Alvaro Lopez Ortega / 2025-06-27 Briefing

Created Fri, 27 Jun 2025 22:44:24 +0000 Modified Sat, 28 Jun 2025 02:44:34 +0000
5386 Words

Today’s top highlights include Meta’s $29 billion AI data center funding push, SteamOS outperforming Windows 11 in gaming on Lenovo devices, and US AI data centers threatening grid capacity with projected 70% power consumption by 2035. Additionally, Germany investigates US hyperscalers’ dominance, and Denmark proposes rights over body features to combat deepfakes.

▶️ Internet Infrastructure

SteamOS 3.7 Outperforms Windows 11 in Gaming on Lenovo Legion Go S

Recent Ars testing indicates that SteamOS 3.7 delivers higher frame rates than Windows 11 on the Lenovo Legion Go S, with improvements up to 36%, due to optimized drivers and Linux graphics performance.

  • Ars testing on Lenovo Legion Go S shows games run at higher frame rates on SteamOS 3.7 than on Windows 11.
  • Performance improvements range from small to significant, with up to 36% better frame rates on SteamOS in four of five tested games.
  • Updated drivers from Asus helped bring Windows performance closer to SteamOS, but benchmarks still favored SteamOS, especially at low graphics settings.

Let’s Encrypt Nears Release of IP SAN Certificates with 6-Day Validity

Let’s Encrypt is close to releasing IP address SAN certificates with 6-day validity in a restricted profile, supporting IP and DNS SANs, using the existing trust chain, but not yet publicly available.

  • Let’s Encrypt is preparing to issue IP address SAN certificates in production under the shortlived profile with a 6-day validity, currently allowlist-only.
  • Certificates will include both DNS and IP SANs, with IP SANs formatted as IP.# in OpenSSL, and the feature is not yet publicly available.
  • The certificates are based on Let’s Encrypt’s existing trust chain and will be publicly trusted once launched.

Meta Seeks $29 Billion in Private Credit for AI Data Centers

Meta plans to raise $29bn via private credit to fund AI data centers, with $3bn equity and $26bn debt, to bolster AI capabilities amid competitive delays.

  • Meta seeks $29 billion from private credit firms, including Apollo, Brookfield, and Pimco, to finance AI data centers in the US.
  • The funding plan involves raising $3 billion in equity and $26 billion in debt, with structuring options under discussion for one of the largest private financings.
  • Meta aims to accelerate its AI development amid lagging rivals, investing in data centers, AI models like Llama 4, and acquiring energy resources such as a nuclear plant in Illinois.

Redwood Energy Repurposes EV Batteries for AI Data Center Storage

Redwood Materials introduced Redwood Energy, repurposing used EV batteries into 12 MW/63 MWh energy storage for AI data centers, reducing emissions and costs amid rising power demand.

  • Redwood Materials launched Redwood Energy to repurpose used lithium-ion batteries for AI data center energy storage.
  • Processes over 20 GWh of used batteries annually, representing 250,000 EVs and 90% of North American lithium-ion battery recycling.
  • First deployment: 12 MW power and 63 MWh storage system powering a 2,000-GPU data center for Crusoe, the largest second-life battery project globally.

Microsoft to Discontinue Default Outbound Internet Access for Azure VMs in 2025

Microsoft plans to disable default outbound internet access for Azure VMs in September 2025, requiring organizations to implement explicit network configurations to maintain connectivity and security.

  • Microsoft will retire default outbound internet access for Azure VMs in September 2025.
  • Current defaults allow VMs to access the internet without additional configuration, posing security and visibility risks.
  • Workarounds include assigning fixed public IPs, NAT gateways, Azure Load Balancers, or using Azure Firewall, with cost implications.

Bank of England Boosts Data and Cloud Budget to Support Modernization Efforts

Bank of England expanded its data and cloud framework budget by £26.7 million to £86.7 million, supporting SAP, multi-cloud interoperability, and modernization after revising its data strategy.

  • Bank of England increased support services framework budget from £60 million ($82 million) to £86.7 million ($118 million), a £26.7 million ($36 million) rise in June 2025
  • The expansion supports modernization of SAP (notably Business Warehouse), multi-cloud interoperability, and on-premises and shared data center integration
  • The revised strategy aims to enhance cloud migration, data architecture, and analytics capabilities, following a 2023 on-premises data stack reliance and a 2024 modernization plan

Czech Expert Urges Reducing Microsoft Dependence to Enhance Security and Sovereignty

Czech security expert Miloslav Homer argues that reducing dependence on Microsoft can lower security and strategic risks, supported by data on market share, incidents, and geopolitical factors.

