AI Glossary
AI terms explained simply. Clear, authoritative definitions of the most important artificial intelligence concepts — from LLMs and RAG to prompt engineering and fine-tuning.
A/B Testing (AI)
A/B testing in AI is an experimental method where two or more model versions, prompts, or configurations are compared by randomly assigning real users to different variants and measuring which performs better on predefined metrics like engagement, accuracy, or user satisfaction.
Read full definition →Accuracy (AI Metric)
Accuracy is a fundamental machine learning metric that measures the percentage of correct predictions a model makes out of the total number of predictions, calculated as (correct predictions) / (total predictions).
Read full definition →Active Learning
Active learning is a machine learning approach where the model iteratively selects the most informative or uncertain data points from an unlabeled pool and requests human labels for those specific examples, maximizing learning efficiency with minimal labeled data.
Read full definition →Agentic AI
Agentic AI refers to AI systems designed to operate with a high degree of autonomy, capable of planning multi-step strategies, using external tools, making decisions, and taking actions to accomplish complex goals with minimal human intervention.
Read full definition →AI Agent
An AI agent is an autonomous artificial intelligence system that can perceive its environment, make decisions, plan multi-step actions, and execute tasks to achieve specific goals with minimal human oversight.
Read full definition →AI Aggregator
An AI aggregator is a platform that provides unified access to multiple AI models from different providers — such as OpenAI, Anthropic, Google, and Meta — through a single interface, allowing users to switch between or compare models without managing separate accounts.
Read full definition →AI Alignment
AI alignment is the field of research and engineering focused on ensuring that artificial intelligence systems understand and act in accordance with human intentions, values, and goals, rather than pursuing unintended or harmful objectives.
Read full definition →AI Benchmark
An AI benchmark is a standardized test, dataset, or evaluation framework used to measure and compare the performance of AI models on specific tasks, enabling objective assessment of model capabilities across dimensions like accuracy, reasoning, coding, and safety.
Read full definition →AI Bias
AI bias refers to systematic and unfair patterns in AI model outputs that discriminate against certain groups or individuals, typically arising from biased training data, flawed model design, or unrepresentative evaluation methods.
Read full definition →AI Governance
AI governance refers to the frameworks, policies, laws, standards, and organizational practices that guide the responsible development, deployment, and oversight of artificial intelligence systems to ensure they are safe, fair, transparent, and accountable.
Read full definition →AI Hallucination
AI hallucination occurs when an artificial intelligence model generates information that is factually incorrect, fabricated, or nonsensical, yet presents it with the same confidence as accurate information.
Read full definition →AI Model Temperature
Temperature is a parameter in AI language models that controls the randomness of the output — lower temperature values (e.g., 0.1) produce more deterministic, focused responses, while higher values (e.g., 1.0) produce more creative, varied, and sometimes unpredictable outputs.
Read full definition →AI Safety
AI safety is a multidisciplinary field focused on ensuring that artificial intelligence systems operate reliably, behave as intended, and do not cause unintended harm to individuals, organizations, or society.
Read full definition →Annotation (AI)
Annotation in AI refers to the process of adding structured metadata, labels, or markings to raw data — including bounding boxes on images, entity tags in text, or timestamps in audio — to create the labeled datasets required for training machine learning models.
Read full definition →Attention Mechanism
The attention mechanism is a neural network component that allows a model to dynamically weigh and focus on the most relevant parts of its input when producing each part of its output, enabling context-aware processing of sequential data.
Read full definition →AutoML
AutoML (Automated Machine Learning) is a set of techniques and tools that automate the end-to-end process of building machine learning models, including data preprocessing, feature engineering, model selection, hyperparameter tuning, and deployment, making ML accessible to non-experts.
Read full definition →Autoregressive Model
An autoregressive model is a type of generative model that produces output sequentially, one element at a time, where each newly generated element is conditioned on all previously generated elements, creating a chain of dependent predictions.
Read full definition →Beam Search
Beam search is a decoding algorithm used in AI text generation that maintains multiple candidate sequences (beams) at each step, expanding the most promising ones to find a higher-probability overall output than greedy decoding would produce.
