Read Tech News With Digest

After sixteen years of loyalty, I am finally leaving Gmail because its aggressive AI features feel disrespectful. Unsolicited summaries, auto-generated replies, and constant prompts to 'improve' my writing make me feel incapable. Instead of a helpful tool, the software feels like it is actively trying to inflate metrics at the expense of user dignity. I am now migrating to Fastmail to reclaim control over my communication.

I discovered a critical vulnerability in VSCode's webview security model that allows attackers to steal GitHub tokens with a single click. By exploiting how keyboard events are handled across origins, malicious scripts can simulate user actions to install rogue extensions. This flaw compromises private repositories and grants full write access, highlighting a severe risk in the browser-based github.dev environment.

We are excited to introduce MAI-Code-1-Flash, a new lightweight coding model built specifically for real-world developer workflows. Designed to integrate seamlessly with GitHub Copilot and Visual Studio Code, this model delivers high-quality assistance while significantly reducing token usage. By focusing on adaptive thinking and production-ready training, we help engineering teams write better code faster without the usual trade-off between speed and accuracy.

I used advanced CT scanning technology to peer inside a BYD electric vehicle, exposing the intricate engineering hidden beneath the surface. This detailed look reveals the construction quality and component layout of a car that remains unavailable to American consumers, offering a unique glimpse into the manufacturing standards of this global leader.

I built nbd-vram to help laptop users with soldered memory by turning unused NVIDIA GPU VRAM into fast swap space. Instead of relying on slow SSDs, this tool uses the CUDA driver to create a block device that absorbs memory overflow before hitting zram or disk. It works on consumer RTX cards without needing kernel modules, effectively tripling addressable memory on my test setup.

Our groundbreaking study reveals that law professors overwhelmingly prefer AI-generated answers over responses from their peers. In a blind evaluation of nearly 3,000 comparisons, AI won 75% of matchups on complex contract law questions. We found that AI tutors can offer high-quality, on-demand support that complements classroom instruction, challenging assumptions about the role of artificial intelligence in legal education.

Oracle co-founder Larry Ellison predicts a future where AI, drones, and ubiquitous cameras create an aggressive surveillance state. He argues that the constant recording and reporting of all activities will inevitably force citizens to maintain their best behavior, echoing Orwellian fears about the erosion of privacy in the 21st century.

After a month of using Clojure to build a static site generator, I've found it more cohesive than Common Lisp and more practical than Scheme. While the syntax initially felt overwhelming compared to other Lisps, the ergonomic data structures and robust JVM ecosystem have won me over. Despite some friction with Java interop and bracket-heavy code, I plan to stick with Clojure for its power and fun factor.

President Trump finally signed a scaled-back executive order on artificial intelligence after weeks of internal debate and a last-minute rejection of a stricter draft. The new policy mandates a voluntary 30-day review for powerful AI models before public release, aiming to balance cybersecurity concerns with the need to maintain U.S. competitiveness against China. While the industry celebrated the reduction from a proposed 90-day delay, critics worry the voluntary nature of the review may be insufficient to address rapid AI advancements.

I share a profound three-part rule my father taught me about making a living. It reveals that lying to those who want deception leads to riches, while telling the truth to those seeking it earns a modest living. However, the most dangerous path is telling the truth to people who prefer lies, which inevitably leads to financial ruin. This simple wisdom highlights the harsh reality of human nature in business.

Meta, Google, and Apple are colluding to build a new browser feature called Attribution Level 1 that tracks ad effectiveness without user consent. This system bypasses privacy regulations for Big Tech while forcing others to comply, ultimately skewing data to favor search and social ads. I argue this cartel harms the open web, undermines local journalism, and turns every browser into a lobbyist against real privacy protections.

We are officially retiring the X11 session in Plasma 6.8 to dedicate all our efforts to the Wayland future. While XWayland ensures your legacy applications still run smoothly, this shift allows us to eliminate conflicting code paths and unlock significant performance improvements. With over 95% of recent users already on Wayland, we are ready to move forward, though we still welcome feedback on any remaining issues before the final transition.

I am excited to share that Microsoft has released Coreutils for Windows, a native package bringing familiar UNIX commands like ls, cp, and grep directly to your system. This project aims to make moving between Linux, macOS, WSL, and Windows frictionless by ensuring your existing scripts and pipelines work without translation. While currently in preview, it offers a unified experience for developers, though users should be aware of specific shell conflicts and Windows caveats like CRLF line endings and the lack of POSIX signals.

