Concept Glossary: H¶

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

hallucination¶

• Meaning: Hallucination is an error in which a generative AI system produces unsupported or false content in a fluent and confident form. The key point is that hallucination is not limited to obviously strange output; a polished sentence can still be empty, outdated, or wrong when evidence is missing.
• Why it matters: Hallucination is one of the central risks of generative AI because it separates fluency from correctness. Understanding it helps readers distinguish problems that can be reduced by better prompting from problems that require retrieval, source checking, review, or human confirmation.
• Related concepts: confabulation, evidence, review, factuality, limit of prompting
• Core Section: P1-12.3
• Appears in: P1-10.3

hand-crafted features¶

• Meaning: Hand-crafted features are input cues or attributes that people design before a model learns. Examples include edges, color distributions, textures, lengths, or other measurements chosen because a human believes they will matter for the task.
• Why it matters: This concept shows why earlier image-recognition workflows depended heavily on human feature design, and why representation learning became an important shift. It helps readers compare features selected by people with representations learned inside a model.
• Related concepts: feature, image recognition, learned representation
• Core Section: P1-9.1

harness¶

• Meaning: A harness is an execution structure that wraps model and tool runs so inputs, steps, logs, and evaluation results can be repeated and compared. It is not the model itself, but the outer structure that makes the model measurable under consistent conditions.
• Why it matters: Agentic workflows need to know what ran, where it failed, and whether a change improved behavior under the same conditions. A harness turns one-off demonstrations into repeatable measurement by collecting inputs, execution rules, traces, and evaluation outputs in one frame.
• Related concepts: agent, log, evaluation, trace
• Core Section: P6-15.2
• Appears in: P1-14.5, P1-15.1, P6-14.1, P6-15.1, P6-16.1, P6-17.2, P6-18.2

hash table¶

• Meaning: A hash table is a data structure that uses a hash-based rule to map keys to likely storage locations so values can be found quickly. Python dictionaries are a familiar example for mappings such as user ID -> profile or setting name -> setting value.
• Why it matters: Hash tables explain why key-based lookup is different from scanning a list in order. They are useful when a program needs to retrieve or update the value attached to a name, identifier, or configuration key without checking every item one by one.
• Related concepts: key, data structure, operation
• Core Section: P2-9.4
• Appears in: P2-9.1

heuristic¶

• Meaning: A heuristic is an experience-based rule that does not guarantee the best answer but helps search more promising candidates first. It is closer to a practical guide for limited time than to a proof that always returns the optimal result.
• Why it matters: Many AI problems have too many paths or combinations to inspect exhaustively. Heuristics help readers distinguish a guaranteed solution method from a strategy that tries to find a useful answer quickly enough for the situation.
• Related concepts: search, state, goal
• Core Section: P4-3.1
• Appears in: P1-2.2, P1-7.2, P4-3.2

heuristic function¶

• Meaning: A heuristic function turns the promise of a candidate state into a score or estimate that can guide search order. It is not usually a formula for the final answer; it is a rule for deciding which candidate to inspect first.
• Why it matters: Heuristic search is not just intuition. A heuristic function makes the ordering rule computational, so candidates can be compared and prioritized even when the score is only an approximation.
• Related concepts: heuristic, search, good-enough solution
• Core Section: P1-7.2
• Appears in: P1-2.2, P1-7.1, P1-7.3

heuristic score¶

• Meaning: A heuristic score is a relative number used to compare how promising candidates look. It should usually be read as a ranking signal, not as a calibrated probability that a candidate is correct.
• Why it matters: Numbers in AI workflows do not all mean probability. This concept helps readers avoid treating a larger heuristic score as a guaranteed truth, and instead ask what the score is designed to order or prioritize.
• Related concepts: heuristic, heuristic function, probability estimate
• Core Section: P1-7.3
• Appears in: P1-7.2

