Researchers Discover Glassworm Malware Hidden in Invisible Unicode Aikido Security Warns.

The invisible threat we’ve tracked for nearly a year has re-emerged. While the PolinRider campaign compromised hundreds of GitHub repositories, we now see a sharp rise in glassworm activity impacting GitHub, NPM, and VS Code.  

Last October, we urgently warned about how hidden Unicode characters compromised GitHub repositories, a method unmistakably linked to glassworm. Now the situation is critical. Glassworm has resurfaced this month, and high-profile repositories are already affected, including those from Wasmer, Reworm, and OpenCode-Bench from anomalyco, the team behind OpenCode and SST.  

A Year Tracking the Invisible Code Campaign 

  • In March 2025, Aikido first finds malicious NPM packages that hide payloads using PUA Unicode characters.  
  • In May 2025, we will publish a blog post explaining the risks of invisible Unicode and how attackers can use it in supply chain attacks.  
  • On October 17, 2025, we found compromised extensions on OpenVSX that use the same technique.  
  • On October 31, 2025, we discovered that attackers had started targeting GitHub repositories.  
  • In March 2026, a new large-scale attack compromised hundreds of GitHub repositories, and NPM and VS Code were also affected.  

A Quick Reminder 

Before we uncover just how widespread this alarming new wave is, let’s quickly review how the attack works. Even with months of warnings, it continues to catch developers and tools off guard.  

The attack exploits invisible Unicode characters that escape detection in nearly every editor, terminal, and code review tool. Attackers conceal dangerous payloads within what appears to be empty strings. When the JavaScript runtime executes cold code, a disorder instantly extracts the real bytes and sends them straight to eval(), unleashing the full threat.  

Cybersecurity researchers have identified three new extensions linked to the Glassworm campaign, indicating continued targeting of the Visual Studio Code (VS Code) ecosystem.  

These extensions remain active threats and can still be downloaded right now. They are:  

  • AI-driven-dev.ai-driven.dev 3402 Downloads  
  • Adhamu.history-in-sublime-merge 4057 downloads  
  • Yasuyuky.transient.emacs 2431 Downloads  

Glassworm was first reported by Koi Security late last month. Attackers are exploiting VS Code extensions from both the Open VSX Registry and Microsoft Extension Marketplace to steal Open VSX, GitHub, and Git credentials. They actively drain funds from 49 cryptocurrency wallet extensions and install extra remote access tools, escalating the threat to urgent levels.  

This malware is particularly dangerous because it hides its code using invisible Unicode characters in code editors, stolen credentials fuel a self-replicating infection cycle that rapidly spreads across systems, making it difficult to stop the worm-like attack.  

Based on this evidence, Open VSX said it had found and removed all malicious extensions and had changed or revoked related tokens as of October 21, 2025. However, Koi’s Security’s latest report shows the threat has returned, using unusable Unicode characters to avoid detection.  

The attacker submitted a new Solana blockchain transaction that updated the C2 endpoint for malware downloads, according to security researchers Idan Dardikman, Yuval Ronan, and Luton Sery. This shows the resilience of blockchain-based C2 infrastructure. Even if servers shut down, the attacker can post a cheap transaction, and all infected machines get the updated location.  

The security vendor also found an exposed endpoint on the attackers’ server, revealing a partial victim list across the US, South America, Europe, Asia, and a major Middle East government entity.  

Further analysis found keylogger data that seems to come from the attacker’s own machine. This has provided some indications about where glassworms come from. The attacker is believed to be Russian-speaking and uses an open-source browser extension C2 framework called Redext as part of their setup.  

These are real organizations and real people whose credentials are being harvested now, whose machines may be serving as criminal proxy infrastructure and whose internal networks could be compromised at any moment, Koi Security said.  

The alarming news follows reports from Aikido Security that Glassworm is actively targeting GitHub, with stolen credentials being used to push malicious commits and cause immediate harm.

Sources: GlassWorm Malware Discovered in Three VS Code Extensions with Thousands of Installs 

Glassworm Is Back: A New Wave of Invisible Unicode Attacks Hits Hundreds of Repositories

Amazon Pushes ML-KEM Quantum-Safe Encryption for Future Internet, AWS Blog Reveals

Amazon Web Services (AWS) has deployed the latest hybrid post-quantum key agreement standards for TLS for 23 AWS services. AWS Key Management Service (AWS KMS), AWS Certificate Manager (ACM), and AWS Secrets Manager endpoints now support the lattice-based Key Encapsulation Mechanism (ML-KEM) for hybrid post-quantum key agreements in non-FIPS endpoints across all AWS regions. The AWS Secrets Manager Agent, built on the AWS SDK for Rust, now provides optimal support for hybrid post-quantum key agreement. This allows customers to use end-to-end post-quantum–enabled TLS when bringing data into their applications.  

These three services were selected because they are security-critical and require the highest level of post-quantum confidentiality. They previously supported Crystals Kyber, which ML-KEM now replaces. Crystals Kyber will continue until 2025, but will be removed from all AWS service endpoints in 2026 as ML-KEM becomes the standard.  

Our Migration to Post-Quantum Cryptography 

AWS is following its post-quantum cryptographic migration plan as part of this. AWS will add MLKM support to all services with HTTPS endpoints over the next few years. Customers need to update their TLS clients and SDKs to use ML-KEM when connecting to AWS HTTPS endpoints. This helps protect against future threats from quantum computing. AWS endpoints will select ME, ML-KEM when clients offer it.  

Our hybrid pAWS can negotiate hybrid post-quantum key-agreement algorithms thanks to AWS LibCrypto and AWS LC. Our open-source FIPS 143-validated cryptographic library and S2N TLS, our open-source TLS implementation. AWS LC has received several FIPS certificates from NIST: 434631, 4759, and 4816, and was the first open-source cryptographic module to include MLKM in a FIPS 140-3 validation.  

ML-KEM on TLS Performance 

Migrating from an elliptic curve Diffie-Hellman (ECDH) only key agreement to an ECDH plus ML-KEM hybrid key agreement necessarily requires that the TLS handshake send more data and perform more cryptographic operations. Switching from a classical to a hybrid post-quantum key agreement will transfer approximately 1,600 additional bytes during the TLS handshake and will require approximately 80 to 150 microseconds more compute time to perform ML-KEM cryptographic operations. This is a one-time TLS connection startup cost or amortized over the lifetime of the TLS connection across the HTTP requests sent over it.  

AWS is working to provide a smooth migration to hybrid post-quantum key agreement for TLS. This work includes benchmarking example workloads to help customers understand the impact of enabling hybrid post-quantum key agreements with ML-KEM.  

Using the AWS SDK for Java v2, AWS measured how many AWS KMS GenerateDataKey requests per second a single thread can send between an Amazon EC2 C6i bare metal client and the public AWS KMS endpoint, both in the US West 2 region. Classical TLS connections used the P-256 elliptic curve, while hybrid post-quantum TLS connections used the X25519 elliptic curve with ML-KEM-768. Your results may vary depending on your environment, including instance type, workload, parallelism, number of threads, and network setup. The tests measured HTTP request rates with TLS connection reuse enabled and disabled. The handshake is never amortized, and every HTTP request must perform a full TLS handshake. Enabling hybrid post-quantum TLS reduces transactions per second (TPS) by about 2.3%, from 108.7 TPS to 106.2 TPS.  