  • Czech researcher Miloslav Homer advocates for reducing reliance on Microsoft to mitigate security and strategic risks, citing incidents like account blocking and US government influence.
  • Homer discusses the financial and operational costs of dependency, including Microsoft 365 usage, Android’s market share, and Google account ties, emphasizing the importance of digital sovereignty.
  • He highlights risks such as potential service refusal, geopolitical influence, and the difficulty of changing entrenched corporate and decision-maker beliefs about Microsoft dominance.

Microsoft Replaces Blue Screen with Black Screen in Windows 11 24H2

Microsoft will replace the traditional blue BSOD with a black screen in Windows 11 24H2, enhancing system resilience and security incident handling through code reengineering and kernel security modifications.

  • Microsoft will replace the Windows Blue Screen of Death (BSOD) with a black screen later this summer as part of the Windows Resiliency Initiative (WRI)
  • The redesign ensures backward compatibility for the BSOD acronym and coincides with code reengineering to improve system security incident management and recovery
  • The new black screen will debut on all Windows 11 version 24H2 devices, following prior black screen references in Windows 3.1 and previous Windows versions

Cisco Patches Critical API Flaws in ISE and ISE-PIC for Remote Code Execution

Cisco issued patches for two critical API vulnerabilities (CVE-2025-20281 and CVE-2025-20282) in ISE and ISE-PIC, enabling unauthenticated remote code execution and root access, affecting versions 3.3 and 3.4.

  • Cisco released patches for two critical vulnerabilities in Identity Services Engine (ISE) and ISE Passive Identity Connector (ISE-PIC)
  • CVE-2025-20281 (severity 9.8) affects ISE and ISE-PIC versions 3.3 and 3.4, allowing unauthenticated remote code execution via API due to insufficient input validation
  • CVE-2025-20282 (severity 10) affects only version 3.4, enabling unauthenticated attackers to upload malicious files and execute code as root due to lack of file validation checks
  • No active exploits are known; patches are available for immediate application, with recommended upgrades to version 3.3 patch 6 or 3.4 patch 2 depending on the CVE

US AI Data Centers Could Consume 70% of Power by 2035 Threatening Grid Capacity

Deloitte warns US AI datacenter energy needs could reach 123 GW by 2035, risking grid capacity shortages due to slow infrastructure growth, supply chain issues, and reliance on gas power.

  • US AI datacenter power demand may increase over 30-fold in a decade, reaching 123 GW by 2035, up from 4 GW in 2024, potentially consuming 70% of total US power (176 GW).
  • Existing US datacenters consume approximately 33 GW, with some new facilities under construction exceeding 2 GW; 50,000-acre campuses could demand up to 5 GW.
  • Power infrastructure development faces a seven-year grid connection delay, supply chain tariffs, and reliance on gas-fired generation, which currently meets most growth needs despite clean energy goals.

AI Boosts Datacenter Networking Costs Amid Growing Demand for Scale-Out Technologies

AI workloads are driving up datacenter networking costs, potentially exceeding 20-30%, as scale-up and scale-out networks like NVLink and InfiniBand expand, prompting AMD and others to develop competing standards.

  • AI workloads are increasing costs of datacenter networking, with masked costs from high-bandwidth memory sharing networks like NVLink and NVSwitch.
  • Networking infrastructure, including scale-up networks (NVLink, NVSwitch), intra-GPU die-to-die, chip-to-chip interconnects, and scale-out networks, now comprise up to 30% of AI system costs.
  • Market data shows datacenter system spending grew by 1.84X from 2020 to 2024, with Nvidia’s InfiniBand revenues increasing 8X, driven by AI scale-out networking demands.

Broadcom Expands AI Infrastructure with Interconnects Switches and Chiplet Tech

Broadcom is quietly expanding into AI infrastructure via licensing interconnect fabrics, high-radix Ethernet switches, co-packaged optics, and multi-die chiplet tech, targeting large-scale AI systems.

  • Broadcom has developed interconnect technologies spanning Ethernet fabrics, chip-to-chip communication, and multi-die packaging for AI infrastructure.
  • The company offers merchant silicon, licensing its IP to clients like Google (TPUs) and Apple for AI server chips.
  • Broadcom’s high-radix switches include Tomahawk 5 (51.2Tbps) and Tomahawk 6 (102.4Tbps), supporting large-scale GPU clusters with fewer switches.

