Demystifying Generative AI

Robert Barton is a Cisco Distinguished AI Engineer with Cisco’s AI Software Engineering Group. A graduate of the University of British Columbia in Engineering Physics, he has extensive expertise in networking, cybersecurity, and AI. Rob has authored books on AI, Wi-Fi networks, quality of service, and the Internet of Things (IoT). He has also co-authored numerous peer-reviewed research papers and holds patents in areas such as cybersecurity, cloud networking, and AI/machine learning. As the leader of Cisco’s AI research program, which collaborates with top universities around the globe, Rob is helping drive both research and innovation for academia and industry. He is also a sought-after public speaker at international AI and computer networking conferences and events.Jerome Henry is a Distinguished Engineer at Cisco Systems. A lead researcher in the CTO group, he started embracing AI and generative AI when they were conversation topics only between likeminded peer researchers, in years when access to powerful-enough GPUs was available only to elite groups. By developing new techniques to make AI applicable to several fields of physics and communications, Jerome has contributed to making AI and GenAI mainstream. He holds more than 500 patents, many of them in innovative AI and GenAI schemes, and has authored multiple books on topics ranging from networking, to IoT, to AI. He is based in Research Triangle Park, North Carolina.

• PrefacePart I The Foundations of Generative AIChapter 1 Ten Breakthroughs That Made Generative AI PossibleBreakthrough 1: The Turing MachineBreakthrough 2: The Artificial NeuronBreakthrough 3: The Dartmouth ConferenceBreakthrough 4: The PerceptronThe Rise of Symbolic Reasoning (1960s)The First AI Winter (Early 1970s to Early 1980s)Breakthrough 5: Neural Networks and BackpropagationBreakthrough 6: Recurrent Neural NetworksThe Second AI Winter (Late 1980s to Mid-1990s)Breakthrough 7: Invention of the GPUBreakthrough 8: Reinforcement LearningBreakthrough 9: Language ModelingBreakthrough 10: The TransformerSummaryReferencesChapter 2 The Machinery of LearningTypes of LearningSupervised LearningUnsupervised LearningReinforcement LearningThe Machine Learning Family TreeWhat Is a Model?How Models Are TrainedTraining, Validation, and Test DatasetsInference ModelsHow to Measure Model AccuracyHyperparametersSummaryChapter 3 Foundational AlgorithmsLinear Regression: One Stroke to Represent the DataDescribing a LineLoss Functions and Other HyperparametersClassificationSupport Vector MachinesDiscovering Structures in DataK-Means, the Clustering KingDBSCAN and Growing ClustersSummaryChapter 4 An Introduction to Neural NetworksNeural Networks Key ConceptsANNs: General Structure and TerminologyTraining a Neural NetworkTraining Models and Overcoming ChallengesThe Importance of Clean DataLabeled Data: The Backbone of Supervised LearningAvoiding the Pitfalls: Overfitting and UnderfittingScaling Up TrainingSummaryChapter 5 Neural Network ArchitecturesFeedforward Neural NetworksTraditional FFNsConvolutional Neural Networks (CNNs)Traditional Generative ModelsGenerative Adversarial Networks (GANs)Variational Autoencoders (VAEs)Diffusion ModelsRecurrent ModelsRecurrent Neural Networks (RNNs)Long Short-Term Memory Networks (LSTMs)SummaryChapter 6 Reinforcement Learning: Teaching Machines to Learn by Trial and ErrorAn AI That Learns Like UsKey Concepts of Reinforcement LearningThe Markov Decision Process (MDP)The Bellman EquationModel-Based Versus Model-Free SystemsOn-Policy Versus Off-Policy Learning: Two Paths to LearningMonte Carlo Reinforcement LearningTemporal Difference (TD) LearningQ-LearningDeep Reinforcement LearningSummaryReferencesPart II The Generative AI RevolutionChapter 7 Language Modeling: The Birth of LLMsAn Introduction to LLMsFoundations of Language ModelingNext-Word Prediction From Words to TokensWord Embedding: Turning Tokens into Numbers How Word Embeddings Are LearnedSemantic Relationships in the Embedding SpaceThe Semantics of LanguageSummaryReferenceChapter 8 Attention