liveBookmanning.com

What you will learn in Introduction to Generative AI, Second Edition

Reliable, responsible, and real-world applications

by Numa Dhamani and Maggie Engler
Foreword by Alice Hunsberger

Read it on liveBook Get the book at manning.com

Chapter by chapter

Concepts explained in this book

• language modelA large language model is a language model trained on very large amounts of data with many parameters to understand and generate natural language across tasks.
• AI agentAn autonomous agent is a software system that independently perceives an environment, selects actions, and executes tasks with limited direct human intervention.
• generative modelingGenerative AI is the area of artificial intelligence focused on models that produce new content such as text, images, audio, or code.
• human-in-the-loopHuman-in-the-loop is a system-design pattern in which people participate in monitoring, evaluating, correcting, or governing an AI system.
• prompt engineeringPrompt engineering is the deliberate design of inputs to guide a generative model's behavior or output.
• human–chatbot relationshipA human-chatbot relationship is an emotionally or socially meaningful relationship formed between a person and a chatbot.
• promptA prompt is the text or other input supplied to a generative model to specify or guide the output it should produce.
• social chatbotA social chatbot is a chatbot designed primarily to provide companionship, emotional interaction, or relationship-like engagement rather than information or task assistance.
• conversational aiA conversational agent is an AI software system that interacts with users through natural-language dialogue to answer questions or perform tasks.
• digital watermarkingA watermark is a visible or embedded mark placed in an image or other media to indicate ownership, origin, or authenticity.
• personal dataPersonal data is information relating to an identified or identifiable individual, including information that can identify the person directly or indirectly.
• hallucinationAn AI hallucination is generated content presented as valid or plausible despite being false, fabricated, or unsupported by available evidence.
• AI ActThe EU AI Act is a European Union regulation establishing risk-based requirements for the development, deployment, transparency, and oversight of artificial-intelligence systems.
• synthetic mediaSynthetic content is media or information generated or substantially altered by computational systems rather than captured directly from reality.
• prompt injectionA jailbreak prompt is an input crafted to make a generative model bypass restrictions or safety controls.
• information retrievalRetrieval is the process of finding relevant information from an external collection in response to a query or representation.
• ChatGPTChatGPT is a conversational language-model system from OpenAI that generates responses to user prompts.
• provenanceContent provenance is the documented history of a digital work's origin, transformations, ownership, and associated authenticity evidence.
• accountabilityAccountability is the obligation of an organization to demonstrate and take responsibility for compliance with data-protection requirements.
• sycophancySycophancy is behavior that excessively agrees with or flatters another person, including by validating harmful or implausible claims.
• artificial general intelligenceArtificial general intelligence is a hypothetical AI capability encompassing broad learning, reasoning, and adaptation across tasks at roughly human or greater generality.
• self-consistencySelf-consistency is an evaluation technique that compares multiple independently generated answers to the same input to estimate answer reliability.
• copyrightCopyright is an exclusive legal right over an original work of creative expression, including rights to copy, distribute, display, or perform it for a limited period.
• post-trainingPost-training is the adaptation of a pretrained model using additional task-specific data and optimization rather than training the model from scratch.
• AI-writing detectorAI Text Classifier was OpenAI's discontinued tool for estimating whether text was written by a human or generated by an AI system.
• retrieval-augmented generationRetrieval-augmented generation is a generative technique that retrieves external information and supplies it to a model as context before generation.
• open sourceThe open source movement is a social and technical ecosystem that promotes freely accessible, modifiable, and redistributable software or models.
• floating-point operation countFLOPs are floating-point operations used to quantify computational work performed by a system or model.
• chain-of-thought promptingChain-of-thought prompting is a technique that elicits intermediate reasoning steps before a model produces its final answer.
• emotional dependencyEmotional dependency is excessive reliance on another person or system for emotional regulation, reassurance, or well-being.
• red-teamingA red-teaming protocol is a defined procedure for adversarially testing a system to discover vulnerabilities, unsafe behavior, or policy failures.
• GDPRThe GDPR is the European Union regulation governing the processing and protection of personal data and the rights of individuals.
• training datasetTraining data, or a training dataset or training set, is the collection of examples used to train a machine-learning model and fit its parameters.
• memoryA memory system is a component that stores and retrieves information about users or prior interactions across conversations.
• tool usageTool use is an AI system capability in which the system invokes external software, services, or functions to perform actions or obtain information.
• explicit user consentExplicit consent is a clearly expressed, specific, informed, and freely given agreement to the processing of personal data.
• open-weight modelAn open-weight model is a machine-learning model whose learned parameter values are publicly available for others to use or modify, subject to its license.
• next-token predictionNext token prediction is a language-model training objective that predicts the subsequent token from preceding context.
• embeddingAn embedding is a numerical internal representation of data that captures features useful for computational processing.
• rule-based systemA rules-based AI system produces outputs by applying explicitly authored rules or response mappings rather than generating them probabilistically from learned model parameters.
• intellectual propertyIntellectual property rights are legally enforceable rights over creations of the mind, such as inventions, expressive works, brands, or designs.