Cisco Promotes Secure, AI-Ready Networks with Embedded Security and New Infrastructure

Cisco advocates integrating security into network hardware, including switches with AMD DPUs, to support AI and agentic AI, requiring datacenter re-architecture and full-stack security deployment.

  • Cisco emphasizes integrating security into network infrastructure, including Catalyst switches, to support AI and agentic AI workloads.
  • Cisco’s approach involves dedicated compute on switches with isolated workloads, utilizing AMD DPU chips for traffic analysis and security enforcement.
  • Cisco’s unified management tool, Cloud Control, was unveiled at Cisco Live, highlighting the need to re-rack datacenters and rebuild networks for AI agent deployment.
  • Cisco develops its own silicon to enable full-stack security and networking, aiming to embed security into the fabric of the data center infrastructure.
  • The integration of security and networking is intended to unify NetOps and SecOps, potentially causing operational disruption but offering enhanced security capabilities.
  • Cisco’s “agentic ops” approach involves using AI for security and traffic analysis, with smart switches featuring embedded security functions that do not impact switch performance.
  • Cisco claims this architecture is applicable across customer segments and represents a new category requiring customer education.
  • The shift toward agentic AI is expected to necessitate significant infrastructure updates in data centers, with some enterprise customers hesitant to adopt early.

Starlink’s IPv6-only broadband helped Tuvalu reach 59% IPv6 adoption, contributing to eight nations surpassing 50%, with France and India at 73%, and Japan, Puerto Rico, Mexico, and Brazil making significant gains.

  • Eight countries—Brazil, Guatemala, Hungary, Japan, Mexico, Puerto Rico, Sri Lanka, and Tuvalu—passed 50% IPv6 adoption since June 2024, according to ISOC
  • Tuvalu’s IPv6 deployment rose from 0% to 59%, driven by Starlink’s IPv6-only broadband service, which holds 88% market share in Tuvalu
  • France and India both reached 73% IPv6 deployment; Japan increased from 49% to 55%; Puerto Rico from 49% to 53%; and Mexico and Brazil also crossed 50%

Ofcom Launches Map Your Mobile for Detailed UK Coverage Insights

Ofcom’s Map Your Mobile tool provides detailed local mobile coverage and performance data, combining crowdsourced and operator data to aid consumer network choices.

  • Ofcom launched Map Your Mobile, an online tool comparing UK mobile coverage and performance by postcode
  • Data sourced from crowdsourced Opensignal user experiences and predictive data from UK mobile operators
  • Coverage data is detailed at 50-meter squares, with ongoing updates to improve accuracy and consumer decision-making

Vintage Networking: From Baud Rates to Ethernet Dominance

Vintage datacenter networking used proprietary protocols, low baud rates (110-1200), and early packet switching, with Ethernet becoming standard in the late 1970s.

  • Networking in the 1960s-1970s relied on baud rates, with early terminals operating at 110 baud and backup dial-up at 600 baud
  • Proprietary network protocols (e.g., IBM Bisync, DECnet, C01) were used over leased lines, often incompatible across vendors
  • The earliest networks utilized packet switching (1965, NPL) and star configurations, with Ethernet emerging in the late 1970s to dominate LANs

▶️ Open Source

Google Launches Open-Source Gemini CLI for Developers with AI-Powered Terminal Tools

Google’s Gemini CLI is an open-source AI agent that embeds Gemini into the terminal, offering powerful, customizable AI tools for developers with extensive free usage limits.

  • Google announced Gemini CLI on June 25, 2025, as an open-source AI agent integrating Gemini into developers’ terminals.
  • Gemini CLI provides AI capabilities for coding, content creation, problem-solving, and task management with a 1 million token context window.
  • Free usage includes 60 model requests per minute and 1,000 requests per day via personal Google accounts; paid options available through Google AI Studio or Vertex AI.
  • Built on open-source Apache 2.0 license, enabling inspection, contribution, and extension by the developer community.
  • Shares technology with Gemini Code Assist, offering AI-driven coding support in VS Code and terminal, with multi-step reasoning and real-time web context grounding.

Creating a Linux USB Driver for Nanoleaf Pegboard Dock in Rust

A step-by-step guide shows how to develop a minimal Linux USB driver for the Nanoleaf Desk Dock using libusb, including device enumeration, interface management, and interrupt handling, enabling LED control and interaction.

