Global AI Weekly
Issue number: 94 | Tuesday, April 22, 2025
Highlights
Microsoft researchers say they've developed a hyper-efficient AI model that can run on CPUs
Microsoft researchers have unveiled an AI model designed to deliver high efficiency while running on standard CPUs, including Apple’s M2. This innovation eliminates the need for more expensive and power-hungry GPUs, making advanced AI more accessible. The model’s lightweight design could significantly broaden the use of AI across various devices and applications.
techcrunch.com
Announcing the GPT-4.1 model series for Azure AI Foundry and GitHub developers
Microsoft has introduced the GPT-4.1 model series, including GPT-4.1, 4.1-mini, and 4.1-nano, to the Azure OpenAI Service. This new generation of models is designed to empower developers using Azure AI Foundry and GitHub, delivering enhanced capabilities and efficiency. These models open new possibilities for innovation and seamless integration into AI-driven applications.
azure.microsoft.com
Meta exec denies the company artificially boosted Llama 4's benchmark scores
A Meta executive has addressed and dismissed claims that the company manipulated its AI models, specifically Llama 4, to excel on benchmarks while hiding their flaws. The exec clarified that Meta's approach focuses on transparency and advancing AI ethically, countering the rumors of artificially inflating the model’s performance metrics. This comes as AI development and evaluation processes face increased scrutiny across the tech industry.
techcrunch.comResearch
Intuitive physics understanding emerges from self-supervised pretraining on natural videos
This paper explores how intuitive physics understanding can emerge from self-supervised pretraining on natural video data. By training models on raw, unlabeled videos, the research demonstrates that systems can grasp fundamental physical concepts like object permanence and motion prediction. The findings suggest the potential for scalable learning of physical reasoning without the need for extensive labels or direct supervision.
arxiv.org
CoRAG: Collaborative Retrieval-Augmented Generation
This paper page introduces CoRAG, a framework designed to combine collaborative retrieval with augmented text generation. It explores how integrating retrieval techniques enhances the efficiency and accuracy of generative models. The discussion includes the potential applications and benefits of this approach in advancing AI-driven information processing.
huggingface.coVideo
Silicon Minds, Human Hearts - Manash Goswami
In this episode of *Human Hearts Silicon Minds*, Manash Goswami, Principal Group Program Manager at Microsoft, shares his remarkable journey from the device industry to pioneering AI infrastructure. He tackles key challenges like running AI models locally and explores how these technologies are rapidly transforming to fit smaller devices. Manash also highlights the evolving complexities of AI infrastructure in a world dominated by generative AI and offers a glimpse into how these advancements are accelerating deployment on Azure. Plus, he reveals how AI is seamlessly integrated into his everyday life in unexpected ways.
youtube.com
Pre-Training GPT-4.5
Sam Altman joins Amin Tootoonchian, Alex Paino, and Daniel Selsam to talk about the process behind building GPT-4.5. They share insights on the challenges, innovations, and strategies involved in pre-training this advanced model. The conversation offers a behind-the-scenes look at what it takes to push AI capabilities forward.
youtube.com
Humans in AI - Sammy Deprez
This episode of Humans in AI features Sammy Deprez, a Nairobi-based freelance contractor specializing in generative AI and data solutions. Sammy discusses the potential for AI to enhance human capabilities, envisioning a future where technology provides greater freedom and leisure in our everyday lives. It’s an engaging conversation on how AI could transform us into "superhumans" while offering innovative benefits.
youtube.comArticles
Anthropic flips the script on AI in education: Claude's Learning Mode makes students do the thinking
Anthropic has introduced Claude for Education, featuring a unique Learning Mode aimed at fostering critical thinking in students instead of simply offering answers. In collaboration with leading universities, this initiative reimagines the role of AI in education, encouraging a more interactive and thought-provoking learning experience. By shifting the focus to student-led problem-solving, Claude aims to enhance how learners engage with AI tools.
venturebeat.com
Announcing the Agent2Agent Protocol (A2A)
Google introduced the Agent2Agent Protocol (A2A), an open standard designed to help developers create AI systems that can seamlessly work together. By prioritizing interoperability, A2A aims to simplify collaboration between different AI agents, expanding possibilities for innovation and integration. This initiative promises to streamline AI development, making it more adaptable and interconnected.
developers.googleblog.com
Agent mode: available to all VS code users and supports MCP
Agent mode is now open to everyone, offering seamless integration with MCP for a more efficient and streamlined experience. Users can enjoy enhanced functionality and wider accessibility, ensuring smoother workflows and improved performance.
code.visualstudio.comUpcoming Events
AgentCon 2025 - Netherlands
The AI Agents World Tour stops in Utrecht with AgentCon 2025, a dynamic one-day conference tailored for developers shaping the future with AI agents. This event gathers top engineers, researchers, and creators to share insights, strategies, and innovations. The main stage will be live-streamed as well! Be part of the global conversation pushing the boundaries of AI technology!
globalai.communityCode
GitHub - The-Pocket/Tutorial-Codebase-Knowledge: Turns Codebase into Easy Tutorial with AI
This project leverages AI to transform complex codebases into easy-to-follow tutorials, making it simpler for developers to understand and learn from them. By contributing to The-Pocket/Tutorial-Codebase-Knowledge on GitHub, you can help build a resource that bridges the gap between coding and comprehension, streamlining the learning process for programmers of all levels.
github.com
Sample Dotnet concepts for Semantic Kernel: semantic-kernel/dotnet/samples/Concepts
This resource provides straightforward examples to help you integrate advanced Language Model (LLM) technology into your applications using Semantic Kernel in .NET. The samples demonstrate core concepts, making it easier to build intelligent and innovative solutions with minimal effort. Perfect for developers looking to enhance their apps with powerful AI capabilities.
github.comPodcast
Azure AI Innovation Podcast
The Azure AI Innovation Podcast features startups and at-scale born in the cloud businesses, influencers such as venture capitalists, and industry thought leaders on all things innovation.
youtube.com