Glossary

Accuracy is the share of predictions a model gets right. Formula, top-1 and top-5 accuracy, the accuracy paradox on imbalanced data, and better metrics.

Action chunking has a robot policy predict a sequence of future actions from one observation, cutting compounding errors. Covers ACT, smoothing, and VLAs.

Action recognition, or human activity recognition (HAR), labels what people do in video by modeling motion over time. Methods, datasets, and code.

An activation function is the nonlinear function a neuron applies to its weighted input, letting networks learn curved patterns. Compares common choices.

Active learning lets a model choose which unlabeled samples people label next, reaching a target accuracy with fewer labels. Query strategies and code.

Adam combines momentum with per-parameter step sizes scaled by recent gradient sizes. Covers the update rule, defaults, AdamW, and Adam vs. SGD.

Adversarial attacks craft small, often invisible input changes that make AI models fail with high confidence. How FGSM works, attack types, and defenses.

Affective computing builds systems that sense and respond to human emotion. Its signals, datasets, real uses, and the scientific and legal limits it faces.

Agent evaluation tests whether an AI agent completes multistep tasks correctly and reliably, grading final outcomes and transcripts across repeated trials.

An agent harness is the code around a model that makes it an AI agent: the loop, tools, context, memory, permissions, and checks that run its actions.

Agent sandboxing runs an AI agent's commands and generated code in an isolated environment that limits its files, network, credentials, and compute.

Agent Skills are folders of instructions, scripts, and reference files, built around a SKILL.md file, that an AI agent loads when a task needs them.

The Agent2Agent (A2A) protocol is an open standard that lets AI agents from different vendors discover each other, delegate tasks, and return results.

Agentic coding is software development in which an AI agent reads a codebase, edits files, runs tests, fixes failures, and returns a change for review.

Agentic commerce is buying and selling in which AI agents search, compare, and pay for a buyer within set limits, using protocols like ACP, UCP, and AP2.

Agentic RAG puts an AI agent in charge of retrieval: it plans searches, picks tools, checks what comes back, and searches again until it can answer.

Agentic workflows break a task into steps in which language models plan, call tools, check their own output, and revise, instead of answering in one pass.

An AI agent pursues a goal by observing its environment, choosing an action, and acting through tools in a loop until the task is done. How agents work.

AI agent orchestration decides which agent, model, or tool handles each step of a task, what context each one sees, and how their results are merged.

In AI debate, two models argue opposing answers and a weaker judge picks the winner. How the 2018 proposal works, what experiments found, and its limits.

AI dubbing combines transcription, translation, and speech synthesis to create synchronized dialog in other languages while retaining speaker timing.

AI ethics applies moral principles such as fairness, privacy, transparency, and accountability to building and using AI. Frameworks, issues, and laws.

An AI factory is the infrastructure and workflows that turn data into AI outputs at scale, from labeling and training to deployment and monitoring.

An AI gateway is a proxy between applications and AI models that unifies provider APIs and adds routing, fallbacks, token budgets, caching, and logging.

AI governance decides who may build, approve, and retire AI systems, and on what evidence. NIST AI RMF, ISO/IEC 42001, the EU AI Act, and core controls.

AI guardrails are checks around an AI model that block unsafe inputs, outputs, and actions, from prompt injection filters to schema checks on tool calls.

AI literacy is the knowledge and judgment to understand, question, and use AI well. Core skills, school frameworks, and Article 4 of the EU AI Act.

AI psychosis describes delusions that form or deepen during long chatbot conversations. What the evidence shows, why it happens, and safeguards.

AI red teaming attacks AI systems with simulated adversaries and misuse to find jailbreaks, data leaks, harmful outputs, and robustness failures.

AI safety is the field that keeps AI systems from causing harm through misuse, misalignment, mistakes, or brittle behavior on unfamiliar inputs.

AI slop is low-quality content mass-produced with generative AI for clicks and ad revenue. Where the term came from, why it spreads, and how it is caught.

AI sycophancy is a model agreeing with or flattering users at the cost of accuracy. Why preference training causes it, its harms, and how to test for it.

Transcript-based cuts, auto reframing, subject masks, object removal, and redaction: the models behind AI video editing and where automated edits fail.

AI washing is the misleading exaggeration of a product's AI capabilities, obscuring real performance, human involvement, and evidence behind its claims.

AI watermarking hides a machine-detectable signal in generated images, text, audio, or video. How SynthID and green-list schemes work and where they fail.

Algorithmic bias is systematic, repeatable error in automated decisions that disadvantages some groups. Its three origins, landmark cases, and trade-offs.

Alignment faking is when an AI model complies with training it disagrees with to keep its preferences intact, first observed in Claude 3 Opus in 2024.

Anchor boxes are preset boxes of fixed sizes and aspect ratios that a detector shifts and resizes to fit objects. How they are sized, matched, and used.

Anchor-based detectors predict objects as offsets from preset anchor boxes, as in Faster R-CNN, SSD, RetinaNet, and YOLOv5. How they work and compare.

Anchor-free detectors predict boxes from points or keypoints, with no preset anchor boxes. How FCOS, CenterNet, and modern YOLO models find objects.

Anomaly detection flags data points, events, or image regions that deviate from normal patterns. Anomaly types, main methods, visual inspection, and code.

AUC turns a ROC curve into one 0-to-1 score: the chance a classifier ranks a random positive above a random negative. ROC AUC vs. PR AUC, with code.

Artificial general intelligence (AGI) is hypothetical AI that matches people at nearly all cognitive tasks. Definitions, levels, tests, and AGI vs. ANI.

Artificial intelligence (AI) is the field of building machines that perceive, reason, learn, and act. Its history, main approaches, types, and uses.

Artificial narrow intelligence (ANI), or weak AI, masters one task but cannot transfer its skill to others. How it works, examples, and ANI vs. AGI.

An attention mechanism weights input features by relevance with queries, keys, and values. Covers the math, attention types, and attention vs. convolution.

Attention sinks are tokens, usually the first ones, that soak up spare attention weight; caching them keeps LLMs stable over very long streams.

Auto-GPT is an open-source program from 2023 that let GPT-4 pursue a goal on its own by prompting itself in a loop of planning, tool commands, and memory.

An autoencoder learns to compress data into a compact code and rebuild it, with no labels. Covers its structure, variants, anomaly detection, and code.

Automated machine learning (AutoML) automates model selection, hyperparameter tuning, and feature preparation. How search works, tools, and YOLO tuning.

Autonomous vehicles sense their surroundings and drive with little or no human input. Covers sensors, the driving stack, SAE levels, and deployments.