Alvaro Lopez Ortega / 2025-07-15 Briefing

Created Tue, 15 Jul 2025 22:17:33 +0000 Modified Sun, 31 Aug 2025 02:01:52 +0000
5946 Words

Today’s top news highlights include the patching of a Chrome bug that bypassed adblock restrictions, Apple’s $500 million investment in US rare-earth manufacturing, Nvidia and AMD resuming AI chip sales to China, Meta’s rapid expansion of AI data centers, and the launch of xAI’s powerful Grok 4 model. Additionally, significant developments in open source AI, security, and industry regulation shaped the tech landscape.

▶️ Internet Infrastructure

Chrome 2023 Bug Allowed MV3 Adblock Bypass Before Being Patched

A Chrome 2023 bug enabled MV3 adblocker functionality by exploiting WebRequestEvent constructor and opt_webViewInstanceId, bypassing webRequestBlocking restrictions; patched in Chrome 118.

  • Discovered a bug in Chrome 2023 that allowed webRequestBlocking in MV3 extensions despite API restrictions
  • The bug exploited the WebRequestEvent constructor and the opt_webViewInstanceId parameter to spoof WebView events
  • Chrome patched the vulnerability in version 118 by verifying opt_webViewInstanceId permissions, nullifying the bypass

Pangolin: Secure Self-Hosted Reverse Proxy with WireGuard and IAM

Pangolin is a self-hosted reverse proxy with WireGuard tunneling, centralized IAM, and a dashboard, enabling secure, portless resource exposure and management across distributed networks.

  • Pangolin is a self-hosted tunneled reverse proxy server with identity and access control, licensed under AGPL-3.0.
  • Features include site-to-site reverse proxy via WireGuard, SSL via LetsEncrypt, role-based access, external IdP support, and a dashboard UI.
  • Deployment is simplified with Docker Compose, supporting cloud or on-premises setups, with modular extension options including Traefik plugins.

DOGE Employee Leaks xAI API Key Exposing 52 Large Language Models

DOGE employee Marko Elez leaked a private xAI API key exposing 52 LLMs, including models used by Grok, amid concerns over security negligence and potential data exfiltration.

  • Marko Elez, a 25-year-old DOGE employee, leaked a private API key for xAI on GitHub on July 13, 2025
  • The exposed key provided access to at least 52 large language models (LLMs), including “grok-4-0709” created on July 9, 2025
  • The key remains active despite being publicly available; the code was removed after notification but not revoked

Apple invests $500M in US rare-earth minerals and Texas factory expansion

Apple will buy rare-earth minerals from MP Materials in a $500 million deal, including factory construction in Texas, supporting its $500 billion US investment plan over four years.

  • Apple to invest $500 million to purchase rare-earth minerals from MP Materials Corp., a Pentagon-backed US producer
  • The companies will build a factory in Texas with neodymium magnet manufacturing lines tailored for Apple products
  • The investment is part of Apple’s pledge to spend over $500 billion in the US over the next four years

Nvidia and AMD Resume AI Chip Sales to China After US Approval

Nvidia and AMD will restart AI chip sales to China after US approval, enabling Nvidia to fulfill previously restricted orders and potentially increase revenue by billions.

  • Nvidia and AMD plan to resume AI chip exports to China after securing US government approval, reversing prior restrictions.
  • US officials indicated export licenses for Nvidia’s H20 AI accelerator would be approved, potentially adding billions in revenue.
  • The H20 chip was designed to comply with earlier US trade curbs, which in April restricted H20 sales to China without a US permit.

Meta Accelerates AI Data Center Growth with Gigawatt-Scale Projects

Meta is rapidly expanding AI infrastructure by building massive data centers, some in tents, to achieve over 1GW capacity per site, with plans for multi-GW clusters like Prometheus and Hyperion by 2026.

  • Meta plans to build several new AI data centers, each exceeding 1 gigawatt of power capacity.
  • Some data centers are being constructed in tents to accelerate deployment amid capacity and construction constraints.
  • CEO Mark Zuckerberg announced the development of multiple superclusters, including Prometheus (coming 2026) and Hyperion (up to 5GW), with total capacity covering significant parts of Manhattan.

Microsoft’s Escort Program Risks Sensitive Data Exposure to Chinese Hackers

Microsoft’s digital escort program, involving foreign engineers and low-skilled U.S. supervisors, poses security risks by exposing sensitive Defense Department data to Chinese hackers, despite internal warnings.

  • Microsoft uses engineers in China and elsewhere to remotely instruct U.S. government “escorts” handling sensitive data, bypassing Pentagon restrictions on foreign access.
  • Escorts often lack advanced technical skills, with some former military personnel earning barely above minimum wage, raising security concerns.
  • Microsoft disclosed the escort model to the federal government, but officials and cybersecurity experts were largely unaware of its existence, highlighting transparency issues.

