What's new in Google Cloud

Key Concepts

• Foundation Models: Pre-trained AI models from Google, open-source, and partners available on Vertex AI.
• Vertex AI: Google Cloud's AI platform for building, training, and deploying models and agents.
• Optimizer: Vertex AI service that automatically selects the most cost-effective or fastest model for a given request.
• Supervised Fine-tuning: Training a model with labeled data to improve its performance on specific tasks.
• Retrieval Augmented Generation (RAG): Using external data to provide context for model responses.
• In-context Learning: Providing additional data within the prompt to guide the model's response.
• Agent Builder: A suite of tools for building, deploying, and managing AI agents.
• Agent Development Kit (ADK): An open-source framework (Python and Java) for building agents.
• Model Context Protocol (MCP): A standard for agents to interact with tools and data sources.
• Agent-to-Agent Protocol (A2A): A standard for agents to communicate and share capabilities.
• Agent Engine: A fully managed service for running and scaling AI agents.
• Agent Space: An enterprise platform for deploying and accessing AI agents.
• Gemini Code Assist: Google Cloud's AI-powered coding assistance tool.
• Firebase Studio: A platform for building end-to-end applications with AI integration.
• Application Design Center: A tool for visually modeling, deploying, and operating cloud applications.
• Cloud Hub: An app-centric operations dashboard for monitoring application health, cost, and performance.
• Gemini Cloud Assist: An AI-powered tool for troubleshooting and optimizing cloud applications.
• Cloud Run: A serverless platform for running containerized applications.
• BigQuery: Google Cloud's data warehouse and analytics service.
• Looker: Google Cloud's business intelligence and data visualization platform.
• Cloud WAN: A wide area network service built on Google's network backbone.
• GKE Inference Gateway: A gateway for managing and routing traffic to models deployed on Google Kubernetes Engine (GKE).
• Google Distributed Cloud: A platform for running Google Cloud services on-premises or in air-gapped environments.
• TPU: Tensor Processing Unit, Google's custom AI accelerator.

AI Models and Vertex AI

• Foundation Models on Vertex AI: Hundreds of curated models from Google, open-source, and partners are available.
• Gemini Models:
  • Gemini 1.5 Pro and Flash are coming soon.
  • Live API enables real-time interactions.
  • V3 allows for video and audio creation.
  • V2 and Llama 3 are generally available.
• Vertex AI Optimizer: Automatically selects the best model (e.g., Flash or Pro) based on cost and performance requirements.
• Pre-training Options: Vertex AI supports supervised fine-tuning (with Gemini 1.5), RAG (with more models), full fine-tuning, and in-context learning.
• Vertex AI Global Endpoint: Simplifies model deployment by automatically selecting the most available and performant region (excluding context caching).

Agent Builder and Agent Technologies

• Agent Builder Ecosystem: Includes the Agent Development Kit (ADK), Agent Garden, and integration with other tools.
• Agent Development Kit (ADK):
  • Version 1.0 released for Python and Java.
  • Allows building agents with instructions, models, and tools without extensive coding.
  • Supports non-Gemini models and external tools.
  • Enables local testing.
• Model Context Protocol (MCP): Facilitates agent interaction with tools and data sources.
• Agent-to-Agent Protocol (A2A): Enables agents to communicate and share capabilities (adopted by Microsoft). SDK available in Python.
• Agent Engine: A fully managed service for running, scaling, and managing agents on GKE, Cloud Run, and other platforms. New dashboard available.
• Agent Space: An enterprise platform for deploying and accessing agents for internal use, connecting to employee directories, third-party systems, and Google Workspace services.

AI at the Edge

• Gemma 3N: A small, high-performing model for mobile devices.
• Model Support: Increased model support for the Light RT and Hugging Face communities.
• On-device RAG: Enables local retrieval augmented generation on mobile devices.

App Developer Tools

• Gemini Code Assist:
  • Free for individual use.
  • Integrates with Google Docs, GitHub, GitLab, and other systems.
  • Supports code completion, generation, and training on custom codebases.
  • Includes coding agents for code reviews (e.g., GitHub agent).
  • Custom rules for code generation (e.g., always generate unit tests).
  • Chat history for prompt and response tracking.
• Firebase Studio:
  • End-to-end application development platform.
  • Supports AI for text-to-app generation.
  • Provides a full environment for local emulation and testing.
  • Integrates with Figma via Builder.
  • Features Firebase Logic (AI-powered backend).
• Application Design Center:
  • Visually model and deploy cloud applications.
  • Share templates with teams.
  • One-click deployment of application architectures.
• Cloud Hub:
  • App-centric operations dashboard.
  • Provides insights into application health, cost, and performance.
  • Displays the most expensive service in the application.
• Gemini Cloud Assist:
  • AI-powered troubleshooting and optimization tool.
  • Investigates logs and identifies potential issues.
  • Suggests solutions and optimizations.
• Cloud Run:
  • One-click deployment from Google AI Studio.
  • Cloud Run MCP server for integration with Cursor and other tools.
  • One-click deployment of Gemma 3 instances with GPUs.
• Security: Unified security offerings and integrations for Chrome.

Data and Analytics

• MCP Toolbox for Databases: Open-source project for connecting to various data services (Google Cloud and third-party) with a common interface.
• Firestore with MongoDB Compatibility: Provides a MongoDB-like instance with serverless scaling.
• Database Center: Manages fleets of databases (Spanner, Firestore, AlloyDB, Cloud SQL, MemoryStore) with a global view of health and performance.
• BigQuery AI:
  • AI-powered engine for data operations.
  • Natural language support.
  • Gemini integration for code completion (SQL and Python) and code explanation.
  • Multimodal tables for interacting with binary objects and images.
  • Vector search for creating vector stores and accessing embeddings.
  • AI-assisted notebooks for building and optimizing notebooks.
• BigQuery Governance:
  • Tools and dashboards for managing data.
  • Business glossary for common terminology.
  • Metadata export to Cloud Storage.
  • At-scale categorization.
  • Disaster recovery management.
• Third-Party Systems:
  • Apache Iceberg support.
  • Serverless Spark integration.
  • Metastore for centralized data access across BigQuery, Spark, and Iceberg.
• Continuous Queries: Real-time query execution on streaming data.
• History-Based Query Optimization: Automatic tuning based on past query patterns.
• Pipe Syntax: New syntax for querying data.
• Analytics and Geospatial: Enhanced support for maps and geospatial data.
• Looker and BigQuery:
  • Conversational analytics for natural language interaction with data.
  • Spectacles for CI/CD of Looker dashboards.

Infrastructure

• Cloud WAN: Wide area network service built on Google's network backbone.
• Cloud Run GPUs: GPUs available on serverless instances for AI workloads. No quota requests required.
• DNS Armor: DNS protection service.
• GKE Inference Gateway: Manages and routes traffic to models deployed on GKE, enabling intelligent routing and observability.
• Gemini on Google Distributed Cloud: Gemini AI available on-premises (connected or air-gapped).
• VLM on Kubernetes: Project with Red Hat, IBM, and CoreWeave for distributed inference across Kubernetes clusters.
• Ironwood: 7th generation TPU with improved performance and energy efficiency.

Conclusion

The presentation highlights Google Cloud's advancements in AI, application development, data analytics, and infrastructure. Key takeaways include the availability of powerful foundation models on Vertex AI, tools for building and deploying AI agents, AI-powered coding assistance, and innovations in serverless computing and data management. The emphasis is on making AI more accessible and easier to integrate into various applications and workflows, while also providing robust infrastructure and security solutions.