Google I/O '25 Developer Keynote

Key Concepts

Gemini 2.5 Pro/Flash, Stitch (code & design blending), AI Studio, Live API, URL Context, Function Calling, MCP (Model Component Protocol), Google GenAI SDK, Androidify (AI-powered selfie app), Gemini Nano (on-device model), Material 3 Expressive, Live Updates (Android 16), R8 & Baseline Profiles, Compose Adaptive Layouts, Android XR, Jetpack Compose, Gemini Code Assist, Firebase Studio, Gemma (open models), MedGemma (medical models), Unsloth, AI-first Colab, SignGemma, DolphinGemma, Chrome DevTools AI Assistance, Baseline (feature availability), Interest Invoker API, Scroll-driven animations, Scroll button pseudo-elements, Target Current pseudo-class, Scroll marker pseudo.

Stitch: Blending Code and Design

Google Labs is experimenting with a new project called Stitch, which aims to blend code and design using Gemini 2.5 Flash. The goal is to enable developers to go from a prompt to an interface to code in a short amount of time.

• Process:
  1. Start with a design prompt (e.g., "make me an app for discovering California").
  2. Stitch generates design screens based on the prompt.
  3. Users can iterate on the design by modifying parameters like dark mode, lime green color, and corner radius.
  4. Stitch uses Gemini 2.5 Pro/Flash to remix and change the design based on user input.
  5. Users can grab the markup (code) directly from the design and copy it to their IDE.
  6. The designs can also be edited further in Figma.
• Key Features:
  • Generates actual designs, not just static screenshots.
  • Allows for iterative design changes.
  • Provides access to the underlying code.
  • Integrates with Figma.
• Access: Labs.Google/Stitch

Gemini API Updates in AI Studio

Logan Kilpatrick presented updates to the Gemini API in AI Studio, focusing on making it easier to build AI voice agents and other applications.

• Live API Updates:
  • New 2.5 Flash native audio model.
  • Proactive audio controls to ignore stray sounds.
  • Native support for 24 languages.
  • Improved function calling and search counting.
  • New URL Context tool: Enables Gemini models to access and pull context from web pages using a link (up to 20 links at a time). This allows grounding model responses on specific and up-to-date information.
  • Example: Passing a link to Developer docs for function calling and having Gemini provide a TLDR.
• Code Editor Integration:
  • Gemini 2.5 Pro is integrated into AI Studio's native Code Editor.
  • Optimized with the SDK for generating web applications that use the Gemini API.
  • Example: Generating an AI-powered adventure game using Gemini and Imagen.
  • The model reasons about the request, composes a spec for the app, generates the code, and self-corrects errors.
  • The experience is multi-turn and iterative, allowing users to refine their ideas with prompts.
• MCP (Model Component Protocol) Support:
  • The Google GenAI SDK now natively supports MCP definitions.
  • Facilitates building agentic apps with open-source tools.
  • Example: A new app using Google Maps and MCP.

Building Agentic Apps with Gemini: Keynote Companion Demo

Paige Bailey demonstrated building an agentic app using the Gemini API, Google Maps, and MCP, creating a "keynote companion" named KC.

• Functionality:
  • KC listens to the keynote, responds, and dynamically updates its UI based on what it hears.
  • KC counts the number of times "AI" or "Gemini" is mentioned.
  • KC displays Shoreline Amphitheater on a map and finds nearby coffee houses with Wi-Fi.
  • KC provides directions to a randomly selected coffee house (Boba Bliss).
• Key Features Demonstrated:
  • Sliding Context Window: The Gemini Live API supports long-running sessions.
  • Asynchronous Execution: Enables seamless dialogue within a conversation by allowing functions to execute in the background.
  • Improved Structured Outputs: The model conforms to a specific JSON return format for displaying information in the UI.
• Deployment:
  • The app can be easily shared and deployed via Cloud Run directly from AI Studio.
  • The app can also be run and viewed with VS Code.
• Quote: "We're making it easy for you to build agents with Gemini, combining multimodal reasoning with a vast and a growing number of tools."