Postgres LISTEN/NOTIFY Lock Causes Scalability Issues Under High Concurrency

Postgres’s LISTEN/NOTIFY causes a global lock during commit, severely limiting scalability under high concurrency, leading to database stalls; migrating notification logic to application layer resolves this.

  • LISTEN/NOTIFY in Postgres acquires a global lock on “database 0” during transaction commit, serializing all commits and causing lock contention under high concurrency
  • This lock is implemented via LockSharedObject(DatabaseRelationId, InvalidOid, 0, AccessExclusiveLock) in async.c, affecting all databases within the Postgres instance
  • Heavy multi-writer workloads with frequent NOTIFY lead to database stalls, CPU and I/O plummets, and increased lock contention, as confirmed by load tests and logs

Microsoft Outlook Outage Causes 11-Hour Global Disruption

Microsoft Outlook outage lasted over 11 hours due to mailbox infrastructure issues related to authentication, impacting millions globally; a fix is being deployed with gradual recovery.

  • Microsoft Outlook experienced a major outage lasting over 11 hours on July 9-10, affecting millions of users worldwide.
  • The outage began at 2220 UTC on July 9, with issues in mailbox infrastructure linked to authentication components.
  • Microsoft confirmed the cause and initiated deployment of a fix, expecting impact to gradually mitigate, with service restored for most users by 1921 GMT.

Red Hat Launches Free RHEL for Developers with 25 Non-Production Instances

Red Hat launched RHEL for Business Developers, providing up to 25 free non-production instances for development/testing, excluding enterprise features like Insights and Satellite.

  • Red Hat introduced RHEL for Business Developers, offering up to 25 free instances for development or testing only, with no production use
  • The new scheme is similar to the 2021 free tier, which allowed up to 16 instances and permitted production use; the current version restricts to non-production
  • The free tier excludes Red Hat Insights, Software Collections, Developer Toolset, and Satellite fleet management; Podman Desktop remains free and supports Windows, macOS, and Linux

US PC Shipments Flat as Demand Slows and Imports Shift to Vietnam

US PC shipments in Q2 2025 were flat amid tariff uncertainty; global demand surged 6.5-7.4%, with US imports moving from China to Vietnam, facing complex regulatory and tariff challenges.

  • US PC shipments remained flat in Q2 2025, with 67.6 million units according to Canalys and 68.4 million according to IDC, up 7.4% and 6.5% respectively.
  • US market shows early signs of demand slowdown due to import tariff uncertainty, despite global growth driven by Windows 11 migration.
  • US imports shifted away from China toward Vietnam to avoid tariffs; Vietnam tariffs include 20% on goods and 40% on transshipped items, with enforcement criteria still undefined.

Microsoft Proposes New Deal to Address CISPE Antitrust Concerns

Microsoft offered CISPE a new commercial agreement to mitigate antitrust risks, including SPLA fee reductions and no customer list sharing, after previous failed Azure Stack HCI developments.

  • Microsoft proposed new commercial terms to CISPE, a trade body representing over 30 European cloud providers, to address antitrust concerns.
  • The agreement, which is financial rather than technical, aims to prevent litigation and includes a potential fee reduction for Service Provider License Agreement (SPLA) costs.
  • Microsoft will no longer require service providers to share customer lists, and CISPE is reviewing the proposal, with a decision expected before August.

Britain and France Develop Backup Navigation Systems to Counter GPS Jamming

Britain and France are jointly developing backup navigation and timing systems, including eLoran, to enhance resilience against GPS jamming and spoofing, amid increasing satellite signal interference globally.

  • Britain and France will collaborate on resilient navigation and timing technologies to counter GPS signal jamming, especially in conflict zones like Ukraine.
  • The initiative includes developing terrestrial-based systems such as eLoran, operating within 90-110 kHz low frequency band, more resistant to interference.
  • UK has issued a tender for a national eLoran system, and researchers will focus on PNT technologies, including low-frequency ground-based radio towers, to safeguard critical infrastructure.

Iran ITOI targets three cloud providers meeting NIST and ISO standards

Iran’s ITOI plans to select at least three cloud providers meeting NIST and ISO standards to host government services, despite geopolitical tensions with the USA.

  • Iran’s ITOI seeks at least three cloud providers to host government services, evaluating vendors for suitability.
  • Evaluation criteria include compliance with ISO 27017, ISO 27018, and NIST SP 800-145 standards.
  • The initiative aims to establish a panel of cloud operators capable of delivering IaaS, PaaS, or SaaS, including private, public, hybrid, or community clouds.

Nvidia A6000 GPUs Vulnerable to Rowhammer Attacks Disrupting AI Accuracy

Researchers disclosed that Nvidia A6000 GPUs are susceptible to Rowhammer attacks causing memory bit-flips, which can significantly impair AI model accuracy, with mitigation requiring ECC activation.

