/AI News/Issue 94
Microsoft researchers say they've developed a hyper-efficient AI model that can run on CPUs
Published Tuesday, April 22, 2025
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.
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.
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.
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.
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.
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.
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.
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.
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!
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.
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.
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.
The AI links worth your time, curated by the community and always free.