  • The article demonstrates creating a basic Linux USB device driver for the Nanoleaf Pegboard Desk Dock using libusb in Rust.
  • It details device identification via lsusb, detaching kernel drivers, claiming interfaces, and sending data through interrupt endpoints.
  • The driver implementation includes polling for device interrupts and controlling device LEDs by writing specific byte sequences.

Microsoft’s Edit Repository Releases Version 1.2.0 with Cross-Platform Features

Microsoft’s edit repository provides a modern, MS-DOS inspired text editor with features like configurable ICU SONAME, Rust-based build system, and cross-platform support, with version 1.2.0 released on June 11, 2025.

  • The repository hosts the “We all edit.” project by Microsoft, with 10.9k stars and 471 forks.
  • The project is licensed under MIT and includes features like code editing, issues, pull requests, discussions, and security.
  • The latest release is version 1.2.0, released on June 11, 2025, with ongoing contributions from 57 developers.

Libxml2 Maintainer Opposes Security Embargoes, Calls for Immediate Disclosure

Libxml2’s 2025 maintainer rejects security embargoes, advocating for immediate public disclosure of vulnerabilities to address sustainability issues and criticize corporate neglect in OSS maintenance.

  • Libxml2’s maintainer in 2025 rejects security embargoes, citing unsustainable security issue management.
  • The project, originally released in 2000 under MIT license, has seen reduced maintenance and delayed releases, with last significant updates in 2021.
  • The maintainer argues that security reports should be public immediately, criticizing corporate reliance on open-source without contributing back.

Snow: Rust-Based Emulator Brings Classic Motorola Macs to Modern Systems

Snow is an open-source, Rust-based Macintosh emulator that accurately emulates Motorola 680x0 Macs (128K to II) with GUI and debugging, available as bleeding edge builds for Windows, macOS, and Linux.

  • Snow is a hardware-level Macintosh emulator written in Rust, supporting Motorola 680x0-based models including Macintosh 128K, 512K, Plus, SE, Classic, and II
  • It features a graphical user interface and extensive debugging capabilities, aiming to emulate the original hardware as closely as possible without ROM patching or system call interception
  • Current development provides bleeding edge builds for Windows 10+, macOS 11.7+, and Ubuntu 24.04 Linux distributions, with download links for each architecture

Ethiopian Developer Raises 5M for Better Auth Open Source Authentication Framework

Ethiopian developer Bereket Engida’s open source Better Auth framework, built in Ethiopia, raised $5M seed funding to simplify user authentication with on-premise data management, attracting AI startups.

  • Ethiopian self-taught developer Bereket Engida raised $5 million in seed funding from Peak XV, YC, P1 Ventures, and Chapter One for Better Auth
  • Better Auth offers an open source, TypeScript-based authentication framework enabling developers to manage user authentication with on-premise data storage
  • The library has over 150,000 weekly downloads, 15,000 GitHub stars, and a community of 6,000+ Discord members; it supports scalable, customizable authentication flows

Researchers Hack Renault CAN Data to Control SuperTuxKart in Real Time

Pen Test Partners converted a Renault Clio’s CAN bus data into a game controller for SuperTuxKart by decoding signals, mapping controls, and re-architecting CAN communication for real-time input.

  • Pen Test Partners hijacked CAN data from a 2016 Renault Clio to control in-game functions in SuperTuxKart
  • Researchers mapped signals for steering, braking, and acceleration by isolating arbitration IDs and decoding CAN packets
  • Modifications involved using Python to update a state machine instead of injecting key presses, enabling real-time control

EU Governments Embrace Linux for Digital Sovereignty and Data Control

EU governments are shifting to Linux desktops for privacy and sovereignty, with initiatives like GendBuntu in France and Denmark’s Microsoft replacement plans, amid distrust of US-based cloud and software providers.

  • European Union governments are increasingly adopting Linux desktops for digital sovereignty, privacy, and control over data, with countries like France, Denmark, and Germany leading.
  • French police use GendBuntu, a customized Ubuntu Linux distribution, on over 97% of workstations; Lyon city is transitioning from Microsoft Office to Linux and PostgreSQL.
  • Denmark’s largest cities, Copenhagen and Aarhus, plan to replace Microsoft software due to concerns over data control; EU proposes a Fedora-based EU OS for official use.

▶️ Management and Leadership

Feynman on Solving Practical, Humble Problems for Scientific Success

Richard Feynman advocates solving accessible, practical problems over grandiose ones, emphasizing personal contribution and success in scientific work, illustrated by his own diverse, humble research efforts.

