An ordinary chatbot might tell you how to book a trip. An AI agent could potentially search flights, compare hotels, construct an itinerary and make the necessary reservations with your authorization.

Why it matters: Agents could transform AI from something people consult into something that actually performs work.

Related terms: Agentic AI · AI Agent · Tool Use · MCP

Agentic AI

In plain English: AI designed to pursue an objective, make decisions and take multiple actions with limited human direction.

Instead of requiring a person to provide instructions at every step, an agentic system can determine some of the intermediate steps itself.

Why it matters: Agentic AI could transform AI from a tool that waits for instructions into a system capable of performing increasingly complex work independently.

Related terms: Agent · AI Agent · Reasoning Model · Tool Use

AGI — Artificial General Intelligence

In plain English: A proposed form of AI capable of performing intellectual tasks across a very broad range of areas at roughly human-level ability or better.

There is no universally accepted definition or test for determining when AGI has been achieved.

Why it matters: AGI represents one of the AI industry's most important—and most disputed—potential milestones.

Related terms: AI · Artificial Superintelligence · Frontier Model · Reasoning Model

AI — Artificial Intelligence

In plain English: Computer systems capable of performing tasks normally associated with human intelligence, such as understanding language, recognizing images, solving problems, reasoning and making decisions.

Modern AI increasingly learns patterns from enormous amounts of data rather than relying entirely on rules explicitly programmed by humans.

Why it matters: AI is becoming a general-purpose technology affecting software, medicine, manufacturing, science, finance, defense, education and many other industries.

Related terms: Machine Learning · Generative AI · LLM · AGI

AI Accelerator

In plain English: A computer chip specifically designed or optimized to perform artificial-intelligence calculations efficiently.

GPUs, TPUs and NPUs are examples of processors commonly used to accelerate AI workloads.

Why it matters: AI accelerators provide much of the computing power behind modern AI and are central to the economics of AI infrastructure.

Related terms: GPU · NPU · TPU · Compute

AI Agent

In plain English: An AI system capable of using tools, making decisions and performing actions in pursuit of an objective.

An AI agent might search the web, use software, analyze information and complete a sequence of tasks rather than merely return a single answer.

Why it matters: AI agents could automate substantial portions of knowledge work and allow individual humans to supervise increasingly capable collections of digital workers.

Related terms: Agent · Agentic AI · MCP · Reasoning Model

AI Alignment

In plain English: The effort to make artificial-intelligence systems behave in ways consistent with intended human goals, instructions and values.

An aligned AI should attempt to accomplish what its users actually intend without producing dangerous or unintended behavior.

Why it matters: As AI systems become more capable and autonomous, mistakes or objectives that differ from human intentions could have increasingly serious consequences.

Related terms: AI Safety · Reinforcement Learning · RLHF · AGI

AI Data Center

In plain English: A data center designed or heavily equipped to run artificial-intelligence workloads.

These facilities can contain enormous clusters of AI processors along with specialized networking, memory, cooling and electrical infrastructure.

Why it matters: The race to build AI data centers has become one of the largest capital-investment cycles in the technology industry.

Related terms: Compute · GPU · Hyperscaler · AI Factory

AI Factory

In plain English: An industry term for infrastructure that continuously turns computing power and data into AI models, intelligence or AI-generated output.

Think of a conventional factory producing physical products. An AI factory produces digital intelligence.

Why it matters: The term reflects the increasingly industrial scale of the infrastructure being constructed to develop and operate AI.

Related terms: AI Data Center · Compute · Training · Inference

Artificial Superintelligence — ASI

In plain English: A hypothetical form of artificial intelligence that substantially exceeds human intellectual capabilities across most or virtually all important cognitive tasks.

ASI goes beyond AGI. AGI generally refers to broad human-level intelligence; superintelligence refers to intelligence substantially beyond it.

Why it matters: ASI remains hypothetical, but its possibility is central to debates about AI's long-term economic potential, safety and impact on civilization.