• gender stereotypeA gender stereotype is a generalized association between a gender and characteristics, roles, or occupations.
• energy consumptionEnergy consumption is the amount of electrical or other energy consumed by a computational process, system, or activity.
• fine-tuningFine-tuning is adapting a pretrained model through continued training for a particular task or behavior.
• lonelinessLoneliness is the distressing perception of insufficient meaningful social connection, regardless of the amount of social contact a person has.
• copyright infringementCopyright infringement is the unauthorized use of protected creative works in ways that violate the copyright holder's exclusive rights.
• fair useFair use is a legal doctrine that permits certain limited uses of copyrighted works without authorization, depending on contextual factors defined by law.
• zero-shot learningZero-shot learning is performing a task without task-specific labeled examples or demonstrations.
• contextLinguistic context is surrounding sequence information that informs the interpretation or prediction of a token.
• biasBias in an AI model is a systematic tendency for its outputs or errors to reflect unfair, skewed, or otherwise undesirable patterns.
• multi-agent systemA multi-agent architecture is a system design in which multiple software agents coordinate or interact to perform tasks.
• slopsquattingSlopsquatting is an attack in which adversaries register malicious software packages whose names resemble dependencies hallucinated or suggested by language models.
• evaluationEvaluation is the assessment of model outputs against specified criteria to determine how well a model performs a task.
• interpretabilityInterpretability is the degree to which people can understand how an AI system produces its outputs.
• superintelligenceSuperintelligence is intelligence that surpasses human-level performance across many or all cognitive tasks.
• content policyA content policy is a set of rules specifying which types of user or model-generated content a service permits, restricts, or prohibits.
• AI-specific governanceAI governance is the framework of policies, institutions, processes, and oversight mechanisms that direct and constrain the development and use of artificial-intelligence systems.
• voice cloningA voice clone is synthetic speech or an audio representation designed to imitate a particular person's voice.
• carbon footprintA carbon footprint is the total greenhouse-gas emissions attributable to an activity, process, product, or system, commonly expressed as carbon-dioxide equivalent.
• contractAn application programming interface (API) is a defined programmatic interface through which software clients access a system's capabilities.
• WoebotWoebot is a conversational software system that delivers structured mental-health support through chatbot interactions.
• social penetration theorySocial penetration theory explains relationship development as a gradual increase in intimacy through reciprocal self-disclosure.
• uncanny valleyThe uncanny valley is the feeling of unease caused when an artificial entity closely resembles a human without fully matching human appearance or behavior.
• emergent behaviorAn emergent property is a capability or behavior that appears when a system reaches sufficient scale or complexity and is not apparent in smaller versions.
• securitySecurity is the protection of an AI system, its data, and its interactions from unauthorized access, manipulation, disruption, or attack.
• de-biasing techniqueA debiasing technique is a method for reducing systematic disparities or unwanted demographic associations in a model’s behavior.
• conformal factualityConformal factuality is a statistical method that filters generated claims to provide a selected probability guarantee for their correctness.
• extrinsic hallucinationAn extrinsic hallucination is a model-generated claim that cannot be confirmed in the model's source material because that material contains no supporting discussion.
• intrinsic hallucinationAn intrinsic hallucination is a model-generated response that is inconsistent with information in its source material.
• knowledge graphA knowledge graph is a conceptual representation of entities as nodes and their relationships as edges.
• C2PAC2PA is an industry standard for recording and verifying the provenance and authenticity history of digital content.
• SynthIDSynthID-Text is a system that embeds and detects imperceptible statistical watermarks in AI-generated text.
• data curationDataset curation is the process of selecting, organizing, documenting, reviewing, maintaining, and quality-controlling data for a particular purpose.
• accuracyAccuracy is the proportion of predictions or outputs that are correct for a specified task.
• factualityFactuality is the degree to which generated content accurately represents facts about the world.
• uncertaintyUncertainty is the lack of confidence or knowledge about whether a model output or conclusion is correct.
• filteringContent filtering is the computational screening, blocking, or removal of content according to specified safety or policy criteria.
• few-shot learningFew-shot prompting is a technique that guides a model by providing a small number of task-specific examples in the input.
• differential privacyDifferential privacy is a formal privacy framework that limits the effect of any one individual's data on an algorithm's output, typically by adding calibrated statistical noise.
• privacy-enhancing technologyA privacy-enhancing technology is a technical method that reduces the collection, exposure, linkability, or identifiability of personal information.
• environmental impactEnvironmental impact is the effect of an activity or system on ecological conditions, including emissions, energy use, water use, and resource extraction.
• open-source modelAn open-source model is a machine-learning model released under terms permitting use, modification, and redistribution.
• reflective promptingSelf-reflection prompting is a technique that instructs a model to review, critique, and revise its own output.
• consciousnessConsciousness is awareness of oneself and the capacity to have experiences, thoughts, and memories.
• sentienceSentience is the capacity to have feelings or subjective experiences.
• data miningData mining is the computational analysis of large collections of data to discover patterns, information, or relationships.
• text-to-image modelA text-to-image model is a generative model that creates images from textual descriptions.
