Google I/O '25 Developer Keynote - Audio Described

Key Concepts

Gemini 2.5 Pro, Gemini 2.5 Flash, Stitch, Google AI Studio, Project Astra, Function Calling, URL Context, GenAI SDK, Model Context Protocol (MCP), Androidify, Gemini Nano, Material 3 Expressive, R8, Baseline Profiles, Jetpack Compose, Firebase Studio, Gemma, Mad Gemma, Unsllo, AI Collab, Sign Gemma, Dolphin Gemma, Chrome DevTools, AI APIs in Chrome, Multimodal AI, Firebase.

Stitch: Blending Code and Design

• Main Topic: A new labs experimental project called Stitch, designed to blend code and design for rapid prototyping.
• Key Points:
  • Allows users to go from a prompt to an interface to code in minutes.
  • Starts with design first, allowing users to paste a prompt (e.g., "make me an app for discovering California").
  • Generates actual designs, not just static screenshots.
  • Enables iteration on designs with options like dark mode, lime green, and corner radius adjustments.
  • Uses Gemini 2.5 Pro or Flash to remix and change designs.
  • Provides markup that can be copied to any IDE or edited further in Figma.
• Step-by-Step Process:
  1. Paste a prompt describing the desired app.
  2. Generate a design based on the prompt.
  3. Iterate on the design by modifying parameters.
  4. Copy the generated markup to an IDE or Figma.
• Real-World Application: Building an app for discovering California activities and getaways.
• Technical Terms: IDE (Integrated Development Environment), Figma (UI design tool).
• URL: labs.google/stitch

Google AI Studio: Building with Gemini API

• Main Topic: Using Google AI Studio to prototype and build AI voice agents with the Gemini API.
• Key Points:
  • Goal is to help developers determine if they can build something with Gemini and get started.
  • Utilizes the new 2.5 Flash native audio model.
  • Features proactive audio, improved noise handling, and support for 24 languages.
  • Introduces URL context, enabling Gemini models to access and pull context from web pages using links (up to 20 links).
  • Improvements to function calling and search grounding.
  • Native code editor with Gemini 2.5 Pro integration for building web applications.
• Important Examples:
  • Building an AI-powered adventure game using Gemini and Imagine.
  • Using Google Maps and MCP (Model Context Protocol) to build a new app.
• Step-by-Step Process:
  1. Select the 2.5 Flash native audio model in AI Studio.
  2. Provide a URL to developer documentation for function calling.
  3. Prompt Gemini to describe Google's approach to function calling based on the provided details.
  4. Use the native code editor to build a dynamic text adventure game.
• Technical Terms: Function Calling, URL Context, MCP (Model Context Protocol).

Paige's Keynote Companion: A Dynamic Web App

• Main Topic: Building a dynamic web app using the Gemini API and Google AI Studio.
• Key Points:
  • Remixing a maps app to create a keynote companion that listens, responds, and updates its UI dynamically.
  • Using a function called "increment word count" to track specific audio inputs.
  • Leveraging the Gemini live API's sliding context window for long-running sessions.
  • Integrating Google Maps to show locations based on voice commands.
  • Enabling asynchronous execution for seamless dialogue within a conversation.
  • Improving structured outputs with function calling to conform to specific JSON return formats.
  • Deploying the app via Cloud Run directly from AI Studio.
• Important Examples:
  • Counting the number of times "AI" or "Gemini" is said during the keynote.
  • Displaying Shoreline Amphitheater on a map and finding nearby coffee houses with Wi-Fi.
  • Providing directions to a randomly selected coffee house.
• Technical Terms: Synchronous Function Calls, Asynchronous Execution, JSON.

Android Development with AI

• Main Topic: Building excellent Android apps powered by AI, focusing on delight, performance, and adaptability across devices.
• Key Points:
  • Using AI models running in the cloud via Firebase for tasks like image description and generation (Androidify app).
  • Leveraging on-device AI with Gemini Nano for tasks like summarize, rewrite, and image description.
  • Implementing Material 3 Expressive for delightful UI design.
  • Enabling R8 and baseline profiles for improved app performance.
  • Adapting apps for various devices (foldables, tablets, Chromebooks, cars, XR) using Compose adaptive layouts.
  • Using Jetpack Compose for Android UI development.
  • Streamlining the development lifecycle with Gemini in Android Studio.
• Important Examples:
  • Androidify app: Taking a photo and generating an Android robot based on the description.
  • Reddit's app: Improved performance and a full star rating increase after enabling R8 and baseline profiles.
  • Canva: Cross-screen users are twice as likely to use Canva every week.
  • Peacock: Created a strong adaptive experience for their large screen app and gets a nice XR app.
  • K: Easily extended their composer to create multi-sensory mindful experiences only possible with XR.
• Step-by-Step Process:
  1. Use Gemini models via Firebase to get a description of a person in a photo.
  2. Use the Imagine 3 model to generate an Android robot based on the image description.
  3. Implement Material 3 Expressive for UI design.
  4. Enable R8 and baseline profiles for performance.
  5. Use Compose adaptive layouts for cross-device compatibility.
• Technical Terms: On-device AI, Gemini Nano, Material 3 Expressive, R8, Baseline Profiles, Jetpack Compose, Android XR.
• Data: 60% of the top 1,000 apps take advantage of the development speed Compose offers.

