1. Microsoft Unveils Maia 300, Plans to Make In-House Chips Its Primary AI Inference System From 2027 Microsoft has unveiled Maia 300, the third generation of its in-house AI accelerator, and plans to begin deploying it to production workloads in 2027.

2. QGIS merges new rule banning all AI-assisted code submissions to its core repository QGIS is a software project whose core code is maintained in the qgis/qgis repository.

3. WindBorne Wins U.S. Air Force Contract Worth Up to $49 Million for Balloon Data From Specified Regions The U.S. Air Force has awarded WindBorne Systems a contract to buy atmospheric observations from the startup’s autonomous weather balloons.

In Brief

• Meta launches its first frontier models, Muse Wonder and Wonder Full Meta says the new models match leading rivals on several benchmarks and improve long-term information recall. Wonder is available through Meta’s API at $4 per million input tokens and $20 per million output tokens, while Wonder Full is limited to Muse Code subscribers.

• Epic reportedly adopts “AI-first” development for its Disney game Bloomberg reports that Epic Games will have AI models generate all code and guide artwork for an upcoming Disney shooter, with developers managing the output. The project has reportedly been delayed indefinitely; Epic CEO Tim Sweeney confirmed a companywide pivot toward greater AI acceleration.

• Nvidia opens Vera Rubin infrastructure designs to the OCP ecosystem Nvidia contributed Vera Rubin rack and compute-tray designs through its open MGX architecture and released open-source components for configuring networks, validating clusters, managing racks, and provisioning hardware. The MGX 1.2 specification supports deployments of up to 250 kilowatts per rack.

• Raven can reportedly locate homes from a single interior photograph Graylark Technologies demonstrated Raven, an AI geolocation tool sold to law-enforcement agencies, matching interior photos to specific properties using apparent data from rental and real-estate listings. Airbnb said it has no relationship with Graylark and prohibits scraping, while Zillow said scraping violates its terms.

• Microsoft president calls for binding global AI safety rules Microsoft president Brad Smith said voluntary commitments are insufficient and argued that the UK is best positioned to lead international AI regulation. His proposals include mandatory breach notifications, outside evaluation before product releases, and mechanisms to pause systems that escape human control.

• Ambi raises $26 million for an always-listening AI assistant Former Plaud executive Peter Mo’s startup launched Ambi apps for iPhone, Mac, and Apple Watch that can capture conversations and surface information during meetings. Ambient capture is off by default, and the company says it encrypts processed data and does not retain recordings, though users remain responsible for notifying other people.

• Blueprints AI prepares permit-ready building plans from sketches Blueprints AI says its second-generation software can turn a description or sketch into coordinated architectural, structural, mechanical, electrical, plumbing, and interior-design documents while checking changes against local codes. Licensed professionals must still review and seal the plans before permit submission.

• Amazon’s warehouse robot fleet surpasses one million Amazon now operates more than one million warehouse robots alongside more than 1.5 million employees and plans to expand its fleet further. The company says robots handle physically demanding work, while researchers and labor advocates warn that nearby collaborative robots can also dictate employees’ pace.

• Analyst challenges Anthropic’s reported $2 trillion IPO valuation New Constructs estimated Anthropic is worth $150 billion and argued that its reported losses and open-source competition make the proposed valuation untenable. Anthropic has not released its prospectus publicly and did not respond to CNBC’s request for comment.

• SoftwareGym uses synthetic apps to train computer-control agents Researchers generated 182 runnable GUI applications and used interaction trajectories from them to train a computer-use model. They report that its Pass@3 score on the out-of-domain OSWorld benchmark rose from 41.39% to 56.97%.

• Reflex teaches a humanoid robot to catch thrown boxes The Reflex framework combines whole-body control with egocentric RGB-D vision to estimate a moving object’s trajectory. On a real Unitree G1 humanoid, the system caught human-thrown boxes 65% of the time from 2.5 meters and 40% from 5 meters.

Read the full edition →