Podcast
AI Intuition
Hosted by Dan Sarmiento · EN
This is the gold rush era of artificial intelligence. You want to learn quickly so you don't get left behind, but how can you learn about AI without an advanced degree in computer science and mathematics?
You translate all the complicated concepts into plain language and you summarize the relevant news into a podcast you can listen to while you do everything else.
This is the method that helped me speed up my learning and maybe it can help you too.
10episodes
Episodes
Newest firstAll episodes
Microsoft Data Agent Development in Fabric
1w ago00:52:46Tap to summarizeAn introduction to Microsoft Fabric data agents, which are conversational Q&A systems powered by generative AI that allow users to retrieve insights from governed OneLake data using natural language. It provides comprehensive instructions on configuring, validating, and managing these agents through developer-defined settings, deployment pipelines, and a dedicated Python SDK, all while strictly adhering to Microsoft Purview security and compliance policies
Transcribe →Agent Builder by Docker
Sep 600:51:03Tap to summarizecagent, Docker's open-source, multi-agent runtime designed to orchestrate autonomous AI systems by allowing users to build and manage teams of specialized AI agents. cagent uses a declarative YAML configuration for defining agents and their interactions, with a hierarchical structure where a root agent delegates tasks to sub-agents. A key innovation is the Model Context Protocol (MCP), which acts as a universal interface enabling agents to interact securely with external tools and services, supported by Docker's MCP Catalog, Toolkit, and Gateway. This ecosystem, especially the MCP Gateway, emphasizes security through containerization and provides enterprise-grade features for managing and deploying agentic AI applications. Overall, the sources highlight cagent's strategic role in Docker's vision to be a foundational platform for the next generation of AI development, providing a secure, accessible, and scalable environment for agentic AI.
Transcribe →Open Agentic Web Development - Project NANDA (MIT)
Sep 300:39:16Tap to summarizeProject NANDA, an initiative by the MIT Media Lab aimed at creating the foundational infrastructure for the "Open Agentic Web," an internet designed for autonomous AI agents rather than human users. This new architecture addresses the limitations of the current internet for agent discovery, identity, and trust, proposing a system where trillions of AI agents can collaborate seamlessly at machine speed. Project NANDA's core components include the NANDA Index for global agent discovery, AgentFacts for verifiable agent identity and capabilities, and the Adapter SDK for universal protocol interoperability. The project strategically positions itself as a complementary "Layer 0/1" foundation, supporting higher-level communication protocols like the industry-backed A2A and Anthropic's MCP, ensuring its relevance and increasing its potential for widespread adoption. With demonstrated progress on its initial roadmap, NANDA seeks to become the silent, critical infrastructure enabling a future agent-driven digital economy.
Transcribe →AI Startup Failure Analysis
Sep 300:46:39Tap to summarizeexamines the paradox of unprecedented investment in the artificial intelligence sector coexisting with an accelerating rate of startup failures. It identifies a failure rate exceeding 90% for AI startups, significantly higher than the broader tech industry. The analysis categorizes these failures into distinct modalities: Market Failure (lack of product-market fit), Product Failure (technology underdelivers or is unreliable), Execution Failure (poor management or fraud, often exacerbated by excessive funding), Financial Failure (running out of capital, usually a symptom of deeper issues), and Competitive Failure (core technology rendered obsolete by larger foundational models, termed the "Foundational Model Guillotine"). The report offers strategic recommendations for founders to build defensible moats beyond mere algorithms, embrace capital efficiency, and solve urgent customer problems, while advising investors to scrutinize for AI-washing and assess competitive risks.
Transcribe →AI Security - Model Denial of Service
Sep 201:13:46Tap to summarizeModel Denial of Service (Model DoS) attacks, a modern evolution of traditional DoS that targets the computational resources of AI and Machine Learning systems, rather than network bandwidth. It explains how these attacks degrade performance or render AI models unavailable, often by exploiting their processing demands or through tactics like Economic Denial of Sustainability (EDoS), which incurs substantial financial costs for victims. The text outlines the threat landscape, identifying highly vulnerable AI services like Large Language Models (LLMs), and offers a multi-layered framework for detection, prevention, and mitigation, emphasizing architectural, application-level, and operational controls to build resilient AI systems.