Results show that enabling post-quantum TLS has little impact on performance. For most workloads, the maximum DPS rates dropped by just 0.05%. In the worst case, with each request creating a new TLS handshake, the drop was only 2.3%.   

Removing Support for Draft Post Quantum Standards 

AWS Service Endpoints that currently support Crystals Kyber, the predecessor to ML-KEM, will continue to support it through 2025. AWS will gradually phase out Crystals Kyber after customers switch to ML-KEM. If you are using an AWS SDK for Java version that only supports Crystals Kyber, upgrade to the latest version with ML-KEM support. If your code uses a recent AWS SDK for Java V2 release, no changes are needed for the transition from Crystals Kyber to ML-KEM.  

Customers whose clients currently use Crystals Kyber must upgrade their AWS Java SDK v2 to a version that supports ML-KEM before 2026, as Crystals Kyber will be removed in 2026. Clients that have not updated will automatically revert to using classical key agreements to maintain connectivity but will lose post-quantum confidentiality.  

How to use Hybrid Post Quantum Key Agreement 

To enable hybrid post-quantum key agreement in the AWS SDK for Rust, add rustls to your crate and activate the prefer-hybrid-post-quantum feature flag.  

For AWS SDK for Java 2.x, enable hybrid postquantum key agreement by calling .postquantumtlsenabled(true) when building the AWS common runtime HTTP client.  

Step 1: Add the AWS Common Runtime HTTP client to your Java dependencies 

Add the latest AWS Common Runtime HTTP Client to your Maven dependencies. Use version 2.30.22 or higher for ML-KEM support.  

Step 2: Enable Post-Quantum TRS in your Java SDK client configuration 

Select AWSCRTAsyncHTTPClient in your AWS Client setup. Enable post-Quantum TLS.  

Things to Try 

Here are a few ways you can use this client with Post-Quantum support:  

  • Run, Load, Tests, and Benchmarks: AWSCRTAsyncHTTPClient is high-performing and uses AWS LibCrypto on Linux. If you are new to it, compare its performance to the default SDK client. Afterward, enable Post Quantum TLS and check whether it outperforms the default client without it.  
  • Test connections from various locations: Requests may be made via proxies or firewalls that use Deep Packet Inspection (DPI). If blocked, ask your security team to update rules for these TLS algorithms. Share feedback on how your network handles this traffic.  

Conclusion: We’ve added ML-KM Hybrid Key Agreement to 3 AWS Endpoints with TLS connection reuse, enabling hybrid post-quantum TLS, with minimal impact on performance in our tests. We saw only a 0.05% drop in the maximum transactions per second when using AWS KMS-generated data key.

Source: ML-KEM post-quantum TLS now supported in AWS KMS, ACM, and Secrets Manager 

Samsung Begins Shipping HBM4 Memory for Next-Gen AI PCs, Newsroom Confirms

Mass production of HBM4 commences with a consistent transfer speed of 11.7 Gbps with a maximum of 13 Gbps.  

Leading Edge DRAM with a 4nm logic-based die maximizes performance, reliability, and energy efficiency for next-gen data centers.  

Secure Process Technology and Supply Capabilities Strengthen Samsung’s HBM Roadmap Beyond HBM4.  

Samsung Electronics announced it has started mass production of its HBM4 memory and has shipped products to its customers. This makes Samsung the first company in the industry to reach this milestone and take an early lead in the HBM4 market.  

By using its cutting-edge 6th-generation 10-nanometer (NM)-class DRAM Process (1C), Samsung achieved stable yields and top performance from the start of mass production without needing any extra redesigns.  

Instead of using proven designs, Samsung chose the most advanced nodes, such as 1C DRAM and a 4nm logic process, for HBM4, said Sang Joon Hwang, Executive Vice President and Head of Memory Development. By leveraging our process strengths and design optimization, we deliver greater performance and can meet customers’ growing needs for higher performance when they need it.  

Setting the Bar for Maximum Effectiveness and Efficiency 

Samsung’s HBM4 operates at a data rate of 11.7 Gbps, which is approximately 46% faster than the prevailing industry standard of 8 Gbps. This represents a 22% increase over HBM3E’s maximum data rate of 9.6 Gbps. HBM4 can achieve peak speeds of up to 13 Gbps, alleviating bandwidth constraints as AI model sizes increase.  

The aggregate bandwidth per HBM4 stack is now 2.7 times the weight of HBM3E, reaching up to 3.3 terabytes per second (TB/s) across all I/O pins. Samsung’s HBM4 uses a 12-layer 3D stacking approach with capacities ranging from 24 GB to 36 GB per stack. Future 16-layer stacks will support modules up to 48 GB, enabling scalability based on system and customer requirements.  

To support increased power and heat from doubling data I/O from 1024 to 2048 signal pins, Samsung integrated advanced low-power circuitry at the core die. HBM4 achieves 40% higher power efficiency through low voltage, through-silicon vias (TSVs), and power distribution network (PDN) tuning, and realizes a 10% gain in thermal resistance and 30% better heat dissipation versus HBM3E. Samsung’s HBM4 delivers high performance, efficiency, and reliability, helping customers get more from their GPUs and control costs in new data centers.  

Comprehensive Yet Agile Manufacturing Capacities 

Samsung will build on its large-scale manufacturing resources to advance HBM4 and future technologies. The company will roll out enhancements and revisions to the roadmap in upcoming product generations.  

Close collaboration between Samsung’s factory and memory teams through design-technology co-optimization (DTCO) supports maintaining high quality and yield. Their in-house expertise in advanced packaging also shortens production cycles and lead times.  

Samsung also plans to expand its technical partnerships. The company works closely with global GPU makers and hyperscalers on next-generation ASIC development.  

Samsung expects its HBM sales to more than triple in 2026 compared to 2025. The company is expanding HBM4 production with HBM4E sampling to begin in the second half of 2026, following the HBM4 launch. Custom HBM samples will be delivered to customers in 2027 as needed, outlining a sequential roadmap: first launch, then sampling, then HBM delivery.

Source: Samsung Ships Industry-First Commercial HBM4 With Ultimate Performance for AI Computing 

NVIDIA Unveils Vera Rubin AI Platform for Efficiency Gains at GTC 2026

NVIDIA is unveiling the Vera Rubin platform, introducing a new era in AI as seven new chips enter full production to scale the world’s largest AI factories.  

The platform combines an NVIDIA GPU, a Vera CPU, a Rubin GPU, and an NVLink 6 switch. Connectex 9, SuperNic, Bluefield 4 DPU, Spectrum 6 Ethernet switch, and the new Groq 3 LPU. These chips work together as a single AI supercomputer, powering every single stage of AI from large-scale training and testing to real-time agent tech influence.  

Jensen Huang, Founder and CEO of NVIDIA, stated that Vera Rubin marks a generation with seven breakthrough chips, five racks, one large supercomputer, all built to support every phase of AI. According to Huang, the arrival of Agentic AI accelerates the latest infrastructure build-out in history.  

