Loading…
Reference
333 plain-language definitions for the terms you see in AI signals and in research. Open any term for a diagram and, where it helps, a few lines of code.
Showing 60 of 333 terms
Comparing two versions of a system by showing each to different users and measuring the results.
The share of a model's predictions that are correct.
A function in a neural network that decides how strongly a unit's output is passed on, letting the network model nonlinear patterns.
Techniques used in AI models to monitor internal neuron activations in real-time to detect harmful or misaligned knowledge activation and prevent misuse.
An AI system that plans and takes multi-step actions, often by calling tools, to achieve a goal.
A system architecture that enables AI agents to store and recall key facts, user preferences, and workflows to improve task success rates and reduce redundant information retrieval.
AI systems that can plan and carry out multi-step tasks with some autonomy.
An independent examination of an AI system's performance, risks and compliance.
The study of moral questions raised by AI, such as fairness, accountability and privacy.
The policies, laws and institutions that guide how AI is developed and used.
Research and practice aimed at making AI systems reliable, controllable and free of serious harm.
A step-by-step procedure for solving a problem or performing a computation.
The principle that organizations should be answerable for the effects of their automated systems.
Systematic unfairness in an AI system's outputs, often reflecting skewed data or design.
Work to make AI systems behave in line with human intentions and values.
The process of marking up data with labels or notes.
Identifying data points that differ markedly from the norm, such as fraud.
Application programming interface: a way for software to request services, such as a model's output, from another system.
A hypothetical AI that can match human ability across most intellectual tasks.
The field of building computer systems that perform tasks normally associated with human intelligence, such as understanding language, recognizing images or making decisions.
An application-specific integrated circuit: a chip designed for one purpose, including custom AI chips.
A method that lets a model weigh which parts of its input are most relevant when producing each output.
A measure of a classifier's ability to separate classes across all decision thresholds.
A network trained to compress data into a compact form and then reconstruct it.
A vehicle that can drive itself with little or no human input.
A managed service that catalogs and enables runtime discovery of AI agents, tools, and Model Context Protocol servers to support scalable and dynamic agent orchestration.
The algorithm that calculates how each parameter in a neural network contributed to the error, so it can be updated.
A technique that normalizes values across a batch of examples to speed up and stabilize training.
The number of training examples processed together before the model's parameters are updated.
A method of updating the probability of a hypothesis as new evidence arrives.
A decoding method that keeps several candidate sequences in play and picks the best overall.
A standardized test used to compare model performance on a task.
A 16-bit number format that keeps the wide range of 32-bit numbers, widely used in AI training.
A systematic error or skew in a model's outputs, often inherited from the data or design.
A metric that scores machine translation by its overlap with reference translations.
An AI agent that completes tasks by navigating websites.
A common tokenization method that builds a vocabulary by repeatedly merging frequent pairs of characters.
A standard for attaching verifiable provenance information, called content credentials, to digital media.
When a model loses earlier skills while being trained on new tasks.
A technique in which a model works through intermediate reasoning steps before giving a final answer.
A program that holds conversations with people, today often powered by a language model.
A saved snapshot of a model's weights at a point during training.
A finding that, for a fixed compute budget, model size and training data should grow together.
Splitting documents into smaller pieces so they can be indexed and retrieved.
Predicting which category an input belongs to.
Grouping similar data points together without predefined labels.
An AI agent that writes, edits and tests software.
A capability that lets an AI operate a computer by viewing the screen and using the mouse and keyboard.
The field of AI that lets computers interpret images and video.
A change in the relationship between inputs and the correct output over time.
A table showing how many predictions a classifier got right and wrong for each class.
A method in which a model is trained to follow a written set of principles.
A decoding technique that enforces output to conform to a specified schema or format during token generation, ensuring structured and valid responses.
Detecting and handling harmful or disallowed content.
Information about where a piece of content came from and how it was made or edited.
The maximum amount of text, measured in tokens, a model can consider at once.
Learning representations by pulling similar items together and pushing different items apart.
A neural network that uses filters to detect local patterns, widely used for images.
An AI assistant embedded in a tool to help a user as they work.
The property of an AI system that allows humans to correct or shut it down.