Read full definition →Benchmark (AI)
An AI benchmark is a standardized evaluation framework consisting of a fixed dataset and metrics used to measure and compare the performance of different AI models on specific tasks, providing a common yardstick for progress in the field.
Read full definition →BLEU Score
BLEU (Bilingual Evaluation Understudy) score is an automatic evaluation metric that assesses the quality of machine-generated text — particularly translations — by measuring the overlap of n-grams (word sequences) between the generated output and one or more reference texts.
Read full definition →BYOK (Bring Your Own Key)
BYOK (Bring Your Own Key) is a feature offered by AI platforms that allows users to connect their own API keys from providers like OpenAI, Anthropic, or Google, using the platform's interface while paying the API provider directly at their standard rates.
Read full definition →Catastrophic Forgetting
Catastrophic forgetting is a phenomenon in neural networks where learning new information causes the model to abruptly lose or significantly degrade its performance on previously learned tasks, because new training overwrites the weights that stored earlier knowledge.
Read full definition →Chain-of-Thought (CoT)
Chain-of-thought (CoT) is a prompting technique and model capability where the AI generates intermediate reasoning steps before arriving at a final answer, significantly improving performance on tasks requiring multi-step logic, mathematics, and complex reasoning.
Read full definition →Compute (AI)
Compute in AI refers to the total processing power and computational resources required to train and run AI models, typically measured in FLOPS, GPU hours, or monetary cost, and widely considered one of the three key ingredients of AI alongside data and algorithms.
Read full definition →Computer Vision
Computer vision is a field of artificial intelligence that enables machines to interpret, analyze, and understand visual information from digital images, videos, and other visual inputs, allowing computers to derive meaningful insights from visual data.
Read full definition →Constitutional AI
Constitutional AI (CAI) is an AI training approach developed by Anthropic where a model is given a set of written principles (a 'constitution') and trained to evaluate and revise its own outputs against those principles, reducing reliance on human feedback for safety alignment.
Read full definition →Content Moderation (AI)
AI content moderation is the use of machine learning models to automatically detect, classify, and filter harmful content — including hate speech, violence, misinformation, spam, and explicit material — across text, images, video, and audio at scale.
Read full definition →Context Window
A context window is the maximum number of tokens (words and word fragments) that an AI model can process and consider at one time, encompassing both the input prompt and the generated response.
Read full definition →Contrastive Learning
Contrastive learning is a self-supervised training technique that teaches models to create useful data representations by pulling similar examples closer together and pushing dissimilar examples apart in an embedding space.
Read full definition →Convolutional Neural Network (CNN)
A Convolutional Neural Network (CNN) is a deep learning architecture specifically designed for processing grid-like data such as images, using convolutional layers that apply learnable filters to detect spatial patterns like edges, textures, and objects.
Read full definition →Corpus
A corpus (plural: corpora) is a large, structured collection of text or speech data compiled for the purpose of training language models, conducting linguistic research, or building natural language processing (NLP) systems.
Read full definition →Curriculum Learning
Curriculum learning is a training strategy where a machine learning model is exposed to training examples in a structured order — typically progressing from simple to complex — rather than randomly, improving convergence speed and final performance.
Read full definition →Data Augmentation
Data augmentation is a technique that artificially expands training datasets by creating modified versions of existing data through transformations like rotation, flipping, cropping, paraphrasing, or noise injection, improving model robustness and reducing overfitting.
Read full definition →Data Labeling
Data labeling is the process of assigning meaningful tags, categories, or annotations to raw data — such as identifying objects in images, classifying text sentiment, or transcribing audio — to create labeled datasets used for training supervised machine learning models.
Read full definition →Dataset
A dataset is a structured collection of data examples organized for a specific purpose — typically training, validating, or testing machine learning models — that can range from thousands of labeled images to billions of text tokens scraped from the internet.
Read full definition →Deep Learning
Deep learning is a subset of machine learning that uses artificial neural networks with multiple layers (deep neural networks) to automatically learn hierarchical representations of data, enabling the modeling of complex patterns without manual feature engineering.
