6/4/2025 AI Express | Google Cloud Run Enables Serverless GPU Acceleration and Overview of Latest AI Advances

Google Cloud Run now supports serverless GPU acceleration, providing developers with an efficient and flexible way to use compute resources. With this new feature, users can deploy applications such as Gemma on Google Cloud Run with a single command. This not only simplifies the development process, but also provides a flexible billing model with per-second billing, further reducing costs.

In addition, the service supports automatic scaling of instances to zero, ensuring that no additional costs are incurred when demand is low. It also supports fully streaming processing for application scenarios that require real-time data processing. These features make Google Cloud Run an ideal choice for many developers and organizations.

Key features include:

• No need to request a quota
• Single-command rapid deployment
• per-second billing
• Automatic scaling to zero instances
• Full streaming support

Google Cloud Run's serverless GPU acceleration is certainly an attractive option for developers who need high-performance computing but don't want to manage the underlying infrastructure.

OpenAI released two new model snapshots today:gpt-4o-audio-preview-2025-06-03 respond in singing gpt-4o-realtime-preview-2025-06-03, designed to improve command execution, tool invocation, and interrupt behavior. In addition, a new speed parameter has been introduced to allow the user to control the speaking rate of the voice.

These updates further enhance the model's interaction flexibility and user experience. At the same time, OpenAI also introduced the Agents SDK for TypeScript, which supports building voice assistants with real-time agent functionality. The feature can run client-side or server-side, and adds support for the Traces dashboard, which enables developers to visualize the operation of the voice agent, including audio input/output and tool invocation.

These new features not only provide developers with a more robust toolset, but also open up new possibilities for building smarter, more responsive voice applications.

A new short course called DSPy was recently launched.DSPy is a powerful open source framework focused on automatically tuning prompts in generative artificial intelligence (GenAI) applications. The course was developed in collaboration with Databricks and is taught by ChenMoneyQ, co-leader of the DSPy framework.

In a context where many AI developers spend a lot of time manually tuning cues, DSPy is able to automate this process, especially in complex AI workflows, where it is particularly effective for optimizing a small number of sample cues. Additionally, when an application switches to a new language model, performance may degrade if the original cues were optimized for the old model, and DSPy can automatically optimize the entire system for the new language model with a small number of evaluated samples.

This course will provide an in-depth explanation of how DSPy works and its best practice approach. Participants will learn how to write programs using DSPy's signature programming model and debug them using MLflow to gain visibility into the performance of individual parts of the pipeline and the overall system. Ultimately, automatic accuracy improvement is achieved with DSPy Optimizer.

The course covers:

• Basic Concepts and Installation and Configuration of DSPy Framework
• How to define and manage prompt templates
• Methodology for tracing and debugging with MLflow
• Practical case studies and projects

The link to sign up is as follows: https://t.co/bb8uILyepf

Building a Multi-Agent Financial Report Generating Chatbot from Scratch

In @jerryjliu0's workshop last week, he shared a complete example of a LlamaIndex agent-based workflow. This example shows how to build a chatbot from scratch that can automatically generate financial reports.

1. Parsing and Indexing Adobe's 10-K Files
2. Answering Analytical Queries Using Agentic Enhanced Search (agentic RAG)
3. Generate complete reports with descriptions, tables and evaluations through chained agents

This hands-on lab program is not only for financial analysts, but also for any professional looking to improve their data processing and automation skills.

You can try this example at the following link: https://t.co/tOG6hBgSVk

Recently.OpenAIState-of-the-art image generation models forGPT-Image-1A remarkable achievement in the field of text-to-image generation, successfully reaching the top position in the field. This breakthrough not only marks theOpenAIAnother major leap forward in AI research that also has far-reaching implications for the industry as a whole.

According to the latest reviewsGPT-Image-1It outperforms Google's Imagen-3.0 on a number of key metrics, with a lead score of more than 50 points, and is also a quantum leap over the previous DALL-E 3. This indicates thatGPT-Image-1A new level of accuracy and diversity in image generation has been achieved.

OpenAIannounced that, effective immediately, developers can access the API through theGPT-Image-1The powerful image generation function of the This feature has the following characteristics:

• More accurate, high-fidelity image generation capabilities
• Variety of visual styles
• Precise image editing functions
• A wealth of world knowledge
• Consistent text rendering

These characteristics makeGPT-Image-1Become a powerful tool in many areas of creative design, content creation and multimedia applications.

For developers, this means they can utilize this technology to create more innovative and useful image content. At the same time, it opens up more possibilities for a wide range of industries and promotes the widespread use and development of image generation technology.

congratulationsOpenAIteam to achieve this significant result! To learn more about how to develop with GPT-Image-1, see the official guide.

OpenAI recently released a new feature that allows workspace administrators to build custom deep research connectors using the Model Context Protocol (MCP), which is in beta.

With MCP, users can integrate proprietary systems and other applications with ChatGPT to enable search, reasoning and manipulation of these knowledge sources. This technology not only supports access to on-premises data, but also works with web results and pre-built connectors to support teams with more comprehensive information.