Enterprises and developers are using the cloud for increasingly intricate, agentic workflows and mission-critical decisions. “That demands infrastructure that can keep pace,” said Dario Amadei, CEO and co‑founder of Anthropic. NVIDIA’s Vera Rubin Platform gives us the compute, networking, and system design to keep delivering as we advance the safety and reliability our customers depend on.  

Sam Altman, CEO of OpenAI, affirmed that NVIDIA’s infrastructure is the foundation for continued AI advancements. With NVIDIA, Vera Rubin, and OpenAI expect to run more powerful models and agents at scale, delivering faster, more reliable systems to a broad user base.  

Shift to POD Scale Systems 

AI infrastructure is evolving from separate checks to servers to fully integrated rack-scale systems to POD-scale deployments to AI factories to sovereign AI. These changes deliver higher performance and make AI more cost-efficient for organizations of any size. They also enhance accessibility to powerful AI and improve energy efficiency, reducing costs for intensive workloads.  

By closely integrating compute, networking, and storage and working with over 80 NVIDIA MGX partners worldwide, NVIDIA and Vera Rubin deliver the most extensive POD-scale platform. This supercomputer combines multiple AI-focused racks into a single unified system.  

NVIDIA Vera Rubin NVL72, Rack 

The Vera Rubin NL72 delivers key benefits by combining 72 Rubin GPUs and 36 Vera CPUs connected via NVLink 6, along with ConnectX 9, SuperNic, and Bluefield 4 DPUs. It enables streaming large mixture-of-experts models with only a quarter of the GPUs required by the NVIDIA Blackwell platform while delivering up to 10 times higher inference throughput per watt and reducing the cost per token by 90%.  

NVL72 is built for large-scale AI factories worldwide. It works smoothly with NVIDIA Quantum X800 InfiniBand and Spectrum X Ethernet, keeping graphics processing utility unit clusters running efficiently while cutting training time and overall costs.  

NVIDIA Vera CPU Rack 

Reinforcement learning and agentic AI workloads depend on large numbers of CPU-based environments to test and validate results generated by models running on GPU systems.  

The NVIDIA Vera CPU Rack features a lens liquid-cooled setup powered by NVIDIA MyMJX and 256 Vera CPUs. It delivers scalable, energy-efficient performance and top single-core performance, enabling large-scale agentic AI.  

With Spectrum-X Ethernet Networking (a network infrastructure solution), Vera CPU Racks keep CPU environments in sync across the AI factory. Alongside GPU Compute Racks, they form the CPU base for large-scale agentic AI and reinforcement learning. CPUs in this system deliver results twice as efficiently and 50% faster than traditional CPUs.  

NVIDIA Groq3 LPX Rack 

Media, Groq3, LPX, Advances Accelerated Computing with Key Benefits. It is designed for low-latency, large-context, agentic systems and delivers up to 35 times higher inference throughput per megawatt and up to 10 times greater revenue potential for trillion-parameter models when used with Vera Rubin. When scaled up, many LPUs can work together as a single large processor for fast, predictable inference. The LPX rack has 256 LPU processors, 128 GB of on-chip SRAM, and 640 TB/s of bandwidth. Used with Vera Rubin, NVL72, Rubin GPUs, and LPUs, speeds up decoding by computing every layer of the AI model for each output token in parallel.  

The LPX architecture is optimized for Trillium parameter models and million-token contexts, working with Vera Rubin to maximize power, memory, and compute. Its higher throughput per watt and better token performance open up a new level of high-end trillion-parameter inference, creating more revenue opportunities for AI providers. LPX is fully liquid-cooled, built on MGX infrastructure, and will be available in the second half of this year as part of next-generation Vera Rubin AI.3.  

NVIDIA Bluefield 4STX Storage Rack 

The NVIDIA Bluefield-4 STX RackScale System is a storage solution created for AI, extending GPU memory across the POD and powered by Bluefield-4, which combines the NVIDIA Vera CPU and ConnectX-9 SuperNIC. STX provides a high-bandwidth shared layer for storing and retrieving a large key-value cache derived from data from language models and agentic AI workflows.  

NVIDIA DOCA Memos, a new DOCA framework that enhances BlueField for storage, allows dedicated KV Cache storage processing. This boosts inference throughput by up to 5x and greatly improves power efficiency compared to general-purpose storage. As a result, the system provides a broader context for faster multi-term interactions with AI agents, more scalable AI services, and better overall infrastructure utilization.  

Timothee Lacroix, Co-Founder and Chief Technology Officer of Mistral AI, commented that the NVIDIA BlueField for STX rack-scale complex memory storage system delivers the significant performance boost required to expand agentic AI. Lacroix highlighted that by introducing a storage tier designed specifically for AI agents, primary STX helps maintain coherence and speed during complex reasoning over large datasets.  

NVIDIA Spectrum 6 SPX Ethernet Rack 

Spectrum-6 SPX Ethernet is designed to accelerate east–west traffic (data transfer between servers in a data center) in AI factories. It can be set up with Spectrum X Ethernet (networking technology) for NVIDIA Quantum X800 InfiniBand switches (high-speed data connections), providing fast, high-throughput connections between racks at scale.  

Spectrum-X Ethernet Photonics with pro-packaged optics (advanced optical technology for networking) delivers up to 5× better optical power efficiency and 10× more resiliency than traditional pluggable transceivers.  

Improving Resiliency and Energy Efficiency 

NVIDIA and over 200 data center partners announced the NVIDIA DSX platform for Vera Rubin. DSX Max-Q dynamically manages power across the AI factory, enabling data centers to deploy 30% more AI infrastructure within the same power limits. The new DSX Flex software makes AI factories grid flexible, unlocking 100 gigawatts of unused grid power.  

NVIDIA also released the Vera Rubin DSX AI Factory Reference Design, a blueprint for AI infrastructure that maximizes tokens per watt and overall output. This design improves system resiliency (ability to handle outages) and speeds up time to first production.  

By tightly integrating compute, networking, storage, power, and cooling, this architecture improves energy efficiency and supports reliable AI factory scaling under heavy workloads while maintaining high uptime.  

Broad Ecosystem Support 

Vera Rubin based products will become available from partners in the second half of the year. Leading cloud providers. Amazon Web Services, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure along with NVIDIA cloud partners CoreWeave, Cursoe, Lombard, Nebius, Nscale, and Together AI will offer these products.  

Global system manufacturers such as Cisco, Dell Technologies, HPE, Lenovo, and Supermicro plan to offer a variety of servers based on Vera Rubin products. Additional partners include Aivres, Asus, Foxconn, Gigabyte, Inventec, Pegatron, Quanta Cloud Technology (QCT), Wistron, and Wiwynn.  

AI labs and leading model developers, including Anthropic, Meta, Mistral AI, and OpenAI, intend to use the NVIDIA Vera Rubin platform to train larger, more advanced models. Their goal is to deliver long-context, multimodal systems with reduced latency and lower costs than previous GPU generations.

Source: NVIDIA Vera Rubin Opens Agentic AI Frontier 

OpenAI Signals AI Could Soon Run Apps on Your Computer

OpenAI is moving from a commerce system to agentic AI using an operator and its computer with a CUA model. The operator can independently control a computer to complete multi-step tasks, enabling AI to interact with websites and apps on the user’s behalf.  

With this new paradigm in mind, consider the following outline of OpenAI’s vision for autonomous computer tasks.  

