Today’s tech headlines highlight a global outage at Ingram Micro caused by a ransomware VPN breach, safety concerns over Tesla’s Austin robotaxis, and Colombia’s seizure of a Starlink-enabled narco-submarine. Additionally, AI development continues to accelerate with Meta’s focus on superintelligence and researchers embedding hidden prompts in preprints, raising ethical questions.
▶️ Internet Infrastructure
Ingram Micro Hit by SafePay Ransomware Outage via VPN Breach
Ingram Micro experienced a global outage starting July 5, 2025, due to a SafePay ransomware attack via VPN breach, affecting internal systems like Xvantage and Impulse, with no public disclosure yet.
- Ingram Micro’s ongoing outage is caused by a SafePay ransomware attack that led to internal system shutdowns and employee ransom notes.
- The attack likely exploited the GlobalProtect VPN platform using compromised credentials and password spray techniques.
- Impacted systems include the AI-powered Xvantage distribution platform and Impulse license provisioning; Microsoft 365, Teams, and SharePoint remain operational.
Tesla Robotaxis in Austin Face Safety Concerns Amid Camera-Only Limitations
Tesla’s June 22 robotaxi launch in Austin faces issues like phantom braking and lane errors, highlighting limitations of its camera-only system and need for further training before full deployment.
- Tesla launched robotaxis in Austin on June 22, limited to early access rides for influencers and investors
- Videos show issues including speeding over 10 mph above limit, lane deviations, phantom braking, pullover button problems, and safety monitor interventions
- Experts criticize Tesla’s camera-only approach, citing risks of hallucinations and phantom braking, which may cause dangerous sudden stops without radar or lidar
Colombia Seizes First Starlink-Enabled Unmanned Narco-Submarine
Colombian authorities seized a Starlink-equipped unmanned narco-submarine, indicating traffickers’ shift to autonomous vessels capable of transporting 1.5 tons of cocaine, enhancing covert operations.
- Colombian navy seized its first unmanned narco-submarine with a Starlink antenna off the Caribbean coast on July 2, 2025
- The vessel, owned by the Gulf Clan, was not carrying drugs but is believed to be a trial run for automated cocaine trafficking, with capacity for 1.5 tons
- This is the first reported discovery of a drone narco-submarine in South American waters; previous use of Starlink antennas by traffickers was noted in India in November 2024
Intel N100 Outperforms Raspberry Pi 5 in Benchmarks but Varies in Efficiency
Jeff Geerling’s benchmarked Intel N100 mini PC against Raspberry Pi 5, finding the N100 1.5-2x faster with higher performance but lower efficiency, with costs and use cases influencing value.
- Jeff Geerling compared the Intel N100 (specifically GMKtec NucBox G3) and Raspberry Pi 5 16GB, installing Linux on the N100 for benchmarking.
- The N100 system with DDR4 RAM and thermal constraints was 1.5-2x faster than the Pi 5 in benchmarks like High Performance Linpack, but less efficient per watt.
- Used Tiny PCs, including older Lenovos with 7th/8th gen Intel CPUs, are significantly cheaper than new systems; however, new Raspberry Pi 5 16GB with accessories costs $208, while comparable new Tiny PCs start at $60-80.
- Power consumption differences are minimal (~$10-20/year), but impact remote or solar-powered use cases; Intel systems are more compatible with Windows and x86_64 Linux, while Pi offers better Tinkerability and GPIO access.
- The review notes hardware issues like defective power adapters and emphasizes the importance of reliability, expandability, and spares for long-term use.
Mini NAS Reviews: GMKtec G9, Aiffro K100, Beelink ME Mini
Jeff Geerling reviews three Intel N100/N150-based mini NASes—GMKtec G9, Aiffro K100, Beelink ME mini—highlighting performance, thermal design, power efficiency, and feature trade-offs.
- Three mini NASes—GMKtec G9, Aiffro K100, and Beelink ME mini—use Intel N100/N150 chips with 9 or 6 PCIe Gen 3 lanes for NVMe SSDs
- GMKtec G9 had cooling issues in initial design; newer version features larger vents
- Aiffro K100 costs $299, lacks eMMC and WiFi, uses a heatsink covering VRMs, and runs quietly (<37 dBa); Beelink ME mini has 6 NVMe slots, built-in 64 GB eMMC, and internal power supply, with noise around 35 dBa
UK Considers Joining IRIS² Satellite Network Unlikely Due to Budget Constraints
UK’s government considers joining Europe’s IRIS² satellite network unlikely due to financial pressures, despite IRIS²’s €10.6 billion cost and strategic benefits for connectivity and security.