Related terms: AGI · AI Alignment · Frontier Model · AI

Benchmark

In plain English: A standardized test used to measure and compare the capabilities of AI systems.

Benchmarks can test mathematics, coding, reasoning, factual knowledge, image recognition and many other abilities.

Why it matters: Benchmarks help compare models, although excellent benchmark scores do not always translate directly into better real-world performance.

Related terms: Frontier Model · Reasoning Model · Model

Compute

In plain English: The processing power used to train or operate an AI system.

More compute generally means more calculations can be performed, although better algorithms and more efficient hardware can allow AI systems to accomplish more with the same amount of compute.

Why it matters: Access to computing power has become one of the central economic and strategic issues surrounding AI.

Related terms: GPU · Training · Inference · AI Data Center

Context Window

In plain English: The amount of information an AI model can consider at one time while producing a response.

A larger context window allows a model to work with longer conversations, documents, software code or collections of information without losing earlier material.

Why it matters: Larger useful context windows can make AI substantially more capable at complex and long-duration tasks.

Related terms: Token · Prompt · LLM · Model

Deep Learning

In plain English: A form of machine learning that uses multilayered neural networks to learn complex patterns from data.

Much of the modern AI revolution has been driven by advances in deep learning.

Why it matters: Deep learning underlies major breakthroughs in language models, image recognition, speech, robotics and generative AI.

Related terms: Neural Network · Machine Learning · Transformer

Distillation

In plain English: A technique for transferring some of the capabilities of a larger or more capable AI model into a smaller or more efficient model.

One way to think about it is a powerful model acting partly as a teacher for another model.

Why it matters: Distillation can reduce the computing power and cost required to obtain powerful AI capabilities, potentially making advanced AI cheaper and more widely available.

Related terms: Model · Quantization · Training · Inference

Edge AI

In plain English: Artificial intelligence that runs close to where data is created rather than relying entirely on a distant cloud data center.

Examples include AI operating inside phones, vehicles, cameras, robots and industrial equipment.

Why it matters: Edge AI can reduce latency, improve privacy, work without a continuous internet connection and reduce dependence on centralized computing infrastructure.

Related terms: On-Device AI · NPU · Inference · Physical AI

Embedding

In plain English: A mathematical representation that converts words, documents, images or other information into numbers representing aspects of their meaning or characteristics.

Items with similar meanings tend to have mathematically similar embeddings.

Why it matters: Embeddings allow AI systems to search and compare information based on meaning rather than simply matching identical words.

Related terms: RAG · Vector Database · LLM

Fine-Tuning

In plain English: Additional training performed on an existing AI model to make it better at a particular task, behavior or area of knowledge.

A general model might, for example, be fine-tuned to perform a specialized business function.

Why it matters: Fine-tuning allows organizations to specialize existing models without necessarily developing a new foundation model from scratch.

Related terms: Training · Foundation Model · Model · Open Weight

Foundation Model

In plain English: A broadly trained AI model capable of serving as the foundation for many different applications and specialized systems.

Rather than building an entirely new AI model for every task, developers can build many products and services on top of the same foundation model.

Why it matters: Foundation models provide much of the underlying intelligence used by today's generative-AI applications.

Related terms: LLM · Frontier Model · Fine-Tuning · Model

Frontier Model

In plain English: One of the most capable AI models available at a particular point in time.

The precise definition varies, and today's frontier model can quickly cease to be frontier as competitors improve.

Why it matters: Frontier models sit at the leading edge of AI capability and at the center of debates about investment, competition, regulation and safety.

Related terms: Foundation Model · Reasoning Model · Benchmark · AGI

Generative AI

In plain English: Artificial intelligence capable of creating new content such as text, images, video, music, speech or computer code.

Instead of merely classifying or analyzing existing information, generative AI can produce new output in response to instructions.

Why it matters: Generative AI transformed AI from primarily an analytical technology into a technology capable of directly producing useful work.

