要点速览
- Sam Altman backs new U.S. AI executive order as balanced framework for safe AI leadership
- Perplexity Computer rolling out hybrid local-cloud agent inference and health data integration
- Hugging Face launches multiple open-source AI tools including Holo 3.1, Gemma 4 12B, and Ideogram 4.0
- NVIDIA unveils Cosmos 3 physical AI model, DGX Station with GB300, and Windows AI agent stack with Microsoft
- Factory Router cuts AI inference costs by 20-25% via intelligent model routing
一、AI Policy & Leadership Signals
- OpenAI CEO Sam Altman praised the new U.S. presidential executive order on AI, stating it strikes the right balance between advancing cutting-edge AI research, ensuring safety, and empowering trusted defenders with network tools to lead global AI development. — via 1
- Anthropic welcomed the U.S. AI executive order, calling it a critical step to strengthen American AI leadership and expressing intent to collaborate with the White House on implementation. — via 1
二、AI Agent & Product Launches
- Perplexity Computer will soon release hybrid agent inference functionality that distributes tasks between local on-device models and cloud frontier models, preserving private data locally while maximizing token efficiency. The feature will launch first on Windows laptops, with support for running local models on personal hardware to balance privacy, per-watt token efficiency, and access to server-grade GPU models when needed. It also added two new health data import methods: Apple Health sync for sleep, activity, and heart rate variability data, and lab test result integration for Perplexity Health. — via 1 2 3
- OpenAI is expanding Codex plugin support to enable single-install domain-specific professional assistants without additional coding, with access to 62 popular apps and 110 job skills across sales, data analysis, creative production, product design, and public stock investing. — via 1
- NVIDIA launched a tutorial for building persistent workflow agents using NemoClaw and OpenShell, enabling deployment of @nousresearch Hermes Agent connected to Slack, Outlook, GitHub, and NVIDIA developer forums, with reusable skills retained across rebuilds and private data enforced via runtime policies. It also released a one-click DGX Spark NemoClaw deployment path for simplified local AI agent setup without external cloud dependencies, and updated OpenShell v0.0.55 with Google Vertex AI support and improved reliability. — via 1 2 3
- AI model routing tool Factory Router can automatically match tasks to optimal models across providers, reducing token spending by 20-25% while maintaining frontier performance by avoiding overprovisioning with the largest models for all tasks. — via 1 2
三、New Open-Source AI Models & Tools
- Hugging Face launched multiple open-source AI offerings: Holo 3.1, a locally runnable computer-use AI model optimized for GUI understanding and control, outperforming Qwen3.5-397B, Kimi-K2.5, and Sonnet 4.6 across devices including Mac, Windows, and Android; Gemma 4 12B, an Apache 2.0-licensed edge-optimized multimodal model for laptops; and Ideogram 4.0, the best open-source image model available for download and fine-tuning. It also added a Parquet file native viewer for Hugging Face Buckets private dataset storage, and explained the mid-training model development phase. — via 1 2 3 4 5 6 7
- NVIDIA released Cosmos 3, its first physical all-modal AI model, which topped 7 physical AI leaderboards including world generation, robotic motion planning, and industrial vision understanding, and is now available on Hugging Face. — via 1 2
- Runway launched Aleph 2.0 via its API, enabling precise 1080p, 30-second multi-shot video editing for integration into third-party apps, with automatic green screen and clean frame conversion without rotoscoping. — via 1 2
- xAI's Grok speech-to-text and text-to-speech APIs are now available on the enterprise voice AI platform Vapi for building custom voice assistants with natural call flows and compliant key detail capture. — via 1
四、AI Infrastructure & Industry Updates
- NVIDIA and Microsoft launched an end-to-end unified accelerated stack for agent AI development across Windows devices, cloud, and on-premises, including the Windows PC-optimized RTX Spark, native Claude model support on Azure GB300 Blackwell Ultra systems, and NVIDIA-accelerated Microsoft Fabric SQL queries up to 6x faster. NVIDIA also began shipping DGX Station systems with GB300 chips to developers, with partner products from Asus, Dell, Gigabyte, HP, MSI, and Supermicro. — via 1 2
- NVIDIA unveiled Vera, a new CPU designed for AI agent tasks that delivers 80% faster completion than x86 chips, tailored for AI factories and the future of intelligent systems. — via 1
- Stanford research found Gemini 2.5 Pro won 75% of blind legal expert evaluations against human responses, with lower perceived harm, and newer model versions show improved performance. — via 1