Read full definition →Deepfake Detection
Deepfake detection is the use of AI and forensic techniques to identify synthetic or manipulated media — particularly images, videos, and audio — that have been artificially generated or altered to impersonate real people or fabricate events.
Read full definition →Diffusion Model
A diffusion model is a generative AI architecture that creates data (typically images) by learning to gradually remove noise from random static through a step-by-step denoising process, producing high-quality outputs guided by text prompts or other conditioning signals.
Read full definition →Discriminative AI
Discriminative AI refers to machine learning models that learn the boundary between different classes or categories in data, focusing on distinguishing inputs from one another rather than generating new data.
Read full definition →DPO (Direct Preference Optimization)
DPO (Direct Preference Optimization) is a training method that aligns AI language models with human preferences by directly optimizing the model on preference data, eliminating the need for a separate reward model used in traditional RLHF.
Read full definition →Edge AI
Edge AI refers to running artificial intelligence models directly on local devices — such as smartphones, cameras, sensors, and embedded systems — rather than in the cloud, enabling real-time processing, reduced latency, enhanced privacy, and offline functionality.
Read full definition →Embedding
An embedding is a dense numerical vector representation of data — such as words, sentences, images, or other objects — in a continuous vector space, where semantically similar items are positioned close together, enabling machines to understand and compute with meaning.
Read full definition →Encoder-Decoder Architecture
The encoder-decoder architecture is a neural network design pattern where an encoder processes an input sequence into a compressed internal representation, and a decoder generates an output sequence from that representation, enabling sequence-to-sequence tasks like translation and summarization.
Read full definition →Evaluation Metrics (AI)
Evaluation metrics are quantitative measurements used to assess the performance of AI models on specific tasks, providing objective criteria for model selection, comparison, and improvement across different aspects like accuracy, fluency, relevance, and safety.
Read full definition →Explainability (XAI)
Explainability (XAI) refers to the methods and techniques that make AI model decisions understandable, interpretable, and transparent to humans, enabling users, developers, and regulators to understand why a model produced a particular output.
Read full definition →F1 Score
The F1 score is a machine learning evaluation metric that combines precision and recall into a single number using their harmonic mean, providing a balanced measure of a model's accuracy that accounts for both false positives and false negatives.
Read full definition →Fairness Metrics (AI)
Fairness metrics are quantitative measurements used to evaluate whether an AI model's predictions and performance are equitable across different demographic groups, helping identify and quantify discriminatory patterns in model behavior.
Read full definition →Feature Engineering
Feature engineering is the process of using domain knowledge to select, create, and transform raw data variables into meaningful features that help machine learning models make better predictions and learn more efficiently.
Read full definition →Federated Learning
Federated learning is a distributed machine learning approach where a model is trained across multiple decentralized devices or servers, each holding local data that never leaves its source, preserving data privacy while still producing a capable global model.
Read full definition →Few-Shot Learning
Few-shot learning is an AI technique where a model can perform a new task after seeing only a few examples (typically 1-10), either through specialized training or by providing examples directly in the prompt (in-context learning).
Read full definition →Fine-Tuning
Fine-tuning is the process of taking a pre-trained AI model and further training it on a smaller, domain-specific dataset to adapt its behavior, knowledge, or style for a particular task or use case.
Read full definition →FLOPS (Floating Point Operations Per Second)
FLOPS (Floating Point Operations Per Second) is a unit of measurement for computing performance that counts how many floating-point arithmetic operations a processor can execute per second, serving as the primary metric for comparing AI training and inference hardware capabilities.
Read full definition →Foundation Model
A foundation model is a large-scale AI model trained on broad, diverse datasets that serves as a general-purpose base, capable of being adapted (through fine-tuning, prompting, or other methods) for a wide range of downstream tasks and applications.
Read full definition →Function Calling
Function calling is a capability of large language models that allows them to generate structured JSON outputs describing which external function or API to call and with what arguments, enabling AI to interact with real-world systems and data sources.
Read full definition →GAN (Generative Adversarial Network)
A GAN (Generative Adversarial Network) is a deep learning architecture consisting of two neural networks — a generator that creates synthetic data and a discriminator that evaluates whether data is real or generated — that are trained simultaneously in a competitive process to produce increasingly realistic outputs.
