LLM Primer

Епизоди

• 
  Prompt Injection and Jailbreaks 07.07.2026 45мин

  This chapter examines prompt injection and jailbreak attacks, which exploit a language model's inherent inability to distinguish between authoritative developer instructions and untrusted user data. It covers the mechanics of direct and indirect injection, categorises common jailbreaking techniques, discusses the limitations of defensive prompt engineering, and outlines a layered mitigation strategy to better defend production systems.Amazon.com: LLM Primer VII AI Security: Defending LLM Systems Against Prompt Injection, Jailbreaks, and Adversarial Threats: 9798185644065: SHIMODA, SHO: Books
• 
  Data Security and Privacy 07.07.2026 32мин

  This chapter examines data security and privacy throughout the LLM lifecycle. It explores the inherent risks of training data, such as copyright issues, personal information (PII) contamination, and data poisoning. Additionally, it details how models can leak sensitive information through memorization and extraction attacks, and outlines operational defenses for securing systems, including input redaction pipelines, encryption, tenant isolation, and data retention policies.Amazon.com: LLM Primer VII AI Security: Defending LLM Systems Against Prompt Injection, Jailbreaks, and Adversarial Threats: 9798185644065: SHIMODA, SHO: Books
• 
  Threat Modeling for LLM Systems 06.07.2026 54мин

  This chapter adapts traditional threat modeling frameworks (such as STRIDE, PASTA, and attack trees) specifically for the unique vulnerabilities of LLM systems. It guides defenders through identifying AI-specific assets and adversaries, and provides a step-by-step procedure for building a living threat model that can be maintained alongside the system's codeAmazon.com: LLM Primer VII AI Security: Defending LLM Systems Against Prompt Injection, Jailbreaks, and Adversarial Threats: 9798185644065: SHIMODA, SHO: Books
• 
  Why AI Security Is Different 06.07.2026 49мин

  This chapter explains that AI security fundamentally and structurally differs from traditional software security. Instead of finding and patching clear bugs in readable source code, defenders must secure probabilistic models whose behaviors are driven by billions of uninterpretable weights and training data. Consequently, the security focus shifts from ensuring code correctness to managing and restricting an unbounded range of unpredictable inputs and outputs. Amazon.com: LLM Primer VII AI Security: Defending LLM Systems Against Prompt Injection, Jailbreaks, and Adversarial Threats: 9798185644065: SHIMODA, SHO: Books
• 
  2-7-7. Hallucinations and Reliability: Managing Confident Errors 19.02.2026 16мин

  This episode covers Chapter 7, examining why Large Language Models confidently generate false information. We discuss the probabilistic nature of "hallucinations," the dangerous gap between fluency and correctness, and practical strategies like calibration and hybrid verification to align model confidence with reality.Amazon.com: LLM Primer VII AI Security: Design Safe and Robust AI System eBook : SHIMODA, SHO: Kindle Store
• 
  2-7-6. Retrieval-Augmented Generation Risks: Securing the Knowledge Pipeline 19.02.2026 34мин

  This episode covers Chapter 6, focusing on the security implications of connecting models to external data (RAG). We discuss how this introduces new trust boundaries, the dangers of malicious document injection where attackers plant traps in your knowledge base, and the necessity of validating documents before they enter the model's context.Amazon.com: LLM Primer VII AI Security: Design Safe and Robust AI System eBook : SHIMODA, SHO: Kindle Store
• 
  2-7-5. Input Validation and Output Filtering: The Defense Pipeline 18.02.2026 29мин

  This episode covers Chapter 5, detailing how to build disciplined pipelines around an AI model. We discuss strategies for sanitizing user inputs to catch attacks early, the importance of structured prompting to reduce ambiguity, and why output moderation is essential to catch policy violations that slip through earlier defenses.Amazon.com: LLM Primer VII AI Security: Design Safe and Robust AI System eBook : SHIMODA, SHO: Kindle Store
• 
  2-7-4. Prompt Injection and Jailbreaks: Defending the Interpreter 18.02.2026 37мин

