Microsoft is reorganizing its leadership: LinkedIn chief Ryan Roslansky will now oversee Microsoft Teams and lead a new Work Experiences Group. The changes follow the retirement of veteran executive Rajesh Jha, with Charles Lamanna taking on a bigger role overseeing Copilot and platform teams. Microsoft is also offering voluntary retirement to long-serving employees.

The race to build the world’s most powerful AI factories demands networking that keeps pace with the ambitions of AI itself. NVIDIA Spectrum-X Ethernet scale-out infrastructure stands at the forefront of that race as the most advanced AI networking technology available today, deployed by industry leaders including OpenAI, Microsoft and Oracle. Multipath Reliable Connection (MRC), an RDMA transport protocol, distributes traffic across multiple network paths, improving throughput, load balancing and availability for large-scale AI training fabrics. MRC has been successfully deployed in the Blackwell generation and released as an open specification through the Open Compute Project. With multiplanar network designs and hardware-accelerated load balancing, Spectrum-X Ethernet provides an intel…

NVIDIA's Nemotron 3 Nano Omni is a multimodal reasoning model that unifies video, audio, image, and text processing into a single efficient model for agentic workflows, avoiding the lossy pipeline of separate models.

iFlytek's Zhiwen Vision Agent uses a multi-agent system to generate professional PPTs. It breaks down the process into intent understanding, outline building, content refinement, and design rendering—each step editable. Tested across travel guides, marketing plans, academic reports, and industry research, it delivers visually appealing, well-structured, and factually accurate slides, signaling a shift from AI PPT 1.0 (mere generation) to 2.0 (business-grade expression).

ClawTick enables users to set up cron jobs for AI agents with a single command, requiring no infrastructure management.

Google Chrome's AI features can automatically download a 4GB Gemini Nano model file, consuming significant storage without clear notification. Users can disable on-device AI to reclaim space.

AI coding tools speed up code writing, but whether organizations deliver value faster is not obvious. Success depends on engineering maturity such as automated testing, guardrails, and CI/CD. AI amplifies the strengths of high performers and the dysfunctions of struggling ones.

Google is testing Remy, a new AI personal agent for Gemini that can take actions on behalf of users in work and daily tasks. Currently in staff-only testing, Remy is described as a '24/7 personal agent' designed to integrate with Google services and learn user preferences. Privacy controls are managed through Gemini's Privacy Hub, allowing users to oversee data access and activity.

Luma AI launched Uni-1.1 API, developed by a team of fewer than 15 people, ranking third globally on Arena.ai's image generation leaderboard, behind only OpenAI and Google. The model innovatively integrates reasoning and generation in a single model, supports multi-reference image input, and offers sentence-level editing at half the cost of comparable models. Enterprises like Adidas, Mazda, and Publicis Groupe have already adopted it.

Partnership on AI has released a progress report on post-deployment governance practices for foundation models. The report evaluates 13 providers across four practices: sharing usage information, enabling research on societal impacts, reporting incidents and policy violations, and sharing user feedback. Key findings reveal that while leading organizations are defining information-sharing standards, adoption is slow among others, and public impact data remains fragmented.

Google introduces MTP drafters for Gemma 4 family, using speculative decoding to achieve up to 3x faster inference without quality loss. The drafters share KV cache and activations, with special optimizations for edge models.

Terence Tao used Claude Code to handle paper review comments, finishing revisions in just 15 minutes and even catching a reviewer's error, highlighting AI's potential as a co-pilot for mathematicians.

Discover how ChatGPT safeguards your privacy by minimizing personal data in training and giving you control over your conversations.

In May 2025, top law firm Latham & Watkins filed a court declaration containing fabricated citations generated by Claude, the very AI model they were defending in a copyright case. The incident exposes new liability issues for attorneys using AI, including metadata verification challenges and the non-delegable duty under Rule 11.

ExploreYC provides a powerful search and analytics platform for over 5,700 Y Combinator companies, offering AI-powered insights, funding data, hiring boards, and more.

Bargains are disappearing and the cost of gadgets such as MacBooks and PS5s is rising as AI competes for memory chips. The shortage could eliminate budget PCs by 2028.

The Financial Stability Board (FSB) warns that the private credit industry's role in funding the AI boom could backfire, leading to sizable losses if a sharp correction occurs. AI companies increasingly rely on private lenders for data centre infrastructure, accounting for over a third of private credit deals in 2025. The report highlights risks from asset valuation corrections, electricity shortages, and oversupply of data centres. Traditional banks are also exposed through direct lending and partnerships, while recent bankruptcies of private credit-backed firms raise transparency concerns.

