Concept Glossary: A¶

This page lists English glossary entries for this letter. Entry bodies are assembled from term source files.

accountability¶

• Meaning: Accountability is the principle that an AI system's outputs, decisions, and automated actions must have identifiable human or organizational responsibility behind them. It asks who reviewed, approved, explained, corrected, and operated the system when something goes wrong.
• Why it matters: AI results can look fluent or authoritative even when they create risk. In high-impact settings, logs, approval records, review procedures, and ownership boundaries are needed so a problem is not dismissed as something the model simply did. Accountability is not only about blame after failure; it is a design principle for making responsibility visible before deployment.
• Related concepts: human oversight, safety, provenance, transparency
• Core Section: P1-15.1
• Appears in: P1-10.3

accuracy¶

• Meaning: Accuracy is the fraction of predictions that are correct: correct predictions divided by all predictions. It is one of the most intuitive metrics for classification because it quickly summarizes how often a model chose the right answer.
• Why it matters: Accuracy is useful as a first check, but it can hide important failure patterns. With imbalanced data, a model can look accurate while ignoring rare but important cases. Readers should treat accuracy as a starting point and then ask what kinds of mistakes remain, whether some errors cost more than others, and whether other metrics are needed.
• Related concepts: loss curve, model validation, test data
• Core Section: P2-13.3
• Appears in: P2-13.3, P4-4.2, P4-5.1, P4-6.1, P4-6.4

action¶

• Meaning: An action is something an AI agent actually chooses and executes in a given state. In search or reinforcement learning it changes the environment; in a service agent it may mean searching, reading a file, calling an API, editing a document, or reporting a result.
• Why it matters: Actions are where reasoning turns into state change. The available action set controls what the AI agent can accomplish, what risks it can create, and what observations it can gather next. This concept helps separate internal computation from external execution, and it makes clear why tool permissions and action design matter for both capability and safety.
• Related concepts: state, reinforcement learning policy, reward, observation
• Core Section: P1-8.3
• Appears in: P1-7.1, P1-14.3, P3-9.8, P4-2.3

actual target¶

• Meaning: An actual target is the result you truly want to know, reduce, or predict. It may be directly visible in the current data, but in many real settings it appears late or is not directly observable, so it must be distinguished from a proxy target.
• Why it matters: If the actual target is unclear, it becomes hard to tell whether the current prediction problem is solving the original goal or only predicting an intermediate judgment. This concept keeps the original purpose visible when a proxy target is used and clarifies what a later score is really measuring.
• Related concepts: proxy target, target, supervised learning label, target candidate
• Core Section: P3-9.9
• Appears in: P3-9.9

activation function¶

• Meaning: An activation function transforms a weighted sum or intermediate score before it is passed to the next layer or output. It introduces a non-linear response instead of passing a simple linear combination forward unchanged.
• Why it matters: Repeating only linear combinations would not give a deep network much more expressive power. Activation functions let layers form more complex decision boundaries and representations, while also affecting gradient flow and training stability. They explain why deep networks need both stacked layers and non-linear response curves.
• Related concepts: linear combination, activation, hidden layer
• Core Section: P5-1.2
• Appears in: P5-3.1, P5-3.2, P5-3.3, P5-3.4, P5-3.5, P5-3.6

AI agent¶

• Meaning: In an AI-system context, an AI agent is an execution structure that receives a goal, inspects state, chooses actions, observes results, and continues across multiple steps. The general English word agent does not mean AI by itself; in this book, this entry refers to service or system structures that coordinate search, tool use, review, and retries.
• Why it matters: A single chatbot answer and a multi-step AI agent workflow are different things. AI agent quality depends not only on the model, but also on planning, tool selection, state tracking, stopping conditions, and recovery from failed steps. This concept also helps distinguish service AI agents from reinforcement-learning agents.
• Related concepts: tool use, permission, observation, human oversight, reinforcement learning agent
• Core Section: P1-14.3
• Appears in: P1-14.4, P1-14.5, P1-14.6, P6-15.1, P6-15.2, P6-16.1, P6-16.2, P7-6.1, P7-6.2, P4-2.3

algorithm¶

• Meaning: An algorithm is a defined procedure for solving a problem by following a sequence of steps. Given the same input and conditions, it should make the calculation or decision process explicit enough to repeat.
• Why it matters: In AI and machine learning, the procedure that performs the computation and the heuristic that chooses which candidates to inspect first are different levels. The concept helps readers see training, search, and optimization as repeatable procedures rather than only intuition.
• Related concepts: heuristic, search, optimization
• Core Section: P4-3.1
• Appears in:

