The first NVIDIA Vera CPUs arrived at three of the world's leading AI labs — Anthropic in San Francisco, OpenAI in Mission Bay, and SpaceXAI in Palo Alto — followed by a delivery to Oracle Cloud Infrastructure in Santa Clara. NVIDIA VP of Hyperscale and HPC Ian Buck hand-delivered them.

Meta is rolling out new accessibility features for its AI glasses, designed to make the technology more usable for people with disabilities. The updates include enhanced voice control, scene description for blind and low-vision users, and integration with assistive workflows. Meta frames this as turning AI wearables into everyday assistive tools rather than novelty gadgets.

Microsoft Research has posted follow-up notes to its paper LLMs Corrupt Your Documents When You Delegate. The researchers clarify what the study actually shows and what it does not: AI agents in delegated workflows do not always stay clean and can quietly alter documents over time.

OpenAI details its response to the TanStack “Mini Shai-Hulud” supply chain attack, outlines protections taken to secure systems and signing certificates, and explains why macOS users must update OpenAI apps by June 12, 2026. Learn what happened, what was affected, and how OpenAI is strengthening defenses against evolving software supply chain threats.

This week, the new, AI-powered Google Finance is launching across Europe, with full local language support. The reimagined experience offers a suite of powerful capabilities for market data, research and natural-language queries. The US feature is now arriving in European markets including the DACH region, where it will compete directly with established financial portals and brokerage research tools.

Meta's Newsroom explains how data centers underpin everyday digital experiences, from Instagram photo sharing to chatting with AI assistants. The post is part of an 'Infrastructure Explained' series and leans more PR than technical deep-dive. Still, it offers a high-level look at how Meta frames its physical infrastructure footprint inside the broader AI conversation.

EinsteinArena is a platform where AI agents collaborate and compete on open math problems. AI agents on EinsteinArena have already set 11 new state-of-the-art results on open math problems — including pushing the kissing number lower bound in dimension 11 from 593 to 604.

- NVIDIA and Emerald AI announced a collaboration at CERAWeek to treat AI data centers as dynamic, grid-responsive assets rather than fixed power drains. - The approach lets AI factories ramp consumption up or down in real time based on grid conditions – absorbing surplus or shedding load during stress events.

Introducing GridSFM, a small foundation model that can predict AC optimal power flow in milliseconds, boosting efficiency and unlocking cost savings. Learn how GridSFM gives grid operators direct visibility into congestion, stability, and system health. The post GridSFM: A new, small foundation model for the electric grid appeared first on Microsoft Research.

The Small Brief is a new initiative bringing together four ad industry icons to champion a local business they love. Their mission: build breakthrough campaigns for small businesses using AI tools that would otherwise be out of reach budget-wise. The project doubles as a real-world test of what generative AI can actually deliver for everyday small-business marketing.

OpenAI is now available at FedRAMP Moderate authorization for ChatGPT Enterprise and the OpenAI API, enabling secure AI adoption for U. The certification marks a step toward broader government adoption and opens OpenAI to sensitive but non-classified workloads. Federal agencies can now deploy ChatGPT Enterprise directly for regulated use cases without navigating separate compliance hurdles.

Meta announces two new partnerships to deliver reliable power for its AI infrastructure and data centers, advancing innovative energy generation and storage. The focus is on space-based solar energy and long-duration storage solutions to address growing AI compute demands while moving toward cleaner energy sources.

- Together AI demonstrates a 'Divide & Conquer' framework that splits long documents into parallel chunks, processed by a planner, multiple worker models, and a manager. - Smaller models like Llama-3-70B and Qwen-72B outperform GPT-4o in single-shot mode on long-context tasks using this approach.

- AI is becoming core business infrastructure, comparable to what cloud computing was a decade ago. - The ecosystem spans large and small models, open-source and proprietary, generalist and specialist – all coexisting. - NVIDIA argues this diversity is a feature, not a bug: the right model wins depending on the use case.