  • Richard Feynman emphasizes that worthwhile problems are those one can realistically solve or contribute to, regardless of their perceived grandeur.
  • He advises focusing on simpler, humbler problems to achieve success and help others, rather than pursuing overly ambitious or abstract issues.
  • Feynman recounts working on numerous modest problems, such as friction coefficients, crystal elastic properties, neutron diffusion, and turbulence, highlighting that no problem is too trivial if it yields progress.

QEMU Bans AI-Generated Code Contributions Over Licensing Uncertainties

QEMU’s policy now forbids contributions using AI code generators due to licensing uncertainties, with plans to revise as legal and technical issues are clarified.

  • The QEMU project has implemented a policy prohibiting contributions involving AI code generators due to unresolved licensing and legal uncertainties.
  • The policy states that contributions suspected or known to include AI-generated code will be declined, citing concerns over licensing, copyright, and compliance with the Developer’s Certificate of Origin (DCO).
  • The policy is subject to future revision as AI tools mature and legal clarifications are established; exceptions may be considered case-by-case.

Puerto Rico Microgrids Sustain Power During 2025 Outage with Solar and Storage

Puerto Rico’s innovative microgrids, combining solar and storage, maintained power during April 2025 outages, demonstrating resilience amid grid neglect and storm damage, despite federal funding shifts.

  • IEEE Spectrum reports on Puerto Rico’s microgrid resilience during April 2025 blackout, with Adjuntas microgrids remaining powered via solar and storage.
  • Puerto Rico’s aging grid has suffered repeated failures due to neglect, vegetation, and storm damage, with over $20 billion in federal relief efforts hindered by bureaucratic delays.
  • Private solar-plus-storage systems, totaling over 1.14 GW capacity and 2.34 GWh storage, now supply 12.5% of island’s residential electricity, enabling microgrid independence.

Amazon Grocery Leader Slams Bureaucracy Amid Restructuring Efforts

Amazon’s grocery chief Jason Buechel criticized excessive bureaucracy, citing slow decision-making; the company is restructuring, cutting management layers, and streamlining policies to improve efficiency and unify Whole Foods, Amazon Fresh, and Amazon Go.

  • Jason Buechel, Amazon’s VP of Worldwide Grocery and Whole Foods CEO, criticized internal bureaucracy as “ridiculous” during an internal meeting.
  • Amazon’s grocery division recently reorganized leadership and corporate structure to unify teams under the “One Grocery” initiative, aiming to streamline processes.
  • The company is actively reducing management layers, simplifying spending and transaction policies, and addressing overlapping work, including layoffs of at least 125 frontline employees in Fresh grocery.

Denmark Proposes Law Giving Rights Over Body Features to Fight Deepfakes

Denmark will introduce a law granting individuals copyright over their features to combat deepfake misuse, enabling content removal and compensation, with plans to influence broader European regulation.

  • Denmark plans to amend copyright law to grant individuals rights over their body features, voice, and digital imitations, aiming to combat deepfake misuse.
  • The proposed legislation defines deepfakes as highly realistic digital representations of a person’s appearance and voice, allowing for content removal upon consent violation.
  • The law, supported by nine in ten MPs, will enable affected individuals to demand removal from online platforms and seek compensation; it excludes parody and satire.

Supreme Court Limits Nationwide Injunctions, Opening Path for Citizenship Order

US Supreme Court limited federal judges’ power to issue nationwide injunctions, potentially allowing Trump’s order to restrict birthright citizenship to be partially enforced, impacting judicial oversight and constitutional protections.

  • US Supreme Court supported Trump’s effort to limit lower-court nationwide injunctions, ruling they only apply to specific plaintiffs, not entire countries.
  • The 6-3 decision, authored by Justice Amy Coney Barrett, did not address the legality of Trump’s order to end birthright citizenship; its implementation is delayed by 30 days.
  • The ruling could enable partial enforcement of Trump’s 2025 executive order to restrict birthright citizenship, which challenges the 14th Amendment and has been blocked by previous injunctions.

Germany investigates US hyperscalers’ dominance in AI infrastructure and competition risks

Germany’s competition regulator is examining whether US hyperscalers’ control over AI infrastructure and data creates barriers to market entry, amid concerns over monopolistic practices and ecosystem dependencies.