Gemini in Android Studio

• Main Topic: Using Gemini in Android Studio to streamline the development lifecycle.
• Key Points:
  • Using natural language to perform actions and make assertions in end-to-end tests.
  • Automating dependency updates with an AI agent that analyzes modules, checks for updates, and fixes build issues.
  • Accessing Gemini in Android Studio for businesses with Gemini Code Assist, designed for privacy, security, and management needs.
• Important Examples:
  • Testing Androidify features using natural language commands.
  • Updating dependencies to the latest versions with the AI agent.
• Step-by-Step Process:
  1. Write end-to-end tests using natural language.
  2. Use the AI agent to update dependencies.
  3. Review the changes and explanations provided by the agent.

Web Development with Chrome

• Main Topic: New features in Chrome to build better UI, debug sites more easily with DevTools, and create AI features more quickly and cost-effectively with Gemini Nano.
• Key Points:
  • Using CSS primitives to build carousels and other offscreen UI dramatically easier.
  • Using scroll button and scroll marker pseudo-elements for carousel navigation.
  • Using the experimental interest invoker API with anchor positioning and popover APIs to build accessible complex layered UI elements without JavaScript.
  • Using Baseline to show feature availability across all major browsers.
  • Using AI assistance in Chrome DevTools to debug and fix issues with natural language queries.
  • Using Gemini Nano in Chrome for on-device AI processing.
  • Using multimodal built-in AI APIs to create experiences where users can interact with Gemini using audio and image input.
• Important Examples:
  • Pinterest: Switched to using new CSS APIs for carousels, cutting down 2,000 lines of JavaScript into just 200 lines of more performant browser native CSS.
  • Deote: Experimenting with an integration of Chrome's built-in AI APIs right into the Deote engineering platform to improve onboarding and navigation.
  • Building a virtual theater site with carousels, hover cards, and AI-powered seat finding.
• Technical Terms: CSS Primitives, Scroll Button, Scroll Marker, Interest Invoker API, Anchor Positioning, Popover APIs, Baseline, Gemini Nano, Multimodal AI.
• Data: Pinterest saw a 5% improvement in LCP and INP and a 15% improvement in product pin load times.

Firebase Studio

• Main Topic: Using Firebase Studio, a cloud-based AI workspace, to create fully functional apps with AI assistance.
• Key Points:
  • Importing Figma designs into Firebase Studio with the Builder.io plugin.
  • Using Gemini in Firebase Studio to generate code and build app features.
  • Automatically provisioning a backend for apps that need a database or authentication.
• Important Examples:
  • Building a furniture store app with a product grid listing page and a single product detail page.
• Step-by-Step Process:
  1. Install the Builder.io plugin in Figma.
  2. Export the Figma design to Firebase Studio.
  3. Use Gemini in Firebase Studio to generate code for additional features.
  4. Let Firebase Studio provision a backend for the app.
• Technical Terms: Firebase Studio, Builder.io, Figma.

Gemma: Open Models

• Main Topic: Using Gemma, a family of open models, to fine-tune AI for specific use cases.
• Key Points:
  • Gemma 3N: A model that can run on as little as 2 GB of RAM.
  • Mad Gemma: A collection of open models for multimodal text and image understanding in healthcare.
  • Using Unsllo, a library for fine-tuning LLMs like Gemma.
  • Using AI Collab to build UIs for testing and comparing models.
  • Sign Gemma: A family of models trained to translate sign language to spoken language text.
  • Dolphin Gemma: The world's first large language model for dolphins.
• Important Examples:
  • Fine-tuning Gemma to create a personalized emoji translator.
  • Using Mad Gemma to analyze radiology images or summarize patient information for physicians.
  • Using Dolphin Gemma to understand patterns in how dolphins communicate.
• Technical Terms: Gemma, Mad Gemma, Unsllo, AI Collab, Sign Gemma, Dolphin Gemma.
• Data: Gemma has been downloaded over 150 million times, and the community has created close to 70,000 Gemma variants.

Conclusion

The presentation showcased a range of new tools and features designed to empower developers to build innovative applications using AI. From rapid prototyping with Stitch and Google AI Studio to streamlining Android development with Gemini in Android Studio and creating powerful web experiences with Chrome's AI APIs, the focus was on making AI more accessible and easier to integrate into the development workflow. The introduction of Gemma and its various specialized models, such as Mad Gemma and Dolphin Gemma, highlighted the potential for fine-tuning AI for specific domains and use cases. The overall message was clear: Google is committed to providing developers with the tools and resources they need to build the next generation of AI-powered applications across various platforms and devices.