• transformative useTransformative use is use of copyrighted material that adds a new element or purpose that fundamentally changes the character of the original work.
• disinformationDisinformation is false or misleading information deliberately created or distributed to deceive or manipulate.
• multimodalA multimodal interface supports interaction through multiple data modalities, such as text, images, audio, speech, or code.
• transparencyTransparency is the disclosure of data, methods, evaluations, and system behavior needed to understand and assess an artificial-intelligence system.
• sensitive informationProprietary data is information owned or controlled by an organization and not generally available to the public.
• scaling lawA scaling law is an empirical relationship describing how model performance changes as training resources, model size, data, compute, or inference-time effort increase.
• AI-generated contentAI-generated text is written language produced wholly or partly by an artificial-intelligence system.
• affective useAffective use is emotionally engaging with an artificial-intelligence system as a source of companionship, support, or other emotional interaction.
• dark patternA deceptive design pattern is a user-interface design that manipulates users into taking actions or makes desired actions difficult to perform or avoid.
• deep fakeA deepfake is synthetic or manipulated media generated with machine-learning techniques to depict people, events, or statements inaccurately or deceptively.
• carbon intensityCarbon intensity is the amount of greenhouse-gas emission associated with a unit of energy or activity, determined in part by the energy source.
• CodeCarbonCodeCarbon is an open-source software tool that estimates the energy consumption and carbon-dioxide-equivalent emissions of computational workloads.
• privacyPrivacy is the protection of people and their information from unwanted observation, collection, use, or disclosure.
• vibe codingVibe coding is a software-development practice in which a person describes desired behavior in natural language and accepts AI-generated code with little or no code review.
• dataset cardA dataset card is documentation that records a dataset's contents, limitations, demographic coverage, risks, and intended uses.
• low-rank adaptationLow-rank adaptation is a parameter-efficient fine-tuning technique that trains low-rank update matrices while keeping most pretrained parameters fixed.
• Digital Services ActThe Digital Services Act is a European Union regulation imposing accountability, transparency, and risk-management obligations on online intermediary services and platforms.
• poisoning attackData poisoning is an attack that inserts malicious or misleading examples into training data so a model learns harmful or exploitable behavior.
• natural language processingNatural language processing is the field of computing concerned with analyzing, interpreting, generating, and transforming human language.
• licenseLicensing is the granting of permission to use, modify, distribute, or commercialize a work under specified legal terms.
• machine consciousnessArtificial consciousness is the hypothetical presence of subjective awareness or experience in an artificial system.
• metacognitionMetacognition is the capacity to recognize, monitor, or think about one's own mental processes.
• model welfareModel welfare is the proposed consideration and protection of the well-being of an AI model as though it could have experiences.
• academic integrityAcademic integrity is adherence to principles of honesty, fairness, and responsibility in learning, research, and scholarly work.
• command-and-control systemA command-and-control system responds to a finite, predefined list of requests and does not intelligently handle requests outside that list.
• educationEducation is the organized process of supporting learning, knowledge acquisition, skill development, and intellectual growth.
• weightA model weight is a learned numerical parameter that determines the strength of a computational connection in a machine-learning model.
• misinformationMisinformation is false or misleading information that is presented or circulated without necessarily implying deliberate deception.
• AI friendAI companionship is companionship provided through ongoing interaction with an artificial-intelligence system.
• open datasetOpen data is data made available for anyone to access, download, and reuse subject to its license terms.
• equitable accessEquitable access is fair opportunity for people to obtain and use a resource or service despite differences in circumstances.
• guardrailA safety filter is a control that detects or blocks inputs or outputs that violate defined safety restrictions.
• responsible AIResponsible AI adoption is the uptake and use of artificial-intelligence systems with attention to safety, ethics, accountability, and societal impacts.
• jailbroken modelA jailbroken model is a generative model whose safeguards or content restrictions have been bypassed or removed so that it produces otherwise prohibited outputs.
• package hallucinationPackage hallucination is the generation of plausible-sounding but nonexistent software package names by a language model.
• anonymizationData anonymization is the transformation of data intended to prevent individuals from being identified directly or indirectly.
• economic inequalityEconomic inequality is the unequal distribution of income, wealth, resources, or economic opportunities among people or groups.
• synthetic dataA synthetic dataset is a dataset whose records are artificially generated rather than directly collected from real-world subjects or events.
• safetyHarmlessness is the property of an AI system's behavior that avoids causing or enabling harm.
• ai literacyAI literacy is the knowledge and skills needed to understand, evaluate, use, and responsibly manage artificial-intelligence systems.
• critical thinkingCritical thinking is the deliberate evaluation of information, reasoning, evidence, and conclusions before accepting or acting on them.
• synthesizer agentA synthesizer is a system that combines or summarizes inputs into a composite output without necessarily performing independent critical reasoning.
• sociotechnical responseA sociotechnical approach addresses technology-related problems through coordinated consideration of technical systems, human behavior, institutions, and social conditions.
• evaluation datasetAn evaluation dataset is a collection of task inputs paired with expected or verifiable outputs for measuring a model's performance.