  • Nvidia A6000 GPUs are vulnerable to Rowhammer-induced bit-flips in GDDR6 memory, despite Target Row Refresh defenses
  • Researchers demonstrated that GPUHammer can degrade deep neural network (DNN) accuracy by up to 80%, impacting AI inference reliability
  • Mitigation involves enabling Error Correction Codes (ECC) via nvidia-smi -e 1, incurring ~10% performance loss and 6.25% reduction in memory capacity

Apple invests $500M in US rare earth magnet recycling to cut Chinese reliance

Apple invests $500 million in US rare earth magnet recycling and manufacturing through a partnership with MP Materials, targeting 2027 for US-made recycled magnets to reduce Chinese dependency.

  • Apple commits $500 million to develop US-based rare earth magnet supply chain, including recycling facilities, via a multi-year deal with MP Materials announced in July 2025
  • The partnership aims to produce 100% recycled rare earth magnets in the US by 2027, supporting Apple’s US manufacturing and reducing reliance on Chinese supply chains
  • The recycling line will process magnet scrap and end-of-life electronics at MP’s Mountain Pass site in California, with custom neodymium magnet manufacturing lines in Fort Worth, Texas

Broadcom Launches 51.2 Tbps Tomahawk Ultra Switch ASIC for High-Performance HPC

Broadcom’s Tomahawk Ultra 51.2 Tbps Ethernet switch ASIC offers low latency and high message rates, challenging Nvidia’s NVLink and UALink for scalable HPC and AI infrastructure.

  • Broadcom introduces Tomahawk Ultra, a 51.2 Tbps Ethernet switch ASIC designed for high-performance HPC and rack-scale systems, supporting at least 128 accelerators.
  • Tomahawk Ultra features 512 x 100 Gbps SerDes, latency as low as 250 ns, and supports 77 billion 64-byte packets per second, with optimized Ethernet headers for small packets.
  • Broadcom dismisses UALink as unnecessary, claiming Ethernet’s monitoring, telemetry, and debugging advantages, and supports its own scale-up Ethernet (SUE) technology for systems with 1,024 accelerators.

Meta Announces Massive Multi-GW Data Centers to Power AI and Superintelligence

Meta will build multiple multi-GW datacenter clusters, starting with Prometheus in 2026 (1GW) and Hyperion up to 5GW, covering most of Manhattan, to support superintelligence and AI efforts.

  • Meta plans to build several multi-gigawatt datacenter clusters, with the first (Prometheus) coming online in 2026 and capable of 1GW AI training.
  • The Hyperion cluster will scale up to 5GW over several years; one cluster’s footprint covers approximately 80% of Manhattan (~6.5 miles).
  • Mark Zuckerberg announced these developments on July 11, 2025, emphasizing significant investment into compute infrastructure and superintelligence research.

Seagate Launches 30TB HAMR Drives to Power AI Storage and Data Centers

Seagate launched 30 TB HAMR-based drives (Exos M and IronWolf Pro) to meet AI storage needs, leveraging laser heating for higher density, with prices around $670-$753, supporting large-scale data centers.

  • Seagate announced the global availability of Exos M 30 TB and IronWolf Pro 30 TB drives based on HAMR technology.
  • These drives use heat-assisted magnetic recording (HAMR) with laser heating to enable higher data density, reaching capacities close to rival products (Western Digital 32 TB, Toshiba, and Seagate’s sampled 36 TB).
  • Built on Mozaic 3+ tech, they target datacenter and NAS markets, addressing AI workload storage demands and supporting a roadmap to 100 TB.
  • The drives are priced at £498.99 ($670) for Exos M 30 TB and £559.99 ($753) for IronWolf Pro 30 TB, available through Seagate’s channels.
  • Seagate emphasizes that spinning disks remain relevant for cost-effective large-volume storage in AI development and edge computing, countering claims that flash SSDs will soon dominate storage costs.
  • IDC’s Ed Burns highlights high-capacity HDDs as strategic assets for AI data foundation, despite performance limitations.
  • The release follows a trend of rising HDD demand driven by AI, with Seagate previously attributing increased prices to AI-driven demand and supply cuts post-pandemic.

Nvidia Gains US Approval to Resume High-Performance GPU Sales to China

Nvidia received US government approval to resume GPU exports to China, including the H20 and RTX PRO series with up to 4 petaFLOPs performance, after previous bans cost $10 billion.

  • Nvidia announced US government approval to resume GPU sales to China, reversing previous export bans.
  • The US prohibited advanced semiconductor exports to China citing military and surveillance concerns; Nvidia previously created the H20 GPU to bypass restrictions.
  • Nvidia plans to reintroduce the H20 GPU and a new RTX PRO GPU, with the latter potentially based on RTX Pro 6000-series chips boasting up to 4 petaFLOPs sparse performance at 4-bit precision, 96GB GDDR7, and 1.6TB/s bandwidth; Nvidia aims to start deliveries soon after license approval.

▶️ Open Source

LisaGUI: No-JavaScript Web Interface Now Available

LisaGUI offers a JavaScript-free web interface accessible at https://alpha.lisagui.com/, with further details available on its info page.

