AI Glossary

59 AI and machine learning terms, explained in plain English — search or filter by category.

Agent / AI Agent Business & Product: An AI system that can take multi-step actions toward a goal — using tools, calling APIs, or executing tasks — rather than only responding to a single prompt with text.
AI Act / AI Regulation Safety & Governance: Government rules governing how AI systems can be developed, sold, or deployed — the EU AI Act is the most comprehensive example, with other jurisdictions developing their own frameworks.
AI Alignment Safety & Governance: The broad effort to ensure an AI system's behavior and goals actually match what its developers and users intend, rather than pursuing its training objective in unintended ways.
AI-Native Application Business & Product: Software designed from the ground up around AI capability, rather than software that had AI features added on afterward.
API (Application Programming Interface) Infrastructure: A way for one piece of software to send requests to another. AI providers expose APIs so developers can integrate a model into their own apps, priced separately from consumer chat subscriptions.
Artificial General Intelligence (AGI) Core Concepts: A hypothetical AI system with human-level ability across essentially any intellectual task, rather than being specialized for narrow tasks. No current system meets this bar, and there's no consensus on when — or whether — one will.
Artificial Intelligence (AI) Core Concepts: Software that performs tasks normally requiring human judgment — recognizing patterns, generating text or images, making predictions — without being explicitly programmed for each specific case.
Batch API Infrastructure: A discounted way to submit large volumes of AI requests for asynchronous processing (results delivered later, not instantly), typically around half the price of standard real-time requests.
Bias (in AI) Safety & Governance: Systematic skew in a model's outputs that reflects imbalances or prejudices present in its training data, often producing unfair or inaccurate results for certain groups or topics.
Chain-of-Thought Prompting Prompting: Prompting a model to reason step-by-step before giving a final answer, which often improves accuracy on tasks that require multi-step logic or math.
Context Window Models & Architecture: The maximum amount of text (measured in tokens) a model can consider in a single conversation or request, including both what you send and what it generates back.
Copilot Business & Product: An AI assistant embedded directly into an existing tool or workflow (like code editors or office software) to help a user perform tasks, rather than a standalone chat product.
Custom GPT / Custom Assistant Business & Product: A version of a base chat model configured with specific instructions, knowledge, or tools for a particular use case, built by a user or business without training a new model from scratch.
Deep Learning Core Concepts: Machine learning using neural networks with many layers ('deep' networks), which is what makes modern AI capable of tasks like image recognition and language generation.
Diffusion Model Models & Architecture: The architecture behind most AI image generators (like Midjourney or Stable Diffusion), which works by starting from random noise and gradually refining it into a coherent image.
Distillation Training: Training a smaller 'student' model to mimic a larger 'teacher' model's outputs, producing a faster, cheaper model that retains much of the larger model's capability.
Embedding Models & Architecture: A numerical representation of text, images, or other data that captures meaning as coordinates in space — similar concepts end up as nearby points, which is what powers semantic search and recommendation systems.
Explainability Safety & Governance: The degree to which a model's decision-making process can be understood or justified in human terms — a persistent challenge for deep learning systems, which don't reason in explicitly readable steps.
Few-Shot Prompting Prompting: Including a small number of example input-output pairs in a prompt to show the model the pattern you want, rather than relying on instructions alone.
Fine-Tuning Training: Further training a pretrained model on a smaller, more specific dataset to adapt its behavior for a particular task, domain, or style.
Foundation Model Business & Product: A large, general-purpose model (like GPT-5.6 or Claude) trained on broad data and adaptable to many downstream tasks, as opposed to a model built for one narrow purpose.
Freemium (AI Products) Business & Product: A pricing model offering a limited free tier alongside paid tiers with higher usage limits or more capable models — the standard structure for ChatGPT, Claude, and Gemini.
Generative AI Core Concepts: AI that creates new content — text, images, audio, video, or code — rather than just classifying or analyzing existing content.
GPU (Graphics Processing Unit) Infrastructure: A type of computer chip originally designed for rendering graphics, whose ability to do many calculations in parallel makes it well-suited to training and running AI models. Most large-scale AI compute runs on GPUs.
Guardrails Safety & Governance: Rules, filters, or checks built around an AI system to constrain its outputs — for example, blocking certain content categories or verifying claims before they reach a user.
Hallucination Safety & Governance: When an AI model generates false or fabricated information and presents it confidently as fact — a well-documented limitation of current LLMs, not a rare bug.
Inference Infrastructure: The process of running a trained model to generate a response to a new input — as distinct from training, which is how the model learned in the first place. Inference is what happens every time you send a message to an AI tool.
Large Language Model (LLM) Core Concepts: An AI model trained on huge amounts of text to predict and generate language. ChatGPT, Claude, and Gemini are all built on LLMs.