AI, artificial intelligence¶

• Meaning: Artificial intelligence is a broad field and system category for computational methods that perform functions associated with intelligence, such as rule-based reasoning, search, learning, generation, and planning. It is not the name of one model or one recent technology.
• Why it matters: AI, machine learning, deep learning, generative AI, and LLMs are often mixed together in casual speech. This concept gives readers the top-level category needed to separate the problem area, the learning method, the model family, and the deployed service. It also helps make claims about AI progress more precise.
• Related concepts: machine learning, deep learning, generative AI, large language model
• Core Section: P1-1.1
• Appears in: P1-index, P4-1.1, P1-1.2, P1-summary

AI learning¶

• Meaning: AI learning is the broad idea that performance on a specific task improves after experience with data. It should not be read as proof that a model understands meaning like a person; it is a way to ask whether the model's behavior improved under a given dataset, objective, and performance standard.
• Why it matters: The Korean word for learning can easily cover both learning and training, so the broad result of performance improvement must be separated from the concrete procedure of adjusting internal values. This distinction keeps the phrase the model learned from becoming a vague claim about intelligence, and pushes us to ask what improved under which data, task, and metric.
• Related concepts: model training, model, generalization, inference
• Core Section: P1-5.1

AI alignment¶

• Meaning: AI alignment is the design problem of making a model's behavior fit human intent, safety rules, and policy constraints. It is not only about understanding instructions, but also about avoiding unacceptable outputs and acting cautiously in sensitive contexts.
• Why it matters: A more capable model is not automatically a safer model. AI alignment separates usefulness from acceptable behavior and shows why performance, truthfulness, refusal boundaries, and safety constraints must be considered together. It also separates model behavior from external controls such as permissions and approval workflows.
• Related concepts: instruction tuning, prompt, metric, human oversight
• Core Section: P6-10.2
• Appears in: P6-17.1

ANN, approximate nearest neighbor¶

• Meaning: Approximate nearest neighbor search finds sufficiently close vector candidates quickly instead of always guaranteeing the exact closest item. It intentionally trades perfect optimality for speed at scale.
• Why it matters: Large vector collections make exact comparison expensive. ANN methods explain why real search systems often prefer fast, good-enough candidates over exact answers that arrive too slowly. This concept connects vector search quality with latency, memory use, and service constraints.
• Related concepts: nearest neighbor, search index, vector database
• Core Section: P1-13.4
• Appears in: P6-3.4

Attention¶

• Meaning: Attention is a structure that lets a model weight different positions in an input sequence when producing an output. Instead of treating every input position equally, it gives more influence to the parts that are most relevant to the current computation.
• Why it matters: Attention reduces the bottleneck of compressing a long input into one fixed vector and marks a key transition toward Transformer-based models. It prepares readers for self-attention, parallel attention-head structures, and long-context behavior by showing that a model can redistribute focus depending on the task.
• Related concepts: Transformer, language modeling, self-attention, fixed-length vector
• Core Section: P1-11.2
• Appears in: P1-9.3, P5-12.2, P5-13.1, P5-13.2, P5-14.1, P5-14.2, P6-4.2

automatic differentiation¶

• Meaning: Automatic differentiation is a method where a program records calculation steps and uses that record to compute derivatives automatically. Instead of manually deriving one large formula, it follows the operations that actually ran.
• Why it matters: It is the practical basis that lets deep-learning frameworks run backpropagation. It shows that gradients are organized through recorded computation paths, intermediate values, and local derivative rules.
• Related concepts: computation graph, backpropagation, gradient
• Core Section: P5-5.2

AI risk¶

• Meaning: The possibility that an AI system can cause harm through errors, misuse, unsafe deployment, bias, privacy exposure, or loss of control.
• Why it matters: It helps readers separate useful capability from the conditions under which that capability becomes unsafe or socially costly.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P1-10.3
• Appears in: P1-15.1

attribution¶

• Meaning: The act of clearly indicating the source, author, or origin of material that is quoted, summarized, reused, or transformed.
• Why it matters: It matters because responsible learning material must let readers trace where claims, examples, and borrowed expressions came from.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P1-15.2

AI ethics¶

• Meaning: The study and practice of asking what harms, responsibilities, rights, and social effects must be considered when AI systems are built or used.
• Why it matters: It keeps AI discussion from stopping at technical capability and forces model use, data use, safety, accountability, and human impact to be checked together.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P1-15.1

autoregressive model¶

• Meaning: A model that generates or predicts the next value by conditioning on earlier values in the sequence.
• Why it matters: It is a key way to read language models and sequence models, where each new output depends on what has already been produced.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P1-9.2
• Appears in: P1-9.3