To provide better context, let’s start by focusing on the first key area:  

  • An operator is an AI agent designed to take control of a user’s web browser and eventually their computer to handle repetitive or complex tasks.  
  • The operator uses a computer running the agent CUA model, which combines GPT-4’s Visual Reasoning with Reinforcement Learning. Unlike older automation tools that require API interfaces, CUA can view the screen via screenshots and interact with graphical interfaces, as a person does.  
  • An operator can fill out forms, order groceries, do research, create memes, and schedule appointments.  

The Shift to Agentic AI 

  • Operator denotes a shift from a checkbox that only talks to agents who can take action. It is built to manage long, multi-step tasks with little need for people to step in.  
  • A new ChatGPT Agent feature lets AI use a virtual computer to check calendars, book restaurants, and make slide decks.  
  • Once you set a goal, agents work on their own. For example, it could plan a weekend trip.  

Present Constraints & Safety 

  • The operator is still in the research stage and is primarily available to pro users in the US.  
  • The AI pauses for human approval before any action that can be undone, like sending emails or deleting calendar events.  

The agent can sometimes get stuck on streaky interfaces, capture or password fields, so it may need help from a person.  

Future Outlook discusses the upcoming directions and possibilities for the Operator platform and agentic AI. 

  • OpenAI plans to expand the operator to the Plus team and Enterprise users.  
  • OpenAI positions Customer Agents as a Foundation for Progress Towards Artificial General Intelligence (AGI).  
  • The aim is to move from a single tool to an ecosystem in which agents work independently across multiple systems seven times.  

The move to agentic AI is part of a broader trend in 2025, with companies like Anthropic and Google building similar capabilities.  

At the start of this year, OpenAI CEO Sam Altman predicted 2025 would be pivotal for AI agents—tools that automate tasks and act on users’ behalf.  

Building on this vision, OpenAI is now making its first real move in this area.  

OpenAI has announced a research preview of Operator, an AI agent that controls a web browser and autonomously performs tasks. It will initially be available to US users with ChatGPT’s Pro subscription and will expand to Plus, Team, and Enterprise plans, with dates to be announced.  

Operators will be available in other countries soon, though a specific launch date has not been announced. OpenAI CEO Sam Altman said during a live stream on Thursday that Europe will, unfortunately, take a while.  

Currently, the research preview is at operator.chatgpt.com. OpenAI plans to add Operator to all ChatGPT clients soon. The operator promises to automate tasks such as booking travel, making reservations, and shopping. The interface offers categories such as shopping, delivery, dining, and travel for different automations.  

When users activate Operator in ChatGPT, a dedicated web browser opens, allowing the agent to complete tasks and explain its actions. Users still control their own screen, as Operator operates in its own browser.   

OpenAI explains the browser runs on a computer using an agent or CUA model, combining GPT-4o’s vision skills with advanced reasoning. The CUA attempts to interact directly with website interfaces, bypassing developer APIs.  

This allows the CUA to click, navigate menus, and fill forms on web pages much like a person.  

OpenAI says it’s collaborating with companies like DoorDash, eBay, Instacart, Priceline, StubHub, and Uber to ensure operators comply with their terms of service.  

The CUA model is trained to ask for user confirmation before finalizing tasks with external side effects, for example, before submitting an order or sending an email, so that the user can recheck the model’s work before it becomes permanent. Open-air rights in materials provided for death crimes. It has already proven useful in a variety of cases, and we aim to extend that dependability across a wider range of tasks.  

But OpenAI warns the CUA isn’t perfect. The company says it doesn’t expect the CUA to perform reliably in all scenarios just yet.  

Currently, the operator cannot consistently handle many complex or specialized tasks. OpenAI adds support for tasks such as creating detailed slide shows, overseeing intricate calendar systems, or interacting with highly customized or non-standard web interfaces.  

To be extra careful, OpenAI requires users to supervise certain tasks, such as banking transactions, even though the CUA and operator could handle them on their own. For example, users must enter credit card information themselves. OpenAI also says the operator does not color or screenshot any data.  

On particularly sensitive websites, such as email, the operator requires active user supervision, guaranteeing users can directly catch and handle any potential mistakes the model might make, OpenAI says in its support materials.  

This does limit what an operator can do, but it also helps prevent mistakes like the agent accidentally spending your mortgage payment on edgy accent chairs. Google has taken a similar approach with its Project Marina AI agent, which also avoids entering sensitive information such as credit card numbers.  

Limitations 

The operator does have some important limitations.  

There are both daily and task-based rate limits. OpenAI says an operator can handle seven tasks at once, but there are dynamic limits on how many. There is also an AU total-usage limit that resets each day.  

For security reasons, the operator will not perform certain tasks at this stage, such as sending emails or deleting calendar events, even though the CUA can. OpenAI says this may change in the future, but there is no timeline yet.  

An operator can also get stuck if it encounters a complex interface, a password field, or a captcha. When this happens, it will prompt the user to take over.  

An Agentic Future 

Compared with competitors like Rabbit, Google, and Android, OpenAI has taken longer to develop an AI agent. This may be due to the technology’s safety risks.  

When an AI system can take actions on the web, it opens the door to much more dangerous use cases from nefarious actors. You could automate AI agents who orchestrate phishing scams or DDoS attacks or have them snatch up tickets to a concert before anyone else can. Especially for a tool as widely used as ChatGPT, it’s important that OpenAI takes steps to prepare for such exploits.  

OpenAI believes the operator is safe enough to release now, at least as a result review.  

The operator employs tools that seek to limit the model’s susceptibility to malicious prompts, consent instructions, and prompt injection. OpenAI explains on its website that a monitoring system triggers action if suspicious activity is detected, while automated and human-reviewed pipelines continuously update safety balances.  

Operator is OpenAI’s most ambitious effort so far to create an AI agent. Lastly, OpenAI launched Tasks, which gave ChatGPT basic automation features such as creating reminders and scheduling prompts to run at specific times each day. Tasks added some familiar but important features to ChatGPT, making it as practical as Siri or Alexa. However, the Operator introduces capabilities that earlier virtual assistants could not offer in AI. After ChatGPT, a new technology that will change how people use the internet and their PCs. Instead of simply delivering and processing information, agents can, in theory, take actions and actually do things.  

Now that OpenAI has released its first real AI agent, we will soon see how realistic this vision actually is.

Source: OpenAI launches Operator, an AI agent that performs tasks autonomously 

Microsoft Develops AI Agents to Automate Office Tasks, Report Reveals

When Generative AI emerged, It Jolted Our IT Engineering Team.  

We felt excitement, curiosity, skepticism, and had questions about what this technology meant for IT’s future.  

At Microsoft Digital, we didn’t start with a major transformation plan. We recognized AI wasn’t just another tool, but a big shift in engineering. This realization led us to a gradual, thoughtful approach.  

For years, our IT teams prioritized skill, reliability, and excellence. These priorities persisted. What shifted was our potential.  

Engineers could now generate code in seconds, digitize complex systems rapidly, or automate tasks that once took hours. This gave us a chance to grow our team’s AI expertise.  

This shift in capabilities prompted us to pause and consider tougher questions.  

How do you help thousands of engineers see how AI can perfect their daily work? How do you build trust after trying new things? And how do you use AI in a way that supports, not weakens, core engineering skills?  