  • Germany’s Bundeskartellamt convened 14 European AI industry representatives to assess competition concerns.
  • US hyperscalers—Amazon Web Services (AWS), Google Cloud, and Microsoft Azure—are identified as gatekeepers due to their control over compute infrastructure and access to large datasets.
  • The discussion highlighted potential barriers to entry, reliance on dominant cloud providers, and risks of monopolistic practices, with ongoing investigations into anti-competitive behaviors.

Kaseya CEO Cites Fragmented Data and Poor Tools as Barriers to AI Adoption

Kaseya’s CEO highlights slow AI adoption due to fragmented data, ineffective tools, and change management issues; emphasizes connecting data and improving software to accelerate enterprise and MSP AI use.

  • Kaseya CEO Rania Succar states AI adoption is slower than expected, with enterprise and SMB uptake being “very nascent”
  • Main obstacles include fragmented business data, poor AI tools, and challenges in change management
  • Academic benchmarks show LLM-based agents perform poorly on CRM tasks; Gartner estimates nearly 30% of proof-of-concept AI projects will be abandoned by end of 2025

CISA and NSA Urge Adoption of Memory-Safe Languages to Reduce Vulnerabilities

CISA and NSA recommend adopting memory-safe languages like Rust and Go to reduce security risks, citing that 70% of serious vulnerabilities in Chromium and 90% in Android stem from memory safety flaws.

  • CISA and NSA issued guidance urging adoption of memory-safe programming languages (MSLs) to mitigate memory vulnerabilities.
  • Languages supporting automated memory management or compile-time checks include Rust, Go, C#, Java, Swift, Python, and JavaScript.
  • The report highlights that C and C++ are not memory-safe by default and importing unsafe libraries can break safety guarantees; Google reduced Android memory vulnerabilities to 24% by 2024.

Nigel’s Experience Managing Office Systems During Hurricane Ivan 2004

Nigel recounted managing critical office systems during Hurricane Ivan in 2004, including system shutdowns and emergency responses, highlighting risks of untested backup power and safety challenges.

  • Nigel, a system administrator, managed office systems during Hurricane Ivan in 2004, with untested diesel generator and UPS systems.
  • The company implemented a plan to conserve generator fuel by shutting down systems at night and restarting in the morning.
  • Nigel’s first night went smoothly until alarms blared early morning; he discovered the emergency battery was faulty, causing false alerts.
  • During shutdown, Nigel encountered police, firefighters, and startled officers questioning his presence during the storm.
  • The alarm system’s emergency battery failure led to false paramedic calls; Nigel restored systems in the morning, and his employer paid a “danger money” bonus.

Palantir and TNC Collaborate on Real-Time Nuclear Operating System

Palantir will embed engineers within TNC to develop NOS, enabling faster, safer nuclear plant construction through real-time data monitoring, amid ongoing nuclear project delays and rising AI power demands.

  • Palantir partnered with The Nuclear Company (TNC) in June 2025 to develop a Nuclear Operating System (NOS) for data-driven nuclear plant construction.
  • The NOS aims to monitor supply chains, project calendars, regulatory compliance, and on-site issues in real time.
  • TNC, founded in 2023, plans to deploy 6 GW of nuclear power across multiple projects, focusing on existing designs and licensed sites to accelerate development and reduce costs.

AI Penetration Tester XBOW Tops US HackerOne Leaderboard with 1,060 Vulnerabilities

XBOW, an AI-powered autonomous penetration tester, achieved top US HackerOne ranking by submitting over 1,060 validated vulnerabilities, including critical flaws like a GlobalProtect VPN bug affecting 2,000+ hosts.

  • XBOW, an autonomous AI-driven penetration tester, reached the top position on the US HackerOne leaderboard on June 24, 2025
  • Conducted over 1,060 fully automated vulnerability submissions, with 130 confirmed by program owners; identified vulnerabilities include RCE, SQLi, XXE, SSRF, XSS, and more
  • Discovered a previously unknown vulnerability in Palo Alto’s GlobalProtect VPN affecting over 2,000 hosts; reported findings are being published in upcoming technical blog posts

▶️ Technology

GPU Memory Bandwidth Limits Performance; Tiling and Fusion Boost AI Efficiency

GPU performance is limited by memory bandwidth and compute throughput; optimizing matrix multiplication with tiling and fusion increases arithmetic intensity beyond 13 FLOPs/Byte, approaching peak compute.