Latency Infrastructure: The time between sending a request to a model and receiving a response — a key factor in whether an AI feature feels instant or sluggish.
Machine Learning (ML) Core Concepts: A method for building AI systems by having them learn patterns from data, rather than hand-coding every rule. Nearly all modern AI, including large language models, is built this way.
Mixture of Experts (MoE) Models & Architecture: An architecture where a model has many specialized sub-networks ('experts') but only activates a small subset for any given input, reducing compute cost without shrinking total model capacity.
Model Card Business & Product: A document a provider publishes alongside a model describing its capabilities, limitations, training data at a high level, and intended uses.
Model Context Protocol (MCP) Infrastructure: An open standard for connecting AI models to external tools, data sources, and applications in a consistent way, reducing the custom integration work needed for each connection.
Multimodal Model Models & Architecture: A model that can process and/or generate more than one type of content — for example, understanding both text and images in the same request.
Neural Network Core Concepts: A model architecture loosely inspired by connections between neurons in the brain, made of layers of simple mathematical units that combine to recognize complex patterns.
Open-Weight Model Models & Architecture: A model whose trained parameters are published for anyone to download and run, as distinct from 'open source,' which would also require releasing training code and data.
Overfitting Training: When a model learns its training data too specifically — including its noise and quirks — and performs worse on new data it hasn't seen before.
Parameters Models & Architecture: The internal numerical values a model learns during training. Roughly, more parameters mean more capacity to capture complex patterns — though not always better real-world performance.
Pretraining Training: The initial, most expensive phase of training an LLM, where it learns general language patterns from a massive, broad text dataset before any task-specific fine-tuning.
Prompt Prompting: The text input given to an AI model to elicit a response — a question, instruction, or piece of context.
Prompt Caching Infrastructure: A provider feature that reduces cost and latency for repeated prompt content by reusing prior processing instead of reprocessing the same text from scratch on every request.
Prompt Engineering Prompting: The practice of deliberately structuring a prompt — its wording, examples, and format — to reliably get better or more predictable output from a model.
Prompt Injection Prompting: A security exploit where malicious instructions are hidden inside content an AI system processes (like a webpage or document), attempting to override its original instructions.
Red Teaming Safety & Governance: Deliberately probing an AI system for weaknesses, unsafe outputs, or exploitable behavior before or after release, similar in spirit to penetration testing in cybersecurity.
Reinforcement Learning from Human Feedback (RLHF) Training: A training technique where human raters rank model outputs, and the model is adjusted to produce more of what raters preferred — a major part of what makes chat models like ChatGPT feel conversational and helpful rather than just predicting text.
Retrieval-Augmented Generation (RAG) Infrastructure: A technique where a model's response is grounded by first retrieving relevant documents or data (often from a private knowledge base) and including that content in its context, rather than relying solely on what it memorized during training.
Small Language Model (SLM) Models & Architecture: A language model with far fewer parameters than a flagship model, trained to run efficiently on modest hardware or even on-device, trading some general capability for speed and lower cost.
Synthetic Data Training: Training data generated by an AI model rather than collected from real-world sources, often used to supplement scarce or expensive real data.
System 2 / Reasoning Model Prompting: A model variant that takes extra internal steps to work through a problem before answering, generally trading speed for improved accuracy on complex tasks like math or coding.
System Prompt Prompting: Instructions set by the application (not the end user) that shape a model's behavior across an entire conversation, such as its persona, tone, or rules it must follow.
Temperature Prompting: A setting that controls how random or deterministic a model's output is — low temperature gives more predictable, focused answers; high temperature gives more varied, creative ones.
Token Models & Architecture: The basic unit of text a model reads and generates — often a word or part of a word, averaging roughly 4 characters in English. API pricing and context limits are measured in tokens, not words or characters.
Token Limit / Rate Limit Business & Product: A cap providers place on how many tokens or requests an account can use in a given period, used to manage cost, fairness, and system load.
Training Data Training: The text, images, or other content a model learns from during training. Its scope and quality heavily shape what the model knows and how it behaves.
Transformer Models & Architecture: The neural network architecture behind nearly every modern LLM, introduced in 2017. Its key innovation is 'attention,' which lets the model weigh the relevance of every word to every other word at once.
Turing Test Core Concepts: A proposed test where a human judge tries to distinguish an AI from a human through conversation alone. Passing it is a philosophical benchmark, not a technical certification any modern model formally holds.
Vector Database Infrastructure: A database optimized for storing and searching embeddings, commonly used to power semantic search and RAG systems by quickly finding content similar in meaning to a query.
Watermarking (AI Content) Safety & Governance: Embedding an invisible or subtle signal in AI-generated content (text, images, audio) to make it identifiable as AI-generated later.
Zero-Shot Prompting Prompting: Asking a model to perform a task with instructions only, no examples included — relying entirely on what it learned during training.