Related terms: LLM · Multimodal AI · Prompt · Transformer

GPU — Graphics Processing Unit

In plain English: A type of processor originally developed largely for computer graphics that proved exceptionally useful for the massive parallel calculations required by modern AI.

Thousands of GPUs can be connected together to train and operate extremely large AI models.

Why it matters: GPUs became the computational engines of the AI boom and one of the most valuable resources in the AI economy.

Related terms: Compute · HBM · AI Accelerator · AI Data Center

Hallucination

In plain English: When an AI system generates information that sounds plausible but is incorrect, unsupported or invented.

The model may present the false information confidently because it generates probable outputs rather than inherently knowing whether every statement is true.

Why it matters: Hallucinations are one reason important AI-generated information should be independently verified rather than automatically trusted.

Related terms: RAG · LLM · AI Alignment · Model

HBM — High-Bandwidth Memory

In plain English: Extremely fast computer memory designed to move large quantities of data quickly between memory and high-performance processors such as AI accelerators.

Why it matters: Powerful AI processors cannot operate efficiently if they spend too much time waiting for data. Memory speed and capacity therefore become critical parts of AI performance.

Related terms: GPU · AI Accelerator · Compute · AI Data Center

Hyperscaler

In plain English: A technology company that operates computing and cloud infrastructure at enormous scale.

The term generally refers to companies capable of operating vast global networks of data centers and rapidly adding computing capacity.

Why it matters: Hyperscalers are among the largest purchasers of AI chips and the largest investors in AI infrastructure.

Related terms: AI Data Center · Compute · AI Factory · GPU

Inference

In plain English: What happens when a trained AI model actually goes to work.

Training teaches the model. Inference is the model using what it learned to answer a question, generate an image, write software, recognize an object or perform another task.

Why it matters: As billions of people, businesses and machines use AI, the amount of computing power devoted to inference could become enormous.

Related terms: Training · Compute · Token · GPU

LLM — Large Language Model

In plain English: An AI model trained on enormous amounts of language data so that it can understand and generate human-like language.

Modern LLMs can increasingly perform other tasks as well, including reasoning, writing software, analyzing documents and interacting with tools.

Why it matters: LLMs provide much of the underlying intelligence behind today's AI assistants and many emerging AI agents.

Related terms: Transformer · Foundation Model · Token · Multimodal AI

Machine Learning

In plain English: A branch of artificial intelligence in which computer systems learn patterns from data rather than having every rule explicitly programmed by humans.

Why it matters: Machine learning is one of the foundational technologies behind modern AI and allows systems to improve at tasks using examples and experience.

Related terms: AI · Deep Learning · Neural Network · Training

MCP — Model Context Protocol

In plain English: An open standard for connecting AI applications to external tools, data sources and software systems in a consistent way.

One way to think about MCP is as a standardized connection layer that helps AI systems interact with the digital world around them.

Why it matters: Common connection standards can make AI agents substantially easier to integrate with the software, information and tools they need to perform useful work.

Related terms: Agent · AI Agent · Agentic AI · Tool Use

Mixture of Experts — MoE

In plain English: An AI architecture containing multiple specialized components, called “experts,” while activating only some of them for a particular input or task.

Why it matters: Mixture-of-experts architectures can allow very large models to operate more efficiently because every part of the model does not need to be used for every request.

Related terms: Transformer · Model · Inference · LLM

Model

In plain English: The trained mathematical system at the heart of an AI application.

A model learns patterns and relationships during training and then uses what it has learned to generate predictions, decisions or other outputs.

Why it matters: The model is essentially the engine supplying intelligence to an AI application.

Related terms: Training · Parameter · Foundation Model · Inference

Multimodal AI

In plain English: Artificial intelligence capable of working with more than one type of information, such as text, images, audio and video.

A multimodal system might look at a photograph, discuss what it sees, listen to spoken instructions and respond using speech.

Why it matters: Multimodal AI allows machines to interact with information in ways that more closely resemble how humans experience the world.