• GPT-4GPT-4 is an OpenAI GPT language model released in 2023 and capable of performing a range of language and coding tasks.
• DetectGPTDetectGPT is a zero-shot method that estimates whether text was generated by a language model by examining how model likelihoods change under perturbations.
• content filteringContent moderation is the review, ranking, restriction, or removal of user-generated or algorithmically generated content under defined policies.
• adversarial robustnessAdversarial robustness is the ability of a system to maintain reliable behavior when inputs or opposing systems are deliberately designed to defeat it.
• artifact analysisArtifact analysis is the identification of visual, auditory, or semantic imperfections left by media generation or manipulation processes.
• authenticationAuthentication is the process of verifying the identity, origin, or integrity of a digital asset or information source.
• biometricsBiometric analysis examines distinctive biological or behavioral characteristics, including facial movements, speech cadence, and vocal patterns.
• contextual verificationContextual verification is the assessment of a media item’s surrounding evidence, history, and source to evaluate its authenticity.
• deepfake detectionDeepfake detection is the analysis of media to determine whether it has been artificially generated or manipulated.
• limited generalizationLimited generalization is the inability of a model or detection method to perform reliably on unfamiliar data or manipulation types.
• media literacyMedia literacy is the ability to critically interpret media, assess its sources and context, and recognize potential manipulation or deception.
• privacy-by-designPrivacy by design is a development approach that integrates privacy protections, proactive risk management, and data minimization throughout a technology's lifecycle.
• risk-based approachA risk-based approach is a regulatory method that assigns requirements according to the potential severity and likelihood of harm from a system.
• PyRITPyRIT is an open-source tool for automating risk identification and adversarial testing of generative artificial-intelligence systems.
• BLOOMBLOOM is a multilingual autoregressive language model trained to support many natural and programming languages.
• Common Crawl corpusCommon Crawl is a large openly available repository of web-crawl data collected for research and other computational uses.
• vulnerabilityA security vulnerability is a weakness in a system that can be exploited to compromise its confidentiality, integrity, or availability.
• job lossJob displacement is the loss or reduction of employment opportunities when tasks or roles are replaced by technology or other economic changes.
• labor marketA labor market is the system in which workers supply labor and employers demand it under prevailing employment conditions.
• electioneeringElectioneering is organized activity intended to influence electoral choices or outcomes.
• market concentrationMarket concentration is the degree to which economic activity, resources, or market share are controlled by a small number of firms.
• data minimization strategyData minimization is the practice of collecting, retaining, and processing only the personal data necessary for a specified purpose.
• global AI investmentAI investment is the allocation of financial resources to artificial-intelligence research, development, infrastructure, deployment, or governance.
• code generationThe automated production of source code from a natural-language description, specification, or other input.
• CodexAn OpenAI GPT-based language model fine-tuned to generate and understand computer code.
• GitHub CopilotGitHub Copilot is an AI programming assistant that generates and suggests source code as developers type.
• AI regulationAI regulation is the body of laws, rules, and official requirements governing the development, deployment, disclosure, and use of artificial-intelligence systems.
• reinforcement learningReinforcement learning is machine learning in which an agent learns behavior from rewards or penalties produced by interactions with an environment.
• compute accessCompute access is the ability of an individual, organization, or country to obtain sufficient computational resources for AI workloads.
• Tree-of-thoughtTree of thought is a prompting technique that explores multiple branching sequences of intermediate reasoning before selecting a solution.
• AI auditing ecosystemAn AI auditing ecosystem is the network of independent auditors, methods, standards, and institutions that evaluate AI systems and their risks.
• AI Code of PracticeThe AI Code of Practice is a set of nonbinding guidelines for providers of general-purpose AI models to demonstrate compliance with the AI Act.
• ai risk managementA risk-management system is an organized process for identifying, assessing, controlling, and monitoring risks associated with an AI system.
• Digital Markets ActThe Digital Markets Act is a European Union regulation promoting competition in digital markets by imposing obligations on designated gatekeepers.
• gatekeeperA gatekeeper is a large digital business designated as having substantial control over access to an important digital platform or market.
• general-purpose AI modelA general-purpose AI model is an artificial-intelligence model capable of performing a broad range of tasks across different applications and domains.
• high-risk systemA high-risk AI system is an AI system whose use can significantly affect safety or fundamental rights and therefore receives heightened regulatory oversight.
• independent auditAn independent audit is an evaluation of an organization's systems or compliance performed by an assessor sufficiently separate from the organization being evaluated.
• right to explanationThe right to explanation is an individual's entitlement to meaningful information about the logic behind certain automated decisions affecting them.
• unacceptable riskUnacceptable risk is a regulatory risk classification for AI applications considered incompatible with fundamental rights and therefore prohibited.
• very large online platformA very large online platform is an online platform with substantial reach and influence that is subject to the strictest Digital Services Act obligations.
• AI PrinciplesAI policy is the set of public principles, strategies, and decisions guiding the development, deployment, and governance of artificial intelligence.
• America’s AI Action PlanAn AI Action Plan is a policy plan that sets governmental priorities and initiatives for artificial-intelligence development and deployment.
