Global AI Weekly
Issue number: 99 | Tuesday, May 27, 2025
Highlights
At Google I/O, Google claims it will build the first AGI
Google co-founder Sergey Brin made an unexpected appearance at Google I/O, where he confidently announced that "Gemini will be the very first AGI." This bold statement highlights contrasting approaches within the company, as DeepMind CEO Demis Hassabis advocates for a more measured and cautious approach in the rapidly advancing AI race. The surprise declaration added a layer of excitement and philosophical debate to the event.
venturebeat.com
News publishers call Google’s AI Mode ‘theft’
The News/Media Alliance, representing major US news publishers, has strongly criticized Google's AI Mode, accusing it of unfairly using their content. They claim the tool is reducing traffic and revenue for publishers by compiling information without driving readers to the original sources. This concern highlights the growing tension between tech platforms and traditional media over content usage and earnings.
theverge.com
Mistral AI launches Devstral, powerful new open source SWE agent model that runs on laptops
Mistral AI has unveiled Devstral, a cutting-edge open-source software engineering agent model designed to operate seamlessly on laptops. With impressive performance capabilities and portability, it stands out as a practical solution for developers. Its Apache 2.0 license also makes it an attractive choice for commercial applications, offering flexibility and accessibility for businesses and individuals alike.
venturebeat.comResearch
Chain-of-Thought May Not Be a Window into AI's Reasoning
Anthropic's new study sheds light on the limitations of chain-of-thought reasoning in AI systems, challenging the assumption that these step-by-step processes accurately reflect how AI "thinks." The research highlights hidden gaps where models might appear logical but fail to truly understand the problems they address. This raises questions about the reliability of such reasoning methods in evaluating AI performance.
marktechpost.com
MMaDA: Multimodal Large Diffusion Language Models
This paper introduces MMaDA, a framework that combines large diffusion models with multimodal language capabilities. It explores how integrating text, images, and other modalities can enhance language model performance and generate richer outputs. The discussion highlights the advancements and potential applications of this approach in various fields.
huggingface.coVideo
Humans in AI - Bill Ayers
Bill Ayers shares his perspective on AI as more than just a technological shift but a revolution in how we think and work. From his experience in training individuals on Microsoft 365 and AI in Derbyshire, UK, he explores how AI can free us from mundane tasks while introducing new challenges. This thought-provoking discussion highlights the transformative potential of AI in shaping the future of work and human creativity.
youtube.com
Highlights from Microsoft Build 2025: Satya Nadella's Keynote Recap
Microsoft Build 2025 kicked off with Satya Nadella exploring the concept of the agentic web and the vast opportunities it presents for developers across the tech stack. The opening keynote highlighted pivotal moments and key announcements, setting the stage for new possibilities in technology and innovation. The event emphasized Microsoft's commitment to empowering developers and shaping the future of digital experiences.
youtube.com
LangGraph Platform: MCP Support For Your LangGraph Agents
This video highlights how the LangGraph Platform integrates with the Model Context Protocol (MCP), enabling LangGraph agents to instantly act as MCP servers without additional coding or infrastructure. It explains how the built-in MCP support simplifies deployment and accelerates integration with other clients using MCP's streamable HTTP specification. The step-by-step overview covers creating deployments, accessing MCP documentation, configuring clients, and utilizing features like native tracing and multi-agent support. Ideal for developers and enterprises, this feature streamlines the process of deploying intelligent agents efficiently.
youtube.comArticles
Falcon-Arabic: A Breakthrough in Arabic Language Models
The Technology Innovation Institute shares insights on Falcon-Arabic, a cutting-edge Arabic language model designed to enhance natural language processing capabilities for the Arabic-speaking world. This advanced model focuses on improving linguistic accuracy and context understanding across diverse Arabic dialects, setting a new standard for AI applications in the region. The post highlights its potential impact on industries like education, healthcare, and media, enabling better communication and technology integration.
huggingface.co
Introducing Claude 4
Claude 4 brings advanced AI capabilities designed to handle complex tasks with greater reliability and ease. Whether you're navigating professional challenges or enhancing your learning experience, this AI offers more accurate and interpretable support to make your work smarter and more efficient.
anthropic.comUpcoming Events
June: AgentCon 2025
The AI Agents World Tour, part of AgentCon 2025, is a global series of one-day events tailored for developers working with autonomous AI agents. This June we will be bringing together top engineers and researchers. These conferences will take place in major cities like San Francisco and Singapore. It's a unique opportunity to connect, learn, and shape the future of AI technology.
agentcon.devCode
Tiny Agents: an MCP-powered agent in 50 lines of code
Learn how to create a minimalist agent in just 50 lines of code using the Hugging Face Machine Learning Core Programming (MCP) tools. This quick tutorial demonstrates the power and simplicity of MCP for building efficient AI agents. Perfect for those looking to dive into AI development with streamlined, open-source principles!
huggingface.co
Techniques for improving text-to-SQL
Discover key techniques to enhance text-to-SQL performance, including strategies like context building, effective table retrieval, and utilizing LLM-as-a-judge. Explore methods for crafting better prompts for language models and refining outputs through post-processing. Gain insights into optimizing these processes for more accurate and efficient database querying.
cloud.google.com