We found our answers by taking a step-by-step approach focused on people, culture, and ongoing learning.  

Phase 1: Building Awareness and Giving Access 

It may seem unexpected for an engineering team, but our earliest challenge was not technology; it was understanding.  

When generative AI became a topic, most engineers noticed the news and tested some tools, but few understood its impact on their work. Some felt excited; others, cautious; and many were unsure where to start. Bridging the gap between knowing about AI and practical use was our first challenge.  

We realized ordering engineers to use AI without context would create skepticism. Instead, we focused on something both simple and difficult: exposure.  

We started by making AI visible and accessible in the tools engineers already used. GitHub Copilot, Microsoft 365 Copilot, and Early Copilots are embedded directly into engineering workflows. The goal wasn’t immediate productivity gains; it was familiarity, letting engineers see first-hand what AI could and couldn’t do.  

We also made sure to discuss AI’s limitations openly.  

AI wasn’t perfect; it sometimes invented facts or confidently made mistakes. Being honest was the key. By describing AI as an assistant, V-Rime reminded everyone that engineering judgment remained crucial. Engineers kept control. They just needed to know how.  

We also made it safe for engineers who experiment.  

New quotas, not First Adoption Matrix. Engineers were encouraged to try AI on low-risk tasks such as summarizing documentation, generating test cases, or examining unfamiliar code bases. Minor victories built confidence, confidence built curiosity, and curiosity drove organic adoption.  

As more people experimented, their mindset started to change.  

We encouraged the use of tools so people would experiment with AI, says Mukul Singhal, partner group engineering manager. Once they did, they saw value. The mindset shifted from “AI, replace me” to “AI can be my companion.”  

Over time, the conversation changed from asking “Should we use AI?” to “Where does AI help most?”  

Engineers began sharing prompts, tips, and lessons with one another. What began as individual exploration soon became community learning. Awareness turned into momentum.  

Ultimately, Phase 1 focused on giving people access to explore, ask questions, and learn. With this foundation in place, we were ready to build on what we had learned.  

Phase 2: Culture Shift 

When people gained access to AI, their increased use led to observable outcomes. Curiosity grew as individuals experimented, and we began to measure concrete ways AI enhanced productivity and problem-solving.  

As more engineers tried AI, we observed some teams working faster and more efficiently, while others initially slowed. Our analysis indicated that mindset, not technical skill, was the differentiator.  

We needed people to view AI as a standard part of modern engineering, not as an experiment.  

This meant making AI a trusted part of the engineering process.  

Leaders were key. Rather than call AI a shortcut, they described it as a way to improve basic engineering, clarify design talks, produce better documentation, provide faster feedback, and give more time for real problem-solving. The message was steady: using AI meant rethinking work, not shortcuts.  

We also had to deal with an early fear that using AI meant people would be repressed, not empowered.  

People have shifted from the mindset of “will AI work?” to “AI is working for me,” says Veera Mamilla, a Principal Growth Engineering Manager at Microsoft Digital. I think that was a transformative shift, and I believe many engineers in the organization began to believe in AI.  

How we talked about AI made a difference.  

As engineers used AI, success focused on results. Did AI help you learn systems, reveal risks, or free up time for critical work?  

Over time, AI use became routine. Observed outcomes included users setting new standards, peer-to-peer learning, and shared team success. Teams began discussing optimal ways to use AI, focusing less on adoption and more on effectiveness.  

Phase 3: Upskilling and Role Evolution 

Once AI was deployed, building new skills was essential.  

We deliberately chose upscaling and rescaling rather than replacement. The goal: invest in our current workforce.  

That choice influenced everything we did next.  

At first, upscaling focused on practical basics, including terminal learning the tools and observing how corporates and early agents worked in real-world situations. We encouraged every engineer to try these, no matter their specialty.  

But just developing new skills wasn’t enough. Evolving job roles changed how engineers contributed and worked together.  

In software, service, and cloud network engineering, work shifted from hands-on tasks to more oversight and coordination. Engineers learn to guide AI, review results, and decide on automation.  

As things changed, we looked at how the industry was changing engineering jobs. We just compared new job descriptions from the market with Microsoft’s own words. There was no official AI engineer job yet. Instead of creating a new title, we focused on updating expectations for current roles.  

The idea of an AI-native engineer became more about mindset than job title.  

An AI-native engineer still knows systems, architecture, and risk. Defense and routine tasks go to AI, while judgment, design, and responsibility stay with people. Engineers now oversee AI-assisted work instead of doing it themselves.  

Your title might be Software Engineer or Principal Engineer, says Ragini Singh, a Partner Group Engineering Manager in Microsoft Digital. But if you are acting like an AI engineer, what does that actually mean? That question helped us start defining how these roles were evolving.  

That evolution looked different across disciplines. Software engineers focused on AI-assisted coding, test generation, and spec-driven development. Service engineers relied on AI for incident response, knowledge capture, and out-of-the-ordinary decision support. Good cloud network engineers began moving from manual operation toward intelligent orchestration and agent-assisted troubleshooting. The common thread wasn’t identical to any of them; it was a shift toward higher-order work and robust code.  

Phase 4: Embedding AI Across the Engineering Life Circle 

At this stage, we realized that enabling individuals to be more productive would lead to greater benefits. Most AI users showed up in familiar places: code suggestions, document summaries, quick answers useful but fragmented. The bigger opportunity emerged when we stepped back and asked a harder question: What would it look like if AI were embedded across the entire engineering life cycle, not just used at isolated moments?  

We stopped thinking in terms of tools and started thinking in terms of flow, design, build, test, deploy, operate, and improve. AI needed to show up across all of it in ways that stringent engineers already worked.  

In Software Engineering, this meant using AI earlier to write requirements, explore design choices, and review code with a broader view. Coding help remains important, but is not the main focus.  

AI improved testing and quality, too, by creating tests, finding defects, and reviewing code to reduce repetitive work and spot problems earlier. This lets engineers focus more on quality and design.  

In service engineering, AI for incident management means summarizing findings, gathering information, and analyzing signals. In cloud network engineering, it enables simpler coordination and troubleshooting. The goal: AI should streamline processes.  

Expanding this approach showed that adding AI was more than technical. It changed the entire system.  

If AI shows up at one step, you don’t get the full value, says Sudhakar Sadasivuni, a principal growth engineering manager at Microsoft Digital. The real influence comes when it’s integrated across the life cycle, enabling engineers to design, build, operate, and learn faster as a system.  

With AI as part of daily work, engineers ensure results and next standards through checks, tests, and validations. Using AI raised expectations for judgment and integrity; responsibility and governance became even more important.  

Over time, these changes brought even more benefits.  

Faster design cycles, reduced network data testing, fewer operational problems, improved insights, and accelerated recovery. AI became a core engineering system component, not just an analytical tool, accelerating outcomes.  

Every AI story faces the same question. Does it really make engineers work better? For us, the answer showed up quickly when we saw less tedious work.  

At Microsoft, digital enemies have always had to do work that was needed but tiring. This included manual troubleshooting, repeating diagnostics, examining logs, and other routine tasks that kept things running but did not help the company move ahead.  

AI gave us an opportunity to change this.  