Related terms: LLM · Generative AI · Physical AI · Model

Neural Network

In plain English: A computing architecture loosely inspired by networks of neurons in the brain.

Neural networks learn relationships and patterns from data rather than requiring programmers to explicitly define every rule.

Why it matters: Neural networks form the technological foundation of modern deep learning and many of today's most capable AI systems.

Related terms: Deep Learning · Transformer · Machine Learning · Parameter

NPU — Neural Processing Unit

In plain English: A processor specifically designed to perform neural-network and artificial-intelligence calculations efficiently.

NPUs are increasingly appearing in phones, personal computers, vehicles and other devices.

Why it matters: NPUs help move AI away from centralized data centers and directly onto devices at the edge.

Related terms: Edge AI · On-Device AI · GPU · AI Accelerator

On-Device AI

In plain English: Artificial intelligence that runs directly on a user's device rather than sending every task to a remote cloud server.

Examples include AI models operating on smartphones, laptops, vehicles and other intelligent devices.

Why it matters: On-device AI can provide faster responses, greater privacy, offline capability and potentially much lower reliance on centralized data centers.

Related terms: Edge AI · NPU · Inference · Quantization

Open Source AI

In plain English: AI software made available under licenses that allow others to inspect, use, modify and redistribute its underlying source code according to the terms of the license.

The term is sometimes used too loosely and should not automatically be treated as synonymous with “open weight.”

Why it matters: Open-source AI can accelerate innovation and competition by allowing large communities of developers to build upon existing technology.

Related terms: Open Weight · Foundation Model · Fine-Tuning · Model

Open Weight

In plain English: An AI model whose trained numerical weights are made available for others to use.

Open weight does not necessarily mean fully open source, because training data, source code, training methods or other components may remain proprietary.

Why it matters: Open-weight models can allow developers and organizations to run and customize powerful AI systems without depending entirely on the company that created them.

Related terms: Open Source AI · Model · Fine-Tuning · Foundation Model

Parameter

In plain English: One of the numerical values inside a neural network that changes during training as the model learns patterns from data.

Large AI models can contain billions or even trillions of parameters.

Why it matters: Parameter count became a common measure of model scale, although a model with more parameters is not automatically more capable or efficient.

Related terms: Model · Training · Neural Network · LLM

Physical AI

In plain English: Artificial intelligence that perceives, interacts with or controls the physical world rather than operating entirely inside software.

Examples include robots, autonomous vehicles, drones and intelligent industrial equipment.

Why it matters: Physical AI could extend the AI revolution from information and knowledge work into manufacturing, transportation, logistics and physical labor.

Related terms: World Model · Edge AI · Multimodal AI · Agent

Prompt

In plain English: The instruction, question or information given to an AI system to guide what it does next.

A prompt can range from a simple question to a lengthy collection of instructions, examples and background information.

Why it matters: What a user tells an AI—and the context supplied with it—can materially affect the quality of the result.

Related terms: Prompt Engineering · Context Window · LLM · Token

Prompt Engineering

In plain English: The practice of designing instructions and context to help an AI system produce more useful or reliable results.

Why it matters: Better prompting can materially improve AI performance, although increasingly capable models may reduce the amount of specialized prompt engineering required over time.

Related terms: Prompt · Context Window · LLM · Agent

Quantization

In plain English: A technique that reduces the numerical precision used by an AI model so that it requires less memory and computing power.

The objective is generally to make the model smaller and more efficient while preserving as much capability as possible.

Why it matters: Quantization can allow powerful models to operate more cheaply and, in some cases, directly on personal computers, phones and other edge devices.

Related terms: Distillation · Edge AI · Inference · On-Device AI

RAG — Retrieval-Augmented Generation

In plain English: A technique that allows an AI system to retrieve relevant information from external sources before producing its answer.

Instead of relying entirely on information encoded during training, the system can search a collection of documents or other information and use what it finds when responding.