Building Excellent Android Apps with AI

Diana Wong and Florina Muntenescu discussed building excellent apps powered by AI on Android, focusing on delight, performance, and cross-device compatibility.

• Androidify App:
  • A new app that uses AI to create an Android robot based on a selfie.
  • Uses Gemini models in the Cloud via Firebase.
  • Process:
    1. Get a description of the person in the photo using Gemini's multimodal capabilities (text prompt + image input).
    2. Generate an Android robot based on the image description using the Imagen 3 model.
  • The Androidify sample app is available on GitHub.
• On-Device AI with Gemini Nano:
  • Gen AI APIs powered by Gemini Nano offer APIs for common tasks like summarize, rewrite, and image description.
• Delightful Apps:
  • Material 3 Expressive: An update to the Material 3 design system with new features and improvements.
  • Live Updates (Android 16): Allows showing time-sensitive updates for navigation, deliveries, or rideshares using the new ProgressStyle template.
• Performant Apps:
  • Enable R8 and Baseline profiles for improved performance.
  • Example: Reddit's app improved so much they got a full star rating increase within two months with R8 and Baseline profiles.
• Adaptive Apps:
  • Android 16: API changes to no longer react to orientation, resizability, and aspect-ratio restrictions.
  • Enhanced desktop windowing capabilities in Android 16 for more powerful productivity workflows (collaboration with Samsung DEX).
  • Compose Adaptive Layouts library: New features like pane expansion.
  • Cross-screen users in app categories like music, entertainment, and productivity have a 2 to 3 times increase in engagement.
  • Cross-screen users are twice as likely to use Canva every single week.
• Expanding to More Devices:
  • Android apps are being brought to more devices automatically, including cars and XR.
  • Android XR: An extended reality platform built together with Samsung.
  • Developer Preview 2 for Android XR SDK: New material XR components, updated emulator support in Android Studio, and spatial video support for Play Store listings.
• Productivity Tools:
  • Jetpack Compose: 60% of the top one thousand apps take advantage of the development speed Compose offers.
  • CameraX and Media3 Compose libraries are being released.
  • Rebuilt Jetpack Compose navigation library: Simpler, more intuitive, and powerful for managing screens in a stack, retaining state, enabling seamless animations and adaptive layouts.
  • Gemini and Android Studio: Streamline the entire development life cycle, from refactoring to testing and even fixing crashes.
• Gemini in Android Studio Demos:
  • Natural Language End-to-End Testing: Use natural language to perform actions and make assertions.
  • AI-Assisted Dependency Updates: Gemini can help with version updates by analyzing the modules, checking for library updates, building the project, and using Gemini to fix problems.
• Gemini Code Assist for Businesses:
  • Designed to meet the privacy, security, and management needs of businesses.

Building a More Powerful Web with Chrome

Una Kravets and Addy Osmani presented new features in Chrome to help developers build better UI, debug sites more easily, and create AI features more quickly and cost-effectively.

• Building Engaging UI:
  • Leveraging HTML and CSS to create beautiful, accessible, declarative, cross-browser UI.
  • Carousel Example:
    • Using new CSS primitives to build carousels and other types of off-screen UI dramatically easier.
    • CSS classes for positioning items, setting overflow, and requiring snapping at the center.
    • Scroll button pseudo-elements for navigation buttons.
    • Scroll marker pseudo for navigation dots (indicators).
    • Target Current pseudo-class that manages the active marker classes.
    • Pinterest switched to using new CSS APIs for carousels, cutting down around 2,000 lines of JavaScript into just 200 lines of more performant, browser-native CSS, improving product pin load times by 15%.
  • Interest Invoker API:
    • Used with the existing anchor positioning and popover APIs to build accessible, complex, layered UI elements without JavaScript.
    • Example: Adding a seat preview for each section in a virtual theater layout.
    • The browser handles state management, event listeners, ARIA labeling, and more.
• Baseline:
  • Shows feature availability across all major browsers.
  • Available in IDEs, linters, and analytics tools.
  • VS Code: Baseline status and browser availability are shown in the tooltip when hovering over CSS additions.
  • ESLint can be configured to warn about anything that doesn't match the targeted baseline version for HTML and CSS files.
• AI in Chrome DevTools:
  • AI Assistance: Baked into the panel to help debug and fix issues.
  • Example: Using natural language to ask Gemini how to fix a misaligned button.
  • AI assistance can create and apply a fix directly from DevTools.
  • Redesigned Performance Panel:
    • Highlights layout shift culprits.
    • Ask AI button to get help understanding and fixing problems.
    • Example: Asking Gemini how to prevent layout shifts on a page.
• Gemini Nano in Chrome:
  • Seven AI APIs across various stages of availability.
  • Backed by on-device models, from Gemini Nano to Google Translate.
  • Data never leaves the device.
  • Deloitte is experimenting with an integration of Chrome's built-in AI APIs into their engineering platform to improve onboarding and navigation.
• Multimodal Capabilities from Gemini Nano:
  • Built-in AI APIs let users interact with Gemini using audio and image input.
  • Example: Helping people find their seat in a theater using a photo of their ticket.
  • Partnered with Gemini and Firebase to offer a hybrid solution that works everywhere, on-device or in the Cloud.

