What This AI Glossary Covers
This AI glossary is a plain-English reference to the vocabulary that actually turns up in AI work right now: 120+ terms covering agents, models, prompting, retrieval and the tooling built around them. Every entry is one short definition, written so you can use the term in a meeting the same day you read it.
The list runs A to Z, so you can scan it like a dictionary or search the page for the one term you came for. Because the field moves quickly, this AI glossary gets updated as the language shifts. The newest additions include context engineering, Model Context Protocol, vibe coding, reasoning models and evals, for example, none of which were everyday vocabulary two years ago.
How To Use This AI Glossary
If you are new to the field, start with the foundational entries in this AI glossary, such as artificial intelligence, machine learning, large language model, prompt engineering and hallucination. If you already build with these tools, however, read context engineering, MCP, agentic RAG, structured outputs and evals first, because those five describe how AI systems get assembled today.
Entries sit under their first letter rather than by theme, so related ideas sometimes land far apart in this AI glossary. Where one definition leans on another, the text names it, so you can jump across and fill the gap.
A
Adapter
Adapters are an advanced method for making pre-trained AI models adaptable to new tasks without complete retraining. These modules save time, money, and resources by efficiently repurposing existing models for different tasks in areas like natural language processing, computer vision, and robotics.
Agentic AI
Agentic AI refers to AI systems designed to autonomously pursue complex goals and workflows with limited direct human supervision.
Agentic RAG and GraphRAG
Agentic RAG and GraphRAG move retrieval past naive chunk-and-search. Rather than pulling a handful of semantically similar passages, the system traverses a knowledge graph or runs multi-step retrieval, and judges at each step whether the evidence it has gathered is actually enough to answer the question.
AI Assistant
An AI assistant is a conversational interface that uses large language models to support users in various tasks and decision-making processes across multiple domains within an enterprise environment.
AI Plugin
AI plugins are specialized software components that allow AI systems to interface with external applications and services.
AI Search
AI search allows users to find information using natural language queries instead of keyword-based searches.
Annotation
Annotation is the process of labeling data with additional information to help machine learning algorithms understand and learn.
Artificial General Intelligence (AGI)
Artificial General Intelligence (AGI) refers to an AI system that possesses a wide range of cognitive abilities, much like humans, enabling them to learn, reason, adapt to new situations, and devise creative solutions across various tasks and domains, rather than being limited to specific tasks as narrow AI systems are.
Artificial Intelligence (AI)
The simulation of human intelligence in machines that are programmed to think and learn like humans. Example: A self-driving car that can navigate and make decisions on its own using AI technology.
Associative Memory
Associative memory refers to a system’s ability to store, retrieve, and process related information based on connections between elements, enabling it to efficiently identify and use relevant data for decision-making.
Attention Mechanism
Attention mechanisms are AI techniques that enable models to focus on the most relevant parts of input data when processing information.
Automatic Speech Recognition (ASR)
Automatic speech recognition (ASR) is a technology that transcribes spoken language into text.
Automation
Automation refers to the use of technology to perform tasks with minimal human intervention.
B
Benchmarking
Benchmarking is the process of evaluating and comparing products or systems using standardized tests to gauge performance and capabilities.
C
Chain-of-Thought (CoT) and Test-Time Compute (TTC)
Chain-of-thought prompting asks a model to reason step by step before committing to an answer. Test-time compute is the broader idea behind it: letting a reasoning model spend more tokens thinking at inference time to improve accuracy on hard problems, instead of scaling the model itself. The approach was formalized in the original chain-of-thought research paper.
Chatbot
A user-friendly interface that allows the user to ask questions and receive answers. Depending on the backend system, it can range from basic pre-written responses to a fully conversational AI that automates issue resolution.
ChatGPT
A chat interface built on top of GPT-3.5 — a large language model developed by OpenAI that is trained on a massive amount of internet text data and fine-tuned to perform a wide range of natural language tasks such as language translation, text summarization, and question answering.