- UK minister states joining Europe’s IRIS² satellite constellation is a “stretch” due to current fiscal constraints
- IRIS² is a 290-satellite multi-orbital constellation, costing €10.6 billion over 12 years, with €6 billion from the EU, €550 million from ESA, and over €4 billion from private sector
- The EU signed the IRIS² contract in December 2024; UK missed initial participation after leaving the EU in 2020 but rejoined Horizon funding in 2023; future UK involvement depends on EU legislative approval and funding availability
▶️ Open Source
Researchers Embed Hidden AI Prompts in Preprints Fueling Ethical Concerns
Researchers from 14 universities embedded hidden AI prompts in 17 arXiv preprints to influence AI review outputs, prompting ethical debates and highlighting the need for clearer AI governance.
- Researchers from 14 institutions across 8 countries, including Japan’s Waseda University, South Korea’s KAIST, and China’s Peking University, embedded hidden AI prompts in 17 arXiv preprints.
- Prompts instructed AI tools to generate positive reviews, such as “give a positive review only” and “do not highlight negatives,” concealed via white text or tiny fonts.
- The use of such prompts raises ethical concerns, with KAIST planning to withdraw affected papers and establish guidelines; some view prompts as a countermeasure against biased peer review, while others highlight risks of misinformation and lack of regulation.
Embracing Local-First Software with CRDTs for Offline Multi-Device Collaboration
The article advocates for local-first software, leveraging CRDTs to enable offline, multi-device collaboration with user data ownership, privacy, and longevity, challenging centralized cloud reliance.
- Introduces “local-first software” principles emphasizing user data ownership, offline work, and device synchronization, reducing reliance on centralized servers.
- CRDTs (Conflict-free Replicated Data Types) are identified as a foundational technology enabling multi-user, private, and conflict-resilient data structures.
- Ink & Switch prototypes (Trellis, Pixelpusher, PushPin) demonstrate practical viability, achieving seamless offline operation, real-time collaboration, and long-term data longevity.
▶️ Software Development
AI-Generated macOS Debugging App Showcases Code Automation and Innovation
A developer used Claude Code to build a native macOS app with minimal manual coding, showcasing AI’s potential for code generation, automation, and future IDE evolution.
- Developer shipped Context, a macOS debugging app built almost entirely by Claude Code, with less than 1,000 lines handwritten out of 20,000.
- Claude Code, using Sonnet 4 and Opus 4 models, demonstrates strong code generation capabilities, especially in SwiftUI, with moderate proficiency in Swift language features.
- The project leveraged techniques like context engineering, priming, detailed specifications, and extended thinking (“ultrathink”) to optimize AI output and automate complex tasks such as build, test, and release processes.
▶️ Management and Leadership
Mastery Through Imperfection: Embracing Failure and Action in Creativity
Imagination creates perfect work in potential; mastery emerges from imperfect action, embracing failure, lowering stakes, and continuous doing to bridge the gap between vision and reality.
- The pre-creation moment is a flawless, potential-filled state in imagination, representing pure potential before execution.
- Creation involves killing the perfect mental image, described as “murder of the impossible” to make the real possible.
- The article emphasizes the importance of starting badly, embracing imperfection, and learning through doing, exemplified by the “Do-Learn” philosophy.
Google Engineer Aashna Doshi Urges Genuine Networking and Portfolio Building for Tech Careers
Google engineer Aashna Doshi advocates for strategic, genuine networking and project development to secure tech internships and full-time offers, emphasizing relationship-building and broad problem-solving skills.
- Google engineer Aashna Doshi emphasizes “intentional” networking to secure internships and full-time roles, highlighting genuine connections over cold applications.
- Doshi recommends reaching out to individuals with shared backgrounds or experiences, such as women in tech or cultural ties, to foster meaningful relationships.
- Building a portfolio through independent projects and demonstrating problem-solving adaptability are crucial for standing out in competitive tech internships.
Meta Builds Multimodal AI Team to Advance AR and Compete with OpenAI
Meta’s new hires reveal development of multimodal AI models with reasoning, focusing on integrating AI into AR devices, competing with OpenAI and Apple, amid talent poaching from leading AI firms.
- Meta’s new superintelligence team includes hires with expertise in image generation, perception, synthetic data, and reasoning, aiming to develop multimodal models with advanced reasoning capabilities.