Firebase Studio Updates

David East presented updates to Firebase Studio, a Cloud-based AI Workspace where you can create a fully functional app with a single prompt.

• Figma Integration:
  • Bring Figma designs to life in Firebase Studio with help from Builder I/O.
  • Install the Builder I/O plugin in Figma, click to export to Firebase Studio, and it translates all the component code.
  • Example: Importing a Figma mock of a furniture store app.
• Gemini in Firebase Studio:
  • Ask Gemini for advice on where to begin.
  • Example: Asking Firebase Studio to build a single-product detail page using the existing component system and sample data.
  • Gemini breaks changes into multiple steps, making it easier to review each change.
  • Gemini can update placeholder data and generate descriptions for products.
• Backend Generation:
  • Firebase Studio will detect when your app needs a backend and provision it for you if your prompt includes a database or authentication.
  • Firebase Studio will set up the configuration for the backend services and generate code to authenticate users and save data to a database.
  • Firebase Studio will provision those backend services and deploy to Firebase App Hosting.

Gemma Open Models Updates

Gus Martins presented updates to Gemma, Google's family of open models.

• Gemma 3n:
  • A model that can run on as little as two gigabytes of RAM.
  • Shares the same architecture as Gemini Nano.
  • Engineered for incredible performance.
  • Much faster and leaner on mobile hardware compared to Gemma 3.
  • Added audio understanding, making it truly multimodal.
  • Available in preview on Google AI Studio and with Google AI Edge.
  • Coming to open-source tools like Hugging Face, Ollama and Unsloth.
• MedGemma:
  • A collection of open models for multi-modal medical text and image understanding.
  • Works great across a range of medical, image, and text applications.
• Demo: Fine-Tuning Gemma with Unsloth and AI-First Colab:
  • Using Unsloth, a library for fine-tuning LLMs like Gemma.
  • Fine-tuning Gemma to create a personalized translator for an emoji language.
  • Using the new AI-first Colab to generate a UI for comparing the custom model against the original one.
• Gemmaverse:
  • Tens of thousands of model variants, tools, and libraries created by the developer community.
  • Gemma has been downloaded over 150 million times.
  • Close to 70,000 Gemma variants have been created.
  • Gemma is available in over 140 languages and is the best multilingual open model on the planet.
• SignGemma:
  • A new family of models trained to translate sign language to spoken language texts.
  • Best at American Sign Language in English.
  • The most capable sign language understanding model ever.
• DolphinGemma:
  • The world's first large language model for Dolphins.
  • Fine-tuned on data from decades of field research to help scientists better understand patterns in how dolphins communicate.

Conclusion

The presentation showcased a range of new tools and updates designed to empower developers to build innovative and impactful applications using Google's AI technologies. From simplifying UI development with CSS and AI-powered DevTools to enabling on-device AI with Gemini Nano and fostering open-source collaboration with Gemma, the focus was on making AI more accessible, efficient, and scalable for developers across various platforms and use cases. The demos and examples highlighted the potential of these tools to transform industries and create new possibilities for human-computer interaction.