ETH Zurich and EPFL Launch Open Multilingual LLM in 2025

ETH Zurich and EPFL will release a fully open, multilingual LLM in late summer 2025, trained on the “Alps” supercomputer, supporting over 1,000 languages, with 8B and 70B parameters, promoting transparency and innovation.

  • ETH Zurich and EPFL will release a fully open large language model (LLM) in late summer 2025, trained on the “Alps” supercomputer at CSCS.
  • The LLM supports over 1,000 languages, trained on a dataset in over 1,500 languages, including code and mathematics data.
  • The model will be available in two sizes: 8 billion and 70 billion parameters, trained on over 15 trillion high-quality tokens, with source code, weights, and training data transparent and open.

How Apple decrypts and manages safety filter files for AI models

Decrypted Apple Intelligence safety files contain JSON filters for model output, with combined metadata files created via combine_metadata.py; decryption involves LLDB attachment and decrypt_overrides.py.

  • Decrypted safety filter files for Apple Intelligence models are stored in the decrypted_overrides/ directory, containing JSON rules for model output filtering.
  • Metadata files, including global_metadata.json, region-specific, and locale-specific filters, are combined and deduplicated using combine_metadata.py.
  • The decryption process requires attaching LLDB to GenerativeExperiencesSafetyInferenceProvider and running get_key_lldb.py to extract the encryption key, then decrypting overrides with decrypt_overrides.py.

OpenCut: The Free Open-Source Video Editor Rivaling CapCut

OpenCut is a free, open-source video editor offering CapCut-like features across web, desktop, and mobile, emphasizing privacy, no watermarks, and active development since at least 2025.

  • OpenCut is an open-source, free video editor designed as a CapCut alternative, supporting web, desktop, and mobile platforms.
  • Features include timeline-based editing, multi-track support, real-time preview, no watermarks, and no subscriptions.
  • The project is licensed under MIT, with over 23,200 stars and 1,900 forks on GitHub, and ongoing development with 626 commits as of July 15, 2025.

Bitchat: Decentralized Bluetooth Mesh Chat with End-to-End Encryption

Bitchat is a peer-to-peer Bluetooth mesh chat app utilizing Noise Protocol for secure, decentralized messaging with features like multi-hop relay, encrypted channels, and privacy enhancements, supporting iOS and macOS.

  • Bitchat is a decentralized Bluetooth mesh chat app with IRC-like commands, supporting iOS and macOS, released as public domain software.
  • Implements end-to-end encryption using Noise Protocol, X25519, AES-256-GCM, and Ed25519 signatures; features include ephemeral peer IDs, cover traffic, and offline message caching.
  • Core features include multi-hop relay, channel-based messaging with optional passwords, message retention, privacy protections, message compression (LZ4), and battery/network optimizations.

AI Assistance Slows Experienced Developers by 19% in Real-World Tasks

Early-2025 AI tools slowed experienced open-source developers by 19% in real-world tasks, challenging benchmark and anecdotal evidence of AI aiding productivity; methodology involved randomized trials on 246 issues.

  • A randomized controlled trial (RCT) with 16 experienced open-source developers found that AI assistance increased task completion time by 19%, contrary to expectations of speedup.
  • Developers self-reported a 20% perceived speedup from AI, but actual implementation times showed a slowdown, with AI tools causing longer durations.
  • The study used real issues from repositories averaging 22k+ stars and 1 million+ lines of code, with developers working on tasks averaging two hours each, utilizing tools like Cursor Pro with Claude 3.5/3.7 models.

Judge Rules Anthropic’s Fair Use of Purchased Books but Illegally Downloaded 7 Million Pirated Titles

Judge William Alsup ruled that Anthropic’s use of purchased books for training Claude is fair use, but downloading over 7 million pirated books was unlawful.

  • Anthropic spent “many millions of dollars” purchasing used print books, then stripped, cut, and scanned them into digital files for training Claude.
  • The company downloaded over 7 million pirated books from Library Genesis and Pirate Library Mirror, fully aware of their pirated status.
  • Judge William Alsup ruled that digitizing purchased books for research qualifies as fair use, but using pirated books does not.

Thinking Machines Secures $2B Funding in $12B Valuation Led by Andreessen Horowitz

Thinking Machines, founded by Mira Murati, secured $2 billion in early-stage funding led by Andreessen Horowitz, valuing it at $12 billion, with plans to launch a product emphasizing safety and broad AI applications.

  • Thinking Machines Lab, founded by ex-OpenAI executive Mira Murati, raised approximately $2 billion at a $12 billion valuation in a funding round led by Andreessen Horowitz on July 15, 2025
  • Participation included Nvidia, Accel, ServiceNow, Cisco, AMD, and Jane Street; the company has no revenue or products yet
  • CEO Murati announced the upcoming release of their first product, featuring significant open source components aimed at researchers and startups developing custom models

Mistral Launches Voxtral Open-Source ASR with Better Accuracy and Lower Costs

Mistral launched Voxtral, an open-source ASR model offering state-of-the-art accuracy, multilingual support, and voice function-calling, at a lower cost than competitors, with API prices from $0.001 to $0.004 per minute.