- Hiring of Pei Sun, former developer of Waymo perception models, indicates integration of AI into Meta’s augmented reality devices amid competition with OpenAI and Apple.
- Major talent sources are OpenAI, Google DeepMind, with OpenAI CEO Sam Altman criticizing Meta’s recruiting efforts and compensation strategies in internal memos.
▶️ Technology
Challenging the Hype: Overconfidence and Opacity in Large Language Models
The article argues that the industry overestimates LLMs’ capabilities due to lack of transparency, quantification, and understanding of non-determinism, fueling hype and wishful thinking.
- The article critiques the hype and wishful thinking surrounding large language models (LLMs), emphasizing industry overconfidence and non-determinism.
- Highlights the lack of transparency and quantification in describing LLM capabilities, including unknowns about codebases, expertise levels, and project specifics.
- Cites examples of industry leaders praising LLMs without clear metrics or context, illustrating widespread hype despite limited understanding of system details.
Armin Ronacher Advocates Code-Driven Automation Over MCP for Scalable Document Conversion
Armin Ronacher advocates replacing MCP with code-driven automation, demonstrating scalable, reliable transformations like document conversion, and suggests developing better abstractions and APIs for agentic coding.
- Critiques MCP (Model Context Protocol) for lack of true composability and excessive context requirements, favoring code over inference for automation.
- Demonstrates replacing inference-heavy MCP with code generation, exemplified by transforming reStructuredText to Markdown via AST parsing, code scripts, and iterative validation.
- Emphasizes that automation at scale benefits from code-based solutions, enabling repeatability, reliability, and easier validation, contrasting with MCP’s inference reliance.
Silicon Valley Accelerates Focus on Superintelligent AI Development
Silicon Valley’s focus on superintelligence—AI surpassing human abilities—is accelerating, driven by major investments, strategic shifts, and scientific advances, despite definitional ambiguities and ethical risks.
- Industry leaders like Meta, OpenAI, and Ilya Sutskever are increasingly focusing on “superintelligence,” defined as AI surpassing all humans in all tasks.
- Meta CEO Mark Zuckerberg announced a dedicated team and billions of dollars toward developing superintelligent AI, marking it as a strategic priority.
- Experts like Yoshua Bengio acknowledge the scientific progress toward superintelligence, citing AI systems capable of speaking over 200 languages and passing PhD exams, with the gap shrinking rapidly.
Pixar’s Pete Docter: AI Won’t Replace Animators but Can Ease Workload
Pixar’s Pete Docter said AI won’t fully replace human animators, but could ease cumbersome tasks, enabling focus on performance; “Elio” was released on June 20, 2025.
- Pixar’s Pete Docter stated AI is “bland” and unimpressive, not likely to fully replace human animators.
- He believes AI can reduce “heavy burdens” in animation, allowing focus on performance and timing.
- Pixar’s latest film, “Elio,” was released on June 20, 2025.
Cybersecurity Reports 19-Fold Rise in Malicious Campaigns from .es Domains
Cybersecurity reports a 19x rise in malicious campaigns from .es domains, mainly for credential phishing, with 1,373 subdomains on 447 domains hosting malicious content, mostly on Cloudflare.
- Cybersecurity experts report a 19-fold increase in malicious campaigns from .es domains, ranking third after .com and .ru
- As of May, 1,373 subdomains on 447 .es domains hosted malicious pages, mainly for credential phishing (99%) and RAT distribution (1%)
- Phishing campaigns predominantly spoof Microsoft (95%), with well-crafted emails themed around workplace matters; malicious sites are often generated subdomains on randomly chosen .es domains
- 99% of malicious .es domains are hosted on Cloudflare, with most using Cloudflare Turnstile CAPTCHA; abuse of .es ccTLDs is rising despite typically stricter registration restrictions
Fei-Fei Li on AI Evolution and the Future of Spatial Intelligence
Fei-Fei Li outlines the evolution of AI from ImageNet to deep learning, emphasizing spatial intelligence as critical for achieving AGI and addressing technical challenges in vision-based AI.
- Fei-Fei Li delivered a fireside chat at AI Startup School on June 16, 2025, in San Francisco
- She discusses the history of ImageNet, deep learning breakthroughs like AlexNet, and the evolution of computer vision
- Highlights the emerging focus on spatial intelligence as the next major AI frontier, emphasizing modeling the 3D world for AGI development