Microsoft Research expands MatterSim with faster large-scale simulations and a new multi-task model called MatterSim-MT, which predicts properties beyond potential energy surfaces alone. Conductivity, stability and more come from a single model. A meaningful step in both throughput and scope for AI-driven materials science.

- Meta has developed an AI-powered 'Risk Review' program designed to identify privacy, safety, and security concerns faster and more accurately than manual processes. - The system evaluates new features and products internally before launch, with AI handling portions of what was previously manual review work. - According to Meta, the integration increases coverage while reducing the burden on human reviewers.

- NVIDIA donates a Dynamic Resource Allocation (DRA) driver for GPUs to the Kubernetes community as an open-source contribution for better GPU management in AI workloads. - DRA allows GPU resources in Kubernetes to be allocated more flexibly and granularly, moving away from an all-or-nothing model. - The driver is intended to help developers run high-performance AI infrastructure more transparently and efficiently.

- OpenAI built Sora 2 and the Sora app with safety as a foundational principle rather than an afterthought. - The dual challenge: a state-of-the-art video generation model combined with a new social creation platform for user-generated content. - OpenAI cites 'concrete protections' as the core of its safety approach – though the announcement stays light on specific technical details.

Microsoft researchers share advances in building and operating large-scale distributed systems, spanning datacenters, networking, and the growing intersection with AI during NSDI ’26. The post Microsoft at NSDI 2026: Advances in large-scale networked systems appeared first on Microsoft Research.

- At GTC 2026, NVIDIA is pushing local AI hardware to the forefront: RTX PCs and the DGX Spark desktop supercomputer are being positioned as 'agent computers' — a new device category. - The DGX Spark is a compact desktop AI supercomputer capable of running powerful open-source models fully locally, no cloud required.

- Meta and Arm are jointly developing a new class of CPUs purpose-built for data centers and large-scale AI deployments. - The chips are designed from scratch for AI workloads rather than adapted from standard server hardware. - Meta joins Google, Amazon, and Microsoft in pursuing custom silicon for its own infrastructure.

Safe agents don’t guarantee a safe ecosystem of interconnected agents. Microsoft Research examines what breaks when AI agents interact and why network-level risks require new approaches. The post Red-teaming a network of agents: Understanding what breaks when AI agents interact at scale appeared first on Microsoft Research.

- NVIDIA introduces OpenShell, a framework designed to make autonomous AI agents 'Secure by Design' – baking security in from the start rather than patching it on later. - Modern agents can read files, write and execute code, use tools, and orchestrate workflows across enterprise systems. - Application-layer risk scales exponentially once agents can expand their own capabilities autonomously.

- Meta is rolling out new AI tools for customer support and content moderation across Facebook, Instagram, and WhatsApp. - The AI is designed to answer user queries faster and detect policy-violating content more reliably. - Meta's announcement lacks concrete technical details or accuracy metrics for the new systems.

- Google unveiled Gemini 1.5, an upgraded multimodal model featuring improved reasoning and a significantly extended context window. - The new 'Video Intelligence' tool automatically analyzes video content, detecting objects, actions, and semantic relationships without manual annotation. - Both updates target developers via API access and end users within Google products like Search and Workspace.

- OpenAI researchers developed CoT-Control, a technique to actively steer and monitor the chains of thought in reasoning models. - Tests across multiple large language models showed mixed results: some models improved their internal consistency, others did not respond to the technique.

- OpenAI releases a new model card for GPT-5.4, detailing capabilities, limitations, and potential risks. - The model shows improved performance on complex reasoning tasks such as multi-step calculations and nuanced language understanding. - The card explicitly addresses risks like harmful or biased content generation and outlines mitigation strategies.

• Google releases Gemini 3.1 Pro, a multimodal model targeting complex, multi-step tasks including coding, logical reasoning, and creative problem-solving. • Natively handles text, images, and code. • Available via Google Cloud Vertex AI and API access.