Using AI does not always deliver better results. We shifted our focus and began to ask what’s different now that our engineers wield AI?  

This shift changed how we measured success. Instead of just tracking tool usage, we looked at the bigger picture. We saw faster design cycles, earlier defect detection, less time on rotator tasks, quicker incident resolution, clearer documentation, fewer end-offs, and less work.   

These weren’t just abstract numbers. We witnessed these improvements every day in our work.  

We made sure not to impose a single definition of value on everyone. Software engineers, service engineers, and cloud network engineers all feel the impact in different ways. What was most important was that each team could see real improvements in how their work flowed. This mindset reshaped how our leader’s described success.  

Adoption was always the starting point, says Ullas Kumble, a principal group software engineering manager at Microsoft Digital. But we are clear from the beginning that usage isn’t the destination. The real goal is impact. Semicolon, more time for engineers to focus on the work that really matters. Over time, this approach elevated our conversations. Instead of debating whether AI working teams pinpointed where it helped and where it fell short, measurement shifted to a tone for learning and setting priorities.  

Gazing Forward 

As we look forward to the future, one thing is clear: this journey isn’t finished. We have new challenges and opportunities ahead.  

AI tools will improve, agents will get smarter, and engineering roles will evolve. So we must stick to the principles that guide us, invest in people, focus on the basics, use AI in real workflows, and be honest about what works.  

We did not try to build an AI-driven animating organization all at once. We built it bit by bit.  

We met the engineers who transformed our culture before redefining roles. We moved through the life cycle rather than simply hiding in town. We eliminated liquidity war and quantified impact where it mattered most.  

In summary, our key takeaways are: focus on real impact, adapt to evolving roles, invest in people, and remain honest about results by building step by step. By embedding AI throughout, we achieve better engineering powered by AI and guided by human insight.

Source: Powering the new age of AI-led engineering in IT at Microsoft 

Samsung Gauss 2 Supreme Model Ramps To Automate Enterprise SDK Workflows

These tiers show Samsung’s goal to improve code, language, and image workloads across settings. Early adoption has led to noticeable productivity gains. Developer use of its assistant grew by 4x after switching to Gauss 2. Many technical details remain undisclosed. Analysts await independent proof.  

This article unpacks Gauss 2’s specifications, strategy benefits, and unanswered questions for enterprise buyers. To set the context, it first situates Samsung’s Live within the wider enterprise Gen.AI model landscape shaping 2025. With this perspective, readers gain concrete data points and applicable considerations for future AI roadmaps. Professionals may also explore certification paths to guide successful project deployment. Let us explore the core developments powering Samsung’s latest AI statement.  

Samsung Gauss 2 Model Overview. 

Building on the introduction, Gauss 2 is Samsung’s second internal formation model following Gauss. This project highlights Samsung researchers’ growth in AI. The enterprise GenAI model comes in three versions: Compact, Balanced, and Supreme, each for different tasks. Compact runs directly on devices for offline help with Galaxy phones and appliances. Balanced operates in Samsung data centers to enable broader consumer services, balancing speed and scale.  

Supreme uses a mixture of experts for complex inference and training. Samsung includes a custom tokenizer that supports 9 to 14 languages, depending on the setup, enabling faster multilingual processing than top open-source options. All versions support multimodal input — text, code, and images making Gauss 2 a flexible corporate content platform. In short, Samsung offers a range of options within a single enterprise Gen AI model family, informing enterprise adoption strategies.  

Strategic Enterprise Gen AI Move 

Samsung’s shift aligns with the world’s goal to use AI across 90% of its business areas. Leaders see Gauss 2 as the main engine for this change. By building its own platform, Samsung can control data location, privacy, and how the model works. It also saves on ongoing API costs to outside providers. Experts note that Samsung’s chip expertise helps it improve both the model and the hardware. It runs on. Competitors rely on third-party hardware and unclear messages. Gauss 2 also gives Samsung more power when working with telecom and cloud partners. These benefits support the company’s investments. Still, keeping funding and top talent is key to achieving long-term success. This context leads to a closer look at multimodal features.  

Multimodal Capabilities in Depth 

Multimodality refers to the ability to use multiple input types (text, code, images, and language translation) within a single system. For example, users can upload screenshots or design drafts and receive code suggestions tailored to the context. Developers can have the model update old scripts while viewing visual layouts. Call center agents get quick language summaries from recorded calls. Samsung says response crafting is now three times faster with those tools. The supreme version also improves knowledge in graphs, meaning it connects answers to real product facts. This reduces errors and improves productivity for support teams. Most open models require separate tools for each input type, but Gauss-2 combines them. These features set the stage for performance analysis.  

Performance And Adoption Data 

HUD numbers remain limited, yet Samsung shared several adoption metrics. According to the firm, usage of the coding assistant increased within months of Gauss 2 integration. Moreover, about 60% of Device-experience developers access the assistant weekly. The enterprise Gen AI model backs these gains by delivering 1.5 to 3 times faster processing. Samsung compared Balanced and Supreme against unnamed open-source baselines on internal benchmarks. However, the company has not released full datasets, tasks, or details on statistical significance as independent topics. Therefore, treat the figures as marketing claims awaiting third-party validation.  

Analysis of these performance data would not be complete without considering transparency and validation. This natural progression leads to broader consideration of benefits and challenges for stakeholders evaluating the platform.  

Benefits For The Samsung Ecosystem. 

The Gauss 2 rollout benefits more than just developers. On-device processing means tasks run directly on devices, reducing cloud latency and improving privacy. Galaxy phones with the compact version can transcribe or capture images offline, offering faster language translation and keeping data on the device. The balance-term and supreme versions help service teams by summarizing information and routing tickets efficiently, reducing support costs. Samsung fine-tunes the enterprise Gen AI model for business needs using its own data (instead of third-party data), which is harder to do on generic platforms. Organizations considering Gauss 2 should keep these key benefits in mind:  

  • Cost control through reduced external API calls.  
  • Unified handling of software, language, and image data.  
  • On-device experiences boosted buyer interest.  
  • Scalable architecture matching workload size.  

Together, these benefits make a strong case for Samsung’s AI platform. However, to provide a balanced view, before adopting Gauss 2, organizations should consider potential challenges and questions.  

Challenges And Open Questions. 

Like any proprietary platform, Gauss 2 comes with some risks. Samsung has not shared specifics such as parameter counts (number of model settings) or training sources (datasets used for learning), making it hard for analysts to compare it to models like GPT-4 or Gemini. There is also limited information on safety testing (risk evaluation), bias controls (methods to reduce bias in outputs), and governance (policies overseeing AI use). The Enterprise Gen AI model does not yet have a public API, meaning external developers cannot easily access its features, and there is no pricing information for planning integrations. By contrast, open-source models on Hugging Face are easier to try out right away. Ongoing maintenance, especially for on-device updates, is another concern. Though Samsung’s hardware expertise may help reduce some costs, professionals can improve oversight by earning the AI Project Manager certification. These problems show there are still important unknowns, so reviewing the roadmap is essential.  