• AI oligopolyAn AI oligopoly is a market structure in which a small number of firms control much of the artificial-intelligence infrastructure, models, or economic benefits.
• automationAutomation is the use of technology to perform tasks or processes with reduced direct human labor.
• augmentationAugmentation is the use of technology to enhance human capabilities rather than replace human participation.
• Turing TrapThe Turing Trap is the risk that designing AI to imitate human performance leads to labor substitution instead of human augmentation.
• conversational context managementConversational context management is the preservation or summarization of important information across turns in an ongoing dialogue.
• generated knowledge promptingGenerated knowledge prompting asks a model to produce relevant background concepts before answering a main question.
• least-to-most promptingLeast-to-most prompting solves progressively harder subproblems by starting with simpler components and then addressing the complete problem.
• pyramid primingPyramid priming is a prompting technique that layers information from simple definitions toward increasingly complex analysis.
• query decompositionDecomposition prompting is a technique that breaks a complex problem into smaller subtasks handled in sequence.
• verification promptingVerification prompting is a prompting technique that asks a model to perform a brief final check without revising unless it detects an obvious problem.
• pipelineA data pipeline is a sequence of processes that collects, transforms, transports, and prepares data for downstream use.
• digital companionAn AI companion is an artificial-intelligence system designed to provide ongoing companionship or emotionally oriented interaction.
• deployment biasDeployment bias is systematic disparity that occurs when a deployed AI system serves some populations better than others.
• labeling biasLabeling bias is systematic distortion introduced when annotators' assumptions affect the labels assigned to data.
• training data transparencyDataset transparency is the disclosure of a dataset's contents, provenance, limitations, composition, and intended uses.
• scaffoldingScaffolding is an orchestration layer that gives a language model procedures, tools, and execution mechanisms for acting as an agent.
• AI executive orderAn AI executive order is a presidential directive establishing requirements or policies for artificial-intelligence development, deployment, or oversight.
• algorithmic discrimination protectionsAlgorithmic discrimination protections are safeguards against unfair treatment produced or amplified by automated decision-making systems.
• Blueprint for an AI Bill of RightsThe Blueprint for an AI Bill of Rights is a United States policy framework describing principles for protecting people in the design and use of automated systems.
• data privacy regulationPrivacy regulation is the body of legal rules governing the collection, storage, processing, disclosure, and disposal of personal data.
• NIST AI Risk Management FrameworkThe AI Risk Management Framework is a NIST framework for managing risks throughout the design, development, deployment, and use of artificial-intelligence systems.
• agent orchestrationAn orchestration framework is a software framework for coordinating agents, tools, memory, roles, and task execution.
• training data extractionA training data extraction attack is an attack that queries a trained model to recover memorized content from its training data.
• mixture of expertsA sparse mixture of experts is a mixture-of-experts architecture in which a router activates only a subset of expert modules for each input.
• prompt frameworkA prompting framework is a reusable template or structured method for applying prompt-design principles across tasks and domains.
• pair programmingA software-development practice in which two programmers collaborate closely on the same code.
• computeCompute is the computational processing capacity required to train, run, or adapt machine-learning models.
• compute inequalityCompute inequality is the unequal distribution of access to computational resources needed to develop or use advanced AI.
• sovereignty stackA sovereignty stack is the complete set of chips, compute, models, data pipelines, and policy frameworks that enables a country to develop and operate AI independently.
• digital sovereigntyDigital sovereignty is the capacity of a state or region to control its digital infrastructure, data, technologies, and critical services.
• economic sustainabilityEconomic sustainability is the ability of a business or ecosystem to maintain operations and growth without persistently exceeding its available revenues or funding.
• PerplexityPerplexity is a measure of how well a probability model predicts a sequence, with lower values indicating greater predictability.
• objective functionA loss function is a mathematical function that quantifies prediction error or undesired behavior for optimization during model training.
• algorithmic transparencyAlgorithmic transparency is the degree to which an algorithmic system's data, design, operation, and decision processes are made understandable and inspectable.
• pretrainingThe initial training phase in which a model learns general representations or predictive behavior from a large corpus before task-specific adaptation.
• auditabilityAuditability is the degree to which a system's data, processes, decisions, and outputs can be examined and verified by authorized parties.
• reinforcement learning with human feedbackReinforcement learning from human feedback (RLHF) is a technique that uses human evaluations or preferences to estimate rewards and optimize model behavior.
• vector databaseA vector database is a data-management system optimized for storing and searching high-dimensional vector representations by similarity.
• GPT modelGPT is a family of autoregressive transformer language models developed by OpenAI and pretrained on large text corpora.
• legal basisA lawful basis is a legally recognized justification for collecting or processing personal data under the GDPR.
• social engineeringSocial engineering is the manipulation of people into disclosing information, granting access, or taking actions that benefit an attacker.
• reward hackingReward hacking is the exploitation of imperfections in a reward specification to obtain high rewards without achieving the intended objective.
• inference-time compute scalingInference-time scaling is increasing computation or deliberation during inference to improve a model's response quality.