Morningstar analysts value SpaceX at $780 billion, significantly below its $1.75 trillion IPO target. They cite uncertain economics for the company's AI ventures like xAI and potential hurdles for Starlink. Despite high market enthusiasm, the firm warns that SpaceX is currently overvalued and suggests investors might find better entry points after the stock debut.

I built a perfect, accessible website in seconds using an AI agent, proving that traditional clean code is now a museum piece. In this new era, complexity is the valuation, and burning millions in tokens is the actual business model. While static sites are for those who can still read, the future belongs to agentic slop that prioritizes attention over functionality, leaving open source to drown in machine-generated garbage.

The legendary HP-16C programmer's calculator returns as a Collector's Edition, blending its classic RPN design with modern speed. This updated model offers up to 100x faster performance, enhanced programmability, and the new ability to save and load programs. It remains an essential tool for engineers and computer scientists who value bitwise operations and base conversions.

We are introducing MAI-Thinking-1, a medium-sized reasoning model trained from the ground up on clean, licensed data without distillation. Despite its smaller footprint, it matches leading models on SWE-Bench Pro and achieves top scores in mathematical reasoning. In blind human evaluations, users preferred MAI-Thinking-1 over Claude Sonnet 4.6 for its clarity and helpfulness. This model represents a key step toward Humanist Superintelligence, designed to serve organizations with advanced capabilities while maintaining full control over the training stack.

I explore FidoNet, a massive store-and-forward email network connecting over 20,000 nodes globally via modems. Originally built for MS-DOS, it now runs on UNIX and Apple systems, driven by a need to minimize costly phone time. From its unique addressing hierarchy to experimental tunnels through the Internet, this network proves how private individuals built a resilient communication infrastructure before the web took over.

We discovered that sending raw images to models at query time is too expensive and often impossible at scale. Instead, we describe every image once during indexing using a cheap vision model, storing the text descriptions alongside our documentation. This approach cuts per-query costs to just 1% to 6% over text-only while significantly improving answer quality, turning static visuals into actionable insights without the heavy vision bill.

We are building the Open Repair Data Standard to help groups worldwide share and combine repair data for small electronics. By standardizing how we record product details, repair outcomes, and session information, we can identify global trends in device failures. Our goal is to make this open data easily accessible so communities can learn from each other and improve repair success rates everywhere.

I've observed a growing national revolt against data centers, driven by concerns over noise, water, and energy use. However, I believe this backlash is a proxy for a deeper dread of AI and an uncertain future. Instead of blocking facilities one by one, we need a broader national debate on regulating AI to ensure it expands human agency rather than diminishing it.

I am excited to introduce the GitHub Copilot app, a new desktop experience designed for agent-driven development. This tool allows you to manage your entire development lifecycle, from picking up an issue to merging code, all within a single interface. You can run multiple parallel agents across repositories, review diffs, and even let an agent close the loop for you. Built to adapt to your workflow, it supports automation and custom skills via MCP servers.

We tackled the software hurdles to run DeepSeek-V4-Flash on AMD MI300X, addressing unique FP8 dialect mismatches and missing attention kernels in AITER. By fixing HIP graph capture issues and optimizing sparse paths, we proved this underappreciated accelerator can deliver competitive performance at half the cost of NVIDIA hardware, despite its initial incompatibilities.

I am thrilled to announce QBE 1.3, our most significant release since version 1.0. We have introduced a new IL matching algorithm, added Windows ABI support, and enabled position-independent code generation. Thanks to new optimizations, we now achieve over 63% of commercial compiler performance on benchmarks, with even greater gains for Hare users. This release proves that teamwork drives innovation in compiler development.

Microsoft has introduced Scout, a new autonomous AI agent built on the OpenClaw framework designed to act as an 'autopilot' within Microsoft 365. Unlike traditional tools, Scout operates independently around the clock to manage tasks like scheduling and identifying workflow risks across Teams, Outlook, and other apps. Currently available as an experimental release, this always-on agent aims to reduce mundane office work while maintaining enterprise-grade security controls.

I report on how Meta repeatedly refuses to engage with the Appeals Centre Europe regarding wrongful account bans. Despite reviewing 4,600 cases, Meta provided evidence in fewer than 100, leaving users frustrated and unable to appeal. The report also reveals that major platforms often fail to remove hate speech, with TikTok and Instagram showing high rates of non-compliance with their own policies.

