I gave a Gemma 4 AI agent a sandbox and it taught itself physics | The Agent Factory Podcast

Gemma 4 → https://goo.gle/4tFKxGE Gemma 4 in Google Cloud → https://goo.gle/4mpWI8f Deploy Gemma 4 in Cloud Run → https://goo.gle/48mIizK Can an open source model running entirely offline on a smartphone outperform the cloud based APIs of last year? In this episode of The Agent Factory, we dive deep into the release of Gemma 4, the latest family of open models from Google DeepMind. We answer the critical question for developers: how do you balance high intelligence with low hardware overhead and data sovereignty? You’ll see Omar Sanseviero from Google DeepMind explain the architectural breakthroughs that allow Gemma 4 to deliver massive "intelligence per parameter." We don’t just talk about the specs; we show them in action. You’ll watch a food tour agent navigate Seattle via an MCP server, and witness a stunning demo where Gemma 4 reasons through laws of physics to code a bouncing ball animation from scratch in a Python sandbox. We also break down Gemma 4 adopting the Apache 2.0 license and what that opens up for you. Whether you are a cloud architect or an independent developer, you will walk away with a clear roadmap for self hosting these models. You'll learn how to leverage new features like native function calling, structured JSON output, and multimodal understanding to build agents that are faster, cheaper, and more private than ever before. Chapters: 0:00 - Introduction 0:50 - Meet Omar Sanseviero from Google DeepMind 1:40 - What makes Gemma 4 a "major milestone" for agents? 1:53 - Demo: Running Gemma 4 on an Android phone 2:45 - Agent skills running on a phone! Function calling, structured output, thinking, and multi-modal understanding. 4:05 - The move to Apache 2.0: Why the license change matters 5:29 - ADK demo: Building a Food Tour agent with Google Maps 8:03 - Demo: How Gemma 4 uses physics and gravity to create animations using Python code 12:05 - Why 256k context on a local device is a game changer 12:34 - The explosion of code execution agents 15:10 - Cost analysis: Self hosting vs. API- based models 17:23 - Architectural deep dive: MoE and per-layer embeddings 18:30 - Variable aspect ratio in multimodal understanding 19:51 - Gemma 4 vs. Gemini: When to use which? 22:07 - Final thoughts and how to get started More resources: Food tour agent → https://goo.gle/3OinTFh Coding agent with Gemma 4 → https://goo.gle/4dBDNEY Fine-tuning Gemma 4 → https://goo.gle/4chRlDj BentoRun, Python sandbox in Cloud Run → https://goo.gle/4voDPX2 Watch more of The Agent Factory → https://www.youtube.com/watch?v=qBOvM7SiDa4&list=PLIivdWyY5sqLXR1eSkiM5bE6pFlXC-OSs 🔔 Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech Speakers: Omar Sanseviero, Vlad Kolesnikov, Shir Meir Lador Products Mentioned: Gemma 4, Cloud Run, Agent Development Kit, Gemini 3, Android, Google Maps MCP Server