An AI glossary you can use.

The broad field of systems performing tasks such as recognition, prediction, language generation and decision support.

Methods that fit patterns from data, rather than specifying every decision through hand-written rules.

Machine learning methods that use layered neural networks.

Models used to generate content such as text, images or audio.

A research concept concerning broadly transferable capabilities across many kinds of task; definitions and evaluation standards vary.

A model trained on language-related patterns and used to generate or transform text.

An encoded unit processed by a language model; it may be a word, part of a word or punctuation.

The instructions and information made available to a model for a particular generation.

Using a trained model to produce outputs for supplied inputs.

The process of adjusting model parameters using data and an optimization objective.

A pattern that retrieves relevant material and supplies it as context for a generated answer.

Additional training used to adapt a model’s parameters for a particular behavior or task.

A numerical representation used to compare or process items such as text passages.

A system in which a model selects some steps or tool calls toward a goal.

A designed sequence of operations; it may include models without giving them control over the whole sequence.

A generated claim that is unsupported or incorrect in the context of the task.

An attempt to make a system follow instructions from untrusted input in ways that conflict with its intended task or controls.

Connecting a response to relevant external information; the connection still needs to be checked.

A repeatable comparison of system behavior against defined task requirements and test cases.

A reference approach, such as the current process or a simple rule, used to judge whether a proposed change helps.

For defect detection, the share of flagged items that are actually defective: TP divided by TP plus FP.

For defect detection, the share of actual defects detected: TP divided by TP plus FN.

An acceptable item incorrectly flagged as defective, when defect is the positive class.

A defective item incorrectly passed, when defect is the positive class.

Information from evaluation examples influencing training or development, making reported performance less independent.

A change in data or conditions that can alter how useful a model or system is.

A bounded experiment to investigate whether an approach can work under stated conditions.

Methods that analyze images or video for tasks such as classification, inspection and counting.