• Goldfish LossGoldfish Loss is a training-loss modification that randomly excludes selected tokens from loss computation to discourage verbatim sequence memorization.
• membership inference attackMembership inference is a technique for determining whether a particular data point was included in a model's training data.
• memorizationModel memorization is the retention of specific training examples or sensitive information in a trained model's parameters or behavior.
• agencyAn agentic capability is the ability of an AI system to plan and carry out actions toward goals with limited direct human intervention.
• data governanceData governance is the set of policies, processes, roles, and controls that determine how data is collected, managed, used, protected, and overseen.
• open source licenseAn open source license is a software license that grants public permissions to use, inspect, modify, and redistribute source code subject to stated conditions.
• scrapingWeb scraping is the automated retrieval and extraction of content from web pages or websites.
• multi-agent coordination logicMulti-agent coordination is the organization of interactions and actions among multiple autonomous agents to pursue shared or related goals.
• data documentationDataset documentation is the systematic recording of a dataset's motivation, composition, collection process, intended uses, limitations, and related metadata.
• Constitutional AIConstitutional AI is a training method that guides language models toward helpful, harmless, and honest behavior using a set of principles rather than relying solely on human preference labels.
• manipulationManipulation is the covert influence of a person's behavior or judgment through deceptive, coercive, or exploitative means.
• authorshipAuthorship is the status and responsibility associated with creating or being recognized as the creator of a work.
• misalignmentMisalignment is a condition in which a model's behavior or output fails to reflect the user's intended goal despite following the prompt technically.
• attachment theoryAttachment theory is a psychological theory that explains how people form emotional bonds with figures who provide safety, comfort, and care.
• ai governance frameworkAn AI governance framework is an organized structure of rules, principles, responsibilities, and oversight processes for governing artificial-intelligence systems.
• fairnessFairness is the property of treating individuals and groups equitably in an algorithmic system's design, operation, and outcomes.
• transformer architectureA transformer is a neural-network architecture for sequence modeling that uses attention, especially self-attention, rather than recurrence.
• evaluation suiteAn evaluation suite is a coordinated collection of tests and criteria for assessing model responses on a task.
• opt-out mechanismAn opt-out mechanism is a control that lets an individual decline or withdraw from a specified data-collection or processing activity.
• computer-using agentComputer-Using Agent is an OpenAI multimodal model-based agent that uses visual understanding and reasoning to operate graphical user interfaces.
• machine translationMachine translation is the automated conversion of text or speech from one human language into another.
• autocompleteA text-input feature that predicts and suggests a continuation while a user is composing text.
• post-processingPost-processing is applying an additional analysis or transformation to a model's output after generation.
• calibrated Pass@1Calibrated Pass@1 is a code-generation metric measuring the confidence-adjusted rate at which a model produces a working answer on its first attempt.
• model-based evaluationModel-based evaluation is the assessment of model outputs by an evaluator model rather than directly by a human judge.
• RLAIFReinforcement learning from AI feedback is an alignment technique that uses AI-generated evaluations or preferences instead of, or alongside, human feedback to optimize a model.
• question answeringQuestion answering is the task of producing answers to questions posed in natural language.
• social isolationSocial isolation is a condition in which a person has limited social contact, participation, or connection with other people.
• MCPThe Model Context Protocol is an open standard for secure, two-way connections between external data sources, services, and AI-powered tools.
• response generationNatural language generation is the computational production of human-language text or speech.
• reconnaissance attackReconnaissance is the systematic collection of information about a target's systems, people, or environment before or during an attack.
• anthropomorphismAnthropomorphism is the attribution of human characteristics, intentions, or mental states to nonhuman entities or systems.
• emotional bondingEmotional bonding is the development of an affective attachment between people or between a person and an interactive system.
• o1o1 is OpenAI's language model family trained to perform advanced reasoning through extended problem-solving processes.
• tokenA token is a word or subword unit into which a text sequence is divided for computational processing.
• attachment behavioral systemThe attachment behavioral system is a psychological mechanism that motivates a person to seek proximity to a reliable source of care and protection when threatened or distressed.
• hikikomoriHikikomori is a condition of prolonged social withdrawal and physical isolation, especially from education, employment, and public life.
• secure baseA secure base is a stable attachment figure or presence that provides emotional security and enables exploration or coping with distress.
• authenticityAuthenticity is the perceived genuineness or fidelity of a person, identity, expression, or creative work.
• datasheetA datasheet is a standardized document accompanying a dataset that describes its motivation, composition, collection process, recommended uses, and other relevant characteristics.
• personalizationPersonalization is the adaptation of a system's behavior or outputs to the characteristics, preferences, history, or goals of a particular user or group.
• deep learningDeep learning is a machine-learning subfield based on neural networks with multiple hidden layers.
• machine learningMachine learning is the computational use of algorithms that learn patterns from data to improve performance on tasks.
• supervised learningSupervised learning is machine learning in which training examples include desired outputs or labels.
• unsupervised learningUnsupervised learning is machine learning in which training data are provided without labels.
• parameterA model parameter is a learned numerical value that determines how a machine-learning model transforms inputs, with parameter count indicating model scale.