We are introducing Search as Code (SaC) to replace outdated monolithic search pipelines with an agentic architecture. By exposing atomic search primitives via an SDK, we empower AI agents to generate custom retrieval strategies through code. This approach grants models fine-grained control over how context is retrieved and processed, solving critical bottlenecks in complex, long-running tasks that traditional systems cannot handle.

Uber has implemented strict monthly spending limits of $1,500 per employee for AI tools like Claude Code and Cursor after exhausting its entire annual budget in just four months. This shift follows a previous strategy that aggressively encouraged staff to use AI as much as possible, even tracking usage on leaderboards. The move highlights growing industry skepticism about the tangible return on investment for rapid AI adoption.

I argue that the AI industry has built a bubble on lies, deliberately hiding the true costs of token usage behind subsidized subscriptions. As companies like Microsoft and Anthropic shift to token-based billing, enterprises are finally facing the reality that they cannot measure ROI or control expenses. This shift reveals that AI is becoming more expensive, not cheaper, forcing businesses to confront the uneconomical nature of their adoption strategies.

I recently received a notification from Amazon stating that my paid music subscription will soon include advertisements and no longer allow offline downloads. This unexpected change to the service terms feels like a significant downgrade for paying customers who expected an ad-free experience with full access to their music library.

Leaked internal documents show Microsoft executives explicitly planning to make users addicted to their new Scout AI assistant. The strategy outlines a three-phase rollout starting with building daily dependency before expanding features. While the tool, originally piloted as ClawPilot, aims to automate workflows for knowledge workers, the candid language regarding addiction has sparked concern among employees about ethical product design.

I tested Elon Musk's Vegas Loop, expecting a revolutionary underground highway but finding a confusing, underwhelming experience. With only nine stations, human-driven Teslas, and a broken booking system, the project feels like a gimmick rather than a solution. From waiting in a parking garage to navigating single-lane tunnels, the reality of this ambitious transit system falls far short of its promise.

Anthropic is scaling Project Glasswing, leveraging its powerful Claude Mythos AI to secure critical infrastructure in over 15 countries. This expansion adds 150 new partners, including major players in power, water, and healthcare sectors, aiming to prevent catastrophic attacks affecting millions. The initiative follows Anthropic's confidential IPO filing and a massive funding round, as the company races to establish global security safeguards before rival AI models catch up.

Amazon is facing a class action lawsuit alleging that Ring's Familiar Faces feature collects facial data from passersby without their consent. Despite pushback from privacy advocates and Senator Ed Markey, the company launched the AI tool in December. This legal challenge highlights ongoing concerns about Ring's privacy record, including a past FTC settlement and controversial partnerships with law enforcement.

The European Central Bank reveals a historic shift in global finance where gold has overtaken US Treasuries to become the world's leading reserve asset. This change signals growing distrust in traditional paper currencies and highlights a strategic pivot by central banks toward tangible assets for long-term stability.

After suing ICE for documents regarding its $2 million contract with Paragon, a spyware firm, we received heavily redacted files. The documents reveal Paragon's Graphite tool can hack encrypted apps like Signal and WhatsApp, while explicitly promising to help the agency avoid detection. Despite public claims about fighting fentanyl, ICE is withholding crucial details about this controversial surveillance capability, raising serious civil rights concerns.

New York police are probing bizarre reports of groups climbing out of the city's sewer systems late at night. Videos circulating on social media show individuals in waders entering manholes and resurfacing hours later, prompting the NYPD and the Environmental Protection Department to inspect the tunnels. While no damage or nefarious items were found, investigators suspect the intruders might be scavenging for valuables lost in the sewage.

I am willo2, the largest holder of YES on the MicroStrategy sells Bitcoin by May 31st market, and I was just scammed for $500K by Polymarket. Despite my significant position, the platform's handling of the event left me with massive losses. I am sharing the full story of what happened to ensure others understand the risks involved in these prediction markets.

As President, I am cutting bureaucratic red tape to unleash American AI innovation while prioritizing national security. My administration is working with the private sector to harden government systems, protect intellectual property, and deploy advanced AI tools rapidly. We will establish a voluntary framework for secure frontier model deployment and crack down on criminal actors using AI, ensuring the United States maintains global dominance in this critical technology.