Read full definition →Generative AI
Generative AI refers to artificial intelligence systems that can create new content — including text, images, audio, video, code, and 3D models — by learning patterns from training data and producing novel outputs that resemble but are not copies of that data.
Read full definition →GPT (Generative Pre-trained Transformer)
GPT (Generative Pre-trained Transformer) is a family of large language models developed by OpenAI that use the transformer architecture to generate coherent, human-like text based on input prompts.
Read full definition →GPU Cluster
A GPU cluster is a high-performance computing system consisting of multiple GPUs (Graphics Processing Units) interconnected via high-speed networks, working in parallel to train and serve large AI models that require more memory and compute than any single GPU can provide.
Read full definition →Greedy Decoding
Greedy decoding is a text generation strategy where the AI model always selects the single most probable token at each step, producing deterministic output without any randomness.
Read full definition →Guardrails (AI)
AI guardrails are safety mechanisms — including input filters, output validators, topic restrictions, and behavioral constraints — implemented around AI models to prevent them from generating harmful, inappropriate, biased, or off-topic content.
Read full definition →Large Language Model (LLM)
A Large Language Model (LLM) is a type of artificial intelligence trained on vast amounts of text data that can understand, generate, and reason about human language with remarkable fluency and versatility.
Read full definition →Latent Space
Latent space is the multi-dimensional mathematical space where an AI model internally represents data in a compressed, abstract form, capturing essential features and relationships that enable the model to understand, generate, and manipulate complex data.
Read full definition →Leaderboard (AI)
An AI leaderboard is a public ranking system that orders AI models by their performance on one or more benchmarks or evaluation criteria, providing the community with a transparent comparison of model capabilities.
Read full definition →LoRA (Low-Rank Adaptation)
LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning technique that adapts pre-trained AI models by injecting small, trainable low-rank matrices into the model's layers, enabling customization without modifying the original model weights.
Read full definition →LSTM (Long Short-Term Memory)
LSTM (Long Short-Term Memory) is a specialized recurrent neural network architecture that uses gated memory cells to selectively remember or forget information over long sequences, solving the vanishing gradient problem that plagued standard RNNs.
Read full definition →Machine Learning
Machine learning (ML) is a branch of artificial intelligence where computer systems learn patterns from data and improve their performance on tasks over time without being explicitly programmed with rules for every scenario.
Read full definition →MCP (Model Context Protocol)
MCP (Model Context Protocol) is an open standard developed by Anthropic that provides a universal protocol for connecting AI models to external data sources, tools, and services, enabling structured two-way communication between AI assistants and the systems they interact with.
Read full definition →Mixture of Agents
Mixture of Agents (MoA) is an AI framework where multiple language models collaborate on tasks by generating responses independently and then having aggregator models synthesize and refine these responses into a superior final output.
Read full definition →Mixture of Experts (MoE)
Mixture of Experts (MoE) is a neural network architecture that divides the model into multiple specialized sub-networks (experts) and uses a gating mechanism (router) to selectively activate only the most relevant experts for each input, enabling massive model capacity with efficient computation.
Read full definition →MLOps
MLOps (Machine Learning Operations) is a set of practices that combines machine learning, DevOps, and data engineering to automate and streamline the entire lifecycle of ML models — from development and training to deployment, monitoring, and retraining in production.
Read full definition →MMLU (Massive Multitask Language Understanding)
MMLU (Massive Multitask Language Understanding) is a widely used AI benchmark that evaluates language models across 57 academic subjects — including STEM, humanities, social sciences, and professional fields — using multiple-choice questions to measure general knowledge and reasoning capability.
Read full definition →Model Collapse
Model collapse is a phenomenon where AI models trained on data that includes outputs from previous AI models progressively degrade in quality and diversity, losing the ability to represent the full distribution of the original training data and converging on a narrow, repetitive subset.
Read full definition →Model Quantization
Model quantization is a compression technique that reduces the size and computational cost of AI models by converting their numerical weights from high-precision formats (like 32-bit floating point) to lower-precision formats (like 8-bit or 4-bit integers), with minimal loss in quality.