hidden layer¶

• Meaning: A hidden layer is a neural-network layer between the input layer and the output layer. It builds intermediate representations that are not directly read as the final answer but help transform inputs into useful internal features.
• Why it matters: Hidden layers explain why deep learning is more than a direct input-output mapping. They let a network reshape data through several internal stages, so later layers can use more useful representations than the raw input alone.
• Related concepts: perceptron, multilayer neural network, activation function
• Core Section: P5-2.1
• Appears in: P5-2.2, P5-7.1

hidden state¶

• Meaning: Hidden state has two common meanings in this book. In recurrent neural networks, it is the internal state that carries information from earlier inputs into later computation. In notebook execution, it can mean runtime values or settings that are not visible on the page but still affect later cells.
• Why it matters: The same phrase can describe model memory or execution-environment state. Understanding the distinction helps readers explain why the same visible input may produce different results when unseen internal state has changed.
• Related concepts: RNN, recurrent neural network, execution order, runtime state
• Core Section: P1-11.2
• Appears in: P2-10.3

histogram¶

• Meaning: A histogram divides numeric values into bins and shows how many values fall into each bin. Instead of inspecting every point, it reveals where values are concentrated, sparse, skewed, or spread out.
• Why it matters: A mean alone can hide the shape of data. Histograms help readers inspect distribution, outliers, and multiple clusters, while also noticing that the chosen bin width can change the visible pattern.
• Related concepts: distribution, outlier, plot
• Core Section: P2-13.2
• Appears in: P2-13.3, P4-16.1

hold state¶

• Meaning: Hold state is a run state in which execution pauses because an external condition is not yet open, such as missing approval, a pending human response, or an unmet permission boundary. It is not the same as a tool failure.
• Why it matters: Treating hold state as ordinary failure obscures the next action. This concept helps readers distinguish retrying a broken step from waiting for approval, reporting a pause, or preserving state until a safe next step is allowed.
• Related concepts: approval policy, approval, state, retry, run record
• Core Section: P7-6.3
• Appears in: P7-6.1, P7-6.2

host¶

• Meaning: A host is the outer application or execution environment that a user interacts with. In an MCP context, it owns one or more clients, manages requests, applies tool-use policy, and shapes the user-facing workflow.
• Why it matters: A host is not the same thing as a server. The distinction helps readers separate the app the user sees, the client that opens a connection, and the server that provides tools or resources.
• Related concepts: client, server, Model Context Protocol, MCP, approval
• Core Section: P1-14.4

human evaluation¶

• Meaning: Human evaluation is an assessment method in which people judge output quality by reading context, tone, usefulness, factuality, safety, and possible misunderstanding. It often uses a rubric so multiple reviewers can compare results more consistently.
• Why it matters: Automatic metrics can miss subtle quality and real-use problems, especially in generative AI. Human evaluation supplies a broader judgment, but it also needs criteria and agreement procedures because human judgment is not automatically consistent.
• Related concepts: evaluation, automatic evaluation, review, factuality
• Core Section: P6-16.2
• Appears in: P6-17.1

human oversight¶

• Meaning: Human oversight is a control structure that lets people review AI output, intervene, stop execution, reject results, or require changes when needed. It is stronger than a formal approval step that has no real power to alter the outcome.
• Why it matters: In high-risk automation, responsibility requires meaningful intervention points. This concept helps readers distinguish the mere presence of a human from a design in which a human can actually inspect, stop, and correct the system.
• Related concepts: safety, accountability, harness, approval
• Core Section: P1-15.1
• Appears in: P1-10.3

hyperparameter¶

• Meaning: A hyperparameter is a setting chosen before training that controls how a model is trained or shaped. Examples include learning rate, number of layers, tree depth, and other values adjusted from outside the learned model parameters.
• Why it matters: Hyperparameters are different from parameters learned from data. The distinction helps readers understand tuning, validation, and whether a performance change came from learned internal values or from the conditions under which training was run.
• Related concepts: parameter, tuning, validation
• Core Section: P4-9.1
• Appears in: P4-9.2, P4-9.3, P4-14.2, P4-16.2