Why it matters: RAG can give AI access to more current or specialized information and can improve the factual grounding of its answers.

Related terms: Embedding · LLM · Hallucination · Model

Reasoning Model

In plain English: An AI model designed to devote additional computational effort to working through difficult problems before producing an answer.

Rather than immediately generating a response, a reasoning model can perform additional internal computation to improve its result.

Why it matters: Reasoning approaches have substantially improved AI performance in mathematics, science, coding and other complex tasks.

Related terms: Frontier Model · Inference · Compute · Agent

Reinforcement Learning

In plain English: A method of training AI through feedback based on the results of actions or outputs.

Behaviors producing better outcomes receive stronger reinforcement, helping the system learn strategies that are more likely to succeed.

Why it matters: Reinforcement learning has become an important technique for improving AI reasoning, decision-making and autonomous behavior.

Related terms: RLHF · Training · Agent · AI Alignment

RLHF — Reinforcement Learning from Human Feedback

In plain English: A training technique that uses human judgments about AI responses to help a model learn which behaviors and answers people prefer.

Humans might compare several AI answers, for example, and indicate which is better.

Why it matters: RLHF played an important role in making modern AI assistants more useful, conversational and responsive to human instructions.

Related terms: Reinforcement Learning · AI Alignment · Training · Model

Scaling Law

In plain English: An observed relationship showing how AI performance tends to change as resources such as computing power, training data or model size increase.

Why it matters: Scaling laws helped motivate the enormous investments in larger models, more computing power and increasingly massive AI data centers.

Related terms: Compute · Training · Parameter · Frontier Model

Sovereign AI

In plain English: The idea that a country should possess or control important parts of its own artificial-intelligence capability and infrastructure.

That can include domestic computing capacity, models, data, energy resources and semiconductor supply.

Why it matters: AI is increasingly viewed not merely as commercial technology but as strategic national infrastructure.

Related terms: AI Data Center · Compute · Frontier Model · AI Accelerator

Synthetic Data

In plain English: Artificially generated information used to train, test or improve AI systems rather than information collected directly from the real world.

AI systems themselves can generate synthetic training data.

Why it matters: Synthetic data could help overcome shortages of useful training material and allow developers to create large quantities of specialized examples relatively cheaply.

Related terms: Training · Model · Reinforcement Learning · LLM

Token

In plain English: A small unit of information processed by a language model.

A token can represent a whole word, part of a word, punctuation or another piece of text.

Why it matters: AI companies frequently measure model usage, context capacity and pricing in tokens, making tokens something like units of consumption in the AI economy.

Related terms: LLM · Context Window · Inference · Prompt

TPU — Tensor Processing Unit

In plain English: A specialized processor developed by Google for machine-learning and artificial-intelligence workloads.

Why it matters: TPUs demonstrate how major technology companies can design their own specialized AI processors rather than relying exclusively on commercially available GPUs.

Related terms: GPU · NPU · AI Accelerator · Compute

Training

In plain English: The process through which an AI model learns patterns, relationships and capabilities from data and feedback.

Training advanced models can require enormous amounts of computing power, electricity, data and time.

Why it matters: Training creates the capabilities inside AI models and has driven much of the industry's demand for massive computing clusters.

Related terms: Inference · Model · GPU · Parameter

Transformer

In plain English: A neural-network architecture that enabled much of the modern large-language-model revolution.

Transformers use a mechanism called attention to identify relationships between different pieces of information, even when those pieces are far apart.

Why it matters: The transformer became one of the foundational technological breakthroughs behind modern generative AI.

Related terms: LLM · Neural Network · Deep Learning · Model

World Model

In plain English: An AI system or component designed to build an internal representation of how an environment works and predict what may happen when actions are taken.

In simple terms, a world model attempts to understand how things relate, change and respond to actions.

Why it matters: World models could become particularly important for robotics, autonomous systems and AI operating in the physical world.

Related terms: Physical AI · Agent · Multimodal AI · Reinforcement Learning