  • GPU performance hierarchy: NVIDIA A100 achieves 19.5 TFLOPS compute but only 1.5 TB/s memory bandwidth, creating an imbalance.
  • Performance regimes are memory-bound (AI < 13 FLOPs/Byte) and compute-bound (AI > 13 FLOPs/Byte); optimization aims to increase AI.
  • Matrix multiplication AI depends on data reuse strategies like tiling and operator fusion; naive models yield AI far below the ridge point.

Google Launches Gemma 3n, a Multimodal AI Model for Edge Devices

Google announced the full release of Gemma 3n, a mobile-first, multimodal AI model with 5B/8B parameters, innovative architecture, and optimized for on-device use, supporting multiple input types and achieving high benchmark scores.

  • Gemma 3n, fully released on June 26, 2025, builds on the over 160 million downloads of the original Gemma model and supports multimodal inputs.
  • Features include native multimodal support (image, audio, video, text), optimized sizes (E2B and E4B) with 5B and 8B parameters respectively, operating with as little as 2GB and 3GB memory.
  • Incorporates innovative architecture such as MatFormer for elastic inference, Per Layer Embeddings (PLE) for memory efficiency, KV Cache Sharing for faster long-context processing, advanced audio encoder based on USM, and MobileNet-V5-300M vision encoder, achieving high performance on edge devices.

Speeding Up Audio with ffmpeg Cuts Transcription Costs by Up to 33%

Speeding audio with ffmpeg at 2x or 3x before OpenAI transcription reduces token costs by up to 33%, significantly decreasing time and expenses while maintaining acceptable accuracy.

  • Speeding up audio with ffmpeg at 2x or 3x reduces transcription costs and time with minimal quality loss
  • OpenAI charges for transcription based on audio duration (whisper-1) or tokens (gpt-4o-transcribe)
  • 3x speed can decrease input tokens by approximately 33%, lowering costs from $0.09 to $0.07 for a 40-minute file

Build Static Websites Using XML and XSLT for Browser-Based Transformation

The article promotes using XML and XSLT as a simple, client-side build system for static websites, transforming XML data into styled HTML directly in web browsers without frameworks.

  • The article advocates using XML and XSLT as a native, client-side build system for static websites, eliminating complex frameworks.
  • Demonstrates how browsers natively support XSLT transformations to convert XML data into styled HTML, enabling dynamic, static site generation without JavaScript.
  • Highlights that XML with XSLT can serve as a flexible, human-readable data format and build process, leveraging web standards and browser capabilities for static site rendering.

Rethinking Tokenization: BLT and New Architectures for Language Models

The article analyzes the limitations of tokenization in LLMs, explores architectures like BLT that learn or eliminate tokenization, and demonstrates BLT’s superior scaling and downstream results, especially on character-level tasks.

  • The article discusses the impending shift away from tokenization in language models, highlighting its fragility and limitations.
  • It reviews recent architectures like Byte Latent Transformer (BLT) and related models (CANINE, Charformer, Hourglass, Megabyte) that aim to learn or bypass tokenization.
  • The Byte Latent Transformer (BLT) employs a Patcher for dynamic byte patching, local encoder, global transformer, and local decoder, achieving improved scaling curves and downstream performance, especially on character-level tasks.

Anthropic Launches In-App AI App Development and Sharing with Claude

Anthropic launched in-application development and hosting of AI-powered apps with Claude, enabling API-based artifacts, code editing, and instant sharing, with usage billed to end-users and no deployment needed.

  • Anthropic introduced the ability to build, host, and share interactive AI-powered apps directly within the Claude app on June 25, 2025
  • Developers can create artifacts that interact with Claude via API, with user API usage billed to their own subscriptions and no cost to developers
  • Features include code generation, debugging, customization, and instant sharing without deployment; limitations include no external API calls, no persistent storage, and text-based API only
  • Available in beta for Free, Pro, and Max plan users

Microsoft’s Copilot Fails to Compete as Amgen Switches to ChatGPT

Microsoft struggles to sell Copilot as OpenAI’s ChatGPT gains enterprise traction, with Amgen switching from Copilot to ChatGPT within 13 months, highlighting ChatGPT’s growing market dominance.

  • Microsoft’s Copilot AI assistant, initially adopted by Amgen for 20,000 employees, faces competition from OpenAI’s ChatGPT in enterprise use
  • Thirteen months after Microsoft’s endorsement, Amgen employees predominantly use ChatGPT instead of Copilot
  • OpenAI’s ChatGPT is gaining enterprise market strength, causing concern for Microsoft, its partner and largest investor

Microsoft to Assess AI Usage in Employee Performance Evaluations

Microsoft is integrating AI usage into employee performance evaluations, with plans to include it as a formal metric, amid rising competition for GitHub Copilot from Cursor and external AI tools.

