Google for Developers

Large language models are trained in two main phases: pre-training and post-training. Pre-training uses next-token prediction on massive scale data, while post-training aligns the model through supervised fine-tuning and reinforcement learning. Evaluation methods for each phase are also discussed.

Topics:
- Pre-training: self-supervised learning at scale
- Post-training: supervised fine-tuning (SFT) and reinforcement learning (RL)
- Evaluating model quality: next-token prediction vs. alignment metrics

Anyone interested in the practical workflow of training LLMs, from data preparation to alignment techniques, will gain a clear overview of the entire process.

The Gemini Interactions API is now generally available, offering a unified interface for all Gemini models and agents with support for synchronous and background execution, media generation, and a stateful design. This video demonstrates building a photo editing app using the API, highlighting its iterative capabilities with previous interaction IDs.

■ Gemini Interactions API Features
- Unified interface for all Gemini models and agents
- Support for synchronous and background execution
- Stateful API allowing previous interaction IDs
- Integrated media generation (images, music, speech)

■ Demo: Building a Photo Editing App
- Setting up the API with the Gemini Interactions API skill in Anti-gravity IDE
- Implementing audio transcription and image editing using the API
- Iterative editing with cached tokens and previous interaction IDs

Developers interested in agent-based applications will learn about the API's capabilities and see a practical example of building an interactive image editor. Try it out using the recommended coding agent skill.

This project demonstrates how to build an AI-driven game development platform using Google technologies, enabling anyone to generate and deploy arcade games through natural language prompts. It showcases the concept of modular AI skills for consistent game output and a robust cloud-native deployment pipeline.

■ Project Overview
- AI-generated game creation at Google I/O using Anti-Gravity and Gemini
- Backed by Cloud Run, Firestore, Firebase Authentication, and Eventarc

■ AI Skills for Consistent Output
- Modular skill bundles for arcade visuals, input handling, and retro styling
- Dynamic loading prevents context window bloat and maintains quality

■ Deployment Pipeline
- Custom Go CLI tool with setuid for secure API token handling
- Automated build, upload, deduplication, and validation via Cloud Run
- Firestore-based game moderation for instant playability on arcade cabinets

Developers interested in AI agents, cloud infrastructure, or creative coding will understand how to leverage Google's ecosystem to build vibe coding environments and deploy AI-generated applications.

This video highlights Google-organized events where developers and tech enthusiasts gather to explore and experiment with AI-powered Google tools. The focus is on hands-on learning, community engagement, and fostering innovative solutions to real-world problems.

■ Event Focus: Hands-on Learning
- Workshops for developers to use and experiment with Google’s AI tools
- Training on AI agents and their role in the future

■ Community Engagement: Diverse Perspectives
- Opportunities to meet experts and industry peers
- Collaborative learning across different backgrounds

■ Impact: Social Problem Solving
- Spaces where new solutions to social issues are developed

Ideal for developers and tech enthusiasts interested in leveraging Google’s AI offerings and connecting with a like-minded community to gain practical skills and insights.

In this tutorial, you will learn how to build a custom Android app from scratch using AI Studio, with no technical or Android development experience required. The process is entirely prompt-driven, and you will see an example of a to-do and fitness tracker app that includes an AI chat feature powered by the Gemini API.

■ Build the app by describing it
- Describe your app (e.g., to-do and fitness tracker with AI chat powered by Gemini API)
■ Customize design (optional)
- Select from preview designs while building
■ Install on your Android phone
- Enable developer mode and USB debugging, connect via USB, and install
■ Iterate and update
- Add features or fix errors by describing changes and reinstalling

This guide is perfect for beginners who want to create functional Android apps quickly, and it illustrates how easy it is to prototype, test, and iterate using only natural language descriptions.

We cover how to create custom agent skills for on-device generative AI using the Google AI Edge stack and LiteRT LLM APIs. This workshop explains the benefits of on-device AI, the structure of agent skills, and provides a hands-on exercise to build a task planner and timer skill.