Is All You Need: The Foundation of Generative AIA New Architecture Begins to Take ShapeAttention Is All You NeedFrom Sequential to Parallel ProcessingPositional EncodingThe Self-Attention MechanismSummaryReferencesChapter 9 Attention Isn’t All You Need: Understanding the Transformer ArchitectureThe Encoder BlockThe Multi-Head Attention LayerThe Add and Norm Layers and Residual ConnectionsThe Feedforward Network (FFN) LayerLayers Upon Layers of Encoder BlocksHow Encoders Are TrainedThe Decoder BlockThe Decoder’s Output ClassifierHow Decoders Are TrainedWhat Type of Machine Learning Is Involved in Training LLMs?Case Study: The GPT-3 TransformerFuture DirectionsSummaryReferencesPart III Living with Generative AIChapter 10 Making Models Smarter: Prompt and Context EngineeringPrompt and Context WindowsPrompt Engineering TechniquesShot-Based ApproachesChain-Based ApproachesSelf-Ask ApproachesPrompt Engineering LimitationsContext EngineeringTypes of Contexts in LLM WorkflowsTools and ProtocolsContext Design TechniquesSummaryChapter 11 Retrieval-Augmented Generation The Need for RAGCommon Applications of RAGRAG Trends and PracticesThe RAG PipelineQuery FormulationRetrieval FilteringWorking with Knowledge DatabasesLoading DocumentsChunking: Splitting DocumentsEmbedding and Storing SegmentsRetrieving SegmentsSummaryChapter 12 Fine-Tuning LLMsThe Need for Fine-TuningComparing Fine-Tuning and RAGInference Hyperparameter Tuning for LLMsTemperatureTop-K SamplingTop-P (Nucleus) SamplingRepetition PenaltyPrinciples of Fine-Tuning with New Data Fine-Tuning for Model Types and ObjectivesSupervised Fine-Tuning (SFT)Transfer LearningParameter-Efficient Fine-Tuning (PEFT) MethodsRetrieval-Augmented Fine-Tuning (RAFT)Reinforcement Learning from Human Feedback (RLHF)Benchmarking Model PerformanceSummaryReferencesChapter 13 Securing LLMs from AttackWhat Makes AI Security DifferentThe Emergence and Importance of AI Security FrameworksNIST AI Risk Management FrameworkThe OWASP Top 10MITRE ATLASThe ISO/IEC Suite of AI StandardsA Comparison of the AI Security FrameworksAI Vulnerabilities and Attack VectorsDirect Prompt Injection AttacksPrompt Injections with JailbreakingIndirect Prompt Injection AttacksExtraction and Inversion AttacksAI Supply Chain ThreatsDefending Models from AttackExtending the Guardrail SystemArchitectural SafeguardsContinuous Monitoring and Detection SystemGenerative Adversarial Defense TechniquesSummaryReferencesChapter 14 AI Ethics and Bias: Building Responsible SystemsBias and Ethical Risks in GenAIThe Biased Data That Shapes GenAIThe Difficulty of Stopping GenAI BiasWhen Generative AI Goes Wrong: Unethical and Harmful OutputsHallucination and MisinformationSynthetic Media, Deepfakes, and DisinformationOwnership, Consent, and CopyrightTransparency, Explainability, and TrustBuilding Responsible Generative AIAlignment and AI SafetyPractical Responses to Ethical AI ChallengesSummaryReferencesChapter 15 The Future of AI: From Generative to General IntelligenceWhere We Stand: A Snapshot of Today’s CapabilitiesCurrent Limitations and Known Pain PointsThe Emergence Question: Are We Seeing Sparks of AGI?What Makes AGI Different?What Is AGI?What Is Intelligence Anyway?Do Reasoning LLMs Really Reason?Predicting Versus UnderstandingAre We Already on the Path to AGI?Paths to AGIThe Scaling HypothesisThe Modular HypothesisThe Embodied System HypothesisHybrid ModelsIs AGI the End of Humanity?The Alignment Problem RevisitedBlack Boxes and Loss of InterpretabilityThe Singularity and the Skynet ProblemControlling Existential RisksIs AGI Helping or Hurting Society?Will AI Take Your Job?Education in the Age of Generative AISocietal Identity and StabilityThe Evolving Voice of Generative AIFrom Single-Goal Prompting to Multimodal PartneringResponsive InterfacesRedefining CreativityThe Future We ChooseScenario A: The Co-creative SocietyScenario B: The Automated PresentScenario C: The Disrupted PathSummaryReferencesAppendix A The History of AIAppendix B A Summary of Neural Network Model ArchitecturesGlossary 9780135429419 TOC 12/19/2025