  This episode explores Chapter 4, detailing how attackers manipulate model behavior through crafted inputs like instruction overrides. We discuss why prompt injection is an inherent property of instruction-following systems rather than a standard bug. The episode covers jailbreaking techniques like role-playing and obfuscation, and why defense requires architectural layers rather than just better prompts.Amazon.com: LLM Primer VII AI Security: Design Safe and Robust AI System eBook : SHIMODA, SHO: Kindle Store
• 
  2-7-3. Data Security and Privacy: The AI Lifecycle 18.02.2026 25мин

  This episode breaks down Chapter 3, tracking data risks from training to deployment. We discuss how models can memorize sensitive training data, the subtle dangers of leakage through generated outputs, and the critical importance of treating user prompts and logs as sensitive assets.Amazon.com: LLM Primer VII AI Security: Design Safe and Robust AI System eBook : SHIMODA, SHO: Kindle Store
• 
  2-7-2. Threat Modeling for LLM Systems: A Step-by-Step Guide 18.02.2026 29мин

  This episode covers the systematic approach of Chapter 2, moving beyond vague security worries to concrete risk analysis. We discuss how to identify unique AI assets—like prompts, logs, and retrieval indexes—and map the expanded attack surface of API-based systems to build durable defenses.Amazon.com: LLM Primer VII AI Security: Design Safe and Robust AI System eBook : SHIMODA, SHO: Kindle Store
• 
  2-7-1. The Probabilistic Shift: Why AI Security is Different 18.02.2026 36мин

  This episode dives into Chapter 1, exploring why traditional security measures fail when applied to Large Language Models. We discuss the fundamental shift from deterministic code to probabilistic behavior, how LLMs expand the attack surface from endpoints to context, and why security must be designed into the architecture rather than patched on later.Amazon.com: LLM Primer VII AI Security: Design Safe and Robust AI System eBook : SHIMODA, SHO: Kindle Store
• 
  2-1-12. The System Architect — Building Your Own LLM System 17.02.2026 38мин

  In this episode, we bring every previous concept together to answer the ultimate practical question: How do you actually build a complete LLM system from scratch? We move beyond the model itself to construct the full production environment—from legal compliance to user interface—required to turn a neural network into a working product.Join us as we:• Secure the Foundation: We tackle Datasets and Licensing, explaining why data governance, provenance tracking, and legal compliance are the non-negotiable starting points of any system.• Engineer the Pipeline: We break down the Training Pipeline, detailing the operational discipline required to automate preprocessing, manage distributed training, and ensure reproducibility.• Define Success: We construct Evaluation Frameworks, moving beyond simple accuracy metrics to build systematic testing for robustness, safety, and bias mitigation.• Orchestrate the Stack: We explore the Integrated Application Stack, visualizing how inference APIs, vector databases, caching layers, and security modules must coordinate to serve users reliably.• Learn from Reality: We review Case Studies & Best Practices, synthesizing lessons from real-world deployments to highlight why modular design and observability are critical for long-term maintenance.This episode serves as the comprehensive blueprint for engineers ready to integrate data, algorithms, and infrastructure into a unified, scalable system.
• 
  2-1-11. The Research Frontier — Cutting-Edge Research 17.02.2026 29мин

  In this episode, we look beyond the current generation of models to explore the experimental architectures and learning paradigms that will define the future of AI. We analyze how researchers are redesigning the Transformer to overcome its fundamental limitations: computational cost, static knowledge, and isolation from the physical world.Join us as we:• Scale Efficiently: We break down Sparse Models and Mixture of Experts (MoE), explaining how "gating mechanisms" allow models to scale to trillions of parameters while only activating a small fraction of them for each specific task.• Unlock Memory: We discuss the shift from static "parametric memory" (fixed weights) to Dynamic Retrieval and Memory Mechanisms, where models can update their knowledge without expensive retraining.• Unify the Senses: We explore Multimodal Models, examining how text, vision, and audio are being mapped into shared representation spaces to create systems that can "see" and "hear" as well as they read.• Learn Continuously: We tackle the challenge of Continual Learning and Catastrophic Forgetting, looking at techniques that allow models to learn incrementally over time rather than being frozen after a single training run.This episode is a roadmap for understanding how AI is evolving from static text generators into dynamic, efficient, and multi-sensory systems.
• 
  2-1-10. The Trust Architecture — Safety, Ethics, & Trust 17.02.2026 37мин

