AI, in plain words.

An AI system that takes actions, uses tools, and adapts its next steps to work toward a goal.

A group of AI agents that coordinate their work toward a shared goal, often tackling different parts in parallel.

AI with broad abilities to learn and perform intellectual tasks across many domains, commonly described in relation to human capability.

Research and practices aimed at preventing harmful AI behavior and outcomes.

The work of making AI systems behave in ways that match human intentions, values, and constraints.

Attributing human qualities, feelings, or intentions to an AI system.

The field of building systems that perform tasks associated with intelligence, such as learning, reasoning, and perception.

A systematic tendency in data or a system’s behavior that can skew results or produce unfair outcomes.

Intermediate reasoning steps generated as part of working toward an answer.

The amount of information a model can work with in a single inference context.

A branch of machine learning that uses neural networks with multiple layers to learn patterns from data.

Training one model to reproduce useful behavior from another, often larger, model.

A numerical representation that places items in a space where useful relationships can be compared.

A measure of uncertainty in a probability distribution: higher entropy means less predictable outcomes.

Tests that measure how well an AI system performs specific tasks.

A structured way to measure whether an AI model or system does what matters for a task.

Continuing a pretrained model’s training to adapt its behavior to a particular task or style.

A broadly trained model that can be adapted or used as the basis for many different tasks.

AI systems at the leading edge of capability at a given time.

AI that produces content such as text, images, audio, video, or code.

Designing the connections between tasks, agents, tools, and state so an AI system can coordinate its work.

Safeguards that help constrain an AI system’s outputs and actions.

An AI output that presents false or unsupported information as if it were established.

The software around a model that manages its tools, context, execution, and constraints.

Running a trained model on an input to produce a prediction or other output.

Methods for understanding how an AI model produces its behavior.

A large neural model trained on language patterns that can generate and process text.

A repeated cycle in which an AI system acts, observes the result, and decides what to do next.

Methods that learn patterns from data to make predictions or guide decisions.

A mismatch between an AI system’s behavior or objectives and intended human goals or constraints.

A computational system whose structure and learned parameters map inputs to outputs.

Documentation describing a model’s intended uses, performance, and limitations.

An entity that can bear moral obligations and, on many accounts, responsibility for its actions.

An entity owed moral consideration for its own sake, whether or not it can bear responsibility.

Able to work with more than one kind of information, such as text, images, or audio.

A model made of connected computational units that learn useful transformations of data.

The adjustable numerical values a model learns during training.

The input that gives a model instructions, context, examples, or material to work on.

An attempt to make an AI system treat untrusted content as instructions that override its intended task.

Giving a model retrieved information to help ground the answer it generates.

A process in which an AI system improves itself, then uses the improved version to make further improvements.

A way of learning through feedback in which an agent improves its choices by seeking higher cumulative reward.

Using human feedback to help train a model toward preferred behavior.

An isolated environment that limits what code or an AI agent can access and change.

Capable of subjective experiences, such as feeling pain or pleasure.

A hypothetical future turning point when superhuman intelligence drives rapid change that becomes difficult for humans to predict.

Agents assigned focused pieces of a larger task by another agent or coordinating system.

A sampling setting that changes how strongly a model favors its most likely next outputs.

A unit of information a model processes, often a word, part of a word, or punctuation.

External capabilities an AI system can call to retrieve information, run calculations, or take actions.

Adjusting a model’s parameters using data and a learning objective.

The computational work used to train an AI model.

A neural-network architecture that uses attention to connect information across a sequence.

Learned numerical values that control how signals are combined inside a model.

Software that adds an interface or features around an existing AI model.