Liam Fedus, VP of Post-Training Research at OpenAI, is leaving to start an AI company focused on materials science. He plans to leverage his physics background and collaborate with OpenAI, which will invest in his venture. The new firm will compete with Google DeepMind and Microsoft in using AI for material discovery, though some experts question AI's ability to drive true scientific innovation.

Google plans to collaborate with MediaTek for its next-generation TPU, an AI chip used internally. The production may shift partially to MediaTek next year, focusing on input/output modules, while Google continues its partnership with Broadcom. TPUs, specialized ASICs optimized for TensorFlow, enhance machine learning tasks and efficiency in cloud and edge computing. The latest Trillium TPU offers significant performance improvements over its predecessor, providing Google a competitive edge in A

Intel's new CEO Lip-Bu Tan is reportedly planning significant changes in chip manufacturing and AI strategies, including cutting middle management roles to steer the company back on track. He rejoins after leaving Intel's board in 2024.

xAI, founded by Elon Musk, acquired Hotshot, a startup developing AI-powered video tools similar to OpenAI's Sora. Hotshot's CEO Akshay Sastri announced the deal, highlighting their foundational video models: Hotshot-XL, Act One, and Hotshot. Originally focused on AI photo tools, Hotshot shifted to text-to-video models and attracted investors like Lachy Groom and Alexis Ohanian. The acquisition may indicate xAI plans to compete with Sora and Google's Veo 2 through a video model for Grok. Hotshot

Google expands access to Gemini 2.0 Flash AI model for developers, featuring image generation and editing capabilities. Concerns arise as users find it can accurately remove watermarks and add images of real individuals like Elon Musk. While SynthID tags replace copyright marks, potential misuse exists. These features are currently restricted to AI Studio users, with no official response on safeguards against watermark removal.

Roblox launched its first AI model, Cube 3D, which aims to improve 3D creation efficiency. Trained on licensed, public, and ecosystem data, Cube 3D tokenizes shapes to predict and generate 3D objects. Future updates will enable image-based generation and multimodal input processing, integrating with existing AI tools. Plans include making it open source.

YouTube Music's AI-driven "Ask Music" radio generator, initially launched on Android, is now available on iPhone and iPad. The feature lets users create personalized radio stations by inputting moods, activities, or preferences, with voice input support. It integrates with Music Tuner and requires a YouTube Music subscription. Currently in select countries like the US and Canada, it will expand globally soon. This update highlights YouTube Music's focus on intelligent, tailored music experiences

Zoom's AI companion will get a "delegate-style" upgrade to perform tasks like scheduling meetings and generating docs/videos. New features include voice recording, real-time note-taking, and customizable AI avatars. Most features are free for Zoom Workplace users, with a $12/month Custom AI Companion add-on planned.

Some Windows 11 users found the Microsoft Copilot app was uninstalled due to the latest monthly update. Microsoft acknowledged the issue and is working on a fix, advising users to reinstall it from the Microsoft Store. This bug may be welcomed by users uninterested in Copilot. Microsoft is rolling out an updated local version with a better UI and press-to-talk function. More updates are planned for Copilot, to be revealed at Microsoft's 50th-anniversary event on April 4, including new functional

Xiaomi's large model team achieved a 64.5% accuracy in audio reasoning, topping the MMAU benchmark using reinforcement learning inspired by DeepSeek-R1. They surpassed GPT-4o and Gemini2.0Flash, open-sourcing the technology for broader collaboration. Human-level performance at 82.23% remains the target.

Sony's leaked test demo used AI to recreate Horizon series protagonist Aloy, sparking debate among players and developers. Voice actress Ashly Burch, who was informed in advance, expressed concerns about the future of game performance as an art form. The AI version showcased functional gameplay elements but raises worries about voice actors' rights amid ongoing SAG-AFTRA strikes. Sony has not commented on integrating such tech into PlayStation products.

Amazon is ending the local processing option for Alexa requests, requiring all voice data to be uploaded to its cloud servers. The company informed users with the "Don't Send Voice Recordings" feature enabled that it will discontinue on March 28th due to the integration of generative AI capabilities. This move comes as Amazon launches Alexa+ and addresses past privacy concerns, including a $25 million FTC settlement in 2023 over children's privacy issues.

OpenAI and Elon Musk agreed on a December trial date over OpenAI's shift to a for-profit entity. Musk, a co-founder, sued claiming deviation from its nonprofit mission; OpenAI defends the move as essential for funding and competitiveness. Musk's $97.4 billion acquisition offer was rejected. OpenAI also proposed AI policies to the White House, emphasizing U.S. leadership, federal prioritization, and export controls. Legal and policy challenges ahead will test OpenAI's future.

Claude's upcoming Harmony feature will allow full access to local directories for content manipulation, file scanning, editing with change tracking, keyword searching, and IDE-like functionality. The new web app includes tools like Memory, Tasks, and Create Sub-agent, expanding beyond Web Search and Fetch. Additional features remain undisclosed.

Baidu launched two new AI models, Ernie X1 and Ernie 4.5, enhancing its AI capabilities. Ernie X1 focuses on reasoning, matching DeepSeek's R1 performance at half the cost, while Ernie 4.5 excels in multimodal understanding and emotional intelligence. Baidu plans to open-source the Ernie models later this year, amid growing competition from firms like Tencent and Alibaba.

Google's Gemini 2.0 Flash AI can remove watermarks from images, including those from Getty Images, raising copyright concerns. While labeled as experimental and not perfect with semi-transparent marks, its capability to fill gaps after removal is notable. Unlike rivals Claude 3.7 and GPT-4o, which refuse watermark removal due to ethical issues, Gemini lacks such restrictions. Removing watermarks without consent is illegal under U.S. law. Google has yet to comment.

OpenAI accused Chinese AI firm DeepSeek in a policy proposal, suggesting the US government ban models from DeepSeek and other China-backed AI operators due to security and privacy risks. OpenAI claims DeepSeek's models, especially its R1 reasoning model, could threaten user privacy under Chinese law. However, there is no evidence that DeepSeek is directly controlled by the Chinese government. The proposal has sparked criticism, highlighting double standards and concerns over stifling competition

NVIDIA's GTC conference kicks off next week in San Jose, with a keynote focusing on AI, digital twins, cloud technology, and sustainable computing. CEO Jensen Huang will highlight NVIDIA's accelerated computing platform and address market trust amid evolving AI landscapes. Although gaming advancements will be mentioned, the event primarily emphasizes AI innovations to engage enthusiasts and investors, showcasing NVIDIA's strategic shift towards future technological demands.

Cohere Inc. launched Command A, a high-performance LLM for commercial use with reduced hardware needs. It outperforms models like GPT-4o and DeepSeek-V3 in business tasks, running on just two GPUs for private deployments. Command A offers 156 tokens per second, a context window of 256,000 tokens, and integrates with Cohere's AI agent platform North for enterprise data utilization. The model supports scalable AI agent operations and agentic AI trends.

Sesame launched CSM-1B, a 1-billion-parameter open-source model powering its realistic voice assistant Maya. Built on Meta's Llama series, the model generates RVQ audio codes from text or audio and supports multiple voices with limited non-English capabilities. Despite lacking robust safety measures, Sesame urges responsible use. The company, co-founded by Brendan Iribe, is backed by Andreessen Horowitz and explores AI glasses.