CLI Agents
CLI agents are terminal-first autonomous coding agents that run locally with shell access. They execute test suites, read build logs, debug failures and commit git diffs on their own, rather than handing you code to paste somewhere.
Codebase Indexing and AST-Aware Retrieval
Instead of embedding a repository as plain text chunks, AST-aware retrieval feeds the model the structure of the code itself: the abstract syntax tree, call graphs and module boundaries. The agent gets the shape of the codebase, not just text that happens to look similar.
Collective Learning
Collective learning is an AI training approach that leverages diverse skills and knowledge across multiple models to achieve more powerful and robust intelligence.
Context Engineering
Context engineering is the successor to prompt engineering. Prompt engineering asks how to phrase a request; context engineering asks what the model needs to see in order to answer it well. That covers chunking, retrieval design, tool schemas, prompt caching and selective memory injection across the whole context window. We broke the shift down in Prompting For AI Is Changing.
Context Window Management
Context window management is the practice of deciding what stays in the context window and what gets dropped, summarized or routed elsewhere. The named patterns are context routing, which splits work across agents, isolation, which gives each agent only its own slice, and reduction, which compresses history as a session grows.
Controllability
Controllability is the ability to understand, regulate, and manage an AI system’s decision-making process, ensuring its accuracy, safety, and ethical behavior, and minimizing the potential for undesired consequences.
Conversational AI
A subfield of AI that focuses on developing systems that can understand and generate human-like language and conduct a back-and-forth conversation. Example: A chatbot that can understand and respond to customer inquiries in a natural and human-like manner.
Cost of Large Language Models
The cost of large language models primarily stems from their size and complexity, which demand significant computational power, storage, and resources for training and deployment. These factors can result in substantial expenses for building, maintaining, and using such models, sometimes amounting to several dollars per conversation or thousands of dollars per month.
D
Data Augmentation
Data Augmentation is a technique used to artificially increase the size and diversity of a training set by creating modified copies of the existing data. It involves making minor changes such as flipping, resizing, or adjusting brightness of images to enhance the dataset and prevent models from overfitting.
Deep Learning
A subfield of ML that uses neural networks with multiple layers to learn from data. Example: A deep learning model that can recognize objects in an image by processing the image through multiple layers of neural networks.
Deterministic Model
A deterministic model follows a specific set of rules and conditions to reach a definite outcome, operating on a cause-and-effect basis.
Direct Preference Optimization (DPO) and RLVR
DPO is an alignment method that tunes a model directly on pairs of preferred and rejected responses, skipping the separate reward model that traditional RLHF needs. RLVR, or reinforcement learning with verifiable rewards, trains on tasks where the answer can be checked automatically, such as math problems or passing unit tests.
Discriminative Model
Discriminative models are algorithms designed to directly model and learn the boundary between different classes or categories in a dataset.
E
Enterprise AI
Enterprise AI refers to the strategic integration and deployment of AI within an organizational framework to enhance various business processes, decision-making, and overall operational efficiency.
Explainability
Explainability refers to techniques that make AI model decisions and predictions interpretable and understandable to humans.
Extensibility
Extensibility in AI refers to the ability of AI systems to expand their capabilities to new domains, tasks, and datasets without needing full retraining or major architectural changes.
Extraction
Extraction is the ability of generative models to analyze large datasets and extract relevant patterns, trends, and specific pieces of information.
F
Few-Shot Learning
Few-shot learning is a machine learning approach where models can learn concepts from just a few labeled examples, often 5 or less per category.
Fine Tuning
The process of adapting a pre-trained model to a specific task by training it on a smaller dataset. For example, an image classification model trained on all intersection pictures can be fine-tuned to detect when a car runs a red light.
Foundation Model
Foundation models are a broad category of AI models which include large language models and other types of models such as computer vision and reinforcement learning models. They are called “foundation” models because they serve as the base upon which applications can be built, catering to a wide range of domains and use cases.