The feature is now available to Team, Enterprise, and Education administrators, and Professional Edition users are able to start using it as well.

To help developers better understand and apply this feature, a detailed official tutorial document (link: https://t.co/uq4DgFvZ8h) is available, which includes specific steps for setting up a remote MCP server:

1. First, make sure your environment meets the requirements of the MCP protocol.
2. Next, follow the guide to configure the server endpoint.
3. Finally, connect your MCP server to ChatGPT via the API interface.

Through the above steps, enterprises can not only make full use of existing information resources, but also further expand the boundaries of the capabilities of AI assistants.

Google recently open-sourced a set of Google Gemini Fullstack LangGraph is a framework designed to rapidly build production-grade DeepResearch systems. By combining Gemini's powerful inference capabilities with LangGraph's state management capabilities, developers can easily create an AI assistant capable of executing complex, multi-step, iterative research processes.

The entire Agent workflow, including the steps of dynamic query generation, network research, reflective analysis, iterative optimization, and answer synthesis, is encapsulated in the framework. In addition, LangGraph provides flexible interfaces that allow developers to replace the APIs in it as needed to meet the demands of specific application scenarios.

The framework is especially suited for scenarios that require deep research, such as academic research, market analysis, and technological innovation. Gemini Fullstack LangGraph is a good choice for developers looking to build a Research Agent or looking for a DeepResearch open source alternative.

At the same time, this project also provides an opportunity for AI developers with a certain LangGraph foundation to learn and learn from it, and stand on Google's shoulders to better understand and master the relevant technology.

GitHub repository address:https://t.co/QwmWzJhWuF

Hugging Face has officially launched a new free course, MCP Course, which aims to help developers fully grasp the theory and practice of Multi-Modal Pretrained Models (MCP) from scratch.

The course covers three main modules: MCP theory foundation, practical development and deployment application. Through systematic learning, students will not only gain a deep understanding of the core concepts, architecture and components of MCP, but also master the skills of practical development and deployment applications, and ultimately obtain the Hugging Face official certification.

The course outline includes:

• Getting Started Guide: Introduces the tools and platforms used to prepare for the course;
• MCP Basics, Architecture, and Core Concepts: Explains the core concepts, architecture, and components of MCP and demonstrates simple use cases using MCP;
• MCP End-to-End Hands-On: Building a Simple End-to-End MCP Application;
• MCP Deployment in Action: Build a deployable MCP application and learn how to deploy it to a production environment.

Each chapter is designed to be completed in 1 week, requiring approximately 3-4 hours of study time per week.

Suitable for students with some basic AI and programming knowledgeUpon completion of the course, you will be able to familiarize yourself with the MCP SDK and framework for development.

MCP technology has a wide range of applications in many fields such as natural language processing, computer vision, etc., and is one of the important directions of current AI research and development. Through this course, students can not only improve their technical skills, but also apply what they have learned in real projects.

Course address: https://t.co/cvOQNAdpOk

The Aria Gen 2 eyewear marks a significant leap forward in wearable technology, offering enhanced functionality and performance to meet the needs of a wider range of applications and the needs of researchers. The eyewear excels in areas such as machine perception, contextual AI, and robotics, significantly accelerating research efforts in industry and academia.

Equipped with advanced sensors and high-resolution cameras, Aria Gen 2 glasses support real-time data acquisition and processing, providing researchers with high-quality data support. In addition, its built-in computing unit and efficient algorithm optimization enable a significant increase in data processing speed, helping to improve research efficiency.

Key application scenarios include:

• Machine vision and perception technology
• Situational Intelligence and Environmental Understanding
• Human-computer interaction and collaboration

The Aria Gen 2 glasses are undoubtedly a powerful tool for researchers working to develop the next generation of intelligent systems. For more information and to pre-order, please visit: https://t.co/aEyjaw0rge.

If you're using ClaudeCode, try theclaude-traceThis powerful logging tool. It logs all ClaudeCode requests, including user prompt words, and saves this information in an HTML file for easy viewing.

The tool works in a very clever way, by starting itself first, then modifying the global.fetch API of Node.js, and then starting ClaudeCode. this way, when ClaudeCode sends a request, theclaude-traceIt is then possible to intercept and record the complete request, eventually generating a detailed report.

Application Scenarios:

• Developers can use these logs to debug and optimize code.
• Researchers can analyze these logs to understand the behavior and performance of the model.
• Users can use these logs to better understand how the model works.

There are three main models involved when using the Claude Max subscription service:

1. claude-3-5-haiku-20241022 for preprocessing user input.
2. claude-opus-4-20250514, the master model used by default, is responsible for the actual code writing and tool invocation.
3. claude-sonnet-4-20250514 as a backup when the default model quota runs out.

With claude-trace, you can get a deeper understanding of the working mechanism behind ClaudeCode to develop and research more efficiently.