Vanishing Gradients

Bryan Bischof, Head of AI at Theory Ventures and long-time data scientist, explains why real-world data challenges reveal agent failures. He recounts a hackathon testing agents on SQL, logs, and 750,000 PDFs. Topics include failure funnels and binary checkpoints, why unlimited submissions encourage hill-climbing, DocETL for document extraction, MCP as a semantic layer, and when simple coding agents beat heavy frameworks.

Samuel Colvin, creator of Pydantic and lead of the Pydantic Stack, explains building durable AI agents with engineering-grade reliability. He discusses agentlets as small specialized building blocks, using Temporal for robust workflow durability, separating deterministic workflows from stochastic model calls, and making observability and type-safe validation central to production AI.

Alex Strick van Linschoten, ML engineer and curator of the LLMOps database, tracks real-world production AI deployments. He recounts a $50K infinite-loop cost and warns about silent failures. They discuss ripping out and rebuilding agent systems, extreme low-latency voice agents that toss context, three-tier agent architectures, the 100-to-1 token noise problem, and when simple tools beat complex stacks.

Alison Hill, product leader focused on multimodal UX and community learning. Marcel Kornacker, CTO and co-creator of Apache Parquet, expert in multimodal data systems. Wes McKinney, creator of pandas and data tooling veteran. They discuss agent ergonomics and language tradeoffs, adversarial AI code review, making image/video operations native to data platforms, and why taste and schemas matter for multimodal workflows.

Join search guru Doug Turnbull, who shaped systems at Reddit and Shopify, and John Berryman, the brain behind GitHub Copilot, as they dive into the future of agentic search. They explore the evolution from traditional search to agentic retrieval, spotlighting John's five-level maturity model for AI adoption. Learn why understanding user intent is paramount and discover practical steps to create your own agentic loops. They also share insights on avoiding common pitfalls in search design, emphasizing the importance of real user feedback.

Eleanor Berger and Isaac Flaath, co-founders of Elite AI Assisted Coding, delve into the future of software development with AI. They explore how agents can maintain living documentation using simple markdown files, enhancing ongoing learning. Discover the power of specification-first planning to define success, while automated tech debt audits keep projects in check. The duo emphasizes the importance of accountability and clear communication in teamwork. With insights on using agents for routine tasks, they reveal how to save developers hundreds of hours!

Eric Ma, a research data science leader at Moderna specializing in AI for biotech, discusses the surprise finding that Moderna's systems are built on reliable workflows rather than autonomous agents. He emphasizes the importance of mapping permissions in regulated environments and the risks of data leaks from LLM execution traces. Offloading “janitorial” tasks to AI can improve efficiency, but Eric advises starting with simpler tools to reduce risks. He astutely highlights the need for evaluation rigor that matches the stakes involved in biotech applications.

Jeff Huber, CEO and co-founder of Chroma, dives into the fascinating world of agentic search, transforming how we approach AI and information retrieval. He discusses the importance of 'context engineering' for reliable AI systems and how context rot complicates this. Huber explains the concept of the 'agent harness' for advanced tools and the necessity of hybrid search to maintain balance. With insights on best practices for builders and the challenges in agent evaluation, this conversation illuminates the evolving landscape of AI search.

In this discussion, Michael Kennedy, a seasoned Python developer and educator known for his insights on AI and software practices, tackles the myth of complexity in tech. He shares how to simplify production Python systems, emphasizing the importance of the 'Docker barrier' for cost-effective self-hosting. The conversation explores how Agentic AI is shifting development mindsets and enhancing efficiency. Michael also stresses the value of struggling through learning and the need for complementary skills in navigating the evolving tech landscape.

Ravin Kumar, a researcher at Google DeepMind specializing in generative models and LLM products, joins to discuss the groundbreaking Gemini 3. They illustrate how models can 'self-heal' and adapt, reshaping software development. Topics include the transition from basic tool calling to advanced agent harnesses, the contrast between deterministic workflows and high-agency systems, and the importance of robust evaluation infrastructures. Ravin also shares insights on the evolution of productive features like Audio Overviews and the future of multimodal agents.