Full-Stack Artifacts and Ephemeral Sandboxes
These are instant, throwaway execution environments, usually micro-VMs or in-browser WebContainers, that spin up to run and preview AI-generated code in real time and then disappear. They are what makes describe an app and watch it run possible in a chat window.
G
Generation
Generation is the ability of a generative model to create brand new, original content such as text, images, audio, or video from scratch.
Generative Adversarial Networks (GANs)
GANs are a powerful type of neural network capable of generating new, never-seen-before data that closely resembles the training data.
Generative AI
Generative AI models create new data by discovering patterns in data inputs or training data. For example, creating an original short story based on analyzing existing, published short stories.
Generative Pre-Trained Transformer (GPT)
Generative pre-trained transformers (GPT) are neural network models trained on large datasets in an unsupervised manner to generate text.
GPT-3
GPT-3 is the 3rd version of the GPT-n series of models. It has 175 billion parameters with weights to make predictions. ChatGPT uses GPT-3.5, which is another iteration of this model.
GPT-4
GPT-4 is the latest model addition to OpenAI’s deep learning efforts and is a significant milestone in scaling deep learning. GPT-4 is also the first of the GPT models that is a large multimodal model, meaning it accepts both image and text inputs and emits text outputs.
Grounding
Grounding is the process of anchoring AI systems in real-world experiences, knowledge, or data to improve the AI’s understanding of the world, so it can effectively interpret and respond to user inputs, queries, and tasks in a context-aware way.
H
Hallucination
Hallucination refers to a situation wherein an AI system, especially one dealing with natural language processing, generates outputs that may be irrelevant, nonsensical, or incorrect based on the input provided. This often occurs when the AI is unsure of the context or lacks a proper understanding of the subject matter.
Human-in-the-Loop (HITL) and Guardrails
HITL and guardrails are the policy layer wrapped around an autonomous agent: allow-lists, approval gates and confirmation steps placed in front of high-impact actions such as financial transactions, shell commands or database writes.
I
Instruction-tuning
Instruction-tuning is an approach where a pre-trained model is adapted to perform specific tasks by providing a set of guidelines or directives that outline the desired operation.
Intelligence Amplification
Intelligence amplification refers to empowering human capabilities through synergistic combinations of AI systems and traditional tools.
Intelligence Augmentation
Intelligence augmentation refers to enhancing human capabilities through synergistic combinations of AI systems and traditional tools.
Interpretability
Interpretability refers to how inherently understandable or explainable an AI model is based on its architecture, logic, and behavior.
K
K-Shot Learning
K-shot learning is a machine learning approach where models learn from only k labeled examples per class, where k is a small number like 1–5.
Knowledge Generation
Knowledge generation involves training models on extensive datasets, allowing them to analyze data, discover patterns, and craft new insights.
Knowledge Graph
Knowledge graphs are data structures that connect information in a web of relationships, allowing AI systems to navigate and understand complex datasets.
L
Large Language Model (LLM)
A type of deep learning model trained on a large dataset to perform natural language understanding and generation tasks. Famous LLMs include BERT, PaLM, GPT-2, GPT-3, GPT-3.5, and GPT-4. These models vary in size, breadth of tasks, and what they’re trained on.
Latency
Latency refers to the time delay between when an AI system receives an input and generates the corresponding output.
LLM Evals and Synthetic Evaluation
Evals are automated test suites for AI output. They measure faithfulness, hallucination rate and tool-call accuracy using LLM-as-a-judge scoring, unit tests and red-teaming, so prompts get tuned against a fixed set of real cases rather than by impression.
Long-Term Memory and Reflection
Memory layers such as vector stores, graph memory and episodic memory that let an agent recall past decisions and carry them across sessions. Reflection is the companion step, where the agent reviews what it did and updates its approach before the next attempt.
Low-code
Low-code is a visual approach to software development that enables faster delivery of applications through minimal hand-coding.