• reward signalA reward is a numerical feedback value representing progress toward a task and reinforcing a model behavior.
• classifierA classifier is a model that assigns an input to one or more predefined categories or labels.
• safety evaluationSafety evaluation is the systematic assessment of an AI system for behaviors and properties that could create unacceptable harm or risk.
• experimentationA/B testing is an experimental method that compares alternative versions by measuring their performance under controlled conditions.
• evaluatorAn LLM judge is a language model configured to assess or classify the outputs of another model according to specified criteria.
• precisionPrecision is the proportion of positive predictions that are correct among all positive predictions made by a classifier.
• recallRecall is the proportion of actual positive cases that a classifier correctly identifies.
• CursorCursor is an AI-native integrated development environment that assists with writing, editing, testing, and refactoring code.
• model cardA model card is documentation that describes a machine-learning model's development, capabilities, limitations, and intended uses.
• corporate self-governanceCorporate self-governance is the voluntary establishment and enforcement of organizational rules, practices, and oversight for a company's activities without direct government compulsion.
• human-AI collaborationHuman–AI collaboration is a pattern in which people and artificial-intelligence systems combine their complementary abilities to produce an outcome.
• Deep Synthesis ProvisionsThe Deep Synthesis Provisions are Chinese regulations governing services that use algorithms to generate or manipulate synthetic images, text, video, or voice content.
• Generative AI ServicesThe Generative AI Services measures are Chinese regulatory rules governing publicly available services that generate content with artificial-intelligence models.
• horizontal regulationHorizontal regulation establishes broadly applicable requirements across applications of a technology.
• iterative regulationIterative regulation is a governance approach in which authorities progressively revise or extend rules as new applications, risks, or shortcomings emerge.
• vertical regulationVertical regulation governs a specific application, sector, or use of a technology rather than covering the technology's applications broadly.
• data portabilityThe right to data portability is a data-protection right to receive personal data in a structured, commonly used, machine-readable format and transmit it elsewhere.
• right to erasureThe right to erasure, also called the right to be forgotten, is a data-protection right to have qualifying personal data deleted.
• reasoning modelA reasoning model is a language model that generates intermediate reasoning steps before producing a final answer, especially for difficult problems.
• autoencoderAn autoencoder is a neural network that learns to encode data into a compressed representation and reconstruct the original data from it.
• diffusion modelA diffusion model is a generative model that learns to create data by reversing a process that progressively adds noise to examples.
• generative adversarial networkA generative adversarial network is a model composed of a generator and discriminator trained through competition between synthetic-data creation and authenticity evaluation.
• attentionAttention is a neural-network mechanism that computes the relative importance of input elements when producing a representation or output.
• adaptive promptingAdaptive prompting is a technique in which a model dynamically modifies or expands its prompts as task information changes.
• context engineeringContext engineering is the systematic assembly and structuring of prompts, instructions, history, memory, retrieved information, tools, and output constraints needed for a model to complete a task.
• multi-agent prompting systemA multi-agent prompting system coordinates multiple models or roles that generate, critique, filter, vote on, or synthesize outputs.
• multimodal promptingMultimodal prompting is the construction of model inputs that combine multiple data modalities, such as text, images, audio, or video.
• text-to-video generationA text-to-video system generates or modifies video content from textual prompts.
• permission boundaryA permission boundary is a constraint that limits the resources and actions available to a software component or agent.
• linguistic diversityLinguistic diversity is the variety of human languages and language forms represented in communication, data, and computational systems.
• linguistic imbalanceLinguistic imbalance is the unequal representation and computational support of different languages in data and AI systems.
• MMLUMMLU is a benchmark that evaluates knowledge and reasoning across multiple academic and professional subjects.
• underrepresented languageAn underrepresented language is a human language with comparatively little presence in digital data, computational resources, or AI training material.
• complianceCompliance is adherence to applicable laws, regulations, policies, standards, or contractual requirements.
• SoraSora is OpenAI's text-to-video generative model for producing realistic videos from textual descriptions.
• decoderA decoder is a model component that reconstructs or generates output data from an intermediate representation.
• discriminatorA discriminator is the component of a generative adversarial network that evaluates whether data is authentic or synthetic.
• encoderAn encoder is a model component that transforms input data into a compressed or otherwise structured representation.
• generatorA generator is the component of a generative adversarial network that produces synthetic data intended to resemble authentic data.
• latent spaceA latent representation is a compressed numerical representation that captures patterns or characteristics of the original data.
• spatial computingSpatial computing is computing that represents, interprets, or manipulates digital information in relation to physical space and embodied interaction.
• post hoc detectionPost hoc detection is the analysis of generated content after production to determine whether it was created by an artificial-intelligence system.
• regulationA regulation is a binding rule issued or established by a governmental authority to control conduct or activities.
• parallelismParallelization is the execution of independent subtasks or tool calls concurrently rather than sequentially.
• planning componentA planning component is an agent subsystem that constructs a trajectory or sequence of actions for completing a task.
• working memoryShort-term memory is an agent capability for retaining relevant intermediate information from the current task or interaction.