Roadmap and Industry Impact 

Samsung plans to add GALF to most of its products over the coming years. The supreme version targets cloud systems while the compact one powers wearables and home devices. Adding knowledge graphs will make information more precise and customized. Experts, Apple, Google, and Xiaomi are expected to respond with updates. Samsung’s move may also drive demand for better mobile AI chips and push job providers to reveal more about costs and performance. Companies will need to balance vendor independence with ecosystem benefits. The choice of a foundation model will depend on openness, transparency, and cost-effectiveness. Those tools’ roadmap could reset buyer expectations for AI. These points lead us to our final thoughts.  

Gauss 2 shows that Samsung wants to shape its own AI features. The platform brings together software, language, and image processing into a single system. Early results point to real productivity gains and faster service. However, the lack of technical transparency means buyers need to do careful research. Companies should ask for clear benchmarks, safety information, and governance policies. As for the competition, Samsung will likely disclose more details soon. Professionals can help guide these decisions by earning the AI Project Manager certification. Now is the time to align your strategy with the fast-changing world of enterprise Gen.AI.

Source: Samsung Gauss2 Enterprise GenAI Model for Multimodal Workflows 

OpenAI, GPT-5.4 Reasoning Effort Parameter Scales Autonomous Decision Logic

GPT 5.4 is our most advanced model so far. It enables faster, more accurate results in the API and Codex, helping people and teams make better decisions, increase productivity, and streamline processes.  

In most cases, GPT-5.4 is the default choice for general tasks and coding, chosen to simplify complex workflows, save time on software engineering, enhance reasoning, improve writing quality, and open tools, all with one model.  

This article presents the standard features of the GPT-5 models and shows practical ways to make the most of GPT-5.4.  

Key Improvements 

GPT 5.4 offers several improvements over the previous GPT 5.2 model:  

  • Experience sharper coding, better document understanding, smarter audio, and more reliable instruction following.  
  • Enhanced image perception lets users analyze visuals more accurately. It also helps manage multimodal workflows more easily.  
  • Users can complete long-running tasks faster than before. They can also execute multi-step agent workflows more reliably.  
  • More efficient token use reduces costs and improves end-to-end performance for heavy tool-based workloads.  
  • Faster, smarter web search uncovers hard-to-find information, saving time and simplifying research.  
  • Streamlining the handling of many documents or spreadsheets boosts productivity across customer service, analytics, and finance workflows.  

Developers produce production-ready code and polished interfaces faster and more consistently, with fewer prompts for refinement.  

For agent-based tasks, GPT 5.4 completes multi-step processes faster. It often uses fewer tokens and tool calls. This makes agent-based approaches more responsive and reduces the cost of operating complex workflows at scale in API and Codex.  

New Features in GPT 5.4 

Like its predecessors, GPT 5.4 offers flexible tool options, control over explanation detail, and curated tool lists. Now enjoy new features that make building agent systems easier, help manage more information, and ensure reliable automation.  

  • With the API tool search, you can seamlessly browse tools across vast ecosystems. Only what you need. Work smarter with fewer tokens and on-point choices. Discover more in the tool search guide.  
  • 1M token context window: GPT‑5.4 can handle up to 1M tokens. This makes it easier to analyze entire codebases and large sets of documents, or to run agent processes in a single request. You can read more in the “1M context window” section.  
  • Interact directly with software for the first time. Agents can now complete, check, and fix tasks faster in a complete build, run, and verify. Check out the computer use guide for more.  
  • Power through longer processes. Keep vital content thanks to GPT-5.4’s native compaction support.  

Meet the Models 

For most tasks and coding, GPT-5.4 is your new go-to model. It now replaces GPT-5.2. GPT-5.4 Codex and ChatGPT users get GPT-5 chat (latest) by default. Need better answers? GPT-5.4 Pro Raw hardness offers extra compute for data-fest challenges.  

Prefer a compact model try GPT-5 Mini for streamlined performance.  

Ready to choose and weigh these trade-offs to find your perfect match:  

Variant  Best for  
GPT 5.4  General purpose work including complex reasoning, broad word knowledge and code-heavy or Code heavy multistep agentic tasks  
Gpt 5.4 Pro  Tough problems that may take longer to solve and need deeper reasoning  
GPT 5 mini  Cost-optimized reasoning and chat; balancesspeed, cost, and capability.  
GPT 5 nano  High-throughput tasks, especially straightforward instruction-following or classification  

Lower Reasoning Effort 

The reasoning effort setting determines how many reasoning tokens the model uses before responding. Older models like O3 only offered low, medium, and high options. “No” meant faster, less thoughtful responses, while I meant longer, more reasoned answers.  

From GPT 5.2 on, the lowest setting is called NUM, which enables faster responses. This is now the default in GPT 5.2 and later; to increase model reasoning, raise the setting to medium and observe the changes.  

When reasoning effort is set to none, prompts become more important. For better reasoning, even at the default setting, ask the model to think or list its steps before answering.  

Verbosity 

Verbosity controls how many output tokens the model produces. Fewer tokens make responses quicker. Reasoning style remains mostly unchanged, but responses will be briefer, which can end or hurt depending on your needs. Use high verbosity for detailed explanations or major code changes. Use low for brief answers or simple code.  

  • High verbosity is useful for detailed document explanations or major code refactoring.  
  • Low verbosity is best for short answers or simple code, such as SQL queries. GPT-5 supports high, medium, and low settings. In GPT 5.4, you can still adjust verbosity, with medium as the default.  

With GPT 5.4, medium and high robustness produce longer, more organized code with explanations. Semicolon, more,e generates shorter code with little extra commentary.  

GPT 5.4 is designed to solve problems by reasoning through them.  

Models like GPT-5.4 solve problems step by step. They create an internal chain of thought as their reasoning; for best results, send these reasoning steps back to the model. This prevents the same reasoning from being repeated and keeps the conversation aligned with the model’s training. In conversations with multiple turns, using previous_response_id will automatically include earlier reasoning steps. This is especially useful when using tools. For example, if a function call needs another wrong group, you can use the previous_response_id, or alternatively, add the reasoning steps directly to the input.

Source: Using GPT-5.4 

Google Cloud Hyperdisk ML Achieves 500,000 IOPS For Model Hot Swapping

As AI models become more complex, serving, scaling, and switching models instantly is now essential. In March 2026, Google Cloud tackled this challenge with its updated Hyperdisk ML storage. Recent benchmarks show that Hyperdisk ML achieves 500,000 IOPS during model hot swapping, establishing a new standard for high-performance AI infrastructure. 

For systems architects, this number is important. It enables “Always-On” generative AI apps that can switch between base models, LoRA adapters, and specialized weights in just milliseconds. 

The Bottleneck: Why IOPS Matter for Model Serving 

Traditional block storage has often quietly limited AI inference performance. Even top GPUs, as well as TPUs, can end up waiting while storage loads large model weights into memory. In production, where “hot swapping” means replacing one active model with another without downtime, IOPS becomes the main bottleneck. 

Loading a 70B parameter model from a regular disk can take seconds or minutes, causing cold-start delays. Hyperdisk ML, using Titanium offload, separates storage processing from the CPU. At 500,000 IOPS, it delivers the random-read performance required for G4 and A3 instances. 

Achieving 500,000 IOPS: The Titanium Advantage 

Hyperdisk ML achieves 500,000 IOPS due to its unique Google Cloud Hypercomputer design. Unlike classic SANs, it’s network-attached yet behaves like a local SSD. 