M
Machine Learning (ML)
A subfield of AI that involves the development of algorithms and statistical models that enable machines to improve their performance with experience. Example: A machine learning algorithm that can predict which customers are most likely to churn based on their past behavior.
Model Chaining
Model chaining is a technique in data science where multiple machine learning models are linked in a sequence to make predictions or analyses.
Model Context Protocol (MCP)
MCP is an open protocol that standardizes how large language models and autonomous agents connect to external tools, local file systems, APIs and databases. It replaces one-off custom integrations with a single interface any compliant client can speak. The specification is published at modelcontextprotocol.io.
Multi-Agent Systems (MAS) and Swarms
Architectures in which several specialized agents, typically a planner, a coder, a tester and a reviewer, collaborate through an orchestration layer to finish a workflow that one agent could not hold in a single context window.
Multi-hop Reasoning
Multi-hop reasoning refers to the process by which an AI model retrieves answers to questions by connecting multiple pieces of information across various sources and systems, rather than directly extracting information from a single passage.
Multimodal Language Model
Multimodal Language Models are a type of deep learning model trained on large datasets of both textual and non-textual data.
N
N-Shot Learning
Zero/Single/Few shot learning are variations of the same concept — providing a model with little or no training data to classify new data and guide predictions. A “shot” represents a single training example.
Natural Language Ambiguity
Natural language ambiguity refers to situations where a word, phrase, or sentence can have multiple meanings, making it challenging for both humans and AI systems to interpret correctly.
Natural Language Generation (NLG)
A subfield of AI that produces natural written or spoken language.
Natural Language Processing (NLP)
A subfield of AI that involves programming computers to process massive volumes of language data. Focuses on transforming free-form text into a standardized structure.
Natural Language Understanding (NLU)
A subtopic of NLP that analyzes text to glean semantic meaning from written language, including understanding context, sentiment, and intent.
Neural Network
A machine learning model inspired by the human brain’s structure and function, composed of layers of interconnected nodes or “neurons.” Example: A neural network that can recognize handwritten digits with high accuracy.
No-code
No-code is an approach to designing and using applications that doesn’t require any coding or knowledge of programming languages.
O
Observability and Tracing
Observability for AI systems means monitoring multi-step agent traces in production: where latency piles up, how many tokens each step burns, which tool calls fail, and at what point a run went off course.
OpenAI
The organization that developed ChatGPT. More broadly, OpenAI is a research company that aims to develop and promote friendly AI responsibly. OpenAI’s GPT-3 model is one of the largest and most powerful language models available for natural language processing tasks.
Optimization
The process of adjusting the parameters of a model to minimize a loss function that measures the difference between the model’s predictions and the true values. Example: Optimizing a neural network’s parameters using a gradient descent algorithm.
Overfitting
A problem that occurs when a model is too complex, performing well on the training data but poorly on unseen data. Example: A model that has memorized the training data instead of learning general patterns.
P
Parameter-efficient Fine-tuning (PEFT)
PEFT is an approach that helps improve the performance of large AI models while optimizing for resources like time, energy, and computational power, by adjusting a small number of key parameters while preserving most of the pretrained model’s structure.
Pre-training
Training a model on a large dataset before fine-tuning it to a specific task. Example: Pre-training a language model on a large corpus of text data before fine-tuning it for a specific natural language task such as language translation.
Probabilistic Model
A probabilistic AI model makes decisions based on probabilities or likelihoods.
Prompt Caching
Prompt caching lets a model provider store the static prefix of a request, such as a long system prompt, a codebase index or a set of API definitions, so it is not reprocessed on every call. On repetitive workloads it cuts both latency and cost substantially. There is more on keeping usage down in How To Prevent Maxing Your AI Tokens While Coding.
Prompt Engineering
Identifying inputs — prompts — that result in meaningful outputs from large language models. Prompt engineering is essential for guiding LLM behavior and includes providing templates and wizards to direct applications.