- Google AI Edge and LiteRT LLM APIs overview: benefits (low latency, privacy, offline, costs), key tools (LiteRT Torch, Quantizer, LiteRT, Portal, Model Explorer)
- Agent skills: definition, advantages (reduce repetition, dynamic capabilities, domain knowledge), skill format (metadata, instructions, function calls, executable code, assets)
- Workshop: step-by-step building a task planner and timer skill using the AI Edge Gallery app and Gemma 2B model, with a provided collab
- Resources and next steps: open-source AI Edge Gallery app, GitHub repository, and ways to extend skills

This workshop is for developers interested in deploying small on-device LLMs and expanding their functionality with agent skills. Viewers will gain practical knowledge to create custom skills and learn about available tools and resources.

This video explains how to implement dynamic callbacks for Google Pay on Android, a feature that enables real-time updates to shipping options, taxes, and total price within the Pay sheet. It covers the benefits of using callbacks for an express checkout experience, such as earlier Pay button placement and in-sheet authorization.

■ Benefits of Dynamic Callbacks
- Enable real-time shipping and tax updates based on user address
- Allow placing the Pay button earlier in the purchase funnel
- Support in-sheet authorization handling and retries

■ Implementation Steps
- Update Play Services Wallet dependency to version 20.0.0
- Extend BasePaymentDataCallback class to handle events
- Create a service extending BasePaymentDataCallbackService with the required permission
- Include callback intents in the PaymentDataRequest JSON

This video is intended for Android developers who want to optimize their checkout flow using Google Pay. After watching, you will understand how to integrate these callbacks and provide a faster, more modern purchasing experience.

Generative AI models from Google DeepMind enable a workflow that transforms a single text source into rich multimedia content, including images, video, music, and speech.

■ Text and Image Generation
- Analyzing the book and extracting character descriptions with structured output
- Generating consistent character and chapter illustrations using Nano Banana

■ Audio, Video, and Integration
- Creating videos from images with VEO
- Composing background music with Lyria
- Generating multi-character speech with TTS
- Combining all media into a final video and preview of Omni

Developers and content creators can gain practical insights for building integrated generative media pipelines.

This session introduces the Gemini API and AI Studio, showcasing how to build agentic applications from idea to production. It covers the new Interactions API, managed agents with remote sandboxes, and highlights key features like multimodal outputs, tool integration, and one-click deployments.

■ AI Studio Enhancements
- Full-stack Node.js app generation, Firebase integration, and Workspace connectivity
- Generative native Android apps via natural language (Kotlin/Jetpack Compose)
- Design previews, focus annotation, and prompt autocompletion
■ Gemini API & Models
- Gemini 3.5 Flash as the go-to model; multimodal inputs and outputs
- New Interactions API for unified agentic workflows with server-side state
■ Managed Agents
- Antigravity agent harness integrated into the API for remote sandboxed execution
- Agents defined via markdown files and skills (e.g., TTS generation, data analysis)
■ Demo: AI Talk Radio Applet
- Builds a radio show using multiple Gemini models for research, script, speech, music, and cover image

Viewers will learn to rapidly prototype with AI Studio, leverage the Interactions API for complex tasks, and deploy agents using managed sandboxes. Ideal for developers building AI-powered applications with Google's ecosystem.

This episode of Release Notes features a conversation between Logan Kilpatrick and Sameer Samat, head of Android, exploring the latest AI advancements in the Android ecosystem. The discussion covers new features like Rambler, Superfill, and App Automation, as well as the broader vision of transforming Android from an operating system to an intelligence system.

■ Key AI Features
- Rambler: AI voice dictation that refines and summarizes speech
- Superfill: Autofill for personal information from Google account
- App Automation: Gemini performs multi-step tasks in a virtual sandbox
- Generative Widgets: Customizable widgets created on demand

■ Platform and Ecosystem
- AI-driven features without overtly marketing AI
- On-device vs server-side AI processing
- Android Halo: status bar hub for agent interactions
- Gemini integration in vehicles for real-world queries

This video is ideal for Android users and developers seeking to understand how AI is being practically integrated into everyday tasks. Viewers will gain insights into upcoming features that enhance productivity, safety, and personalization, and learn about the platform's evolution toward more intuitive, intent-driven interactions.