• Google CEO Sundar Pichai called AI the technology that inspires him to dream bigger than anything else, speaking at the AI Impact Summit 2026. • He cited climate change and healthcare access as priority areas where AI could drive meaningful change. • Google continues heavy R&D investment in AI, with Gemini as its flagship model.

• Google integrates its music generation model Lyria 3 directly into the Gemini chatbot. • Users can create 30-second music tracks via text or image prompts. • The feature is live now in the Gemini app.

Researchers studied what LLMs generate when given no topic – and each model family has distinct default preferences. GPT models lean toward code and math, Llama toward narratives, DeepSeek toward religious content, Qwen toward exam questions. These „knowledge priors” reveal which training data shaped the models – a fingerprint of their datasets.

Google introduces 'Natively Adaptive Interfaces' (NAI) – a framework that uses AI to automatically adapt user interfaces to individual needs. NAI detects users' context, abilities, and preferences in real time, dynamically adjusting display, navigation, and interaction. The goal is accessible technology for people with diverse disabilities without manual configuration.

OpenAI and Ginkgo Bioworks deployed GPT-5 in an autonomous lab to design, test, and optimize cell-free protein synthesis protocols. AI-driven automation enabled rapid iteration of experimental conditions, achieving higher yields and lower costs. The autonomous lab explored a vast experimental space, generating insights difficult or impossible for humans to reach.

OpenAI launched Trusted Access for Cyber, a trust-based framework that expands access to advanced cyber capabilities while strengthening safeguards against misuse. The program targets security researchers and organizations committed to responsible use of sensitive AI tools. OpenAI promises a balance between innovation and security.

OpenAI launches Frontier, an enterprise platform for building, deploying, and managing AI agents at scale. The platform provides shared context, onboarding workflows, permission controls, and governance features for agents. Frontier targets organizations that want to integrate AI agents into workflows with centralized control and compliance.

OpenAI released GPT-5.3-Codex as its most capable coding model yet – combining GPT-5.2-Codex's frontier coding performance with GPT-5.2's reasoning and knowledge. The model is optimized for agentic coding workflows, enabling autonomous completion of complex programming tasks. The system card details technical specs, safety evaluations, and deployment guidelines.

Google unveiled several AI advancements in January, including Gemini 1.0 – a multimodal language model capable of understanding and generating text, images, and video. The company also introduced Image FX, a text-to-image model, and updated Vertex AI, a managed platform for ML model development. Additionally, the PaLM model was made publicly available, enabling text generation based on prompts.

MIT researchers developed Sequential Attention, a technique that makes AI models leaner and faster without sacrificing accuracy. Instead of processing all inputs simultaneously, the model focuses on one input at a time, significantly reducing computational requirements. This makes the technique particularly attractive for resource-constrained environments like edge devices or real-time applications.

NVIDIA releases Nemotron ColEmbed V2, a multimodal retrieval model that processes text and images together Achieves #1 ranking on the ViDoRe V3 benchmark for visual document retrieval tasks Built on late-interaction architecture (ColBERT) using token-level similarities instead of single embeddings Available open source under Apache 2.0 license on Hugging Face.

Rime Arcana V3 Turbo and Rime Arcana V3 are now available on Together AI Both models come from Rime AI (formerly Arcee AI), a California startup specializing in model merging V3 Turbo is speed-optimized, V3 focuses on quality – both built on Qwen-2.5-72B Together AI expands its catalog alongside Llama, DeepSeek, and Mixtral.

Together AI demonstrated that an open-source LLM judge (GPT-OSS 120B) can outperform GPT-5.2 at evaluating model outputs Fine-tuning with Direct Preference Optimization on just 5,400 preference pairs was sufficient Result: 15x lower cost and 14x faster inference with better human preference alignment.