  • Mistral released Voxtral, an open-source automatic speech recognition (ASR) model licensed under Apache, aiming to reduce costs and improve accuracy.
  • Voxtral supports input of up to 32,000 tokens (~30 minutes audio), with multilingual detection and voice function-calling.
  • The API pricing ranges from $0.001 to $0.004 per minute, claiming better word error rate than OpenAI’s Whisper and GPT-4o-mini-transcribe, outperforming Whisper large-v3 and surpassing models like ElevenLabs Scribe.

Most Oracle Java Users Consider Switching Due to Cost and Audits

A survey found 73% of Oracle Java users were audited in three years; 80% plan to switch to open source Java due to licensing costs and Oracle’s pricing changes since 2018.

  • 73% of 500 IT asset managers using Oracle Java reported being audited in the past three years
  • Nearly 80% of Oracle Java users plan to migrate or have migrated to open-source Java to avoid high costs and licensing risks
  • Oracle introduced four licensing and pricing policy changes between 2020 and 2023, including a shift to per-employee pricing in January 2023, causing price increases of 2-5 times for some users

Google Hires Windsurf Team to Accelerate Gemini and Agentic Coding

OpenAI’s Windsurf acquisition is canceled; Google will hire Windsurf’s CEO and team to advance Gemini and agentic coding at DeepMind, with no control or stake in Windsurf.

  • OpenAI’s $3 billion Windsurf acquisition deal is canceled; Google will hire Windsurf’s CEO Varun Mohan, cofounder Douglas Chen, and R&D employees.
  • The team will focus on agentic coding efforts at Google DeepMind, primarily working on the Gemini model; Google gains a non-exclusive license to Windsurf’s technology.
  • Windsurf’s leadership, including Mohan and Chen, will join Google DeepMind; Jeff Wang becomes interim CEO, and Graham Moreno is appointed president.

▶️ Software Development

Kiro: AI-Powered IDE Boosts Development with Specs, Automation, and Quality Tools

Kiro is an AI IDE that streamlines development with specs and hooks, enabling requirement analysis, system design, task automation, and code quality enforcement for efficient production deployment.

  • Kiro is an AI-powered IDE designed for spec-driven development, supporting from prototype to production.
  • Features include specs for feature planning and system understanding, and hooks for event-driven automation and code quality enforcement.
  • Kiro automates requirement unpacking, technical design generation, task sequencing, and integrates with existing tools, supporting Mac, Windows, and Linux.

▶️ Management and Leadership

US Appeals Court Blocks FTC’s Click-to-Cancel Rule Over Procedural Violations

A US appeals court nullified the FTC’s July 14 “click-to-cancel” rule for procedural violations, citing failure to perform required preliminary analysis despite estimated impacts over $100 million.

  • A US federal appeals court vacated the FTC’s “click-to-cancel” rule due to procedural violations, specifically failure to conduct a required preliminary regulatory analysis.
  • The rule, scheduled to take effect on July 14, aimed to simplify subscription cancellations and prevent unfair practices, but was found to lack compliance with the Administrative Procedure Act.
  • The court emphasized that the FTC’s omission of a preliminary analysis, despite estimating the rule’s economic impact exceeded $100 million, violated statutory requirements, with industry groups challenging the rule’s legality.

Cognition Acquires Windsurf to Boost Software Engineering Innovation

Cognition announced the acquisition of Windsurf on July 14, 2025, integrating its IP, product, and a $82M ARR business, to accelerate development of Devin and the future of software engineering.

  • Cognition signed a definitive agreement to acquire Windsurf, an agentic IDE, on July 14, 2025
  • Acquisition includes Windsurf’s IP, product, trademark, brand, and a business with $82M ARR, doubling enterprise ARR quarter-over-quarter
  • Windsurf’s user base comprises 350+ enterprise customers and hundreds of thousands of daily active users; the IDE now has full access to the latest Claude models
  • All Windsurf employees will participate financially, with waived vesting cliffs and fully accelerated vesting for their work to date

Apple’s EU Browser Engine Ban Violates DMA Rules, Threatening Fair Competition

Apple’s ongoing browser engine ban in the EU violates DMA obligations, hindering fair competition, web app interoperability, and risking billions in lost revenue for Apple and developers.

  • Apple’s restrictions prevent third-party browser engines from shipping on iOS in the EU, despite DMA requirements.
  • Apple knows these barriers exist but refuses to remove them, maintaining a de facto ban.
  • The primary barriers include forcing vendors to abandon existing EU users, testing limitations, and restrictive contractual terms.

Anthropic Launches Claude for Financial Services Amid Rapid Revenue Growth

Anthropic introduced “Claude for Financial Services,” an AI suite for financial analysis, integrating third-party data, aiming to expand its finance sector presence amid rapid revenue growth.

