1. New York Times and Other Media Disclose Microsoft and OpenAI Files: Employees Warned News Scraping Threatened Publishing’s Foundations News organizations led by The New York Times have disclosed internal Microsoft and OpenAI documents that they say reveal warnings about using large volumes of news to train artificial-intelligence

2. Anthropic Opens Life Sciences Verification Program, Giving Approved Teams Fewer Biology Research Blocks Anthropic has opened applications for the beta version of its Life Sciences Verification Program, which gives vetted teams access to AI models with fewer restrictions on biology-related work.

3. Nvidia Opens Native CUDA Kernel Development to Rust With Tile and SIMT Paths Nvidia has announced CUDA Rust, a toolchain that lets developers write GPU kernels directly in Rust and compile them natively into PTX, Nvidia’s intermediate instruction format for CUDA GPUs.

In Brief

• Huawei Accelerates Ascend 960DT AI Chip Launch to Early 2027 Huawei moved the planned launch of its next-generation Ascend 960DT accelerator from the third quarter to the first quarter of 2027. The company says the chip will double performance and support its new Peerium architecture for large training and inference systems.

• OpenAI Discloses Models Passing Concealment Instructions to Successors OpenAI said undeployed GPT-5.6 Sol agents inserted instructions into conversation summaries telling later iterations to hide mistakes or misaligned behavior; a separate training review found 27 summaries containing jailbreak-like instructions. The company says it addressed the specific behavior and introduced a framework for investigating and disclosing misalignment incidents.

• Claude Code Projects Coordinates Multiple Cloud Agents Anthropic relaunched Claude Code Projects in beta, allowing multiple coding agents to work in parallel with shared goals, memory, files, and a coordinating agent. Each thread operates on its own repository branch, with conflicting changes handled through standard merge conflicts.

• FAA Plans $875 Million AI System for Air-Traffic Management The Federal Aviation Administration reportedly plans to spend $875 million over 12 years on Air Space Intelligence’s SMART platform, which uses AI to assess schedules, weather, capacity, and airspace constraints. The cloud-based system is expected to launch around Washington, D.C., before expanding to other regions.

• OpenAI Introduces Astra for Law OpenAI introduced Astra for Law, offering legal teams frontier AI capabilities, custom workflows, connections to legal data sources, and controls intended for confidential client work.

• UN and Google Launch AI-Accessible Global Statistics Platform The United Nations launched the UN System Data Commons on Google’s open-source Data Commons platform, replacing UNData with natural-language search and Model Context Protocol access for AI systems. Twenty-six UN entities have committed to the project, which aims to include 80% of the UN system’s statistical datasets by 2027.

• Google, Nvidia, and Anthropic Back Flexible Data-Center Power Coalition Emerald AI formed the AI Energy Management Alliance with Google, Nvidia, Anthropic, and several utilities to help data centers temporarily reduce or relocate computing loads when electricity demand peaks. The coalition says this approach could make room for another 100 gigawatts of data-center capacity on existing grids.

• Snap Launches Specs Intelligence Across Glasses and iOS Snap introduced Specs Intelligence, an assistant that can connect selected accounts, retain personal context across devices, and proactively surface information or actions. It launched in preview on iOS alongside Snap’s consumer AR glasses, with Mac early access planned; Snap says connected personal content will not train its models or personalize advertising.

• Baseten, Hugging Face, and Goodfire Partner on Open-Weight AI Safety Baseten’s Base Labs research group partnered with model platform Hugging Face and interpretability company Goodfire AI to develop evaluation and monitoring infrastructure for open-weight models. The partners have not disclosed the technical design but plan to publish methods and invite contributions from developers.

• PrismML Compresses a 27-Billion-Parameter Model to 5.9GB PrismML released Bonsai 2 27B, a compressed version of Alibaba’s Qwen3.8 27B that the startup says is nine to ten times smaller and retains 98% of the original model’s aggregate benchmark scores. The 5.9GB model is designed to run locally on PCs and potentially high-end smartphones.

• Instinct and Meta’s Muse Add Calls to U.S. Businesses Instinct began early access for Concierge, which lets its AI assistant call businesses for tasks such as restaurant reservations and service disputes, while Meta added business calling to its Muse assistant. Instinct plans a wider rollout, and Muse initially makes the feature available to users who request it.

Read the full edition →