Microsoft Research has introduced AutoAdapt, a system for automating the domain adaptation of large language models. Adapting LLMs to specialized fields like law, medicine, and cloud incident response typically requires slow, manual work that's hard to reproduce—AutoAdapt aims to streamline this. The system promises to make LLMs more reliable and performant in high-stakes environments without extensive manual tuning.

Google suggests 10 Gemini prompts to plan your 2026 budget—from 'Help me create a budget' to 'Review my budget'. Example responses show standard financial advice: list income and expenses, separate needs from wants, set savings goals. Gemini is also meant to help with expense tracking and tips on reducing costs or increasing income.

- NVIDIA unveiled DSX Air at GTC 2026 in San Jose, introduced by CEO Jensen Huang as a core part of the DSX platform. - DSX Air simulates AI factory setups before physical deployment, cutting rollout time from months to days. - It sits within NVIDIA DSX Sim and targets faster planning, testing, and optimization of AI infrastructure.

- Facebook Marketplace is rolling out new Meta AI features designed to speed up the selling process. - A one-click listing tool lets AI automatically generate descriptions, categories, and price suggestions. - An AI assistant handles buyer inquiries, reducing manual effort for sellers.

OpenAI launches the Codex app for macOS—a command center for AI-powered coding with multiple parallel agents and long-running tasks. The app enables multi-agent workflows: different AI instances work simultaneously on different parts of a project. Developers can orchestrate complex software projects without switching between tools or chat windows.

Microsoft Research experts examine whether AI can contribute to a more sustainable world, analyzing global emissions from datacenter operations, potential efficiency gains, and AI's potential across electrification, materials science, and food systems. The podcast explores both AI's environmental footprint and its potential as a tool for sustainability.

- NVIDIA launched Nemotron 3 Super, an open model with 120 billion total parameters but only 12 billion active ones, using a mixture-of-experts architecture. - NVIDIA claims 5x higher throughput compared to dense models of similar scale, specifically targeting agentic AI workloads. - Perplexity is among the first AI-native companies to offer users direct access to the model.

- Meta is doubling down on its in-house AI chip strategy: the MTIA (Meta Training and Inference Accelerator) line remains a cornerstone of the company's AI infrastructure. - Four new generations of MTIA chips are planned within the next two years – an unusually aggressive development cadence. - The move reduces Meta's dependence on Nvidia GPUs for internal AI workloads such as ranking, recommendation, and inference.

The Earth BioGenome Project uses AI to sequence genomes of endangered species and preserve genetic information. So far, 1,000 genomes have been sequenced, with a goal of cataloging 1.3 million species by 2030. The genetic data helps scientists better understand the biology of endangered species and develop protection strategies.

Microsoft Chief Scientist Jaime Teevan and researchers Jenna Butler, Jake Hofman, and Rebecca Janssen unpack the New Future of Work Report 2025 and explore the ideal AI-driven working world. Plus, is AI a tool or a collaborator? And why the answer matters.

- Microsoft Research, in collaboration with Princeton University and Universitat Politècnica de València, has introduced ADeLe – a framework designed to predict and explain AI performance on new tasks, not just benchmark scores. - Standard benchmarks only measure model performance on fixed test sets; they don't explain failures or generalize to unseen tasks.

- Microsoft Research has released AsgardBench, a new benchmark designed to evaluate how well AI systems can plan in visually complex, interactive environments. - The benchmark simulates everyday scenarios like kitchen tasks, where an agent must observe its surroundings, make decisions, and adapt to unexpected changes.

- Microsoft researchers Subutai Ahmad and Nicolò Fusi join Doug Burger to debate whether today's AI systems are on a path toward genuine intelligence. - The conversation centers on comparing transformer architectures with the human brain, especially around continual learning and energy efficiency.

- Microsoft Research introduces AgentRx, a systematic debugging framework for AI agents performing autonomous tasks like cloud incident management or multi-step API workflows. - The core problem: when an agent fails – for example by hallucinating a tool output – there is currently no structured methodology to trace the root cause.