Concurrent Consumption limits matter. In a typical GKE cluster, many inference nodes access the same weights. Hyperdisk ML supports ReadOnlyMany, letting up to 2,500 nodes mount a volume. Google set 500,000 IOPS and 50 GiB/s throughput at the zonal level, supporting scale. 

Enabling Flawless Model Hot Swapping 

Model hot swapping is the next step in AI deployment. For example, a customer service bot may switch from a general language model to a specialized legal or billing model based on user needs. 

With Hyperdisk ML, developers can use “Weight-Streaming.” High IOPS lets the engine load only the required layers or adapters when needed. 

  • Reduced Idle Time: Accelerators, such as GPUs and TPUs, spend more time computing and less time waiting for the “First Token.” 
  • Cost Efficiency: Faster pod startup times let organizations reduce their pool of idle instances, thereby markedly lowering total costs. 
  • With GKE Volume Populator, weights are pre-cached and moved from Cloud Storage to Hyperdisk ML. When a swap is commanded, data is ready in the fast block layer. 

Performance Tuning for 500,000 IOPS 

To maximize the 500,000 IOPS performance on Hyperdisk ML during active model hot swapping and loading, engineers should focus on three key storage settings: 

  1. Use 4 KB I/O blocks for top IOPS. Larger blocks improve throughput, but 4 KB works well for fast, random reads on small LoRA adapters. 
  1. Queue Length: To fully utilize the Titanium pipeline, set the queue depth to at least 256. This lets the system handle many requests at once without waiting for each one to finish. 
  1. Instance Machine Series: Hyperdisk ML works best with the C3, C4, and G4 machine families. These have the hardware needed to connect with the Titanium storage offload engine at full speed. 

The Future of “Zero-Latency” AI 

Google Cloud Hyperdisk ML, which is reaching 500,000 IOPS for model hot swapping, shows that AI infrastructure is moving from “experimental” to “industrial-grade.” Today, even a 100ms delay can lose users, so storage can’t be ignored. 

With the required throughput and IOPS to make model weights appear “always resident” in memory, Google Cloud is making dynamic, multimodal, and highly personalized AI apps possible. For companies looking to go beyond basic chatbots to real-time, context-aware agents, Hyperdisk ML delivers the speed and reliability needed.

Source: High-performance block storage for any use case 

NVIDIA Project GR00T Update Adds Multimodal Tactile Learning for Robots

NVIDIA has introduced several new technologies to accelerate the development of humanoid robots. This includes NVIDIA ISAAC GR00T-N1, described as the world’s first open and fully customizable foundation model. It is a large artificial intelligence system trained on diverse data that can be adapted for many tasks, in this case, general humanoid reasoning and skills.  

Other technologies in the lineup include simulation frameworks and blueprints, such as the NVIDIA ISAAC-GR00T blueprint. A simulation framework is a set of software tools for testing and training robots. In a virtual environment, the blueprint helps generate synthetic training data. There is also Newton, an open-source physics engine developed with Google Brain and Disney Research, designed specifically to simulate real-world physical interactions for building robots.  

Building on these releases, GR00T-N1 is now available. It is the first in a series of customizable models that NVIDIA will share globally to support industries facing workforce shortages.  

The Age of Generalist Robotics is Here, said Jensen Wong, founder and CEO of NVIDIA, with NVIDIA ISAAC GR00T N1 and new data-generation and robot-learning frameworks. Robotics developers everywhere will open the next frontier in the age of AI.  

GR00T-N1 Advances Humanoid Developer Community 

The GR00T N1 Foundation Model uses a dual system design inspired by how people think. It features System-1, which acts quickly and automatically, like human reflexes or intuition, and System-2, which takes a slower, more careful approach to decision making. Dual-system design refers to splitting cognitive processes into fast and slow systems, similar to theories in human psychology.  

System 2 is powered by a vision-language model, a type of AI that understands images and written or spoken commands, reasons about its environment, and the instructions it has received to plan actions. System 1 then translates these actions into precise, continuous robot movements. System 1 is trained with data from both human demonstrations and a large volume of synthetic data generated by the NVIDIA Omniverse platform. The vision-language model enables the robot to interpret both visual and linguistic inputs.  

GR00T-N1 can handle a variety of common tasks, including grasping and moving objects with one or both arms and passing items between arms. It can also perform more complex multi-step tasks that need a longer context and a mix of general scales. These abilities are useful for tasks such as material handling, packaging, and inspection.  

Developers and researchers can further train GR00T-N1 with real or synthetic data to fit their own humanoid robots or tasks.  

During his GTC keynote, Goan showed 1X’s humanoid robot performing household tidying tasks on its own using a policy trained with GR00T-N1. This autonomous ability comes from an AI training partnership between 1X and NVIDIA.  

The future of human arts is concerning adaptability and learning, said Brent Bonich, CEO of One-X Technologies. While we develop our own models and media, GR00T-N1 provides a significant boost to robot reasoning and skills with minimal post-training data. We fully deploy on Neo-Gamma, promoting our mission of creating robots that are more than tools, yet companions capable of assisting humans in valuable, immeasurable ways.  

Other top humanoid developers with early access to GR00T-N1 include Agility Robotics, Boston Dynamics, Mentee Robotics, and Neura Robotics.  

NVIDIA, Google DeepMind, and Disney Research focus on physics.  

NVIDIA is working with Google DeepMind and Disney Research to develop Newton, an open-source physics engine. In this partnership, NVIDIA leads the development, with Google DeepMind and Disney Research contributing expertise, to help robots learn to perform complex tasks more accurately.  

Newton, built on the N-Media Warp Framework, will be optimized for robot learning when compatible with simulation frameworks such as MuJoCo and Isaac Sim. It will also utilize Disney’s physics engine.  

Google DeepMind and NVIDIA are also co-developing MuJoCo-Warp, aiming to accelerate robotics machine learning tasks by over 70 times. Developers will access it via Google DeepMind’s MJX open-source library and the Newton engine, co-developed with NVIDIA.  

Disney Research, as a partner in the Newton project, will be among the first to use the engine to improve its robotic character platform. This platform powers next-generation entertainment robots like the expressive Star Wars-inspired BDH droids that appeared with Huang during his GTC keynote.  

The BDH droids are just the beginning. We’re committed to bringing more characters alive in ways the world hasn’t seen before. This cooperation with Disney Research and Video and Google is a key part of that vision, said Kyle Laughlin, Sr. Vice President of Walt Disney Imagineering Research and Development. This alliance will allow us to create a new generation of robotic characters that are more expressive and engaging than ever before and connect with our guests in ways that Disney can.  

Continuing their collaboration, NVIDIA, Disney Research, and Intrinsic have announced a new partnership. Each organization will collaborate to develop OpenUSD pipelines and best practices for robotics data workflows, with NVIDIA overseeing the technical architecture and Disney and Intrinsic contributing their expertise in robotics and data management.  

NVIDIA has also announced the BGX Spark Personal AI supercomputer at GTC. It gives developers a ready-to-use system to expand GR00T and N1’s capabilities for new robots’ tasks and environments without requiring much custom programming.  

The Newton physics engine will be released later this year.

Source: NVIDIA Announces Isaac GR00T N1 — the World’s First Open Humanoid Robot Foundation Model