Microsoft

AI & Machine Learning News This week’s digest covers major advancements in AI, from NOAA’s new AI-driven weather models to NVIDIA’s open-source tools for autonomous vehicles and Mobileye’s strategic move into humanoid robotics. Sources: NOAA Deploys New Generation of AI-Driven Global Weather Models NVIDIA’s Alpamayo: Open-Source AI for Autonomous Vehicle Reasoning NVIDIA has announced its Alpamayo family of open-source AI models and tools aimed at accelerating the development of safe and reasoning-based autonomous vehicles. The initial release, Alpamayo 1, is a vision-language-action (VLA) model with 10 billion parameters designed to help AVs understand and reason through complex and unpredictable driving scenarios. This “chain-of-thought” reasoning allows the vehicle to explain its driving decisions, a key factor for safety and scalability. Rather than running directly in the vehicle, Alpamayo models will serve as large-scale ’teacher’ models for developers to fine-tune their own AV stacks. The open-source nature of these tools, announced at CES 2026, is intended to foster innovation across the autonomous driving ecosystem. ...

SAP Boosts Enterprise AI with New Tools and Global Skills Initiative SAP has unveiled a suite of AI-powered enhancements for its SAP Build solution, featuring new data integration capabilities and intelligent automation via Joule Agents. These updates aim to accelerate developer workflows and transform how businesses leverage data and AI. Addressing the global demand for AI talent, SAP has also committed to upskilling 12 million people by 2030 through expanded training programs with partners like Coursera. The company introduced its first relational foundation model, SAP-RPT-1, designed for predicting business outcomes. Furthermore, SAP is expanding its Business Data Cloud ecosystem through a new partnership with Snowflake, enhancing existing integrations with Databricks and Google Cloud. ...

AI Breakthrough: ‘Machine Unlearning’ Enables Models to Selectively Forget Data Researchers have made a significant breakthrough in “machine unlearning,” enabling AI models to be trained to selectively forget specific information upon command. This addresses major privacy and ethical concerns in AI by providing a method to remove particular training examples from a neural network without degrading its overall knowledge. The process involves identifying and isolating information pathways within the neural network that correspond to the targeted data and then using gradient-based algorithms to erase them. This capability is crucial for compliance with regulations like GDPR and for enhancing data privacy in AI systems. The technology has wide-ranging implications across various sectors, including healthcare, where it can protect patient confidentiality, and finance, for securing transaction data. ...

Tech News Roundup: Anthropic, Quantinuum, and Kubernetes Updates This week’s technology digest highlights major developments across the industry, from a landmark legal settlement in AI that could reshape how models are trained, to massive funding in quantum computing, and crucial updates for cloud-native infrastructure and open-source software. Anthropic’s $1.5B Settlement Over AI Training Data Sets Industry Precedent AI company Anthropic has agreed to a $1.5 billion settlement in a major lawsuit concerning the use of pirated books to train its AI models. The lawsuit, filed by authors, alleged that the company used their copyrighted works without permission to develop its chatbots. This settlement represents a significant moment for the AI industry and could establish new standards regarding the use of copyrighted materials for training large language models (LLMs). ...

Daily News Feed: AI Security, Quantum Leaps, and Microsoft’s Rust Bet Cloudsmith Launches ML Model Registry for Centralized AI Model Management Cloudsmith, a software supply chain management company, has launched its ML Model Registry. This new platform will serve as a single source of truth for all AI models and datasets that a company utilizes. The registry integrates with the Hugging Face Hub and SDK, allowing developers to efficiently push, pull, and manage their machine learning models. ...

Microsoft Challenges OpenAI with New Proprietary AI Models: MAI-Voice-1 and MAI-1 Preview Microsoft has announced its first two proprietary artificial intelligence models, MAI-Voice-1 and MAI-1 Preview, marking a significant strategic shift towards developing its own foundational AI technologies. MAI-Voice-1 is an advanced speech generation model capable of creating one minute of audio in under a second on a single GPU. The large language model, MAI-1 Preview, is now available for public testing. This move signals Microsoft’s ambition to compete directly with major AI players, including its partner OpenAI, by building a proprietary AI stack for its future product ecosystem. ...

Microsoft Challenges OpenAI with In-House MAI-1 and MAI-Voice-1 AI Models Microsoft has begun public testing for MAI-1, its first internally developed, large-scale AI model, marking a significant strategic move to reduce its reliance on partner OpenAI. The new foundational model is engineered for general-purpose tasks, including answering complex questions and following instructions. Alongside MAI-1, Microsoft also unveiled MAI-Voice-1, a sophisticated model capable of generating natural-sounding, human-like speech. The voice model can reportedly generate one minute of audio in under a second using minimal computing power. These releases signal escalating competition in the generative AI space, especially within multimodal and voice-based applications. ...

Liquid AI Releases LFM2-VL: Open-Weight Vision-Language Models for On-Device AI Liquid AI has released LFM2-VL, a new family of open-weight vision-language foundation models created for low-latency, on-device deployment. The new models are designed to run efficiently on smartphones, laptops, wearables, and other embedded systems without needing to rely on cloud infrastructure. LFM2-VL is available in two versions: LFM2-VL-450M for highly resource-constrained devices and the more powerful LFM2-VL-1.6B. These models can process both text and images, offering up to twice the inference speed on a GPU compared to similar existing models. The models are available on Hugging Face under a license based on Apache 2.0, allowing for free academic and research use, as well as commercial use for smaller companies. ...

OpenAI Releases GPT-OSS, Its First Open-Source Large Language Models In a significant move for the AI and open-source software communities, OpenAI has released GPT-OSS, its first-ever family of open-source large language models. Licensed under the permissive Apache 2.0, this release introduces two powerful models: gpt-oss-120b and gpt-oss-20b. Both are engineered for efficient inference and strong reasoning, leveraging a Mixture-of-Experts (MoE) architecture. The larger gpt-oss-120b model rivals OpenAI’s proprietary o4-mini and operates on a single 80GB GPU. Meanwhile, the compact gpt-oss-20b model, comparable to o3-mini, is optimized for on-device AI applications, requiring as little as 16GB of memory. Complementing the models, OpenAI also launched ‘Harmony,’ an open-source project defining a new prompt template format. ...