Read full definition →Multimodal AI
Multimodal AI refers to artificial intelligence systems that can process, understand, and generate multiple types of data — such as text, images, audio, and video — rather than being limited to a single data type.
Read full definition →Natural Language Processing (NLP)
Natural Language Processing (NLP) is a field of artificial intelligence focused on enabling computers to understand, interpret, analyze, and generate human language in both text and speech forms.
Read full definition →Neural Architecture Search (NAS)
Neural Architecture Search (NAS) is an automated machine learning technique that uses algorithms to explore and discover optimal neural network architectures for a given task, replacing the manual trial-and-error process of designing model structures.
Read full definition →Neural Network
A neural network is a computational model inspired by the structure of the human brain, consisting of interconnected layers of nodes (neurons) that process data by learning patterns through adjusting the strength of connections (weights) during training.
Read full definition →Perplexity (Metric)
Perplexity is a language model evaluation metric that measures how well the model predicts a given sequence of text — lower perplexity indicates the model assigns higher probability to the actual text, meaning it is less 'surprised' and better at modeling the language.
Read full definition →Precision (AI Metric)
Precision is a classification metric that measures the proportion of positive predictions that are actually correct — calculated as true positives / (true positives + false positives) — indicating how trustworthy the model's positive predictions are.
Read full definition →Prompt Engineering
Prompt engineering is the practice of designing and refining text inputs (prompts) given to AI models to elicit more accurate, relevant, and useful responses.
Read full definition →RAG (Retrieval Augmented Generation)
Retrieval Augmented Generation (RAG) is an AI architecture that enhances language model responses by first retrieving relevant information from external knowledge sources, then using that context to generate more accurate and grounded answers.
Read full definition →Recall (AI Metric)
Recall (also called sensitivity) is a classification metric that measures the proportion of actual positive cases that the model correctly identifies — calculated as true positives / (true positives + false negatives) — indicating how completely the model detects positive instances.
Read full definition →Recurrent Neural Network (RNN)
A Recurrent Neural Network (RNN) is a neural network architecture designed for sequential data that processes inputs one step at a time while maintaining a hidden state that carries information from previous steps, enabling it to model temporal patterns and dependencies.
Read full definition →Red Teaming (AI)
Red teaming in AI is the practice of deliberately and systematically attempting to make an AI model produce harmful, incorrect, or unintended outputs through adversarial testing, in order to identify and fix vulnerabilities before the model is deployed to users.
Read full definition →Reinforcement Learning
Reinforcement learning (RL) is a machine learning paradigm where an agent learns to make optimal decisions by interacting with an environment, receiving rewards for desirable actions and penalties for undesirable ones, gradually improving its strategy over time.
Read full definition →RLHF (Reinforcement Learning from Human Feedback)
RLHF (Reinforcement Learning from Human Feedback) is an AI training technique where human evaluators rank model outputs by quality, and those rankings are used to train a reward model that guides the AI toward generating more helpful, accurate, and safe responses.
Read full definition →Self-Supervised Learning
Self-supervised learning is a machine learning approach where models generate their own training labels from the structure of unlabeled data, enabling them to learn rich representations from massive datasets without human annotation.
Read full definition →Semantic Search
Semantic search is an information retrieval approach that uses AI to understand the meaning and intent behind a query, returning results based on conceptual relevance rather than exact keyword matches.
Read full definition →Semi-Supervised Learning
Semi-supervised learning is a machine learning approach that trains models using a small set of labeled data combined with a large set of unlabeled data, leveraging the structure of the unlabeled data to improve performance beyond what labeled data alone could achieve.
Read full definition →SHAP (SHapley Additive exPlanations)
SHAP (SHapley Additive exPlanations) is an explainability method based on cooperative game theory that assigns each input feature a contribution value (Shapley value) for a specific prediction, providing a mathematically principled way to understand how each feature influences the model's output.
Read full definition →Small Language Model (SLM)
A Small Language Model (SLM) is a language model with a relatively modest parameter count (typically under 10 billion parameters) that is optimized for efficiency, enabling deployment on consumer hardware, mobile devices, and edge environments while maintaining practical capabilities for common tasks.
