KP Sawhney from Google DeepMind discusses the internal strategies for scaling agentic AI, including managing token-hungry workflows, curating a 'Darwinian' skills library, and evolving the Deep Research pipeline from large context blobs to a collaborative file system.
An in-depth look at Gemma 4's novel transformer architecture with per-layer embeddings, enabling efficient parameter offloading for on-device inference. The discussion also covers its native multimodality, the state of fine-tuning, text-based diffusion models, and the growing intersection of research and engineering.
Adrian Bertagnoli from Callosum argues that the era of scaling monolithic models on homogeneous GPU clusters is ending. He introduces "heterogeneous intelligence," a new paradigm where model architectures, chip types, and workflows are optimized together. By routing subtasks to the most efficient model and hardware, this approach achieves significant performance gains, as demonstrated by two key results: a 7x cost reduction in recursive reasoning tasks using Cerebras, and state-of-the-art performance on the Video Web Arena benchmark, outperforming leading GPT and Gemini models at a fraction of the cost and time.
RL Nabors makes a case against chat as the permanent UI for agents, arguing it's the "terminal" to the future "iPhone" of rich, interactive experiences. She demonstrates how to build these experiences using web technologies, introducing MCP Apps for in-agent UIs and WebMCP for making existing websites agent-callable, positioning the web platform as the ultimate "infinite canvas".
A comprehensive overview of Google DeepMind's latest advancements, featuring Paige Bailey demonstrating Gemini 1.5 Flash's cost-effective video analysis and AI Studio's single-prompt app generation. Guillaume Vernade showcases a full generative media pipeline, turning a public domain book into an illustrated, animated, and scored project using Gemini, Nano Banana, VO, and LIA. Ian Valentine closes with the power of Gemma 4, demonstrating on-device, multi-agent code generation and debugging without cloud APIs.
This talk presents a container-first methodology for developing, distributing, and managing AI agents. Using a stack of Podman for local development and Kubernetes for scalable deployment, this approach transforms personalized agent setups from messy collections of files into reproducible, secure, and portable container images that can serve as a team-wide baseline. The session covers practical techniques for secrets management, state persistence, and automated setup, highlighted by a real-world example from an Nvidia team using this pattern for model evaluations.