  • Microsoft is evaluating employees on AI usage, with some managers considering formal metrics for performance reviews.
  • Developer Division President Julia Liuson stated, “AI is no longer optional,” emphasizing its integration into all roles.
  • GitHub Copilot faces increasing competition from services like Cursor, which recently surpassed Copilot in key developer market segments; external AI tools meeting security standards are permitted.

MSI Launches AI-Optimized Hardware to Support Growing SMB AI Adoption

MSI’s AI-optimized hardware, including the Cubi NUC AI+ 2M, supports small businesses adopting AI, with 98% using AI software and 40% employing generative tools, driving productivity and growth.

  • 98% of small businesses now use AI-enabled software; 40% employ generative tools like chatbots and image creators to reduce costs and accelerate workflows.
  • MSI’s Cubi NUC AI+ 2M features Intel® Lunar Lake processors with on-chip NPU and GPU cores for efficient on-device AI computing; includes Microsoft Copilot+.
  • 27% of SMBs have significantly accelerated tech spending due to AI needs; MSI’s new desktops come with Windows 11 Pro pre-installed, supporting smoother upgrades from Windows 10.

Meta doubles AI investment to $14.3B in race for superintelligence

Mark Zuckerberg ramped up Meta’s AI spending, investing $14.3 billion, hiring top researchers, and shifting strategies to develop superintelligence amid competitive pressures from OpenAI, Google, and others.

  • Mark Zuckerberg increased Meta’s AI investment following underperformance at April’s AI conference, with new models failing to meet rivals’ standards.
  • Meta invested $14.3 billion in start-up Scale AI, hired Alexandr Wang, and engaged in aggressive talent recruitment, including over 45 researchers from OpenAI.
  • Discussions included “de-investing” in open-source model Llama, considering proprietary models from competitors, and aiming to develop “superintelligence” surpassing human brain capabilities.

MCP Protocol Boosts LLM Agent Development and Ecosystem Growth

MCP is a vendor-neutral protocol for building agents with LLMs, supported by major providers like OpenAI and Deepmind, enabling faster development, broader adoption, and ecosystem growth.

  • MCP (Model Context Protocol) gained popularity in February 2024, despite its November 2023 release by Anthropic
  • It provides a vendor-neutral, shared protocol for defining and accessing tools across LLM platforms
  • MCP tooling includes SDKs in multiple languages, with examples like the Python SDK, enabling rapid tool development and deployment

Google Influences OpenAI to Switch to TPU Chips, Challenging Nvidia Dominance

Google convinced OpenAI to use TPU v4 chips instead of Nvidia GPUs, boosting training efficiency and reducing costs, challenging Nvidia’s dominance in AI hardware.

  • Google persuaded OpenAI to adopt TPU chips over Nvidia GPUs to gain a competitive edge in AI training
  • OpenAI’s switch to TPU v4 chips enhances training efficiency and cost-effectiveness
  • This strategic move aims to reduce reliance on Nvidia hardware, impacting the AI chip market dynamics

Customer Skepticism Grows Amid Enthusiasm for Agentic AI at HPE Discover 2025

HPE’s Discover 2025 revealed customer hesitation toward agentic AI due to privacy, readiness, and risk concerns, despite industry enthusiasm and partial product maturity.

  • HPE’s Discover 2025 showcased skepticism among customers toward adopting agentic AI, with most preferring to wait for full functionality
  • Agentic AI, capable of autonomous workflow actions without human input, is considered the “Next Big Thing” in AI but faces industry concerns
  • Customers cite data privacy, incomplete solutions, and risk aversion as barriers; Aruba networking AI agents are more mature and widely adopted

Powell Warns AI Will Transform Economy and Labor Market Amid Uncertain Timeline

Fed chair Powell predicts AI will cause major economic and labor market shifts, but timeline and impact remain uncertain; the Fed lacks tools to manage social consequences.

  • Fed chair Jerome Powell stated AI will “significantly change” the US economy and labor market, though timing remains uncertain.
  • Powell highlighted that AI’s current economic impact is minimal but has “enormous capabilities” for future transformative effects.
  • He noted that technological implementation phases tend to be slow, with the final leaps taking longer than expected, and emphasized the Fed’s limited tools to address AI-driven social and labor issues.