  In this episode, we address the critical challenge of turning a powerful probabilistic system into a reliable product. We explore why engineering capability must be matched with ethical responsibility, shifting the focus from "what the model can do" to "whether we should trust it."Join us as we:• Confront the Hallucinations: We analyze why models confidently generate false information—not because they "imagine," but because they predict—and discuss mitigation strategies like retrieval grounding and verification layers.• Address the Bias: We explore how models inherit and amplify societal stereotypes from their training data, examining the technical and procedural steps needed to measure and mitigate these harms.• Build the Guardrails: We examine the defense systems—from input filtering to hierarchical system prompts—that prevent malicious use and keep model behavior within safe boundaries.• Demand the Proof: We discuss Explainability and Transparency, distinguishing between interpreting internal neural weights and providing clear, auditable system behaviors for users and regulators.This episode establishes that trust is not a default feature of AI, but an engineered property built through layered safeguards and governance.
• 
  2-1-9. The Cost of Intelligence — Performance, Scaling, and Costs 17.02.2026 31мин

  In this episode, we face the economic and physical realities of deploying AI. A model’s theoretical capability matters little if it is too slow, too expensive, or too power-hungry to run. We explore the "tradeoff triangle" engineers must navigate to turn a research artifact into a sustainable product.Join us as we:• Weigh the Returns: We analyze Model Size vs. Capability, discussing empirical scaling laws and the point of "diminishing returns" where making a model bigger no longer pays off.• Measure the Speed: We distinguish between Latency (how fast a single user gets an answer) and Throughput (how many users the system can handle), explaining why optimizing for one often hurts the other.• Calculate the Bill: We look at the hard costs of Inference, breaking down how context length and token count directly impact memory usage, energy consumption, and cloud bills.• Compress the Math: We explain Quantization, a technique that reduces the numerical precision of a model (e.g., from 32-bit to 8-bit) to drastically cut memory usage without destroying intelligence.• Move to the Edge: We discuss On-Device Deployment, examining the challenges and privacy benefits of running powerful AI locally on phones and laptops instead of the cloud.This episode is a reality check for anyone wondering why the smartest model isn't always the right choice for the job.
• 
  2-1-8. The Engineering Reality — Using LLMs in Applications 17.02.2026 42мин

  In this episode, we step out of the theoretical lab and into the messy reality of production. We explore how a raw Large Language Model is transformed into a reliable product, shifting the focus from "what the model knows" to "how the system behaves."Join us as we:• Architect the Conversation: We analyze Chatbots & Conversational Agents, explaining why memory management, system prompts, and safety guardrails are just as important as the model itself.• Synthesize and Search: We look at Summarization and Search, discussing how LLMs are breathing new life into old information retrieval systems by understanding meaning rather than just matching keywords.• Structure the Chaos: We dive into Knowledge Extraction, showing how businesses are using LLMs not to write poetry, but to turn messy unstructured text into clean, machine-readable JSON data.• Code with Context: We explore Code Assistants, examining how models are integrated into development environments to predict software logic while navigating complex file structures.• Iterate to Success: We discuss Evaluation and Iteration, emphasizing that deployment is just the beginning—and that real reliability comes from A/B testing, human review loops, and continuous monitoring.This episode is a practical guide for builders who need to wrap orchestration logic around probabilistic models to create software that actually works.
• 
  2-1-7. The Hybrid System — Beyond Next-Token Prediction 17.02.2026 30мин

