-to-video AI model capable of generating up to 60-second videos from text descriptions. Creates realistic scenes with complex motion, multiple characters, and accurate physics.
Product
OpenAI's text-to-video model generating minute-long realistic videos from text. SORA by OpenAI
Supervised Learning Machine learning approach using labeled training data where the correct answers are provided, allowing models to learn input-output mappings for prediction tasks.
Learning
Learning from labeled training data with correct answers provided.
Steerability The ability to guide or control an AI system's behavior according to human intentions, including mechanisms to avoid unintended or undesirable outcomes.
Safety
Ability to guide AI system behavior according to human intentions and objectives.
Strong AI AI systems possessing generalized intelligence and cognitive capabilities on par with human cognition across diverse domains and tasks.
Concept
AI with generalized intelligence and human-level cognitive capabilities across domains.
Structured Data Information organized and labeled in standardized formats like databases, tables, or tagged documents, making it easier for AI systems to process and analyze.
Data
Information organized in standardized formats for easier AI processing.
Summarization AI capability to analyze large texts and produce concise, condensed versions that accurately convey core meaning and key points.
Capability
AI ability to create concise summaries that preserve core meaning from large texts.
T
Token Basic units of text that language models process, roughly equivalent to words or subwords. Models have token limits that determine maximum input/output length.
Technical
Basic units of text processed by language models, roughly equivalent to words.
Tokenization The process of breaking text into individual tokens (words or subwords) that can be processed by language models. Critical for how AI understands and processes text.
Technical
Process of breaking text into tokens for AI processing. HuggingFace Tokenizers
Transfer Learning Using knowledge gained from one task to improve performance on related tasks, typically by starting with a pre-trained model and adapting it to new domains.
Learning
Using knowledge from one task to improve performance on related tasks.
Transformer Revolutionary neural network architecture based on attention mechanisms, introduced in 2017. Foundation for most modern language models including GPT, BERT, and T5.
Architecture
Revolutionary neural architecture based on attention, foundation for modern language models.
U
Unsupervised Learning Machine learning approach that finds patterns in data without labeled examples, including techniques like clustering, dimensionality reduction, and generative modeling.
Learning
Finding patterns in data without labeled examples or correct answers.
Unstructured Data Information that lacks a predefined format or organization, such as text documents, images, videos, and audio files. More challenging for AI to process than structured data.
Data
Information without predefined format, like text, images, and videos.
V
Vector Database Specialized databases optimized for storing and searching high-dimensional vectors (embeddings), crucial for RAG systems and similarity search applications.
Infrastructure
Databases optimized for storing and searching high-dimensional vectors.
Vision Transformer (ViT) Neural network architecture that applies transformer mechanisms directly to image patches, achieving state-of-the-art results in computer vision tasks. Revolutionized visual AI by replacing convolutional approaches.
Architecture
Transformer architecture adapted for computer vision, processing image patches like text tokens.
Variational Autoencoder (VAE) Generative model that learns probabilistic representations of data, enabling sampling from learned distributions. Used for generating new images, data augmentation, and representation learning with uncertainty quantification.
Architecture
Probabilistic generative model that learns data distributions for sampling new content.
Voice Processing AI pipeline combining speech-to-text conversion and text-to-speech synthesis to enable voice-based interactions with AI systems.
Application
AI pipeline for speech-to-text and text-to-speech voice interactions.
Voice Synthesis Using AI to generate realistic, expressive computer speech by analyzing and learning from text and audio data patterns.
Application
AI generation of realistic computer speech from text and audio patterns.
W
Weight Learnable parameters in neural networks that determine the strength of connections between neurons. Training adjusts weights to minimize prediction errors.
Technical
Learnable parameters determining connection strength between neurons.
Weak AI AI systems that excel at specific tasks within limited contexts but lack generalized intelligence and adaptability outside their designed domain.
Concept
AI systems specialized for specific tasks without generalized intelligence.
Whisper OpenAI's automatic speech recognition (ASR) system that transcribes spoken language into text with high accuracy across multiple languages.
Model
OpenAI's speech recognition system for transcribing audio to text. OpenAI Whisper
Z
Zero-Shot Learning AI's ability to perform tasks it wasn't explicitly trained on, using general knowledge and reasoning capabilities without task-specific examples.
Learning
Performing tasks without specific training examples, using general knowledge.
X
X-risk (Existential Risk) The potential for highly advanced AI to pose existential threats to humanity through unintended consequences, goal misalignment, or loss of human control.
Safety
Potential existential threats from advanced AI through misalignment or loss of control.
Z
Zero-Shot Learning AI's ability to perform tasks or recognize concepts it has never been explicitly trained on, using knowledge transfer from related domains.
Learning
Performing tasks without explicit training by transferring knowledge from related domains.