This week, Anthropic and OpenAI announced services companies backed by billions, signaling a shift towards 'last mile' enterprise solutions. Meanwhile, GPT-5.5 Instant rolled out, Gemma 4 released multi-token prediction drafters, and coding agents continue to dominate discussions.

ChatGPT is now available as a sidebar add-in for Excel and Google Sheets, powered by GPT-5.5, allowing users to build, edit, analyze, and explain spreadsheets using plain language directly within the tool. The add-in is now free for Free and Go plan users, previously limited to Business, Enterprise, and Pro. Its key differentiator is a transparency mechanism: ChatGPT explains its reasoning, links answers to specific cells, preserves formatting and formulas, and asks for permission before making changes, which users can revert anytime. Features include generating formatted spreadsheets from descriptions, asking questions across tabs and rows, cleaning data, tracing and fixing formula errors, and building scenario tabs and financial models. It targets analysts, operations leads, and finance…

Cross-Dimensional Intelligence (Kuawei Intelligent) launched robot service stations in multiple cities during the May Day holiday, attracting over 100,000 visitors on the first day. The stations use a general-purpose embodied intelligence brain capable of handling diverse products and machines, demonstrating the technology's versatility and commercial potential.

Knowly 1.0 combines Karpathy's LLM Wiki and NotebookLM into a proactive AI for learning. Save anything, and it auto-organizes, builds understanding, and creates personalized learning flows on demand. The AI proactively suggests next steps, eliminating the need for perfect prompts.

At the COFCO Joycome million-head pig breeding base in Changling, Jilin, iFLYTEK Heguang Technology uses large models and multimodal perception to transform frontline breeding experience into intelligent algorithms, achieving cost reduction, quality improvement, and efficiency gains, providing a replicable benchmark for smart livestock farming in China.

In the second week of the Musk vs. OpenAI trial, Musk sent a text message seeking a settlement, which was rejected by Brockman. He then threatened to make Altman and Brockman the most hated people in America. Brockman took the stand and admitted that financial reward was a secondary motive. His private diaries revealed early intentions to profit. Prediction markets show Musk's chances of winning have dropped.

BifrostUMI is a portable, efficient, and robot-free data collection framework for humanoid robots. It uses lightweight VR devices to capture human demonstrations as sparse keypoint trajectories and wrist-mounted visual data, training a high-level policy network to predict future keypoints. A robust retargeting pipeline maps these trajectories to the robot's morphology, enabling seamless transfer of agile behaviors from natural human demonstrations to humanoid embodiments. The framework's effectiveness was validated in two distinct experimental scenarios.

Researchers propose a hybrid hierarchical control framework for reactive dexterous grasping, decoupling high-level spatial intent from low-level joint execution using multi-agent RL planning and GPU-parallelized QP control, achieving robust zero-shot transfer to real robots.

This paper tackles the complex problem of accurate jumping height control for wheeled-bipedal robots. It proposes a novel Wheeled-Bipedal Jumping Dynamical model (W-JBD) and a Bayesian Optimization for Torque Planning method (BOTP) that reduces height error by 82.3% and energy cost by 26.9% in simulation, paving the way for real-world applications.

This paper investigates risk-aware domain randomization (DR) in contact-rich sampling-based predictive control (SPC). Using the Push-T task, it compares average, optimistic, and pessimistic rollout aggregations, finding that DR not only improves robustness to model error but also reshapes the effective cost landscape by altering the basin of attraction around contact-producing actions, opening new possibilities for robust SPC under model uncertainty.

This paper proposes Neural Control, a boundary-control framework for physical AI tasks governed by implicit equilibrium. It computes memory-efficient proxy gradients via an adjoint formulation, avoiding unrolling of solver iterations, and integrates into receding-horizon MPC for robustness. Evaluated on deformable linear object manipulation, it outperforms gradient-free baselines like SPSA and CEM.

The RLDX-1 technical report introduces a general-purpose robotic policy for dexterous manipulation, built on the Multi-Stream Action Transformer (MSAT). It addresses limitations of existing Vision-Language-Action models (VLAs) by integrating motion awareness, memory-aware decision making, and physical sensing. RLDX-1 outperforms frontier VLAs like π0.5 and GR00T N1.6, achieving 86.8% success on ALLEX humanoid tasks compared to ~40% for competitors.

This paper proposes ICODE-MPPI, a robust framework that uses Input Concomitant Neural ODEs (ICODEs) to learn and compensate for unmodeled residual dynamics, enhancing Model Predictive Path Integral (MPPI) control for complex trajectories. High-fidelity simulations show up to 69% reduction in cross-tracking error and significantly smoother steering compared to standard MPPI.