Prompt Injection
Prompt injection is an attack in which instructions hidden inside content the model reads, such as a web page, a document or an email, get treated as commands. Your prompt is effectively a code injection surface, and the problem has not been fully solved.
R
ReAct and Plan-and-Execute Loops
Agentic design patterns that interleave reasoning, action and observation. The model thinks, calls a tool, reads the result and corrects course, repeating until the goal is met rather than answering in one shot. The pattern comes from the ReAct paper.
Reasoning
AI reasoning is the process by which AI systems solve problems, think critically, and create new knowledge by analyzing and processing available information, allowing them to make well-informed decisions across various tasks and domains.
Reasoning Models
Foundation models trained with reinforcement learning on verifiable rewards so they break a problem down internally before answering. They spend extra thinking tokens at inference, trading speed and cost for accuracy on hard tasks.
Recursive Prompting
Recursive prompting is a strategy for guiding AI models to produce higher-quality output. It involves providing the model with a series of prompts or questions that build upon previous responses, refining both the context and the AI’s understanding to achieve the desired result.
Reinforcement Learning
A type of machine learning in which a model learns to make decisions by interacting with its environment and receiving feedback through rewards or penalties. GPT uses reinforcement learning from human feedback.
Responsible AI
Responsible AI refers to the approach of creating, implementing, and utilizing AI systems with a focus on positively impacting employees, businesses, customers, and society as a whole, ensuring ethical intentions and fostering trust.
Retrieval Augmented Generation (RAG)
Retrieval Augmented Generation (RAG) enhances the capabilities of large language models by combining them with external knowledge sources.
RTCF (Role, Task, Context, Format)
A beginner-friendly prompt framework: tell the model who it is, what to do, what it needs to know, and how to shape the output. Context is the part most people under-invest in. We walk through a longer version in How To Structure A Prompt.
S
Self-Consistency and Tree of Thoughts (ToT)
Both techniques sample several reasoning paths instead of trusting one. Self-consistency takes the majority answer across multiple attempts, while Tree of Thoughts explores branches and prunes the weak ones. They help most on logic problems where a single chain can go wrong early and never recover.
Self-learning AI
Self-learning AI refers to systems that can autonomously acquire knowledge and improve their performance over time without explicit programming for each new task.
Semantic Routing
Semantic routing sends each incoming query to the cheapest thing that can answer it, whether that is a small specialized model, a vector search or a frontier reasoning model, based on intent rather than treating every request identically.
Sequence Modeling
A subfield of NLP that focuses on modeling sequential data such as text, speech, or time series data. Example: A sequence model that can predict the next word in a sentence or generate coherent text.
Small Language Models (SLMs) and On-Device AI
Compact models, typically under eight billion parameters and quantized to formats such as GGUF or AWQ, that run locally on a laptop or phone. They trade some capability for low latency, offline use and privacy.
Speech-to-text
The process of converting spoken words into written text.
Stable Diffusion
Stable diffusion is an artificial intelligence system that uses deep learning to generate images from text prompts.
Stacking
Stacking is a technique in AI that combines multiple algorithms to enhance overall performance by blending the strengths of various AI models, compensating for each model’s weaknesses.
Steerability
AI steerability refers to the ability to guide or control an AI system’s behavior and output according to human intentions or specific objectives, through mechanisms like fine-tuning, rule-based systems, and human feedback loops.
Stochastic Parrot
Stochastic parrots are AI systems that use statistics to convincingly generate human-like text, while lacking true semantic understanding behind the word patterns.
Strong AI
Strong AI refers to machines possessing generalized intelligence and capabilities on par with human cognition.
Structured Data
Structured data refers to information that is organized and labeled in a standardized format.
Structured Outputs and Constrained Decoding
Techniques that force a model to emit output matching a strict schema, such as valid JSON or a Pydantic model, by constraining what it is allowed to generate token by token. It turns an unpredictable text generator into something software can call reliably.
Summarization
Summarization is the ability of generative models to analyze large texts and produce concise, condensed versions that accurately convey the core meaning and key points.