Transcribe →AI Security - Training Data Attacks
Sep 200:59:31Tap to summarizeanalysis of training data poisoning, a critical integrity attack against AI and ML systems. It explains how adversaries corrupt the foundational learning phase by manipulating datasets, leading to altered model behavior, ranging from performance degradation to hidden backdoor attacks. The text highlights that large language models (LLMs) and generative AI are particularly vulnerable due to their reliance on vast, often unvetted internet data, and critically notes that larger models can paradoxically be more susceptible to learning malicious behaviors from minimal poisoned data. Finally, it outlines a multi-layered defense strategy, emphasizing data validation, robust model training, and strong operational security controls throughout the MLOps lifecycle, aligned with industry frameworks like NIST and OWASP.
Transcribe →AI Security - Insecure Output Handling
Sep 200:42:57Tap to summarizeanalysis of Insecure Output Handling, a critical application security vulnerability distinct from insecure input handling, emphasizing the need to never trust data sent to an interpreter. It details the diverse and severe consequences of this flaw, including client-side attacks like Cross-Site Scripting (XSS) and server-side threats such as Remote Code Execution (RCE), providing a comparative table to highlight the differences between input and output vulnerabilities. The document then examines the attack surface across various application architectures, from traditional web applications to modern APIs and the emerging risks posed by Large Language Models (LLMs), before presenting statistical data and real-world case studies to quantify its pervasive impact. Finally, it outlines a multi-layered defense strategy, advocating for a zero-trust approach, robust validation and context-aware output encoding, and the integration of both automated and manual testing methodologies throughout the Software Development Lifecycle (SDLC).
Transcribe →AI Security - Prompt Injection
Sep 200:49:58Tap to summarizeanalysis of prompt injection, which is identified as the leading security vulnerability in applications powered by Large Language Models (LLMs). It explains that this threat arises from the inherent architecture of LLMs, which struggle to differentiate between trusted developer instructions and untrusted user input. The text categorizes prompt injection into direct and indirect attacks, detailing various techniques for each, such as jailbreaking and data exfiltration via hidden payloads in external data. Furthermore, it outlines a multi-layered, defense-in-depth strategy for detection and prevention, emphasizing the importance of secure prompt engineering, architectural safeguards like the principle of least privilege, and continuous operational security. The source concludes by stressing that no single solution exists and that a holistic approach is crucial to securing evolving agentic and multimodal AI systems.
Transcribe →Unsupervised ML for Test Suite Reduction - Test Smarter Not Harder
Aug 3100:42:45Tap to summarizeThis research systematically maps literature concerning the application of unsupervised machine learning approaches to test suite reduction (TSR), a critical process for optimizing software testing efficiency. The study, which reviewed 34 papers published between 2013 and 2023, identifies common algorithms and evaluation metrics in this field. It highlights K-Means clustering as the most frequently used algorithm and coverage metrics as the primary means of assessing effectiveness. The findings also point to a significant gap in the literature regarding scalability considerations and a general lack of shared research artifacts. Despite these challenges, the research underscores the broad applicability of unsupervised learning for TSR across various software domains, from web-based applications to embedded systems.
Transcribe →bytedance USO - Unified Style and Subject-Driven Generation via Disentangled and Reward Learning (Image Model)
Aug 3100:49:41Tap to summarizeanalyze USO, a novel generative AI model developed by Bytedance's Intelligent Creation Lab. USO addresses the long-standing challenge of separately controlling style and subject in image generation by proposing a unified framework that synergizes these tasks. The text details USO's conceptual foundations, including cross-task co-disentanglement and style reward-learning, which allow it to effectively separate and recombine content and style information. It further explains the model's architecture, training methodology utilizing a large-scale triplet dataset, and practical capabilities such as combined style-subject generation and low VRAM inference. Finally, the sources position USO within the broader generative AI landscape, comparing it to specialized models like StyleDrop and PhotoMaker, and highlighting its potential as a step towards universal customization models.
Transcribe →