Read full definition →Speech-to-Text (STT)
Speech-to-text (STT), also known as automatic speech recognition (ASR), is AI technology that converts spoken language from audio input into written text, enabling machines to understand and process human speech.
Read full definition →Supervised Learning
Supervised learning is a machine learning paradigm where a model is trained on a dataset of input-output pairs (labeled data), learning to map inputs to correct outputs so it can make accurate predictions on new, unseen data.
Read full definition →Synthetic Data
Synthetic data is artificially generated data that statistically mimics the properties and patterns of real-world data, created using AI models, simulations, or rule-based systems to augment or replace real training data.
Read full definition →Text-to-Image AI
Text-to-image AI is a type of generative AI that creates images from natural language descriptions (text prompts), using deep learning models trained on large datasets of image-text pairs.
Read full definition →Text-to-Speech (TTS)
Text-to-speech (TTS) is an AI technology that converts written text into natural-sounding spoken audio, using deep learning models that can replicate human-like intonation, emotion, and speaking patterns.
Read full definition →Token (in AI)
In AI, a token is the fundamental unit of text that a language model processes — typically a word, subword, or character — and the total number of tokens determines input limits, output length, and API pricing.
Read full definition →Tokenization
Tokenization is the process of converting raw text into a sequence of tokens — discrete units like words, subwords, or characters — that can be processed by an AI model, using algorithms like Byte Pair Encoding (BPE) or SentencePiece.
Read full definition →Tool Use / Function Calling
Tool use (also called function calling) is an AI capability that allows language models to invoke external tools, APIs, and functions by generating structured requests, enabling them to take real-world actions, access current data, and perform computations beyond their native text generation abilities.
Read full definition →Top-P Sampling (Nucleus Sampling)
Top-P sampling (also called nucleus sampling) is a text generation strategy where the AI model considers only the smallest set of most probable next tokens whose cumulative probability adds up to a threshold P, then randomly samples from that set.
Read full definition →TPU (Tensor Processing Unit)
A TPU (Tensor Processing Unit) is a custom-designed AI accelerator chip developed by Google specifically for machine learning workloads, optimized for the matrix and tensor operations that form the core computations in neural network training and inference.
Read full definition →Training Data
Training data is the collection of examples — including text, images, audio, or structured data — that machine learning models learn from during the training process, forming the basis for all the patterns, knowledge, and capabilities the model develops.
Read full definition →Transfer Learning
Transfer learning is a machine learning technique where a model trained on a large general dataset is adapted to perform well on a different, often more specific task, leveraging the knowledge already learned to achieve better results with less data and training time.
Read full definition →Transformer
The Transformer is a neural network architecture introduced in 2017 that uses self-attention mechanisms to process sequential data in parallel, forming the foundation of virtually all modern large language models including GPT, Claude, Gemini, and LLaMA.
Read full definition →Tree-of-Thought (ToT)
Tree-of-Thought (ToT) is an advanced reasoning framework that extends chain-of-thought by having the AI explore multiple reasoning paths simultaneously, evaluate each path's promise, and select or combine the most productive branches to solve complex problems.
Read full definition →VAE (Variational Autoencoder)
A VAE (Variational Autoencoder) is a generative model that learns to encode input data into a structured latent space distribution and decode samples from that space back into realistic data, enabling both data compression and generation of new samples.
Read full definition →Vector Database
A vector database is a specialized database that stores data as high-dimensional mathematical vectors (embeddings), enabling fast similarity-based search that powers AI applications like semantic search, recommendations, and retrieval augmented generation.
Read full definition →Vision Transformer (ViT)
A Vision Transformer (ViT) is a model architecture that applies the Transformer's self-attention mechanism to image recognition by dividing an image into fixed-size patches, treating each patch as a token, and processing them through standard Transformer layers.
Read full definition →Try These AI Concepts in Action
Explore 400+ AI models, RAG chatbots, fine-tuning, and more on Vincony — the unified AI platform. Free to start with 100 credits/month.