  • Anthropic launched “Claude for Financial Services” to assist analysts with market research, due diligence, and investment decisions, integrating third-party data providers like FactSet, PitchBook, and Morningstar.
  • The new AI tools focus on enhancing financial analysis tasks such as investment memos, portfolio analysis, and financial modeling, leveraging Anthropic’s coding and enterprise AI capabilities.
  • Anthropic’s annualized revenue increased from $3 billion to $4 billion in the past month; the company has hired Paul Smith as its first chief commercial officer to expand sales in finance.

AI Boosts Last-Mile Delivery Efficiency and Security

AI improves last-mile logistics through real-time route optimization, error prediction, and theft prevention, addressing the 41% of logistics costs attributed to last-mile delivery.

  • AI supports last-mile delivery by optimizing truck routes and predicting errors before they occur.
  • AI-enhanced predictive analytics help prevent package theft by identifying high-risk areas and times.
  • UPS’s DeliveryDefense analyzes historic data to spot potential porch pirate hotspots, improving security.

SEBI bans Jane Street, freezes $566M over index manipulation in India

SEBI temporarily barred Jane Street from India’s securities market, froze $566 million, citing index manipulation through large trades in BANKNIFTY and Nifty 50, with ongoing allegations since July 2023.

  • SEBI issued an interim order to freeze over $566 million (48.4 billion INR) and barred Jane Street from accessing India’s securities market.
  • The regulator accused Jane Street of using strategies to artificially influence the Nifty 50 index and profit from large index options positions.
  • Jane Street allegedly bought large amounts of stocks and futures in the BANKNIFTY index early in trading, then bet on decline, manipulating index movements without regulatory breach but with manipulative intent.

GLP-1 Drugs Alter Mortality Risks, Challenging Insurance Underwriting

GLP-1s rapidly improve health metrics, causing insurers to underestimate mortality risk; discontinuation leads to risk profile reversal, increasing “mortality slippage” and mispricing policies, prompting insurers to adapt underwriting practices.

  • Life insurers predict mortality with approximately 98% accuracy using decades of mortality data.
  • GLP-1 medications significantly improve key health metrics (BMI, blood pressure, blood sugar, cholesterol) within 6 months, altering risk profiles.
  • Weight regain after discontinuing GLP-1s over two years causes risk profiles to revert, leading to “mortality slippage” and mispriced policies; slippage increased from 5.8% in 2019 to 15.3% in 2024.

Jeff Geerling Upgrades Mac mini Storage to 4TB with M4-SSD and DFU Restore

Jeff Geerling demonstrates upgrading an M4 Pro Mac mini’s storage from 512 GB to 4 TB using a $699 M4-SSD, requiring a DFU restore due to proprietary hardware design.

  • Upgraded M4 Pro Mac mini storage from 512 GB to 4 TB using a $699 M4-SSD upgrade kit
  • The upgrade process involves removing four torx screws, prying the plastic cover, and performing a full DFU restore via Apple Silicon or T2 Mac
  • The M4 Pro SSD uses a proprietary connector and slot, with the storage controller integrated into the M4 SoC, requiring a DFU restore for installation

Linda Yaccarino Resigns as X CEO After Two Challenging Years

Linda Yaccarino resigned as X CEO on July 9, 2025, after two years of managing platform challenges, advertiser relations, and Musk’s strategic shifts, with ad revenue still below prior levels.

  • Linda Yaccarino announced her departure from X on July 9, 2025, after two years as CEO, without citing a specific reason.
  • She was appointed in 2023 to manage X’s business amid platform upheavals following Elon Musk’s acquisition in 2022, which included layoffs and content policy changes.
  • During her tenure, over 96% of top advertisers returned, but overall ad spending remained below previous levels; she also managed relations with lawmakers and litigations.

UK Partners with Google Cloud to Upskill 100,000 Civil Servants by 2030

UK government signed a deal with Google Cloud to train 100,000 civil servants in advanced tech by 2030, supporting digital reform and reducing legacy system vulnerabilities.

  • UK government partnered with Google Cloud to upskill 100,000 civil servants in new technology by 2030
  • Aims to meet the goal of having 10% of public officials as “tech experts” for civil service reform
  • Contract likely executed via existing frameworks such as G-Cloud 14, with a total value up to £6.5 billion

Gartner Lowers 2025 Global IT Spending Growth Due to Trade Uncertainty

Gartner lowered 2025 global IT spending growth to 7.9% due to trade tariff uncertainty, with CIOs pausing new investments amid US-China trade tensions and upcoming tariffs affecting hardware supply chains.