  In this episode, we challenge the idea that Large Language Models are just text generators. We explore how modern AI extends beyond simple prediction to become a reasoning engine capable of searching databases, understanding images, and grounding itself in external facts.Join us as we:• Map the Meaning: We explain Embeddings, the dense vector representations that transform language into geometry, allowing computers to understand that "king" is to "man" what "queen" is to "woman".• Bridge the Gap: We contrast Generation (synthesizing new ideas) with Retrieval (accessing stored facts), showing how hybrid models combine the best of both worlds.• Fix the Memory: We break down Retrieval-Augmented Generation (RAG), a critical architecture that connects frozen models to up-to-date external databases to improve accuracy and reduce hallucinations.• Expand the Senses: We look at Multimodal Extensions, revealing how models are learning to "see" and "hear" by aligning visual and audio data within the same mathematical space as text.This episode reveals how we are moving from closed, static models to open, dynamic ecosystems.
• 
  2-1-6. From Generalist to Specialist — Fine-Tuning & Adaptation 17.02.2026 31мин

  In this episode, we tackle the critical difference between a model that knows "about" everything and one that can actually do a specific job. We explore the adaptation phase, where a raw, pretrained generalist is transformed into a specialized tool capable of following instructions, coding, or offering legal advice.Join us as we:• Define the Shift: We distinguish between Pretraining (building broad linguistic competence) and Fine-Tuning (refining behavior for specific tasks), explaining how reusing existing knowledge saves massive amounts of compute.• Compare Strategies: We contrast Parameter-Level Adaptation (permanently updating model weights) with Prompt-Based Adaptation (steering the model through context without changing its internal structure).• Align the Behavior: We discuss Instruction Tuning, the crucial process of training models on instruction-response pairs so they learn to obey commands rather than just autocomplete sentences.• Specialize the Knowledge: We examine Domain-Specific Tuning, showing how models are recalibrated for high-stakes fields like medicine or finance by immersing them in specialized technical corpora.This episode explains how we bridge the gap between a model that can write fluent English and a system that actually solves your specific problem.
• 
  2-1-5. The Industrial Pipeline — Training Large Models 17.02.2026 31мин

  In this episode, we move from the theoretical blueprint of the Transformer to the operational reality of building a Large Language Model. We explore how an empty mathematical shell is transformed into a capable system through a massive, coordinated engineering process known as training.Join us as we:• Curate the Curriculum: We discuss why "more data" isn't always better, explaining the critical steps of deduplication, filtering, and balancing diverse sources like web text, books, and code.• Minimize the Surprise: We break down the mathematical objective of Cross-Entropy Loss and the optimization algorithm Gradient Descent, revealing how billions of parameters are nudged iteratively to improve prediction accuracy.• Distribute the Load: We examine the physical infrastructure required for training, detailing how strategies like Data Parallelism and Model Parallelism allow engineers to split massive models across thousands of GPUs.• Balance the Learning: We analyze the risks of Overfitting (memorizing data) versus Underfitting (failing to learn patterns), and how regularization ensures a model can generalize to new, unseen text.This episode reveals that training an LLM is not just a math problem, but a large-scale systems engineering challenge.
• 
  2-1-4. The Blueprint of Intelligence — The Transformer Architecture 17.02.2026 44мин

  In this episode, we explore the specific architectural breakthrough that made the current AI revolution possible. We move from general neural network theory to the concrete blueprint of the Transformer, examining the "self-attention" mechanism that allows models to process massive amounts of information in parallel.Join us as we:• Deconstruct the Block: We break down the essential components of a Transformer layer—multi-head attention, feedforward networks, residual connections, and layer normalization—explaining how they stack to refine meaning.• Explain the Mechanics: We visualize how "Queries," "Keys," and "Values" interact to calculate attention scores, allowing words to "vote" on which other words are most relevant to them.• Solve the Order Problem: We discuss Positional Encoding, the clever mathematical trick that injects order into the system so the model can distinguish "the dog chased the cat" from "the cat chased the dog."• Compare the Variants: We clarify the differences between Encoder-only models (like BERT), Encoder-Decoder models (like the original Transformer), and the Decoder-only models (like GPT) that dominate generative AI today.This episode offers the structural deep dive needed to understand not just that these models work, but why they scale so effectively.