The Google I/O Keynote highlights the transition into an “agentic era” of AI, focusing on foundational infrastructure, advanced multi-modal models, personal AI agents, and deep scientific applications.
1. AI Models & Developer Infrastructure
- TPU 8th Generation (AT & ATI): Google introduced specialized custom silicon for training and inference. The training architecture allows Jackson Pathways to distribute workloads across multiple sites, scaling past 1 million TPUs globally [12:15].
- Gemini Omni & Gemini 3.5 Flash: Demis Hassabis introduced Gemini Omni, a world model capable of simulating physical principles like kinetic energy and gravity to generate highly realistic, editable video across inputs [17:20]. Concurrently, Gemini 3.5 Flash launched, optimized to deliver fast, highly cost-efficient agentic actions and coding workflows [23:45].
- Google Anti-Gravity 2.0: A standalone, agent-first desktop application designed for multi-agent orchestration. In a benchmark test, autonomous sub-agents collaborated to build a functioning operating system from scratch within 12 hours for under $1,000 in API credits [28:12].
2. Reimagined AI Consumer Products
- Google Search Evolution: Google search box has been completely upgraded [46:58]. Enabled by Gemini 3.5 Flash and Anti-Gravity, Search can now deploy live containerized code to create real-time custom layouts, interactive visuals, and continuous dashboards (e.g., personalized weekend trackers) on the fly [51:50].
- Gemini App & Neural Expressive Design: The Gemini App has over 900 million monthly active users and has received a fluid, interactive interface upgrade called “Neural Expressive” [01:10:39].
- Gemini Spark: Announced as a persistent personal AI agent running 24/7 on dedicated cloud VMs [35:27]. Spark securely integrates with your inbox, tools, and third-party apps via Model Context Protocol (MCP) to manage heavy background tasks under user boundaries and supervision [01:16:06].
3. Commercial & Shopping Ecosystems
- Agent E-Commerce Building Blocks: Google unveiled the open-source Universal Commerce Protocol (UCP), co-developed with partners like Amazon, Meta, and Stripe, to standardize transaction communications [59:44].
- Agent Payments Protocol (AP2) & Universal Cart: AP2 securely delegates bounded, accountable financial limits to AI agents [01:01:17]. This operates alongside a cross-merchant “Universal Cart” that checks cross-retailer inventory, points out missing hardware incompatibilities, and sources hidden credit card savings automatically [01:03:07].
4. Interactive Media & Hardware Ecosystems
- Creative Tools: Google Pix allows contextual image editing and formatting [01:24:18], Stitch handles prompt-to-UI designs [01:25:33], and Google Flow / Flow Music empowers video asset variations and R&B audio generation [01:28:22].
- Intelligent Eyewear (Android XR): Built in partnership with Samsung, Qualcomm, Warby Parker, and Gentle Monster [01:35:21]. Google announced display glasses alongside audio glasses coming this fall [01:33:40]. These glasses stream hands-free map navigation, read notifications, and utilize background apps (e.g., ordering coffee via DoorDash) straight into your ear [01:40:51].
5. Security & AI for Science
- Synth ID Scaling: To handle deepfakes and transparency, Google expanded invisible watermarking. It is natively integrated across Chrome, Google Search, and adopted by industry partners like OpenAI and ElevenLabs [21:04].
- Gemini for Science & Weather Next: DeepMind continues targeting drug discovery and planetary simulation. Its updated climate model, Weather Next, accurately predicted a Category 5 Jamaican hurricane 3 days early, demonstrating the potential of scaling digital twin technologies to solve real-world crises [01:47:37].