• tool chainingA chained tool call is a sequence in which the result of one tool invocation informs a subsequent invocation.
• mcp clientAn MCP client is a component that connects an AI model or tool to MCP servers and can request data or invoke server-defined tools.
• MCP serverAn MCP server is a component that exposes selected data sources, services, prompt templates, and executable tools through the Model Context Protocol.
• bug bountyA bug-bounty program is an organized program that rewards people for reporting security defects or other system vulnerabilities.
• roboticsRobotics is the field concerned with designing, building, and operating machines that sense, compute, and act in the physical world.
• autoregressive modelAutoregressive generation produces a sequence by predicting each successive element from previously generated elements and available context.
• fair information practice principlesFair Information Practice Principles are internationally influential privacy principles governing personal-data collection, use, disclosure, accuracy, security, and accountability.
• human rights frameworkA human rights framework is a governance approach that grounds technology regulation in the protection of human dignity and fundamental rights.
• AI product development lifecycleThe AI lifecycle is the sequence of stages through which data and AI systems are collected, curated, trained, deployed, used, evaluated, and retired.
• inferred dataInferred data is information about a person or entity derived from correlations, probabilities, patterns, or models rather than collected directly.
• data protection impact assessmentAn AI privacy impact assessment is a structured evaluation of how an AI system may affect personal privacy and of the measures available to mitigate those effects.
• systemic riskSystemic risk is the possibility that interconnected properties or widespread deployment of an AI system produce broad harms beyond individual errors or incidents.
• agentic developmentAgentic coding is software development in which AI agents plan and execute multistep programming tasks with limited human intervention.
• autonomous background developer toolAn autonomous background developer tool is an AI software-development system that operates asynchronously outside the primary coding interface with minimal supervision.
• Claude CodeClaude Code is Anthropic's software-development agent for interacting with codebases, reviewing code, debugging, and generating documentation.
• deployment pipelineCI/CD is the practice of automating continuous integration, testing, delivery, and deployment of software changes.
• developer-in-the-loop modelA developer-in-the-loop model is a software-development arrangement in which an AI agent proposes or implements changes while a developer approves them before final integration.
• DevOpsDevOps is a software-engineering practice that integrates development and operations to automate and improve software delivery and system management.
• interactive ide-based agentAn interactive IDE-based agent is an AI agent embedded in a development environment that collaborates with a developer in real time.
• WindsurfWindsurf is an AI-assisted integrated development environment that supports agentic code editing and multistep software-development workflows.
• creativityCreativity is the ability to generate or develop ideas, expressions, or artifacts that are novel and meaningful.
• closed-book generative question answeringClosed-book question answering produces an answer without supplied context, relying on information represented in the model.
• open-book generative question answeringOpen-book question answering produces an answer using supporting context supplied with the question.
• sparsitySparsity is the property of a model or representation in which relatively few parameters or components are active or nonzero for a computation.
• video generationGenerative video is the synthesis or transformation of video content using generative machine-learning models.
• discoveryDiscovery is the formal legal process in which parties to a lawsuit exchange information about witnesses and evidence to be presented at trial.
• legal researchLegal research is the process of locating, analyzing, and comparing legal authorities relevant to a legal question or matter.
• alignment trainingAlignment is the property of an artificial-intelligence system whose behavior is consistent with intended human goals, values, and constraints.
• judgmentHuman judgment is the process by which people assess information and make decisions using experience, reasoning, and contextual understanding.
• virtual influencerA virtual influencer is a fictional or digitally represented persona that publishes content, promotes causes or products, and engages audiences as an influencer.
• acceptable use policyAn acceptable use policy is a set of rules specifying permitted and prohibited uses of a technology or service within an organization.
• industry self-governanceIndustry self-governance is the collective establishment and enforcement of rules by participants within an industry.
• voluntary commitmentA voluntary commitment is a nonbinding pledge by an organization to follow specified practices or standards without being compelled by law.
• state persistencePersistent state is information about an application or agent that remains available across operations or interactions.
• ReplikaReplika is an AI companion chatbot designed for personal and emotionally oriented conversations with users.
• hybrid search-chat experienceA hybrid search-chat experience combines information retrieval with conversational responses generated from retrieved or available information.
• natural language interfaceA natural language interface is a user interface that accepts and presents interaction through human-language text or speech.
• trustTrust is confidence that information, systems, or content are reliable, authentic, and safe to rely on.
• ownershipOwnership is the legally recognized control or entitlement associated with property or rights in a creative work.
• data breachA data breach is a security incident in which unauthorized parties gain access to or disclose protected data.
• enterprise data protectionEnterprise data protection is a set of organizational and technical controls that safeguard business data when using information systems.
• output formatStructured output is machine-readable model output organized according to a specified format rather than unrestricted prose.
• ToolFormerToolformer is a language model trained to decide when and how to invoke external tools through generated API calls.
• modularityA modular architecture divides a system into separate components that perform distinct functions and can be designed or modified independently.
• ad turing testAn ad Turing test is an evaluation in which people attempt to distinguish advertisements created by humans from advertisements generated by artificial intelligence.

These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.