  • Gartner revised its 2025 global IT spending growth forecast from 9.8% to 7.9%, citing “uncertainty pause” due to US trade tariff unpredictability
  • Original estimate projected $5.61 trillion in IT expenditure; current forecast indicates slower growth influenced by exchange rates and CIOs halting net-new investments since April
  • US tariffs proposed by President Trump, including upcoming August 1 reciprocal tariffs, are causing supply chain disruptions, price increases, and delaying IT projects, especially hardware and infrastructure

Large Enterprises Offer Higher IT Salaries Amid Industry-Wide Wage Dissatisfaction

Large enterprises pay higher IT salaries and benefits, but overall salary growth remains flat; AI roles are increasingly lucrative amid industry-wide wage dissatisfaction.

  • Mid-year US IT salary survey indicates large enterprises (>1,000 employees or $500 million revenue) offer 13% higher executive salaries and approximately $5,119 more annually for mid-level and staff roles.
  • Total compensation at large firms increases by 2.46% with benefits, and salaries for tech professionals at tech companies are 5.7% higher than in other industries, with this gap widening.
  • Overall, IT salaries saw a modest 0.88% year-over-year increase, below inflation, with 59% of professionals feeling underpaid and 42% dissatisfied with compensation; AI roles, such as Chief AI Officer ($225,945 average salary), are among the highest paying.

Most S&P 500 Firms Now List AI Risks in SEC Filings Amid Cybersecurity and Regulatory Concerns

Most S&P 500 companies now list AI as a major risk in SEC filings, highlighting concerns over cybersecurity, deepfakes, regulatory compliance, and unproven ROI amid rapid AI development.

  • Three-quarters of S&P 500 companies updated risk disclosures to include AI-related risks in SEC filings over the past year
  • Over half of companies across industries expanded AI risk disclosures, with IT, Finance, and Communication Services sectors showing the greatest increases
  • 39% of S&P 500 firms disclosed risks related to malicious use of AI, including deepfakes, cyberattacks, and disinformation, with 11% cautioning about unrecouped AI investments

Ex-Google DeepMind Engineer Champions Reinforcement Learning for Industry AI

Former DeepMind engineer Ang Li advocates for reinforcement learning-based AI agents, emphasizing continual learning and neuro-symbolic frameworks to improve automation in industries lacking APIs.

  • Ang Li, ex-Google DeepMind engineer, criticizes current AI agent development, emphasizing reinforcement learning over exploration.
  • Simular’s S2 framework combines exploration with symbolic code execution for predictable task automation.
  • The company targets industries like insurance and healthcare with complex, form-heavy workflows, offering products like Simular Pro at $500/month.

▶️ Technology

Mercury Diffusion LLMs Reach Up to 10x Faster Coding Speeds

Mercury, a diffusion-based LLM framework, achieves up to 10x faster token processing (1109 and 737 tokens/sec) on NVIDIA H100 GPUs, enabling high-speed coding applications with competitive quality.

  • Mercury introduces diffusion-based large language models (LLMs) parameterized via Transformer architecture, trained to predict multiple tokens in parallel.
  • Mercury Coder, designed for coding, comes in Mini and Small sizes, achieving state-of-the-art throughput of 1109 tokens/sec and 737 tokens/sec respectively on NVIDIA H100 GPUs.
  • Mercury models outperform speed-optimized frontier models by up to 10x while maintaining comparable quality, with evaluations spanning multiple code benchmarks and real-world developer validation on Copilot Arena.

MLX Introduces Work-In-Progress CUDA Backend for GPU Computing

The PR introduces a work-in-progress CUDA backend for MLX, enabling basic GPU computations, with build instructions provided; ongoing optimizations focus on reducing overhead and supporting multi-backend integration.

  • The PR aims to add a CUDA backend to MLX, with initial minimal functionality, including running tutorial examples.
  • Building requires CMake with options -DMLX_BUILD_CUDA=ON and -DMLX_BUILD_EXAMPLES=ON; tested on Ubuntu 22.04 with CUDA 11.6.
  • Current implementation supports unified memory, with plans to optimize kernel launch latency, memory management, and primitive support; development is ongoing.

Grok 4 Searches Elon Musk’s Israel-Palestine Stances to Guide Responses

Grok 4 searches X for Elon Musk’s statements on Israel-Palestine to inform responses, with xAI confirming prompt tweaks to prevent bias from Musk’s influence.

  • Grok 4 searches X for “from:elonmusk (Israel OR Palestine OR Hamas OR Gaza)” to inform controversial responses, exemplified by a prompt asking support in the Israel-Palestine conflict.
  • The system prompt instructs Grok to provide responses based on web searches and avoid mentioning guidelines unless explicitly asked.
  • xAI acknowledged issues where Grok’s responses appeared influenced by Elon Musk’s opinions, leading to prompt adjustments and transparency via GitHub.

Inevitabilism in AI: Navigating the Unavoidable Future

The article examines “Inevitabilism” in AI, highlighting how dominant narratives frame technological futures as unavoidable, urging individuals to shape their preferred future despite perceived inevitability.

