Microsoft introduces a pair of in-house AI models (www.engadget.com)

Microsoft has unveiled two new AI models developed entirely in-house, marking a significant step toward greater independence in the AI space. MAI-Voice-1 is Microsoft's first natural speech generation model and is already integrated into features like Copilot Daily and Podcast. The second model, MAI-1-preview, is a text-based foundation model trained end-to-end and is now available for public testing via LMArena, with plans to roll it out in select Copilot applications soon. This dual release showcases Microsoft’s growing capability to produce competitive AI technologies beyond its existing reliance on OpenAI’s GPT models powering Copilot. Technically, these models emphasize efficiency and cost-effectiveness. MAI-Voice-1 runs on just a single GPU, while MAI-1-preview was trained using roughly 15,000 Nvidia H-100 GPUs—significantly fewer than the 100,000+ GPUs used by other large-scale models like xAI’s Grok. Mustafa Suleyman, head of Microsoft’s AI division, highlighted the importance of optimizing training data to avoid wasting computational resources on uninformative tokens, reflecting a nuanced shift toward smarter, leaner model training. This focus on efficiency, paired with Microsoft's commitment to a long-term five-year investment, signals a strategic move to build a self-sufficient AI ecosystem amid industry uncertainties and the potential for an AI market correction.