Supervised Learning
A type of machine learning in which a model is trained on labeled data to make predictions about new, unseen data. Example: A supervised learning algorithm that can classify images of handwritten digits based on labeled training data.
System 1 vs. System 2 Prompting
A borrowed shorthand for two modes of generation: fast, cheap, intuitive answers (System 1) against slow, deliberate reasoning with visible working (System 2). Knowing which mode a task actually needs is what keeps cost and latency under control.
T
Text-to-speech (TTS)
Text-to-speech (TTS) is a technology that converts written text into spoken voice output, allowing users to hear written content being read aloud using synthesized speech.
Tokenization
The process of breaking text into individual words or subwords to input them into a language model. Example: Tokenizing a sentence “I am ChatGPT” into the words: “I,” “am,” “Chat,” “G,” and “PT.”
Transformer
A type of neural network architecture designed to process sequential data, such as text. The transformer architecture is used in models like ChatGPT for natural language processing tasks.
U
Unstructured Data
Unstructured data is any information that isn’t arranged in a pre-defined model or structure, making it tough to collect, process, and analyze.
Unsupervised Learning
A type of machine learning in which a model is trained on unlabeled data to find patterns or features in the data. Example: An unsupervised learning algorithm that can cluster similar images of handwritten digits based on their visual features.
V
Vibe Coding
Vibe coding is building software by describing what you want in natural language and letting an AI assistant write the core syntax. The term came from a single Andrej Karpathy post in February 2025 and was named Collins Dictionary Word of the Year for 2025. Its sequel term, vibe coding hangover, describes what happens when nobody can maintain the resulting code. Common traps are covered in Vibe Coding Mistakes: The 8 Simple Docs You Need First.
Voice Processing
Voice processing in AI refers to the pipeline of speech-to-text conversion followed by text-to-speech synthesis.
Voice Synthesis
Voice synthesis is using artificial intelligence to generate realistic, expressive computer speech by analyzing and learning from text and audio data.
W
Weak AI
Weak AI refers to narrow systems that excel at specific tasks within limited contexts, but lack generalized intelligence and adaptability outside their domain.
Weak-to-Strong Generalization
Weak-to-strong generalization is an AI training approach that uses less capable models to guide and constrain more powerful ones towards better generalization beyond their narrow training data.
What is Big Data?
Big data refers to the vast volumes of structured and unstructured data that are generated daily from various sources, including social media, sensors, transactions, and more.
What is Computer Vision?
Computer vision enables machines to interpret and understand visual information from the world around them.
What is Prompting?
Prompting refers to the art of crafting clear instructions and specific details to guide AI tools towards a desired output.
What is Quantum Computing?
Quantum computing is a computational approach that could be used to dramatically increase processing power, with significant potential to enhance AI capabilities.
Whisper
OpenAI’s Whisper is an AI system developed to perform automatic speech recognition (ASR), the task of transcribing spoken language into text.
X
X-risk (Existential Risk)
X-risk, short for existential risk, refers to the potential for highly advanced artificial intelligence to pose an existential threat to humanity through unintended consequences or goal misalignment.
Z
Zero-Shot Learning
Zero-shot learning is a technique in which a machine learning model can recognize and classify new concepts without any labeled examples.
Zero-to-One Problem
The zero-to-one problem refers to the difficulty of finding an initial solution when addressing complex challenges, which is often disproportionately challenging compared to subsequent progress.
Keeping This AI Glossary Current
AI terminology ages fast. Prompt engineering dominated the conversation for two years, for example, and context engineering has largely replaced it as the skill people hire for. Vibe coding went from a single social post to Collins Dictionary Word of the Year in under twelve months.
So this AI glossary is maintained rather than published once. New terms join the list as they stop being niche, and definitions get rewritten when a meaning drifts. If a term you keep hearing is missing from this AI glossary, it is worth flagging, because the gaps are usually the terms about to matter most.