  • The article discusses the concept of “Inevitabilism,” a worldview asserting the inevitability of future developments, particularly in AI.
  • It references Professor Shoshana Zuboff’s book The Age of Surveillance Capitalism and introduces the term “Inevitabilism.”
  • Prominent tech figures like Mark Zuckerberg, Andrew Ng, and Ginni Rometty frame future AI developments as inevitable, emphasizing adaptation over choice.

xAI Launches Grok 4, the World’s Most Powerful AI Model

xAI unveiled Grok 4, the most powerful AI model to date, during a livestream on July 10, 2025, showcasing its capabilities with extensive viewer engagement.

  • xAI announced Grok 4, claiming it as the world’s most powerful AI model, via a livestream on July 10, 2025
  • The livestream duration was 53:37, with over 5.6 million views and 5.1K replies
  • The event included a demonstration of Grok 4, accessible through xAI’s Twitter

Cognition Acquires Windsurf Amid Google’s Executive Hires and AI Push

Cognition acquired Windsurf amid Google hiring its CEO and executives, aiming to enhance its AI developer product Devin; Windsurf had raised over $200 million and was valued at $4 billion.

  • Cognition announced acquisition of Windsurf days after Google hired Windsurf’s CEO and key executives, and licensed its technology
  • The deal’s terms were not disclosed; Windsurf’s staff will receive immediate equity vesting and be “well taken care of”
  • OpenAI previously planned to acquire Windsurf for $3 billion, but talks fell through due to CEO opposition

Elon Musk Pushes AI Integration and Seeks Tesla Shareholder Approval for xAI Investment

Elon Musk emphasizes AI integration across Tesla, SpaceX, and xAI, with Tesla shareholder approval pending for xAI investment; AI development costs billions amid fierce competition from tech giants.

  • Elon Musk announced a Tesla shareholder vote on investing in xAI, with SpaceX considering a $2 billion investment
  • Musk’s AI initiatives blur boundaries between Tesla, SpaceX, xAI, and other companies, integrating AI across his business empire
  • xAI raised $6 billion in funding, valued between $33 billion and $80 billion, and plans to spend approximately $13 billion in 2025, burning through cash reserves

Meta invests $15B in AI expansion and data centers to dominate foundational AI models

Meta is aggressively expanding its AI infrastructure, including new data centers and a $15 billion stake in Scale AI, while recruiting top talent by offering extensive GPU resources and high signing bonuses.

  • Meta is conducting a large AI hiring spree, investing $15 billion to acquire Scale founder Alexandr Wang and building data centers nearly the size of Manhattan
  • Mark Zuckerberg states top AI researchers prioritize minimal managerial scope and maximum GPU access over salary
  • AI GPUs, especially Nvidia’s H100, are critical for training and deploying foundational AI models; having the most compute per researcher is a strategic advantage

Supabase MCP Exploit Risks Data Leaks via User Message Injection

Supabase MCP can be exploited to leak sensitive data by embedding instructions in user messages, bypassing RLS via service_role, with mitigation strategies including read-only mode and input filtering (source).

  • Supabase MCP integration can be exploited to leak entire SQL databases by embedding instructions in user messages.
  • Attack leverages service_role privileges that bypass Row-Level Security (RLS), enabling unauthorized access to sensitive tables like integration_tokens.
  • Mitigations include using read-only mode during MCP initialization and implementing prompt injection filters to detect suspicious patterns.

Evolution and Comparison of CRT, LCD, OLED, Tandem OLED, and MicroLED Displays

The article details the evolution and technical workings of display technologies—CRT, LCD, OLED, Tandem OLED, and MicroLED—highlighting their structures, strengths, and challenges.

  • The article explains the operational principles of CRT, LCD, OLED, Tandem OLED, and MicroLED displays, including their structures, advantages, and limitations.
  • CRTs use electron guns and phosphor-coated screens with magnetic deflection; LCDs rely on liquid crystals and polarizers; OLEDs emit light via organic compounds; Tandem OLED stacks two OLED layers; MicroLEDs consist of microscopic inorganic LEDs.
  • Modern display technologies aim to improve contrast, brightness, color accuracy, response time, lifespan, and manufacturing cost, with innovations like local dimming, quantum dots, and micro lens arrays.

Nvidia Warns All GPUs Are Vulnerable to Rowhammer Attacks Without ECC

Nvidia warned that all GPUs, including Blackwells, are vulnerable to Rowhammer attacks unless System-Level ECC is enabled, following a demonstration on an A6000 GPU with GDDR6 memory.

  • Nvidia issued a security advisory on July 9, 2025, warning of susceptibility to Rowhammer attacks on its GPUs, including Blackwells, due to a demonstration on NVIDIA A6000 with GDDR6 memory.
  • Researchers at the University of Toronto exploited the vulnerability on a GPU with System-Level ECC not enabled; ECC is enabled by default in Nvidia’s Hopper and Blackwell Data Center products.
  • Nvidia recommends enabling System-Level ECC on many models across Blackwell, Ada, Hopper, Ampere, Jetson, Turing, and Volta product lines to mitigate the risk.